Scientific Computation

Harold A. Sabbagh
R. Kim Murphy
Elias H. Sabbagh
John C. Aldrin
Jeremy S. Knopp

Computational
Electromagnetics
and Model-Based
Inversion
A Modern Paradigm for Eddy-Current
Nondestructive Evaluation
Computational Electromagnetics and Model-Based
Inversion
Scientific Computation
Editorial Board
J.-J. Chattot, Davis, CA, USA
P. Colella, Berkeley, CA, USA
R. Glowinski, Houston, TX, USA
Y. Hussaini, Tallahassee, FL, USA
P. Joly, Le Chesnay, France
D.I. Meiron, Pasadena, CA, USA
O. Pironneau, Paris, France
A. Quarteroni, Lausanne, Switzerland
and Politecnico of Milan, Milan, Italy
J. Rappaz, Lausanne, Switzerland
R. Rosner, Chicago, IL, USA
P. Sagaut, Paris, France
J.H. Seinfeld, Pasadena, CA, USA
A. Szepessy, Stockholm, Sweden
M.F. Wheeler, Austin, TX, USA

For further volumes:

http://www.springer.com/series/718
Harold A. Sabbagh • R. Kim Murphy
Elias H. Sabbagh • John C. Aldrin
Jeremy S. Knopp

Computational
Electromagnetics and
Model-Based Inversion
A Modern Paradigm for Eddy-Current
Nondestructive Evaluation

123
Harold A. Sabbagh R. Kim Murphy
Victor Technologies, LLC Victor Technologies, LLC
Bloomington, IN, USA Bloomington, IN, USA

Elias H. Sabbagh John C. Aldrin

Victor Technologies, LLC Computational Tools
Bloomington, IN, USA Gurnee, IL, USA

Jeremy S. Knopp
Air Force Research Laboratory
(AFRL/RXLP)
Wright-Patterson AFB, OH, USA

ISSN 1434-8322
ISBN 978-1-4419-8428-9 ISBN 978-1-4419-8429-6 (eBook)
DOI 10.1007/978-1-4419-8429-6
Springer New York Heidelberg Dordrecht London
Library of Congress Control Number: 2013937723

© Springer Science+Business Media New York 2013

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed. Exempted from this legal reservation are brief excerpts in connection
with reviews or scholarly analysis or material supplied specifically for the purpose of being entered
and executed on a computer system, for exclusive use by the purchaser of the work. Duplication of
this publication or parts thereof is permitted only under the provisions of the Copyright Law of the
Publisher’s location, in its current version, and permission for use must always be obtained from Springer.
Permissions for use may be obtained through RightsLink at the Copyright Clearance Center. Violations
are liable to prosecution under the respective Copyright Law.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
While the advice and information in this book are believed to be true and accurate at the date of
publication, neither the authors nor the editors nor the publisher can accept any legal responsibility for
any errors or omissions that may be made. The publisher makes no warranty, express or implied, with
respect to the material contained herein.

Printed on acid-free paper

Springer is part of Springer Science+Business Media (www.springer.com)

Preface

A few years ago, a colleague wrote a note in which he reviewed the state of the
art in eddy-current probes that were to address new problems in the nation’s steam
generators. With respect to calibration standards, he stated
The performance of these probes will not be realized unless calibration standards that
accurately simulate the range of expected defects are used. This means a series of axial
and/or circumferential notches, on both the tube od and id must be used to accurately
calibrate these probes. The ASME Section XI standard with flat bottomed holes is a very
poor representation of cracks, particularly for directional probes. In addition, the cable
between the instrument and the probe should be as short as reasonable and should be low
capacitance, low noise and low loss.

The purpose of this book is to address the entirety of issues raised in this quote,
and beyond, and to effectively resolve them favorably through the use of com-
putational electromagnetics and model-based inversion methods as a replacement
for expensive and unreliable standards. Indeed, we hope to demonstrate that our
assertion, that the computer will soon be the most important instrument in eddy-
current nondestructive evaluation (NDE), is only a mild stretch.
The book, like Gaul, is divided into three parts whose intention is to show that
computational electromagnetics is more than simply solving Maxwell’s equations
for various configurations of “anomalies” in hosts. Rather, it becomes a part of
system science of engineering, which is where we hope to elevate the notion
of ‘nondestructive evaluation’. Maxwell, himself, might not have envisioned the
ramifications of his theory, but surely he would have admired the results. (We are
reminded that a well-known aerospace NDE engineer once told us that had he
realized that you could make money solving Maxwell’s equations, he would have
paid more attention in his undergraduate E&M course.)
In the first part, we compute the Green’s dyad by working with the field equations
directly, without the intervening use of potential functions (Heaviside would be
proud!). Indeed, it isn’t until the derivation of the “vector form” of the volume-
integral equations in Chap. 3 that we first see potential functions explicitly stated.
We applied this approach originally to model electromagnetic responses in plane-
layered anisotropic bodies, such as advanced composites made of carbon-fiber

v
vi Preface

reinforced polymers, which will be an important part of the second volume in this
series. This approach works well in cylindrical coordinates, as is demonstrated in
Chap. 9.
The discretization of the volume-integral equation via Galerkin’s method on a
regular grid is certainly well known in the method of moments, and much of the
remainder of the book attempts to show the advantage of this approach in solving
large problems with reasonable computer resources.
Starting in Part I, and continuing throughout the remainder of the book, we have
sought to impress upon the reader the value of using equivalent electric circuits
or networks to interpret the physical response of the volume-integral equation.
This makes eddy-current NDE a subset of electrical engineering and should allow
those familiar with the basic concepts of electrical engineering lead into the further
development of eddy-current NDE.
The development of advanced probe models in Chap. 6 is original, especially in
treating coils, of whatever shape or orientation, as generalized magnetic dipoles
comprising solenoidal currents flowing in closed loops. This has allowed us to
model rectangular coils, or D-shaped coils, in a consistent numerical manner.
We have validated several of these models against benchmark data, as shown in
Chap. 6. What was especially pleasant, however, was the realization that we could
take the dual of the magnetic-dipole approach and model planar spiral-coil probes
assuming that the source is an electric dipole (Chap. 7). This allows us to compute
capacitive effects, as well as the usual inductive effects, thereby enabling a more
efficient design process for high-frequency applications. Problems of this type—
for example, spiral antennas—are usually treated by boundary-integral equations,
assuming perhaps that the metallic traces are perfect conductors. We have applied
the volume-integral approach, treating the spiral traces as an “anomaly” in free space
that is excited by an electric dipole, rather than an anomaly in a host that is excited by
a magnetic dipole, and are able to compute the capacitive and inductive reactances
and resistance of the probe as a function of frequency. Clearly, this approach allows
a single code to solve a greater variety of problems than originally assumed. We
can now treat resonance phenomena rigorously on the computer, without relying on
trial-and-error laboratory mockups.
Because of the increasing use of transmit–receive (T/R) arrays in eddy-current
NDE, we have included a discussion of N-port analysis of T/R arrays in Chap. 8. We
tie N-port theory of microwave networks with chain matrices to derive equivalent
networks for these arrays. This should help in designing and understanding the
behavior of these networks, especially at higher frequencies, where resonance
effects become important.
Part II is really where the fun begins, and computational electromagnetics
becomes a part of NDE system theory. We believe that the application of sophis-
ticated signal-processing and inversion algorithms lies at the heart of the future
development of NDE into a solid engineering discipline that is capable of handling
the challenging problems that new structures and materials introduce. Indeed, it will
lead NDE into the digital age, in which the computer replaces the oscilloscope as
the instrument of choice.
Preface vii

The mathematical algorithms of Part II, of course, were developed independently

of computational EM, but it is here that the advantages of modern developments in
computational EM are manifest. For example, the use of volume-integral equations
leads quickly to the efficient calculation of “surrogate structures,” such as the
interpolation tables that are used in NLSE, the nonlinear least-squares estimator
that is introduced in Chap. 12. NLSE is the workhorse for the rest of the book and
has proven to be quite flexible and robust in solving a large class of problems,
namely, those in which the anomalous region can be defined through the use of
models containing a few parameters. (In the second volume in this series, we will
introduce certain voxel-based inversion algorithms which are more powerful than
model-based algorithms. The use of volume-integral methods will be even more
crucial in the application of these algorithms. The reader will have to take our word
for it, and be patient.)
The iterative algorithm described in Fig. 14.3 of Chap. 14 is reminiscent of the
projection onto convex sets (POCS) that is described in Fig. 14.1, so we refer to the
former algorithm as “POCS” occasionally in the remainder of the book, even though
we have not yet proved that the relevant sets are in fact convex. At other times,
bowing to the purists among us, we simply refer to it as “the iterative algorithm.”
We have plans to resolve the matter in the future, but in any case, the algorithms
cited in Chap. 14 are not there simply for their looks; we will use all of them in this
book or its sequel, especially in connection with voxel-based inversion algorithms.
The high point of the book, from our perspective, is Part III, where we tie
the material of the first two parts into a demonstration of the power of computer-
aided modeling and design in solving realistic problems in eddy-current NDE. The
examples described in this part are taken from real-life problems that the authors
have explored (and continue to explore), principally in the areas of aerospace and
nuclear power. It is here that we hope to demonstrate the future of eddy-current
NDE and make a case for the aforementioned assertion about the supremacy of
the computer in that future. But the computer is useless without algorithms that
are based on fundamental physics, and this is nowhere illustrated better than in
Chap. 19, where we demonstrate that a straightforward application of electromag-
netic theory via Maxwell’s equations solves the problem of ferritic heat-exchanger
tubes in an elegant manner. Prior to this application, the orthodox view of “eddy
currents” was an inspection technique that relied on analog instruments and could
only be applied to isotropic, nonpermeable conductors. This chapter, therefore,
shows that the phrase “eddy-current NDE” really implies “electromagnetic NDE,”
in which “electric” and “magnetic” are unified, as Maxwell envisioned a century-
and-a-half ago.
At times, the book may appear to be a user’s manual, theoretical manual, or
simply an advertisement for Victor Technologies’ proprietary volume-integral code,
VIC-3D R
[40]. There are two reasons for this: first, VIC-3D R
was written to
solve precisely the problems that are described in the book and is the code best
known to us for doing that, and second, we believe that it is important for the reader
and industry to understand that computational electromagnetics is not reserved for
viii Preface

graduate theses and academic papers but is a commercially viable tool for solving
those problems that the industry needs to solve. VIC-3D R
is our contribution to the
list of codes that solve Maxwell’s equations for profit.
This book is not an introductory text; it will require a good background in electric
circuit theory, especially in understanding the concepts of phasors, impedance and
admittance, magnetic fields, magnetic induction and inductances, and magnetically
coupled circuits, as well as electromagnetic field theory, including Maxwell’s
equations. Material on electric circuits is usually covered in the sophomore year in
electrical engineering courses, whereas the required background in electromagnetic
fields is covered in upper-level undergraduate courses.
In teaching courses on electric circuits, we have found Circuit Analysis, by Elias
M. Sabbagh, Ronald Press, 1961, to be excellent preparation. It covers all aspects
listed above. Although long out of print, it is available on the internet at very low
prices. Of more recent vintage is Linear Circuit Analysis, by Raymond A. DeCarlo
and Pen-Min Lin, Oxford University Press, 2001.
When it comes to senior-level undergraduate texts on electromagnetic theory, we
can do no better than to recommend the classic Fields and Waves in Communication
Electronics, by Simon Ramo, John R. Whinnery, and Theodore Van Duzer, John
Wiley & Sons, New York, 1965. It not only gives a precise development of
electromagnetic theory, including Maxwell’s equations and their applications, but
also derives circuit concepts that are consistent with Maxwell’s equations.
If the reader wishes to develop his/her own code for Green’s functions, he/she
will need to become familiar quite quickly with the notion of “special functions.”
The classic reference on this subject, Handbook of Mathematical Functions, edited
by Milton Abramowitz and Irene A. Stegun, National Bureau of Standards, 1970,
has been superseded by NIST Handbook of Mathematical Functions, edited by
F.W.J. Olver, D.W. Lozier, R.F. Boisvert, and C.W. Clark, Cambridge University
Press, 2010. We have also found Table of Integrals, Series, and Products, by I.S.
Gradshteyn and I.M. Ryzhik, Academic Press, 1980, to be useful. Finally, to tie all
of this together in meaningful computer codes, we recommend Numerical Recipes:
The Art of Scientific Computing, by W.H. Press, B.P. Flannery, S.A. Teukolsky, and
W.T. Vetterling, Cambridge University Press, 1986.
We hope that you, the reader, will find this book useful and that you will agree
that it brings us closer to the goal of making eddy-current NDE a systematic branch
of engineering science, resulting in a more reliable system for making products and
materials safer.

Bloomington, IN, USA Harold A. Sabbagh

R. Kim Murphy
Elias H. Sabbagh
Gurnee, IL, USA John C. Aldrin
Wright-Patterson AFB, OH, USA Jeremy S. Knopp
Acknowledgments

The development of VIC-3D R

has been supported by the Departments of Defense
and Energy, and the National Science Foundation through the Small Business In-
novation Research (SBIR) program, and by commercial sources, including General
Electric, Hercules Aerospace, Rolls-Royce, United Technologies Research Center,
and the Electric Power Research Institute (EPRI). More information can be found
at http://www.sabbagh.com.
Our colleagues at Sabbagh Associates Inc., namely David Sabbagh, Jeff Treece,
and Anthony Chan, contributed significantly to the development of algorithms that
made volume-integral methods so effective in eddy-current NDE problems. The
current version of VIC-3D R
contains code based on these algorithms, as well as
a number of other features that they developed. It is a testament to their dedicated
efforts.
Our good friend, Kenji Krzywosz of the Electric Power Research Institute
(EPRI), suggested and supported the problem areas described in Sect. 18.3,
Sect. 19.4, and Chap. 20, which allowed us to get a much better handle on the
problems faced by the nuclear power industry. The material described in Sect. 15.1
was supported by Jim Benson, also with EPRI. The benchmark validation experi-
ment described there was the first for the “modern” version of VIC-3D R
, and the
clutter rejection algorithm that evolved from this validation effort continues to be an
important application in our current research.
Mark Blodgett, Eric Lindgren, and Charlie Buynak of the Air Force Research
Lab have supported our research into the application of inverse methods and
stochastic processes to complex aerospace structures. This support will allow us to
bring our ideas into practice in a much shorter period of time than would otherwise
have been possible. Our colleagues at the University of Dayton Research Institute
(UDRI), Shamachary Sathish, Ryan Mooers, and Matt Cherry, contributed to this
book in a number of ways, from theoretical discussions to carrying out validation
exercises using benchmark data.
This is a good point at which to look back some 30 years and recall the
contributions of two of our late colleagues, Walt Bantz of General Electric Aircraft

ix
x Acknowledgments

Engines and Sue Vernon of the Naval Surface Warfare Center, White Oak Labs.
Both were early supporters of model-based eddy-current NDE, as well as its
manifestation in VIC-3D R
. Walt introduced us to the Split-D probe and informed
us of its importance in inspecting aircraft engines. He encouraged its inclusion
into VIC-3D R
, and we followed his advice (see Sect. 6.6). It continues to be an
important part of our current research. One of our first papers on modeling eddy-
current probes with ferrite cores using volume-integral equations was coauthored
with Sue. She supplied the data, and we supplied the model calculations. The results
were encouraging, so we thought that we were on to something. Sue introduced
us to carbon-fiber advanced composites and asked if we could do anything with
eddy currents to characterize them. This led us to some very interesting problems in
the electromagnetic modeling of anisotropic media and validating the models with
experimental data. Again, the results were positive and will be a significant part of
our second book.
Our men at Springer, Chris Coughlin and Ho Ying Fan, were a joy to work with.
Erik Wallace, a doctoral student in mathematics at Indiana University, is going to
investigate the “POCS-Iterative” issue that we described above (look for an answer
in our next book), but of more immediate import is the fact that he is a LaTeX guru
and assembled all the document preparation files for Ho Ying to pass to his people
at Springer.
Contents

Part I Computational Electromagnetics Background

1 Overview of Methods of Computational Electromagnetics . . . . . . . . . . . . 3

1.1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3
1.2 Oak Ridge National Laboratory . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4
1.3 Method of Moments.. . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4
1.4 Model-Based Parameter Estimation in Electromagnetics . . . . . . . . . . 5
1.5 Other Approaches to Computational Electromagnetics in NDE . . . 5
2 Green’s Dyad for Plane-Layered Media . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 7
2.1 Eigenmodes of Anisotropic Media . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 7
2.2 Green Dyad for Plane-Parallel Layered Media .. . . . . . . . . . . . . . . . . . . . 10
2.2.1 Infinite-Space Green Dyad . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 11
2.2.2 Layered-Space Green Dyad .. . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 16
2.2.3 A Recursion Relation for Stratified Media . . . . . . . . . . . . . . . . 30
3 The Volume-Integral Equations for Plane-Layered Media . . . . . . . . . . . . 35
3.1 Transformation into the Spatial-Domain .. . . . . . .. . . . . . . . . . . . . . . . . . . . 35
3.2 The Electric Differential Volume-Integral Equation .. . . . . . . . . . . . . . . 36
3.3 Vector Form of the Integral Equation . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 39
3.4 The Volume-Integral Equations in Terms of Amperian Currents . . 41
4 Discretization via the Galerkin Method of Moments . . . . . . . . . . . . . . . . . . . 45
4.1 Expansion of the Anomalous Currents .. . . . . . . . .. . . . . . . . . . . . . . . . . . . . 45
4.2 Testing the Integral Equations . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 49
4.3 Solution via Conjugate Gradients . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 51
4.3.1 Efficient Computation of Convolutions and
Correlations . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 51
4.3.2 The Conjugate-Gradient Algorithm . . .. . . . . . . . . . . . . . . . . . . . 53
4.4 Comments and Conclusions . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 54
A.1 Two Theorems of Vector Analysis . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 54

xi
xii Contents

5 Computing Network Immittance Functions from Field

Calculations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 57
5.1 The Classical Bistatic Arrangement .. . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 58
5.2 The Differential Probe . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 60
5.3 Impedance of Ferrite-Core Probes.. . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 62
5.4 Computation of Impedance Changes due to the
Presence of a Flaw . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 63
A.1 Calculation of Circular Coil Impedance.. . . . . . . .. . . . . . . . . . . . . . . . . . . . 64
A.2 A Coupled-Circuit Model of the Volume-Integral Equation . . . . . . . 66
A.2.1 Frequency-Response Loci .. . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 68
6 Advanced Probe Models Based on Magnetic Dipoles
and Ferrite Cores . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 71
6.1 Modeling Nonstandard Probes.. . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 71
6.2 The Incident Field Due to a Ferrite-Core Probe . . . . . . . . . . . . . . . . . . . . 71
6.3 A Solenoidal Current Model .. . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 73
6.4 Response of a Rectangular Coil When Rotated
Relative to a Crack . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 75
6.4.1 The Anomalous (Scattering) Currents .. . . . . . . . . . . . . . . . . . . . 75
6.5 Validation via Benchmark Experiments .. . . . . . . .. . . . . . . . . . . . . . . . . . . . 80
6.6 Modeling a Differential-Receive Ferrite-Core Probe .. . . . . . . . . . . . . . 81
6.6.1 Sizing Surface Cracks: Experimental Test Cases . . . . . . . . . 82
6.6.2 An Interpolation Algorithm and Inverse Problem.. . . . . . . . 85
6.7 Validation Test: Tangent Coil Over a Crack in a Thin Plate. . . . . . . . 87
A.1 Analysis of a Bridge Circuit for a Split-D Probe . . . . . . . . . . . . . . . . . . . 89
7 Advanced Probe Models Based on Electric Dipoles . . . . . . . . . . . . . . . . . . . . 93
7.1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 93
7.2 Development of the Electric-Dipole Mathematical Model .. . . . . . . . 93
7.3 An Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 97
7.4 Another Example .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 98
7.5 Initial Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 98
7.6 Results With a Grid Of 16×16×4 Cells . . . . . . . .. . . . . . . . . . . . . . . . . . . . 104
7.7 Computation of the Divergence of the Currents . . . . . . . . . . . . . . . . . . . . 106
7.8 Three-Turn Spiral Coil . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 109
7.9 A Two-Layered Spiral Coil . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 110
7.10 A Model of a THz Transmitter on GaAs . . . . . . . .. . . . . . . . . . . . . . . . . . . . 111
8 Planar and Conforming Arrays of Probes .. . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 129
8.1 Modeling a Circular Array .. . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 129
8.1.1 Modeling With the T1R4 Configuration .. . . . . . . . . . . . . . . . . . 131
8.1.2 Modeling with the T1R4–T2R3 Configuration . . . . . . . . . . . 134
8.2 Modeling a Planar T/R Array .. . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 135
8.3 N-Port Theory of the T/R Array .. . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 135
8.3.1 Two-Port Parameter Relations . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 138
8.3.2 Calculation of Freespace Chain Matrix at 1 MHz .. . . . . . . . 139
Contents xiii

8.3.3 Application to Measured Data at 1 MHz . . . . . . . . . . . . . . . . . . 140

8.3.4 Application of Two-Port Theory to Measured Data . . . . . . 141
8.3.5 Modeling N-Port Array on an Aluminum
Host: σAl = 2.801×107 S/m . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 143
9 Multilayered Media with Cylindrical Geometries . .. . . . . . . . . . . . . . . . . . . . 145
9.1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 145
9.2 Some Typical Problems in Steam Generator Tubing . . . . . . . . . . . . . . . 145
9.3 Coupled Ferromagnetic Integral Equations . . . . .. . . . . . . . . . . . . . . . . . . . 145
9.4 Discretization: Method of Moments.. . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 148
9.5 Calculation of Matrices . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 151
9.5.1 Bobbin Coil Incident Field and Moments . . . . . . . . . . . . . . . . . 159
9.5.2 Driving-Point Impedance of a Bobbin Coil . . . . . . . . . . . . . . . 161
A.1 Cylindrical Eigenvectors and the Green Function . . . . . . . . . . . . . . . . . . 162
A.1.1 Application to Axisymmetric Systems . . . . . . . . . . . . . . . . . . . . 167
10 Some Special Topics in Computational Electromagnetics .. . . . . . . . . . . . 177
10.1 Spatial Decomposition Algorithm .. . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 177
10.1.1 Summary of Discrete Equations With Transfer
Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 182
10.1.2 Application: Computing the Scattered Fields
in the Vicinity of a Probe. . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 183
10.1.3 Test Results of Field Calculation in Probe Region .. . . . . . . 184
10.2 A Theory of Error Extrapolation . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 190
10.2.1 Application to a Ferrite Core Probe.. . .. . . . . . . . . . . . . . . . . . . . 195
10.2.2 Application to Tilted Notch Benchmark Problems . . . . . . . 197
10.3 Determining Grid Resolution Requirements .. . .. . . . . . . . . . . . . . . . . . . . 202
10.3.1 The Volume-Integral Equation .. . . . . . . .. . . . . . . . . . . . . . . . . . . . 202
10.3.2 Review of Fourier Analysis . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 203
10.3.3 Application to the 20◦ Notch Problem . . . . . . . . . . . . . . . . . . . . 204
10.3.4 The Electromagnetic Scene . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 206
10.4 Modeling Multiscale Problems . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 209
10.4.1 The Multiscale Algorithm .. . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 209
10.5 Notch at a Bolt Hole Benchmark Problems .. . . .. . . . . . . . . . . . . . . . . . . . 213
10.5.1 Sketch of Benchmark Problems 1 and 2 .. . . . . . . . . . . . . . . . . . 214
10.5.2 One-Port Circuit Models . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 215
10.5.3 Model-Based Inversion with NLSE. . . .. . . . . . . . . . . . . . . . . . . . 219

Part II Inversion Algorithms and Signal-Processing

11 Examples of Basic Inverse Problems . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 227

11.1 Overview of the General Inversion Process
for Parameter Estimation.. . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 227
11.2 Thickness Measurements with Eddy-Current Probes.. . . . . . . . . . . . . . 231
11.2.1 Test Case 1 .. . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 234
11.2.2 Noise and Uncertainties in Δt .. . . . . . . . .. . . . . . . . . . . . . . . . . . . . 235
xiv Contents

11.2.3 Test Case 2 .. . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 237

11.2.4 Conclusions .. . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 238
11.3 Conductivity Profile Measurements . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 239
11.4 Conductivity and Permeability Profile Measurements . . . . . . . . . . . . . 242
11.4.1 Multifrequency Reconstruction
of a Two-Layered Ferromagnetic System . . . . . . . . . . . . . . . . . 245
11.5 A Problem in the Semiconductor Chip Industry .. . . . . . . . . . . . . . . . . . . 245
11.5.1 Statement of the Problem . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 246
11.5.2 Method of Approach . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 247
11.6 Noise Analysis: The Covariance Matrix . . . . . . . .. . . . . . . . . . . . . . . . . . . . 248
11.7 Noise Analysis: Monte Carlo Method .. . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 254
A.1 A Generalized Error Analysis . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 255
12 NLSE: Parameter-Based Inversion Algorithm . . . . .. . . . . . . . . . . . . . . . . . . . 257
12.1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 257
12.2 NLSE: Nonlinear Least-Squares Parameter Estimation .. . . . . . . . . . . 257
12.2.1 Overview of the Algorithm: Nonlinear Least-Squares .. . . 257
12.2.2 Stochastic Methods for Global Optimization . . . . . . . . . . . . . 261
12.2.3 Computation of Function Values. . . . . . .. . . . . . . . . . . . . . . . . . . . 261
12.2.4 Application of Statistical Communication Theory .. . . . . . . 262
A.1 Cramer–Rao Lower Bound . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 263
A.1.1 Inverse Method Quality Metrics . . . . . . .. . . . . . . . . . . . . . . . . . . . 264
A.1.2 Optimizing Layer Estimation Using Metrics .. . . . . . . . . . . . . 265
A.1.3 Two Examples from Chap. 11 . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 267
A.2 Selected Bibliography of Inverse Problems
in Eddy-Current NDE .. . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 270
13 Robust Statistical Estimators.. . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 273
13.1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 273
13.2 Robust Estimators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 273
13.3 Least Median of Squares Estimator . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 274
13.4 Scale (S) Estimator .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 274
13.5 An Application of Robust Estimation . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 277
14 Some Special Signal-Processing Algorithms . . . . . . . .. . . . . . . . . . . . . . . . . . . . 281
14.1 Projection Onto Convex Sets . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 281
14.2 Kaczmarz’ Algorithm and the Algebraic
Reconstruction Technique . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 282
14.3 Analytic Continuation with Constraints .. . . . . . . .. . . . . . . . . . . . . . . . . . . . 285
14.4 Reconstructing Network Functions.. . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 287
15 Preprocessing Data and Transformation of Signal Vectors . . . . . . . . . . . 289
15.1 Clutter Modeling and Rejection . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 289
15.1.1 A Benchmark Validation Experiment... . . . . . . . . . . . . . . . . . . . 289
15.1.2 A Ligatured Outer-Diameter Slot at 200 kHz . . . . . . . . . . . . . 292
15.1.3 An Automated Clutter Removal Algorithm . . . . . . . . . . . . . . . 294
15.2 Transformations of Signal Vectors . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 297
Contents xv

15.3 Further Developments and Applications

of Scaling and Transformations . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 298
15.3.1 Modeling Differential-Bobbin Probes .. . . . . . . . . . . . . . . . . . . . 298
15.3.2 A General Virtual Probe Model . . . . . . . .. . . . . . . . . . . . . . . . . . . . 298
15.3.3 Another Virtual Probe Model . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 299
15.3.4 An Example .. . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 302
15.4 A General Transformation Matrix .. . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 303
15.4.1 Application to Probe Characterization.. . . . . . . . . . . . . . . . . . . . 305

Part III Applications

16 Modeling Corrosion and Pitting Problems . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 309

16.1 Introduction and Overview of Approach .. . . . . . .. . . . . . . . . . . . . . . . . . . . 309
16.2 Modeling the Corrosion Topology Problem . . . .. . . . . . . . . . . . . . . . . . . . 310
16.2.1 A Simplified Probe Model . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 310
16.2.2 Data Transformation .. . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 311
16.3 The Inverse Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 312
16.3.1 The Interpolator for NLSE . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 313
16.3.2 Results at 2,200 Hz . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 313
16.3.3 Results at 1,300 Hz . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 316
16.3.4 Frequency Mixing . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 319
16.3.5 Inverting Only Reactance Data . . . . . . . .. . . . . . . . . . . . . . . . . . . . 320
16.4 A New Flaw Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 321
16.5 Low-Frequency Models and Experiments . . . . . .. . . . . . . . . . . . . . . . . . . . 322
16.5.1 Inversion of Measured Data . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 325
16.5.2 Comments and Conclusions.. . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 330
16.5.3 Applying the POCS Algorithm . . . . . . . .. . . . . . . . . . . . . . . . . . . . 331
16.6 Results of a 2D Circumferential Sweep Feature
Extraction Algorithm.. . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 332
16.6.1 Application of the POCS Algorithm .. .. . . . . . . . . . . . . . . . . . . . 333
16.6.2 Validation with Real Corrosion Pits . . .. . . . . . . . . . . . . . . . . . . . 334
17 Applications to Aerospace Structures . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 337
17.1 Inspection of Fastener Sites in Aircraft Structures . . . . . . . . . . . . . . . . . 337
17.2 The Cessna Sandwich .. . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 337
17.2.1 Comparison with Finite-Element Method Results . . . . . . . . 340
17.3 Simulated Studies on Crack Characteristics:
The Cessna Setups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 342
17.3.1 Analysis of the Ring Probe Circuit . . . .. . . . . . . . . . . . . . . . . . . . 343
17.3.2 Model Results. . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 345
17.4 Further Tests with the SDA and “Weak-Host” Layers.. . . . . . . . . . . . . 347
17.4.1 Half-Sandwich No. 1 . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 347
17.4.2 Half-Sandwich No. 2 . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 348
17.4.3 Half-Sandwich No. 3 . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 350
17.5 Comments and Conclusions . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 350
xvi Contents

18 Applications to Nuclear Power .. . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 353

18.1 Modeling Pitting and Corrosion Phenomena
in Heat-Exchanger Tubes . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 353
18.1.1 An Ellipsoidal Model for Pits . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 353
18.1.2 Computing Impedance Signatures with VIC-3D R
....... 355
18.1.3 The Inversion Problem . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 357
18.1.4 Inversion of a Complex Pit . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 357
18.1.5 A Multifrequency Benchmark Inversion Test . . . . . . . . . . . . . 359
18.1.6 Modeling Wall-Thinning Effects . . . . . .. . . . . . . . . . . . . . . . . . . . 360
18.2 Model-Based Inversion of Measured Impedance Data:
The Carderock Problem .. . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 362
18.3 Model-Based Inversion of Measured Instrument Data:
The EPRI ID Pits Benchmark Test . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 367
18.3.1 Scaling the Measured Instrument Data . . . . . . . . . . . . . . . . . . . . 369
18.3.2 Feature Extraction with Clutter Removal
and Scan-Step Correction . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 370
18.3.3 Summary of Multifrequency Results . .. . . . . . . . . . . . . . . . . . . . 372
18.3.4 The Morphology of “Round-Bottom” Pits . . . . . . . . . . . . . . . . 377
A.1 Modeling Probes + Cables . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 379
A.2 A “Layered” Model of Corrosion Pits. . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 382
19 Coupled Problems in Heat-Exchanger Tubes . . . . . . .. . . . . . . . . . . . . . . . . . . . 383
19.1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 383
19.2 Reconstruction of a Semiellipsoidal Wear Scar
and Permeable Crust in a Heat-Exchanger Tube.. . . . . . . . . . . . . . . . . . . 383
19.3 Modeling Tube-Support Rings (TSR) .. . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 385
19.3.1 Statement of the Problem . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 385
19.3.2 Measured Data . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 386
19.3.3 Model Data Before β -Scaling . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 387
19.3.4 Model Data After β -Scaling . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 388
19.3.5 Scaling in the Opposite Direction.. . . . .. . . . . . . . . . . . . . . . . . . . 388
19.3.6 An Inverse Problem.. . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 389
19.4 Modeling Direct and Inverse Problems in Ferritic Tubes . . . . . . . . . . 391
19.4.1 The Model Problem . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 391
19.4.2 A Coupled-Circuit Model . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 393
19.4.3 Characterizing the Tube: Outer-Coil Experiments .. . . . . . . 394
19.4.4 Characterizing the Tube: Inner-Coil Experiments . . . . . . . . 398
19.4.5 Characterizing the Composite Structure .. . . . . . . . . . . . . . . . . . 399
19.4.6 An Improved Low-Frequency Model ... . . . . . . . . . . . . . . . . . . . 403
19.4.7 Comments and Conclusions.. . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 405
A.1 Inverse Method Quality Metrics for the Ferritic Tube .. . . . . . . . . . . . . 407
20 Applications to NDE of Coatings . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 411
20.1 Assessing Thermal Barrier Coatings . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 411
20.1.1 Inversion of Impedance Data. . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 412
Contents xvii

20.1.2 Sample Calculation: White-5 Top . . . . .. . . . . . . . . . . . . . . . . . . . 415

20.1.3 Application of β -Scaling to the Thermal
Barrier Coating Problem . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 417
21 Model-Assisted Probability of Detection .. . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 421
21.1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 421
21.2 MAPOD Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 422
21.3 Case Study for MAPOD Evaluation of NDI of Fastener
Sites for Sub-surface Fatigue Cracks . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 424
21.4 Bayesian Methods for Estimating Uncertain
Parameters in POD/MAPOD Evaluation .. . . . . . .. . . . . . . . . . . . . . . . . . . . 426
21.5 Model-Assisted Approach for Evaluating Localization
and Characterization Capability of NDE Techniques . . . . . . . . . . . . . . 432

References .. .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 435

Index . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 443
 Part I
Computational Electromagnetics
Background
Chapter 1
Overview of Methods of Computational
Electromagnetics

1.1 Introduction

Nondestructive evaluation (NDE) is to materials and structures what CAT scanning

is to the human body—an attempt to look inside without opening up the body.
As in CAT scanning, modern NDE requires sophisticated mathematical software
to perform its function. This is especially true with regard to quantitative NDE,
wherein we attempt to quantify defects, that is, determine their size, location, even
shape, rather than just to detect their presence. Low-frequency electromagnetic
methods using eddy-currents are a traditional mode of doing NDE (approximately
35 % of NDE uses eddy-currents, depending upon the specific application), but the
technology still suffers from a lack of algorithms and software to allow its full
potential to be realized. The electromagnetics code, VIC-3D R
, was developed to

R
alleviate that problem, and in this book we apply VIC-3D to solve forward and
inverse problems in eddy-current NDE.
To put eddy-current NDE into historical perspective, we quote Burrows [1]
Eddy-current testing first received systematic study during World War II in Germany.
Its wider recognition, however, did not come until 1952. In that year Förster and his
associates published the results of their extensive theoretical, experimental and industrial
investigations in the first of an important series of papers [2, 3].
These papers did not, however, include a quantitative theory for flaw detection, and apart
from some useful English language accounts of the Förster papers. . ..no further theoretical
work on flaw detection has appeared.

Burrows thesis, then, filled the gap in quantitative theoretical works on

eddy-current NDE, but even then, his theory of flaw detection was restricted to
flaws that were small compared to the skin depth of the material. This depth
(or, equivalently, diffusion length) is the quantity which in conducting matter
corresponds to wavelength in free space. By assuming that the flaw is very small,
Burrows was able to represent its electrical behavior by an equivalent electrical
dipole. His theory, though important in establishing the preeminence of Maxwell’s

H.A. Sabbagh et al., Computational Electromagnetics and Model-Based Inversion, 3

Scientific Computation, DOI 10.1007/978-1-4419-8429-6 1,
© Springer Science+Business Media New York 2013
4 1 Overview of Methods of Computational Electromagnetics

equations in eddy-current NDE, is not sufficient to solve contemporary problems,

wherein we must compute the response of eddy-currents to flaws of arbitrary size
and shape.
That’s where modern computational electromagnetics comes in. Though
researchers had been solving problems in electromagnetics since the time of
Maxwell, it has only been since the mid-1960s that computational electromagnetics
has evolved to be a distinct discipline. Research in this area was spurred largely by
the US Department of Defense to solve practical problems in the radiation, scat-
tering, and propagation of electromagnetic fields. Other areas were not neglected,
however; considerable effort has been expended since roughly the mid-1970s in
the application of microwaves to treat certain cancers, and this requires accurate
models to compute the interaction of electromagnetic fields with biological media.
The techniques for solving these problems range from formulations involving
integral equations to those involving differential equations.
It is not surprising, therefore, that eddy-current NDE should benefit from research
into computational electromagnetics, and VIC-3D R
is the natural outgrowth of
that research. It is based on the volume-integral mathematical approach to solving
electromagnetics problems. This approach enables VIC-3D R
to solve problems
more efficiently on smaller computers, and with less preprocessing time, than other
approaches, and, more importantly, it allows one to compute the response of flaws
of arbitrary size and shape.

1.2 Oak Ridge National Laboratory

Following Burrows, work at the Oak Ridge National Laboratory included efforts
in applying rigorous electromagnetic theory and developing computer programs
to solve problems in eddy-current NDE, thereby bringing the discipline into the
computer age. The primary authors of many of these papers were C.V. Dodd and
W.E. Deeds. See [4–8].

1.3 Method of Moments

As a prelude to the work on computational electromagnetics that evolved in the

1970s, we should mention the pioneering and seminal work of R.F. Harrington
on the method of moments (MOM) [12]. This book spurred the development and
application of the modern theory of integral equations to electromagnetic fields, and
was extremely influential in the development of VIC-3D R
, as will be described in
this book.
1.5 Other Approaches to Computational Electromagnetics in NDE 5

1.4 Model-Based Parameter Estimation in Electromagnetics

In the 1990s, E.K. Miller developed an analysis technique called model-based

parameter estimation (MBPE) (see [9] and the references therein), whose purpose
is to substitute a system approach for solving electromagnetics problems instead
of the actual solution from first principle models. Thus, MBPE presents a sort of
“surrogate” for Maxwell’s equations. We will see a simple form of this when we
apply interpolation tables to speed up the calculations for inverse problems in Parts
II and III of this book.

1.5 Other Approaches to Computational Electromagnetics

in NDE

Closely related to the volume-integral approach is the boundary-integral method,

in which the boundary of the region of interest is discretized, rather than the entire
volume. The theory of boundary-integrals, and their relation to the volume-integral
approach in eddy-current NDE, is given in [10]. Recent developments in the theory
and application of the boundary-integral approach to NDE are given in [11].
The method of finite-elements is also used for solving eddy-current NDE
problems [104, 117]. We give an example of it in Chap. 17 [“Applications to
Aerospace Structures”] in connection with the Cessna Sandwich (see Fig. 17.4).
The finite-element method is a “field-solver,” in which the unknowns are the
electromagnetic field components, whereas integral equations, whether volume or
boundary, have sources as their unknowns. Fields extend to infinity, so a field-solver
must have a method for truncating the problem region by “approximating the field
at infinity.” This is unnecessary, of course, in integral equations in which the sources
lie within a compact domain of space.
The key to developing a computationally efficient theory of integral equations is
computing a Green’s function, which is the starting point of this book.
Chapter 2
Green’s Dyad for Plane-Layered Media

The Green’s dyad, which is the electric-field response to a delta-function vector

current source, plays a principal role in volume-integral equations, as we shall see
later. In this chapter we develop the theory of the Green’s dyad for plane-parallel
layered media. In Chap. 9 we extend the development to multilayered media with
cylindrical geometry.

2.1 Eigenmodes of Anisotropic Media

We will consider plane-parallel bodies of infinite extent in the (x, y) plane, which are
made up of layers of homogeneous, anisotropic material. To be specific, we consider
magnetic host materials that are characterized by the following biaxial generalized
electrical permittivity matrix:
⎡ ⎤
εx εxy 0
ε h = ⎣ εyx εy 0 ⎦, (2.1)
0 0 εz

where the entries are generalized permittivities ε + σ / jω .

Maxwell’s equations for an electrically anisotropic body are
∇ × E = − jω μh H − jω (μ (r) − μh)H
= − j ω μh H + J m
∇ × H = jωε h · E + jω (ε (r) − ε h ) · E
= jωε h · E + Je, (2.2)

H.A. Sabbagh et al., Computational Electromagnetics and Model-Based Inversion, 7

Scientific Computation, DOI 10.1007/978-1-4419-8429-6 2,
© Springer Science+Business Media New York 2013
8 2 Green’s Dyad for Plane-Layered Media

where Jm and Je are anomalous magnetic and electric currents that account for the
presence of flaws, or anomalies, in the otherwise-uniform host material. From here
on we drop the subscript h on the generalized host permittivity and permeability.
Because of the material anisotropy, it is convenient to work with a matrix
formulation of these equations that has been useful in crystal optics, plasmas, and
microwave devices [13–21]. If the body is homogeneous with respect to (x, y), then
Maxwell’s equations can be Fourier transformed with respect to (x, y) and written
as the following four-vector matrix differential equation in the spectral domain:

∂z ẽ = S · ẽ + U · J̃, (2.3)
ky kx j ˜
Ẽz = H̃x − H̃y + Jez , (2.4)
εz ω εz ω εz ω
−ky kx j ˜
H̃z = Ẽx + Ẽy − Jmz , (2.5)
μω μω μω
where the tilde denotes a function defined in the transform domain (kx , ky ), and
⎡
⎤
J˜ex
⎡ ⎤ ⎢ ˜ ⎥
Ẽx ⎢ Jey ⎥
⎢ ⎥ ⎢ ⎥
⎢ Ẽy ⎥ ⎢ J˜ ⎥
⎢ ez ⎥
ẽ = ⎢ ⎥
⎢ H̃ ⎥; J̃ = ⎢ ⎥. (2.6)
⎣ x⎦ ⎢ J˜mx ⎥
⎢ ⎥
⎢ ˜ ⎥
H̃y ⎣ Jmy ⎦
J˜mz

The subscript e denotes an electric current and m denotes a magnetic current.

The matrices in (2.3) are given by
⎡ ⎤ ⎡ ⎤
0 0 a b 0 0 kx /ωεz 0 1 0
⎢0 0 c d⎥ ⎢ 0 0 ky /ωεz −1 0 0 ⎥
S = −⎢⎣α β 0 0⎦ ; U = ⎣
⎥ ⎢ ⎥.
0 1 0 0 0 −kx /ω μ ⎦
γ δ 0 0 −1 0 0 0 0 −ky /ω μ
(2.7)
The entries of S are given in terms of the entries of (2.1) by
j j
a= kx ky ; α = (− μεyx ω 2 − kx ky )
ωεz ωμ
j j
b= (μεz ω 2 − kx2 ); β = (− μεy ω 2 + kx2 )
ωεz ωμ
j j
c= (− μεz ω 2 + ky2 ); γ = (μεx ω 2 − ky2 )
ωεz ωμ
j j
d=− kx ky ; δ = (μεxy ω 2 + kx ky ). (2.8)
ωεz ωμ
2.1 Eigenmodes of Anisotropic Media 9

When J̃ is a surface current confined to z = z , i.e., J̃ = J̃s δ (z − z ), then

integration of (2.3) produces

ẽ(+) − ẽ(−) = U · J̃s, (2.9)

which is called the equation of discontinuity. The superscript (+) denotes the limit
z approaches z from above and the superscript (−) denotes the limit from below.
Equation (2.9) will be used in the next section to compute the Green’s dyad for a
layered workpiece.
Starting with these equations, Roberts [24] has developed a fairly complete
theory of normal modes of biaxial anisotropic media. This work is based on, and
extends, earlier work performed at Sabbagh Associates [22, 23, 25–29]. From here
on we specialize the theory developed in [24] to the case to be considered here,
in which the media involved are transversely isotropic to the z-coordinate. The
dielectric permittivity tensor, in its principal-axis coordinate system, then takes the
form
⎡ ⎤
εt 0 0
ε = ⎣ 0 εt 0 ⎦ . (2.10)
0 0 εz
The entries in S now become
j j
a= kx ky ; α = (−kx ky )
ωεz ωμ
j j
b= (μεz ω 2 − kx2 ); β = (− μεt ω 2 + kx2 )
ωεz ωμ
j j
c= (− μεz ω 2 + ky2 ); γ = (μεt ω 2 − ky2 )
ωεz ωμ
j j
d=− kx ky ; δ = (kx ky ). (2.11)
ωεz ωμ

Let’s introduce some notation: kx2 + ky2 = kt2 , ω 2 μεt = Ωt2 , ω 2 μεz = Ωz2 , ε =
εt /εz . Then the eigenvalues of S are
 √
λ1 = kt2 − Ωt2 λ2 = −λ1 λ3 = ε kt2 − Ωz2 λ4 = −λ3 . (2.12)

The linearly-independent eigenvectors that correspond to these eigenvalues are:

⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤
− j ω μ0 ky − j ω μ0 ky λ 3 kx λ 3 kx
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ j ω μ0 kx ⎥ ⎢ j ω μ0 kx ⎥ ⎢ λ 3 ky ⎥ ⎢ λ 3 ky ⎥
v1 = ⎢
⎢ λ k
⎥ v2 = ⎢
⎥ ⎢ −λ k ⎥
⎥ v3 = ⎢
⎢ jωε k ⎥
⎥ v4 = ⎢ ⎥
⎢ − jωε k ⎥ .
⎣ 1 x ⎦ ⎣ 1 x ⎦ ⎣ t y ⎦ ⎣ t y⎦
λ 1 ky − λ 1 ky − jωεt kx jωεt kx
(2.13)
10 2 Green’s Dyad for Plane-Layered Media

When kx = ky = 0, the following are linearly-independent eigenvectors:

⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤
1 1
0 0

⎢ 0 ⎥ ⎢ 0 ⎥ ⎢ μ0 ⎥ ⎢ μ0 ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
v1 = ⎢ ⎥ v2 = ⎢ 0 ⎥ v3 = ⎢ εt ⎥ v4 = ⎢ − εt ⎥. (2.14)
⎢ 0

⎥ ⎢
⎥ ⎢ ⎥ ⎢ ⎥
⎣ εt ⎦ ⎣ εt ⎦ ⎣ 1 ⎦ ⎣ 1 ⎦
−
μ0 μ0 0 0

When we substitute v1 , v2 of (2.13) into (2.4), with the source currents set to
zero, we find that Ẽz = 0; hence, v1 , v2 are transverse electric (TE) modes, with
respect to z. Similarly, v3 , v4 are transverse magnetic (TM) modes. Note that the
TE modes are orthogonal to the TM modes. This will facilitate the computation
of the Green’s dyadic. v1 and v3 are downward-traveling waves in the z-direction;
i.e., they represent waves that travel in the negative z-direction. v2 , v4 are upward-
traveling waves (in the positive z-direction). We see from (2.4) and (2.5) that all
modes are TEM (transverse electric and magnetic) with respect to z for kx = ky = 0.

2.2 Green Dyad for Plane-Parallel Layered Media

Though the theory of eigenmodes that was developed in the preceding section
is applicable to the class of transverse-isotropic media, we will apply it only to
isotropic media. Hence, εt = εz = ε in (2.10), which means that λ1 = λ3 in (2.13).
The flaw, or other anomaly, in a workpiece produces an anomalous current which
is to be determined as the solution of a volume-integral equation. The kernel of this
equation is a Green dyad for a plane-parallel layered workpiece, and we turn our
attention to determining this dyad.
Consider the system shown in Fig. 2.1. The point-source of electric or magnetic
current is in region 0, and we want to compute the fields in this region, or in any
other region. This will give us the Green dyad. If the source is an electric current, and
the field electric, then the resulting dyad is called electric–electric. If the source is
a magnetic current, and the field electric, then the dyad is called electric–magnetic,
and so-on.
We write for the system of four-vectors in the ith region
⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤
− j ω μi ky − j ω μi ky λ i kx λ i kx
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ j ω μi kx ⎥ ⎢ j ω μi kx ⎥ ⎢ λ i ky ⎥ ⎢ λ i ky ⎥
v1i = ⎢⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
λ ⎥ v2i = ⎢ −λ k ⎥ v3i = ⎢ jω ε̂ k ⎥ v4i = ⎢ − jω ε̂ k ⎥ .
⎣ k
i x ⎦ ⎣ i x ⎦ ⎣ i y ⎦ ⎣ i y⎦
λ i ky − λ i ky − jω ε̂i kx jω ε̂i kx
(2.15)
2.2 Green Dyad for Plane-Parallel Layered Media 11

Z
(N)
ε ,μ
Ν Ν z
N−1

z
2
ε ,μ (2)
2 2 z
1
ε ,μ (1)
1 1 z
0
ε ,μ z (0)
0 0
z
−1
ε ,μ (−1)
−1 −1
z
−2

z
−M+1
ε ,μ (−M+1)
−Μ+1 −Μ+1
z
−M
^
ε ,μ
−M −M (−M)

Fig. 2.1 Plane-parallel layered workpiece. The source is in region 0

2.2.1 Infinite-Space Green Dyad

First of all, we will compute the infinite-space Green dyad, which is the dyad
associated with an infinite, uniform, unlayered medium. This corresponds to the
situation wherein region 0 of Fig. 2.1 extends to z = ±∞; i.e., z0 = ∞, and z−1 = −∞.
Figure 2.2 shows a source located at z = z in region 0, together with the
appropriate field eigenvectors on each side of z . This choice of the eigenvectors is
consistent with the fact that v1 and v3 travel in the negative z-direction, and v2 and
v4 are positively traveling waves. The superscript, (inc), denotes “incident” fields
due to the source.
If the source is an electric current vector, then the right-hand side of (2.9) consists
of the three electric excitation vectors:
⎡ ⎤ ⎡ ⎤ ⎡ ˜ ⎤
0 0 kx Jez /ω ε̂0
⎢ 0 ⎥ ⎢ ⎥ ⎢ ⎥
(e) ⎢ ⎥ (e) ⎢ 0 ⎥ (e) ⎢ ky J˜ez /ω ε̂0 ⎥
fx = ⎢ ⎥ fy = ⎢ ⎥ fz = ⎢ ⎥, (2.16)
⎣ 0 ⎦ ⎣ J˜ey ⎦ ⎣ 0 ⎦
−J˜ex 0 0
12 2 Green’s Dyad for Plane-Layered Media

(+)
Z
(inc) - λ 0z (inc) - λ0 z
b V20 e + d V e
40

Region 0
z

λ z λ z
V 10 e 0 + c e 0
(inc) (inc)
a V
30

(-)

Fig. 2.2 Showing the eigenvectors used to calculate the infinite-space Green dyad. The source is
located at z = z

and if the source is a magnetic current, then the right-hand side consists of the
magnetic excitation vectors:
⎡ ⎤ ⎡ ˜ ⎤ ⎡ ⎤
0 Jmy 0
⎢ −J˜ ⎥ ⎢ ⎥ ⎢ ⎥
(m) ⎢ mx ⎥ (m) ⎢ 0 ⎥ (m) ⎢ 0 ⎥
fx = ⎢ ⎥ fy = ⎢ ⎥ fz = ⎢ ⎥. (2.17)
⎣ 0 ⎦ ⎣ 0 ⎦ ⎣ −kx J˜mz /ω μ0 ⎦
0 0 −ky J˜mz /ω μ0

Equation (2.9) becomes therefore

   
− a(inc) v10 eλ0 z + b(inc) v20 e−λ0 z − c(inc) v30 eλ0 z + d (inc) v40 e−λ0 z = f , (2.18)

where f denotes one of the six excitation vectors in (2.16) or (2.17). We will return
to this shortly.
Take the dot product of (2.18) with respect to the TE-mode vectors, v10 , v20 ,
and get
− A(i)α11 + B(i)α12 = F1
−A(i) α12 + B(i)α22 = F2 , (2.19)
where we have used the orthogonality of the TE and TM modal vectors to eliminate
c(inc) and d (inc) . Here

A(i) = a(inc) eλ0 z

B(i) = b(inc) e−λ0 z
F1 = v10 · f
2.2 Green Dyad for Plane-Parallel Layered Media 13

F2 = v20 · f
α11 = v10 · v10 = kt2 (λ02 − ω 2 μ02 )
α12 = v10 · v20 = −kt2 (λ02 + ω 2 μ02 )
α22 = v20 · v20 = kt2 (λ02 − ω 2 μ02 ). (2.20)

Upon solving (2.19) for A(i) , B(i) , and then using (2.20) we get

F1 α22 − F2α12 −λ0 z

a(inc) = e
−α11 α22 + α12
2

−F2 α11 + F1 α12 λ0 z

b(inc) = e . (2.21)
−α11 α22 + α122

Hence, the contribution of the TE-modes to the infinite-space Green dyad is

−F2 α11 + F1α12 

v20 e−λ0 (z−z ) , z ≥ z
−α11 α22 + α12
2

F1 α22 − F2α12 
v10 eλ0 (z−z ) , z ≤ z . (2.22)
−α11 α22 + α12
2

To compute the TM-mode contribution, take the dot product of (2.18) with
respect to v30 , v40 . Proceeding as before, we get for the TM-mode contribution

F4 β33 − F3 β34 
v40 e−λ0 (z−z ) , z ≥ z
β33 β44 − β34
2

F3 β44 − F4β34 
− v30 eλ0 (z−z ) , z ≤ z , (2.23)
β33 β44 − β34
2

where
F3 = v30 · f

F4 = v40 · f

β33 = v30 · v30 = kt2 (λ02 − ω 2ε̂02 )

β34 = v30 · v40 = kt2 (λ02 + ω 2ε̂02 )

β44 = v40 · v40 = kt2 (λ02 − ω 2ε̂02 ) . (2.24)

14 2 Green’s Dyad for Plane-Layered Media

Therefore, upon combining (2.22) and (2.23) we get for the composite infinite-
field Green dyad:
 
−F2 α11 + F1 α12 F4 β33 − F3 β34 
v20 + v40 e−λ0 (z−z ) , z ≥ z
−α11 α22 + α122 β33 β44 − β34
2

 
F1 α22 − F2α12 F3 β44 − F4β34 
v −
2 10
v30 eλ0 (z−z ) , z ≤ z . (2.25)
−α11 α22 + α12 β33 β44 − β34
2

The Fs that appear in (2.25) are defined in (2.20) and (2.24). They are computed
by taking the dot product of the eigenvectors of (2.15) (with i = 0) with the
excitation vectors in (2.16) and (2.17). For example, the value of F1 that is associated
(ex) (e)
with electric excitation in the x-direction is given by F1 = v10 · fx = −λ0 ky J˜ex .
Continuing in this manner, we find:

(ex) (mx)
F1 = −λ0 ky J˜ex , F1 = − jω μ0 kx J˜mx
(ex) (mx)
F2 = λ0 ky J˜ex , F2 = − jω μ0 kx J˜mx
(ex) (mx)
F3 = jω ε̂0 kx J˜ex , F3 = −λ0 ky J˜mx
(ex) (mx)
F4 = − jω ε̂0 kx J˜ex , F4 = −λ0 ky J˜mx
(ey) (my)
F1 = λ0 kx J˜ey , F1 = − jω μ0 ky J˜my
(ey) (my)
F2 = −λ0 kx J˜ey , F2 = − jω μ0 ky J˜my
(ey) (my)
F3 = jω ε̂0 ky J˜ey , F3 = λ0 kx J˜my
(ey) (my)
F4 = − jω ε̂0 ky J˜ey , F4 = λ0 kx J˜my

(ez) (mz) λ0 kt2 ˜

F1 = 0, F1 =− Jmz
ω μ0
(ez) (mz) λ0 kt2 ˜
F2 = 0, F2 = Jmz
ω μ0
(ez) λ0 kt2 ˜ (mz)
F3 = Jez , F3 = 0
ω ε̂0
(ez) λ0 kt2 ˜ (mz)
F4 = Jez , F4 = 0 . (2.26)
ω ε̂0

When these are combined with the α ’s and β ’s of (2.20) and (2.24), we get for
the infinite-space Green dyad
2.2 Green Dyad for Plane-Parallel Layered Media 15


[J2 v20 + J4 v40 ] e−λ0 (z−z ) , z ≥ z

[J1 v10 + J3 v30 ] eλ0 (z−z ) , z ≤ z , (2.27)

where
J1 J2 J3 J4

ky J˜ex ky J˜ex jkx J˜ex jkx J˜ex

(ex) :
2λ0 kt2 2λ0 kt2 2ω ε̂0 kt2 2ω ε̂0 kt2

kx J˜ey kx J˜ey jky J˜ey jky J˜ey

(ey) : − −
2λ0kt 2 2λ0kt2 2ω ε̂0 kt2 2ω ε̂0 kt2

J˜ez J˜ez
(ez) : 0 0 −
2λ0ω ε̂0 2λ0 ω ε̂0

jkx J˜mx jkx J˜mx ky J˜mx ky J˜mx

(mx) : − −
2kt2 ω μ0 2kt2 ω μ0 2λ0 kt2 2λ0 kt2

jky J˜my jky J˜my kx J˜my kx J˜my

(my) : − −
2kt ω μ0 2kt ω μ0
2 2 2λ0kt2 2λ0 kt2

J˜mz J˜mz
(mz) : 0 0. (2.28)
2 λ 0 ω μ0 2 λ 0 ω μ0

When we introduce the vectors of (2.15) into (2.27), and then use (2.28), we
obtain an expression for the x- and y-components of the field in terms of the
source components. The z-component of the fields are obtained from (2.4) and
(2.5). For instance, we will compute the x-component of the electric field, due to
the x-component of electric current:
⎧ 
⎪
⎪ − jω μ0 ky2 J˜ex jkx2 λ0 J˜ex −λ (z−z )
⎪
⎪ + e 0 , z ≥ z
⎪
⎪ 2 λ k 2 2 ω ε̂ k 2
⎨ 0 t 0 t

⎪  
⎪
⎪ − jω μ0 ky2 J˜ex jkx2 λ0 J˜ex λ (z−z )
⎪
⎪
⎪
⎩ + e 0 , z ≤ z
2λ0 kt2 2ω ε̂0 kt2

  
k2 e−λ0 |z−z | ˜
= − jω μ0 1 − x2 Jex . (2.29)
k0 2λ0
16 2 Green’s Dyad for Plane-Layered Media

The coefficient of J˜ex in (2.29) is the xx-component of the Fourier-domain,

(ee)
infinite-space, electric–electric Green dyad G̃(0) . The complete expression for the
dyad is
⎡ ⎤
1 − kx2/k02 −kx ky /k02 ± jkx λ0 /k02
(ee) ⎢ ⎥ e−λ0 |z−z |
G̃(0) (kx , ky ; z, z ) = − jω μ0 ⎣ −kx ky /k0 1 − ky /k0 ± jky λ0 /k0 ⎥
 ⎢ 2 2 2 2
⎦ 2λ0
± jkx λ0 /k02 ± jky λ0 /k02 1 + λ02/k02

δ (z − z )
+j az az , (2.30)
ω ε̂0
δ (z − z )
where (+) sign goes with z > z and (−) with z < z . The term j az az
ω ε̂0
is called the “depolarizing” term [34] and follows from the last term in (2.4).
A similar analysis holds for the magnetic–magnetic Green dyad and produces the
same expression as (2.30), except that μ0 is everywhere replaced by −ε̂0 .
The spectral-domain, infinite-space, magnetic–electric Green dyad is given by
⎡ ⎤
0 ±λ0 − jky 
(em) e−λ0 |z−z |
G̃(0) (kx , ky ; z, z ) = ⎣ ∓λ0 0 jkx ⎦ . (2.31)
2λ0
jky − jkx 0

It is not difficult to show that

(me) (em)
G̃(0) = G̃(0) . (2.32)

2.2.2 Layered-Space Green Dyad

The infinite-space dyad that we computed in the previous section serves as the
“incident” field in the layered medium of Fig. 2.1. We are interested in computing
the fields produced when the incident dyad is scattered from the layers.
As before, we focus our attention on region 0, but this time assume that it has a
finite upper-boundary, z0 , and lower-boundary, z−1 , as shown in Fig. 2.3. Region 1,
which lies above region 0, is a half-space, as is region -1, which lies below region 0.
The infinite-field Green dyad is shown in Fig. 2.3. Because of the singularity

of this dyad, we split it into that part, [J2 v20 + J4 v40 ]e−λ0 (z−z ) , which is valid for

z ≥ z , and [J1 v10 + J3 v30 ]eλ0 (z−z ) , which is valid for z ≤ z . The scattered field,
(a0 v10 + c0 v30 )eλ0 z + (b0 v20 + d0v40 )e−λ0 z , is continuous throughout region 0.
2.2 Green Dyad for Plane-Parallel Layered Media 17

(1)

- λ1 z - λ1 z
b V e + d V e
1 21 1 41
z
0
- λ0 (z-z ) - λ0 (z-z )
JV e + JV e
2 20 4 40

-λ z
a V e λ0 + b V e - λ0 + c V e λ0 + d V e 0
z z z (0)
z
0 10 0 20 0 30 0 40

λ (z-z )
J V e λ0 +JV e 0
(z-z )
1 10 3 30

z
-1
λ z
+ c V3,-1 e λ−1
z
a V e −1
-1 1,-1 -1

(-1)

Fig. 2.3 Showing the source, at z , in region 0, together with the incident and scattered fields.
Regions 1 and −1 are half-spaces

Continuity of the fields at z = z0 requires that

 
b1 v21 + d1 v41 − a0v10 − c0 v30 = b0 v20 + d0 v40 + J2 eλ0 z v20 + J4 eλ0 z v40 , (2.33)

where

b1 = b1 e−λ1 z0 , d1 = d1 e−λ1 z0 , a0 = a0 eλ0 z0 , b0 = b0 e−λ0 z0

. (2.34)
c0 = c0 eλ0 z0 , d0 = d0 e−λ0 z0 , J2 = J2 e−λ0 z0 , J4 = J4 e−λ0 z0

b1 , d1 , a0 , c0 are scattered fields at z0 and b0 , d0 , J2 , and J4 are incident upon z0 .
Continuity of the fields at the lower boundary, z = z−1 , requires that
 
a−1 v1,−1 + c−1 v3,−1 − b0v20 − d0 v40 = a0 v10 + c0 v30 + J1 e−λ0 z v10 + J3 e−λ0 z v30,
(2.35)
where now

a−1 = a−1 eλ−1 z−1 , c−1 = c−1 eλ−1 z−1 , a0 = a0 eλ0 z−1 , b0 = b0 e−λ0 z−1
. (2.36)
c0 = c0 eλ0 z−1 , d0 = d0 e−λ0 z−1 , J1 = J1 eλ0 z−1 , J3 = J3 eλ0 z−1

In this case, a−1 , c−1 , b0 , d0 are scattered fields at z−1 and c0 , a0 , J1 , and J3 are
incident upon z−1 .
18 2 Green’s Dyad for Plane-Layered Media

Equations (2.33) and (2.35) include TE- and TM-modes. They are easier to solve
if we separate these modes. Take the dot product of (2.33) with respect to the
TE-modes, v21 and v10 and get

b1 v21 · v21 − a0v10 · v21 = b0 v20 · v21 + J2 eλ0 z v20 · v21

b1 v21 · v10 − a0v10 · v10 = b0 v20 · v10 + J2 eλ0 z v20 · v10, (2.37)

where

v21 · v21 = kt2 (λ12 − ω 2 μ12 )

v10 · v21 = −kt2 (ω 2 μ0 μ1 + λ0 λ1 )
v10 · v10 = kt2 (λ02 − ω 2 μ02 )
v20 · v21 = kt2 (λ0 λ1 − ω 2 μ0 μ1 )
v20 · v10 = −kt2 (ω 2 μ02 + λ02 ) . (2.38)

The solution of (2.37) is

2μ0 λ0 e(λ1 −λ0 )z0 

b1 = (b0 + eλ0 z J2 )
μ0 λ 1 + μ1 λ 0
(μ1 λ0 − μ0λ1 )e−2λ0 z0 
a0 = (b0 + eλ0 z J2 ) . (2.39)
μ1 λ 0 + μ0 λ 1

Next, compute the TM-modes by taking the dot product of (2.33) with v41
and v30 :

d1 v41 · v41 − c0 v30 · v41 = d0 v40 · v41 + J4 eλ0 z v40 · v41

d1 v41 · v30 − c0 v30 · v30 = d0 v40 · v30 + J4 eλ0 z v40 · v30 , (2.40)

where

v41 · v41 = kt2 (λ12 − ω 2 ε̂12 )

v30 · v41 = kt2 (ω 2 ε̂0 ε̂1 + λ0λ1 )
v30 · v30 = kt2 (λ02 − ω 2 ε̂02 )
v40 · v41 = kt2 (λ0 λ1 − ω 2 ε̂0 ε̂1 )
v40 · v30 = kt2 (ω 2 ε̂02 + λ02 ) . (2.41)

The solution of (2.40) is

2.2 Green Dyad for Plane-Parallel Layered Media 19

2ε̂0 λ0 e(λ1 −λ0 )z0 

d1 = (d0 + J4 eλ0 z )
ε̂0 λ1 + ε̂1 λ0
(λ1 ε̂0 − λ0ε̂1 )e−2λ0 z0 
c0 = (d0 + J4 eλ0 z ) . (2.42)
λ1 ε̂0 + λ0 ε̂1

We repeat the analysis for (2.35). First, for the TE-modes:


a−1 v1,−1 · v1,−1 − b0 v1,−1 · v20 = (a0 + J1 e−λ0 z )v10 · v1,−1

a−1 v1,−1 · v20 − b0 v20 · v20 = (a0 + J1 e−λ0 z )v10 · v20 , (2.43)

where

v1,−1 · v1,−1 = kt2 (λ−1

2
− ω 2 μ−1
2
)
v20 · v1,−1 = −kt2 (ω 2 μ0 μ−1 + λ0λ−1 )
v20 · v20 = kt2 (λ02 − ω 2 μ02 )
v10 · v1,−1 = kt2 (λ0 λ−1 − ω 2 μ0 μ−1 )
v20 · v10 = −kt2 (ω 2 μ02 + λ02 ) . (2.44)

The solution of (2.43) is

2μ0 λ0 e(λ0 −λ−1 )z−1 

a−1 = (a0 + e−λ0z J1 )
μ0 λ−1 + μ−1λ0
(μ−1 λ0 − μ0 λ−1 )e2λ0 z−1 
b0 = (a0 + e−λ0z J1 ) . (2.45)
μ−1 λ0 + μ0 λ−1
The TM-mode equations

c−1 v3,−1 · v3,−1 − d0 v40 · v3,−1 = (c0 + J3 e−λ0 z )v30 · v3,−1

c−1 v40 · v3,−1 − d0 v40 · v40 = (c0 + J3 e−λ0 z )v30 · v40 , (2.46)

where

v3,−1 · v3,−1 = kt2 (λ−1

2
− ω 2 ε̂−1
2
)
v40 · v3,−1 = kt2 (ω 2 ε̂0 ε̂−1 + λ0 λ−1 )
v40 · v40 = kt2 (λ02 − ω 2 ε̂02 )
v30 · v3,−1 = kt2 (λ0 λ−1 − ω 2 ε̂0 ε̂−1 )
v40 · v30 = kt2 (ω 2 ε̂02 + λ02 ) , (2.47)
20 2 Green’s Dyad for Plane-Layered Media

have for their solutions

2ε̂0 λ0 e(λ0 −λ−1 )z−1 

c−1 = (c0 + J3 e−λ0 z )
ε̂0 λ−1 + ε̂−1 λ0
(ε̂0 λ−1 − ε̂−1 λ0 )e2λ0 z−1 
d0 = (c0 + J3 e−λ0 z ) . (2.48)
ε̂0 λ−1 + ε̂−1 λ0
We introduce the following transmission and reflection coefficients:

(E) 2 μ0 λ 0 (M) 2ε̂0 λ0

T1 = ;T =
μ0 λ 1 + μ1 λ 0 1 ε̂0 λ1 + ε̂1 λ0

(E) μ1 λ0 − μ0λ1 (M) ε̂0 λ1 − ε̂1 λ0

R1 = ;R =
μ1 λ 0 + μ0 λ 1 1 ε̂1 λ0 + ε̂0 λ1
.
(E) 2 μ0 λ 0 (M) 2ε̂0 λ0
T−1 = ;T =
μ0 λ−1 + μ−1λ0 −1 ε̂0 λ−1 + ε̂−1 λ0

(E) μ−1 λ0 − μ0 λ−1 (M) ε̂0 λ−1 − ε̂−1 λ0

R−1 = ;R = (2.49)
μ−1 λ0 + μ0 λ−1 −1 ε̂−1 λ0 + ε̂0 λ−1

The final expressions for the mode coefficients are given in terms of these scattering
parameters:
(E) (E)  (E) 
R1 R−1 e−λ0 (2T +z ) J1 + R1 e−λ0 (2z0 −z ) J2
a0 = (E) (E)
1 − R1 R−1 e−2λ0 T

(E)  (E) (E) 

R−1 eλ0 (2z−1 −z ) J1 + R1 R−1 e−λ0 (2T −z ) J2
b0 = (E) (E)
1 − R1 R−1 e−2λ0 T

(M) (M)  (M) 

R1 R−1 e−λ0 (2T +z ) J3 + R1 e−λ0 (2z0 −z ) J4
c0 = (M) (M)
1 − R1 R−1 e−2λ0 T

(M)  (M) (M) 

R−1 eλ0 (2z−1 −z ) J3 + R1 R−1 e−λ0 (2T −z ) J4
d0 = (M) (M)
1 − R1 R−1 e−2λ0 T

(E) (E)  (E) 

T1 R−1 e−λ0 (T −z−1 +z ) J1 + T1 e−λ0 (z0 −z ) J2
b1 = (E) (E)
e λ 1 z0
1 − R1 R−1 e−2λ0 T
2.2 Green Dyad for Plane-Parallel Layered Media 21

(M) (M) −λ0 (T −z−1 +z ) (M) 

T1 R−1 e J3 + T1 e−λ0 (z0 −z ) J4 λ1 z0
d1 = (M) (M)
e
1 − R1 R−1 e−2λ0 T

(E)  (E) (E) 

T−1 e−λ0 (z −z−1 ) J1 + T−1 R1 e−λ0 (T +z0 −z ) J2
a−1 = (E) (E)
e−λ−1 z−1
1 − R1 R−1 e−2λ0 T

(M)  (M) (M) 

T−1 e−λ0 (z −z−1 ) J3 + T−1 R1 e−λ0 (T +z0 −z ) J4
c−1 = (M) (M)
e−λ−1 z−1 , (2.50)
1 − R1 R−1 e−2λ0 T

where T = z0 − z−1 is the height of the source region.

In order to determine the scattered fields within the source region, we start with
the expression a0 v10 eλ0 z + b0 v20 e−λ0 z + c0 v30 eλ0 z + d0v40 e−λ0 z and use (2.15):

Ẽx = eλ0 z (− jω μ0 ky a0 + λ0kx c0 ) + e−λ0 z (− jω μ0 ky b0 + λ0kx d0 )

Ẽy = eλ0 z ( jω μ0 kx a0 + λ0 ky c0 ) + e−λ0z ( jω μ0 kx b0 + λ0ky d0 )

δ (z − z ) ˜
Ẽz = jkt2 (c0 eλ0 z − d0e−λ0 z ) − Jez
jω ε̂0

H̃x = eλ0 z (λ0 kx a0 + jω ε̂0 ky c0 ) + e−λ0 z (−λ0 kx b0 − jω ε̂0 ky d0 )

H̃y = eλ0 z (λ0 ky a0 − jω ε̂0 kx c0 ) + e−λ0 z (−λ0 ky b0 + jω ε̂0 kx d0 )

δ (z − z ) ˜
H̃z = jkt2 (a0 eλ0 z + b0 e−λ0 z ) + Jmz . (2.51)
j ω μ0

From here on we will drop the delta-function terms in Ẽz and H̃z , because they will
be associated with the infinite-field Green dyad.
The coefficients a0 , b0 , c0 , d0 that appear in (2.51) have been previously
expressed in terms of J1 , J2 , J3 , J4 in (2.50). When these results are substituted
into (2.51) we get:
 (E) (E) (M) (M)

 −R R j ω μ k J R R λ k J
Ẽx = e−λ0 (2T +(z −z)) −1 −1
1 0 y 1 1 0 x 3
(E) (E)
+ (M) (M)
1 − R1 R−1 e−2λ0 T 1 − R1 R−1 e−2λ0 T
 (E) (E) (M) (M)

−λ0 (2T −(z −z)) −R1 R−1 jω μ0 ky J2 R1 R−1 λ0 kx J4
+e (E) (E)
+ (M) (M)
1 − R1 R−1 e−2λ0 T 1 − R1 R−1 e−2λ0 T
22 2 Green’s Dyad for Plane-Layered Media

 (E) (M)

−λ0 (2z0 −(z +z)) −R1 jω μ0 ky J2 R1 λ0 kx J4
+e (E) (E)
+ (M) (M)
1 − R1 R−1 e−2λ0 T 1 − R1 R−1 e−2λ0 T
 (E) (M)

λ0 (2z−1 −(z +z)) −R−1 jω μ0 ky J1 R−1 λ0 kx J3
+e (E) (E)
+ (M) (M)
1 − R1 R−1 e−2λ0 T 1 − R1 R−1 e−2λ0 T
 (E) (E) (M) (M)

−λ0 (2T +(z −z)) R1 R−1 jω μ0 kx J1 R1 R−1 λ0 ky J3
Ẽy = e (E) (E)
+ (M) (M)
1 − R1 R−1 e−2λ0 T 1 − R1 R−1 e−2λ0 T
 (E) (E) (M) (M)

−λ0 (2T −(z −z)) R1 R−1 jω μ0 kx J2 R1 R−1 λ0 ky J4
+e (E) (E)
+ (M) (M)
1 − R1 R−1 e−2λ0 T 1 − R1 R−1 e−2λ0 T
 (E) (M)

−λ0 (2z0 −(z +z)) R1 jω μ0 kx J2 R1 λ0 ky J4
+e (E) (E)
+ (M) (M)
1 − R1 R−1 e−2λ0 T 1 − R1 R−1 e−2λ0 T
 (E) (M)

−(z +z)) R−1 jω μ0 kx J1 R−1 λ0 ky J3
+ eλ0 (2z−1 (E) (E)
+ (M) (M)
1 − R1 R−1 e−2λ0 T 1 − R1 R−1 e−2λ0 T

(M) (M)
 R1 R−1 jkt2 J3
Ẽz = e−λ0 (2T +(z −z)) (M) (M)
1 − R1 R−1 e−2λ0 T

(M) (M)
 −R1 R−1 jkt2 J4
+ e−λ0 (2T −(z −z)) (M) (M)
1 − R1 R−1 e−2λ0 T

(M)
 R1 jkt2 J4
+ e−λ0 (2z0 −(z +z)) (M) (M)
1 − R1 R−1 e−2λ0 T

(M)
 −R−1 jkt2 J3
+ eλ0 (2z−1 −(z +z)) (M) (M)
1 − R1 R−1 e−2λ0 T
 (E) (E) (M) (M)

(2T +(z −z)) R1 R−1 λ0 kx J1 R1 R−1 jω ε̂0 ky J3
H̃x = e−λ0 (E) (E)
+ (M) (M)
1 − R1 R−1 e−2λ0 T 1 − R1 R−1 e−2λ0 T
 (E) (E) (M) (M)

−λ0 (2T −(z −z)) −R1 R−1 λ0 kx J2 −R1 R−1 jω ε̂0 ky J4
+e (E) (E)
+ (M) (M)
1 − R1 R−1 e−2λ0 T 1 − R1 R−1 e−2λ0 T
2.2 Green Dyad for Plane-Parallel Layered Media 23

 (E) (M)

−λ0 (2z0 −(z +z)) R1 λ0 kx J2 R1 jω ε̂0 ky J4
+e (E) (E)
+ (M) (M)
1 − R1 R−1 e−2λ0 T 1 − R1 R−1 e−2λ0 T
 (E) (M)

λ0 (2z−1 −(z +z)) −R−1 λ0 kx J1 −R−1 jω ε̂0 ky J3
+e (E) (E)
+ (M) (M)
1 − R1 R−1 e−2λ0 T 1 − R1 R−1 e−2λ0 T
 (E) (E) (M) (M)

−λ0 (2T +(z −z)) R1 R−1 λ0 ky J1 −R1 R−1 jω ε̂0 kx J3
H̃y = e (E) (E)
+ (M) (M)
1 − R1 R−1 e−2λ0 T 1 − R1 R−1 e−2λ0 T
 (E) (E) (M) (M)

−λ0 (2T −(z −z)) −R1 R−1 λ0 ky J2 R1 R−1 jω ε̂0 kx J4
+e (E) (E)
+ (M) (M)
1 − R1 R−1 e−2λ0 T 1 − R1 R−1 e−2λ0 T
 (E) (M)

−λ0 (2z0 −(z +z)) R1 λ0 ky J2 −R1 jω ε̂0 kx J4
+e (E) (E)
+ (M) (M)
1 − R1 R−1 e−2λ0 T 1 − R1 R−1 e−2λ0 T
 (E) (M)

−(z +z)) −R−1 λ0 ky J1 R−1 jω ε̂0 kx J3
+ eλ0 (2z−1 (E) (E)
+ (M) (M)
1 − R1 R−1 e−2λ0 T 1 − R1 R−1 e−2λ0 T

(E) (E)
 R1 R−1 jkt2 J1
H̃z = e−λ0 (2T +(z −z)) (E) (E)
1 − R1 R−1 e−2λ0 T

(E) (E)
 R1 R−1 jkt2 J2
+ e−λ0 (2T −(z −z)) (E) (E)
1 − R1 R−1 e−2λ0 T

(E)
 R1 jkt2 J2
+ e−λ0 (2z0 −(z +z)) (E) (E)
1 − R1 R−1 e−2λ0 T

(E)
 R−1 jkt2 J1
+ eλ0 (2z−1 −(z +z)) (E) (E)
. (2.52)
1 − R1 R−1 e−2λ0 T

When the sources, J1 , J2 , J3 , J4 , are expressed in terms of the electric and

magnetic currents, as in (2.28), then we can express the fields in (2.52) in terms
of layered-space Green dyads:
(ee)
Ẽ(kx , ky ; z, z ) = G̃(s) (kx , ky ; z, z ) · J̃e (kx , ky )
(em)
Ẽ(kx , ky ; z, z ) = G̃(s) (kx , ky ; z, z ) · J̃m (kx , ky )
24 2 Green’s Dyad for Plane-Layered Media

(me)
H̃(kx , ky ; z, z ) = G̃(s) (kx , ky ; z, z ) · J̃e (kx , ky )
(mm)
H̃(kx , ky ; z, z ) = G̃(s) (kx , ky ; z, z ) · J̃m (kx , ky ) . (2.53)

The result for the spectral-domain, electric–electric, dyadic Green function is

shown here:
   2 
(ee)(s) c(z − z) −2λ0 T (E) (E) 1 (M) λ0
G̃xx = − j ω μ0 e G1,−1 − kx G1,−1 2 + G1,−1 2 2
2
2λ0 kt k0 kt
   2 
eλ0 (z+z ) −2λ0 z0 (E) (E) 1 (M) λ
+ e G1 − kx2 G1 2 + G1 2 0 2
2λ0 kt k0 kt
  
 2 
e−λ0 (z+z ) 2λ0 z−1 (E) (E) 1 (M) λ0
+ e G−1 − kx G−1 2 + G−1 2 2
2
2λ0 kt k0 kt
  2 
(ee)(s) c(z − z) −2λ0 T (E) 1 (M) λ0
G̃xy = − jω μ0 (−kx ky ) e G1,−1 2 + G1,−1 2 2
2λ0 kt k0 kt
  2 
eλ0 (z+z ) −2λ0 z0 (E) 1 (M) λ0
+ e G1 2 + G1 2 2
2λ0 kt k0 kt
 
 2 
e−λ0 (z+z ) 2λ0 z−1 (E) 1 (M) λ0
+ e G−1 2 + G−1 2 2
2λ0 kt k0 kt
 
(ee)(s) s(z − z) −2λ0 T (M) 1 eλ0 (z+z ) −2λ0 z0 (M) 1
G̃xz = − jω μ0 (− jkx λ0 ) − e G1,−1 2 − e G1 2
2λ0 k0 2λ0 k0


e−λ0 (z+z ) 2λ0 z−1 (M) 1
+ e G−1 2
2λ0 k0
(ee)(s) (ee)(s)
G̃yx = G̃xy
   2 
(ee)(s) c(z − z) −2λ0 T (E) (E) 1 (M) λ0
G̃yy = − j ω μ0 e G1,−1 − ky G1,−1 2 + G1,−1 2 2
2
2λ0 kt k0 kt
   2 
eλ0 (z+z ) −2λ0 z0 (E) (E) 1 (M) λ0
+ e G1 − k y G1 2 + G1 2 2
2
2λ0 kt k0 kt
  
 2 
e−λ0 (z+z ) 2λ0 z−1 (E) (E) 1 (M) λ0
+ e G−1 − ky G−1 2 + G−1 2 2
2
2λ0 kt k0 kt
2.2 Green Dyad for Plane-Parallel Layered Media 25

 
(ee)(s) s(z −z) −2λ0 T (M) 1 eλ0 (z+z ) −2λ0 z0 (M) 1
G̃yz = − jω μ0 (− jky λ0 ) − e G1,−1 2 − e G1 2
2λ0 k0 2λ0 k0


e−λ0 (z+z ) 2λ0 z−1 (M) 1
+ e G−1 2
2λ0 k0

(a)(ee)(s) (a)(ee)(s) (b)(ee)(s) (b)(ee)(s)

G̃zx = G̃xz , G̃zx = −G̃xz
(a)(ee)(s) (a)(ee)(s) (b)(ee)(s) (b)(ee)(s)
G̃zy = G̃yz , G̃zy = −G̃yz
 
(ee)(s) c(z − z) −2λ0 T (M) kt2 eλ0 (z+z ) −2λ0 z0 (M) kt2
G̃zz = − j ω μ0 e G1,−1 2 − e G1 2
2λ0 k0 2λ0 k0


e−λ0 (z+z ) 2λ0 z−1 (M) kt2
− e G−1 2 , (2.54)
2λ0 k0

   
where c(z − z) = eλ0 (z −z) + e−λ0 (z −z) and s(z − z) = eλ0 (z −z) − e−λ0 (z −z) . The
superscript (a) denotes terms that are convolutional (“Töplitz”) in z and z , i.e.,
depend upon z − z , whereas (b) denotes terms that are correlational (“Hankel”) in z
and z , i.e., depend upon z + z . The Gs are defined in terms of the TE and TM-mode
reflection coefficients:
(E) (E) (E) (E)
(E) R1 R−1 (E) R1 (E) R−1
G1,−1 = (E) (E) −2λ T
, G1 = (E) (E) −2λ T
, G−1 = (E) (E) −2λ T
1−R1 R−1 e 0 1−R1 R−1 e 0 1−R1 R−1 e 0
(M) (M) (M) (M)
(M) R1 R−1 (M) R1 (M) R−1
G1,−1 = (M) (M)
, G1 = (M) (M)
, G−1 = (M) (M)
.
1−R1 R−1 e−2λ0 T 1−R1 R−1 e−2λ0 T 1−R1 R−1 e−2λ0 T
(2.55)

The physical origin of the Töplitz and Hankel terms follows from the definitions
in (2.55), and a graphical illustration is shown in Fig. 2.4.
The Töplitz structure, as shown in the top of Fig. 2.4, arises when the path
between the source point, z , and field point, z, includes reflections from both
boundaries. The total z−directed path length between z and z is z − z + 2T for
path A and z − z + 2T for path B. In each case, the length includes the difference
between the z-coordinate of the source and field points.
The Hankel structure, as shown in the bottom of Fig. 2.4, arises when the path
between the source and field points includes reflections from only one of the
boundaries. The total path length between z and z is 2Z0 − (z + z ) for path A and
z + z − 2Z−1 for path B. In each case, the length includes the sum of the source and
field z-coordinates.
26 2 Green’s Dyad for Plane-Layered Media

Z0
B A A
z’
B
B A
T
B
z
A A B Z −1
Toeplitz

Z0
A A
z’
A

B
z
B B
Z−1‘
Hankel

Fig. 2.4 Illustrating the difference between Töplitz (top) and Hankel (bottom) Green’s functions

The spectral-domain, magnetic–magnetic, dyadic Green function is given by the

dual of (2.54):
   2 
(mm)(s) c(z − z) −2λ0 T (M) (M) 1 (E) λ
G̃xx = jω ε̂0 e G1,−1 − kx2 G1,−1 2 + G1,−1 2 0 2
2λ0 kt k0 kt
   2 
eλ0 (z+z ) −2λ0 z0 (M) (M) 1 (E) λ
− e G1 − kx2 G1 2 + G1 2 0 2
2λ0 kt k0 kt
  
 2 
e−λ0 (z+z ) 2λ0 z−1 (M) (M) 1 (E) λ0
− e G−1 − kx G−1 2 + G−1 2 2
2
2λ0 kt k0 kt
   2 
(mm)(s) c(z − z) −2λ0 T (M) 1 (E) λ0
G̃xy = jω ε̂0 (−kx ky ) e G1,−1 2 + G1,−1 2 2
2λ0 kt k0 kt
  
eλ0 (z+z ) −2λ0 z0 (M) 1 (E) λ
2
− e G1 2 + G1 2 0 2
2λ0 kt k0 kt
 
 2 
e−λ0 (z+z ) 2λ0 z−1 (M) 1 (E) λ0
− e G−1 2 + G−1 2 2
2λ0 kt k0 kt
2.2 Green Dyad for Plane-Parallel Layered Media 27

 
(mm)(s) s(z − z) −2λ0 T (E) 1 eλ0 (z+z ) −2λ0 z0 (E) 1
G̃xz = jω ε̂0 (− jkx λ0 ) − e G1,−1 2 + e G1 2
2λ0 k0 2λ0 k0


e−λ0 (z+z ) 2λ0 z−1 (E) 1
− e G−1 2
2λ0 k0
(mm)(s) (mm)(s)
G̃yx = G̃xy
   2 
(mm)(s) c(z − z) −2λ0 T (M) (M) 1 (E) λ0
G̃yy = jω ε̂0 e G1,−1 − ky G1,−1 2 + G1,−1 2 2
2
2λ0 kt k0 kt
   
eλ0 (z+z ) −2λ0 z0 (M) (M) 1 (E) λ
2
− e G1 − ky2 G1 2 + G1 2 0 2
2λ0 kt k0 kt
  
 2 
e−λ0 (z+z ) 2λ0 z−1 (M) (M) 1 (E) λ
− e G−1 − ky2 G−1 2 + G−1 2 0 2
2λ0 kt k0 kt
 
(mm)(s) s(z − z) −2λ0 T (E) 1 eλ0 (z+z ) −2λ0 z0 (E) 1
G̃yz = jω ε̂0 (− jky λ0 ) − e G1,−1 2 + e G1 2
2λ0 k0 2λ0 k0


e−λ0 (z+z ) 2λ0 z−1 (E) 1
− e G−1 2
2λ0 k0

(a)(mm)(s) (a)(mm)(s) (b)(mm)(s) (b)(mm)(s)

G̃zx = G̃xz , G̃zx = −G̃xz
(a)(mm)(s) (a)(mm)(s) (b)(mm)(s) (b)(mm)(s)
G̃zy = G̃yz , G̃zy = −G̃yz
 
(mm)(s) c(z − z) −2λ0 T (E) kt2 eλ0 (z+z ) −2λ0 z0 (E) kt2
G̃zz = jω ε̂0 e G1,−1 2 + e G1 2
2λ0 k0 2λ0 k0


e−λ0 (z+z ) 2λ0 z−1 (E) kt2
+ e G−1 2 . (2.56)
2λ0 k0

Finally, we list the mixed dyadic functions, starting with the electric–magnetic:
⎧ ⎡ ⎤  (E) 
⎨ e−2λ0 T G
(E)
−G
(M) 
eλ0 (z +z) −2λ0 z0 G1 −G1
(M)
(em)(s)  ⎣ 1,−1 1,−1 ⎦
G̃xx = −kx ky −s(z −z) − e
⎩ 2 kt 2 2 kt 2
 (E) ⎫
e−λ0 (z +z) 2λ0 z−1 G−1 − G−1 ⎬
 (M)
+ e
2 kt 2 ⎭
28 2 Green’s Dyad for Plane-Layered Media

 
(em)(s) s(z −z) (M) eλ0 (z +z) (M) 2λ0 z−1 e−λ0 (z +z) (M)
G̃xy = e−2λ0 T G1,−1 +e−2λ0 z0 G1 −e G−1
2 2 2
⎧ ⎡ ⎤  (E) 
⎨ s(z −z) G
(E)
−G1,−1
(M)
eλ0 (z +z) G −G
(M)
−2λ0 T ⎣ 1,−1 ⎦− −2λ
−ky −
2
e e 00 z 1 1
⎩ 2 kt 2 2 kt 2
 (E) ⎫
e−λ0 (z +z) 2λ0 z−1 G−1 −G−1 ⎬
 (M)
+ e
2 kt 2 ⎭

  

(em)(s) −2λ0 T c(z −z) (E) eλ0 (z +z) (E) 2λ0 z−1 e−λ0 (z +z) (E)
G̃xz = − jky e G1,−1 +e−2λ0 z0 G1 +e G−1
2λ0 2λ0 2λ0

 
(em)(s) s(z − z) (M) eλ0 (z +z) (M) e−λ0 (z +z) (M)
G̃yx = −e−2λ0 T G1,−1 − e−2λ0 z0 G1 + e2λ0 z−1 G−1
2 2 2
⎧ ⎡ ⎤  (E) 
⎨ s(z − z) (E) (M)  (M)
2
G
−2λ0 T ⎣ 1,−1
− G1,−1
⎦ eλ0 (z +z) −2λ0 z0 G1 − G1
−kx e + e
⎩ 2 kt 2 2 kt 2
 (E) ⎫
(M) 
e−λ0 (z +z) 2λ0 z−1 G−1 −G−1 ⎬


− e
2 kt 2 ⎭

(em)(s) (em)(s)
G̃yy = −G̃xx
  

(em)(s) −2λ0 T c(z −z) (E) eλ0 (z +z) (E) e−λ0 (z +z) (E)
G̃yz = − jkx −e G1,−1 −e−2λ0 z0 G1 − e2λ0 z−1 G−1
2λ0 2λ0 2λ0

  

(em)(s) −2λ0 T c(z −z) (M) eλ0 (z +z) (M) 2λ0 z−1 e−λ0 (z +z) (M)
G̃zx = − jky −e G1,−1 +e−2λ0 z0 G1 +e G−1
2λ0 2λ0 2λ0

  

(em)(s) −2λ0 T c(z −z) (M) eλ0 (z +z) (M) 2λ0 z−1 e−λ0 (z +z) (M)
G̃zy = − jkx e G1,−1 −e−2λ0 z0 G1 −e G−1
2λ0 2λ0 2λ0

(em)(s)
G̃zz = 0, (2.57)

and ending with the magnetic–electric:

2.2 Green Dyad for Plane-Parallel Layered Media 29

⎧ ⎡ ⎤  (E) 
⎨ −2λ0 T (E)
−G
(M) λ (z +z) (M)
−2λ0 z0 G1 −G1
(me)(s) e G 1,−1 ⎦ e 0
G̃xx 
= −kx ky s(z −z) ⎣ 1,−1
− e
⎩ 2 kt 2 2 kt 2
 (E) ⎫
e−λ0 (z +z) 2λ0 z−1 G−1 −G−1 ⎬
 (M)
+ e
2 kt 2 ⎭

 
(me)(s) s(z −z) (E) eλ0 (z +z) (E) 2λ0 z−1 e−λ0 (z +z) (E)
G̃xy = e−2λ0 T G1,−1 −e−2λ0 z0 G1 +e G−1
2 2 2
⎧ ⎡ ⎤  (E) 
⎨ s(z −z) G
(E) (M)
−G1,−1 
eλ0 (z +z) −2λ0 z0 G1 −G1
(M)
−2λ0 T ⎣ 1,−1 ⎦
− ky 2
e − e
⎩ 2 kt 2 2 kt 2
 (E) ⎫
e−λ0 (z +z) 2λ0 z−1 G−1 −G−1 ⎬
 (M)
+ e
2 kt 2 ⎭
  

(me)(s) −2λ0 T c(z −z) (M) eλ0 (z +z) (M) e−λ0 (z +z) (M)
G̃xz = − jky e G1,−1 −e−2λ0 z0 G1 − e2λ0 z−1 G−1
2λ0 2λ0 2λ0

 
(me)(s) s(z −z) (E) eλ0 (z +z) (E) 2λ0 z−1 e−λ0 (z +z) (E)
G̃yx = −e−2λ0 T G1,−1 +e−2λ0 z0 G1 −e G−1
2 2 2
⎧ ⎡ ⎤  (E) 
⎨ s(z −z) (E)
−
(M) λ0 (z +z) (M)
−2λ0 z0 G1 −G1
G G
−kx 2 − e−2λ0 T ⎣
1,−1 1,−1 ⎦+e e
⎩ 2 kt 2 2 kt 2
 (E) ⎫
e−λ0 (z +z) 2λ0 z−1 G−1 −G−1 ⎬
 (M)
− e
2 kt 2 ⎭

(me)(s) (me)(s)
G̃yy = −G̃xx

  

(me)(s) −2λ0 T c(z −z) (M) eλ0 (z +z) (M) 2λ0 z−1 e−λ0 (z +z) (M)
G̃yz = − jkx −e G1,−1 +e−2λ0 z0 G1 +e G−1
2λ0 2λ0 2λ0

  

(me)(s) c(z −z) (E) eλ0 (z +z) (E) 2λ0 z−1 e−λ0 (z +z) (E)
G̃zx = − jky −e−2λ0 T G1,−1 −e−2λ0 z0 G1 −e G−1
2λ0 2λ0 2λ0

  

(me)(s) −2λ0 T c(z −z) (E) eλ0 (z +z) (E) e−λ0 (z +z) (E)
G̃zy = − jkx e G1,−1 +e−2λ0 z0 G1 + e2λ0 z−1 G−1
2λ0 2λ0 2λ0
(me)(s)
G̃zz = 0. (2.58)
30 2 Green’s Dyad for Plane-Layered Media

(i+1)
η
i+1
T R
ξ i+1(-) i+1(-)
i+1
z
i
(+) (+)
λi R
i(+) η ,ξ
η = T i i
i j ωμ i i(+)

λi (i)
ξ =
i j ωε i

(−) (−)
R η ,ξ
i(-) i i
z
i-1
η
i-1
(i-1)
ξ
i-1

Fig. 2.5 Definition of scattering parameters and intrinsic wave-immittances in layered structures

2.2.3 A Recursion Relation for Stratified Media

The reflection and transmission coefficients of (2.49) are those of a slab surrounded
by a half-space above and below it. We will derive a recursion relation that will
(E,M)
allow the computation of the reflection coefficients, R±1 , that are used in (2.55)
when there are an arbitrary number of layers above and below the source slab [35].
Of course, there will ultimately be half-spaces that terminate the system at ±∞.
Let region i be bounded above by zi and below by zi−1 . Region i + 1 lies
immediately above region i and region i − 1 lies immediately below. The intrinsic
wave-admittance, ηi , and wave-impedance, ξi , for the TE and TM modes in region
i are
λi λi
(TE) ηi = , ξi = (TM). (2.59)
j ω μi jω ε̂i
(E,M)
We define Ri± to be the reflection coefficient in layer i at the interface with layer
(E,M)
i ± 1, and Ti± to be the transmission coefficient, as shown in Fig. 2.5. Then, as
shown in (2.49), for a slab sandwiched between two half-spaces
2.2 Green Dyad for Plane-Parallel Layered Media 31

(E) μi±1 λi − μi λi±1 ηi − ηi±1

Ri± = =
μi λi±1 + μi±1λi ηi + ηi±1

(M) ε̂i λi±1 − ε̂i±1 λi ξi±1 − ξi

Ri± = =
ε̂i λi±1 + ε̂i±1 λi ξi + ξi±1

(E) μi 2ηi
Ti± =
μi±1 ηi + ηi±1

(M) ε̂i 2 ξi
Ti± = . (2.60)
ε̂i±1 ξi + ξi±1

When we have layers above and below region i, we must replace the intrinsic wave-
parameters, ξi±1 , ηi±1 , by equivalent “load parameters,” ηi± , ξi± , which are defined
to be the surface admittance and impedance in layer i at the interface with layer
i ± 1. The surface admittance is defined to be H̃y /Ẽx at the appropriate interface, and
the surface impedance is defined to be Ẽx /H̃y . These ratios are the same regardless
of which side of the interface they are evaluated at, because Ẽx and H̃y are each
continuous at the interface. Hence, we replace (2.60) by

(E) ηi − ηi± (M) ξi± − ξi

Ri± = , R = . (2.61)
ηi + ηi± i±
ξi± + ξi

Let the TE-field in region i be

ai v1i eλi (z−zi ) + bi v2i e−λi (z−zi )

 
(E)
= bi Ri+ v1i eλi (z−zi ) + v2i e−λi (z−zi ) , (2.62)

then the field components at the lower-boundary, z = zi−1 , are

 
(E)
Ẽx = bi Ri+ (− jω μi ky )e−λi Ti − jω μi ky eλi Ti
 
(E)
H̃y = bi Ri+ (λi ky )e−λi Ti − λi ky eλi Ti . (2.63)

The load-admittance at z = zi−1 , therefore, is given by

(−) (+) H̃y

ηi = ηi−1 =
Ẽx
(E)
λi Ri+ e−λi Ti − λi eλi Ti
=  
(E)
− jω μi Ri+ e−λi Ti + eλi Ti
32 2 Green’s Dyad for Plane-Layered Media

(E)
eλi Ti − Ri+ e−λi Ti
= ηi (E)
eλi Ti + Ri+ e−λi Ti
(+)
ηi − ηi
eλi Ti − (+)
e−λi Ti
ηi + ηi
= ηi (+)
ηi − ηi
eλi Ti + (+)
e−λi Ti
ηi + ηi
(+)
ηi + ηi tanh(λi Ti )
= ηi (+)
. (2.64)
ηi + ηi tanh(λi Ti )

When we use v3i and v4i , we find that the components of the TM-field at z = zi−1
are given by
 
(M)
Ẽx = di Ri+ λi kx e−λi Ti + λikx eλi Ti
 
(M)
H̃y = di Ri+ (− jω ε̂i kx )e−λi Ti + jω ε̂i kx eλi Ti . (2.65)

Hence, by an analysis similar to (2.64), we get

(−) (+) Ẽx

ξi = ξi−1 =
H̃y
(+)
ξi + ξi tanh(λi Ti )
= ξi (+)
. (2.66)
ξi + ξi tanh(λi Ti )
Let the TE-field in region i below the source region be given by
ai v1i eλi (z−zi−1 ) + bi v2i e−λi (z−zi−1 )
 
(E)
= ai v1i eλi (z−zi−1 ) + Ri− v2i e−λi (z−zi−1 ) , (2.67)

where we treat the negatively-traveling wave, ai v1i eλi (z−zi−1 ) , as being incident on
the surface z = zi−1 . Therefore, we have
 
(E)
Ẽx = ai − jω μi ky eλi Ti + Ri− (− jω μi ky )e−λi Ti
 
(E)
H̃y = ai λi ky eλi Ti + Ri− (−λi ky )e−λi Ti , (2.68)
2.2 Green Dyad for Plane-Parallel Layered Media 33

for the fields at the upper-boundary, zi , of the ith region. Hence, the load admittance
at z = zi is

(+) (−) H̃y

ηi = ηi+1 = −
Ẽx
(E)
λi Ri− e−λi Ti − λi eλi Ti
=  
(E)
− jω μi Ri− e−λi Ti + eλi Ti
(E)
eλi Ti − Ri− e−λi Ti
= ηi (E)
eλi Ti + Ri− e−λi Ti
(−)
ηi − ηi
eλi Ti − (−)
e−λi Ti
ηi + ηi
= ηi (−)
ηi − ηi
eλi Ti + (−)
e−λi Ti
ηi + ηi
(−)
ηi + ηi tanh(λi Ti )
= ηi (−)
. (2.69)
ηi + ηi tanh(λi Ti )
Similarly, for regions below the source region
(−)
(+) (−) ξi + ξi tanh(λi Ti )
ξi = ξi+1 = ξi (−)
. (2.70)
ξi + ξi tanh(λi Ti )

In order to avoid large numbers in the computation of the hyperbolic tangent, we

write it as

eλi Ti − e−λiTi 1 − e−2λiTi

tanh(λi Ti ) = = .
eλi Ti + e−λiTi 1 + e−2λiTi

Equations (2.64), (2.69), (2.66), and (2.70) are the iterations that produce the
immittances that go into the expressions for the reflection coefficients, (2.61).
The iterations are started at the interface of the last slab with an infinite half-space,
for which
+
ηN−1 = ηN+ = ηN , ξN−1
+
= ξN+ = ξN
− − − −
η−(M−1) = η−M = η−M , ξ−(M−1) = ξ−M = ξ−M . (2.71)
Chapter 3
The Volume-Integral Equations
for Plane-Layered Media

3.1 Transformation into the Spatial-Domain

Although the development of the infinite-space and layered-space dyadic Green

functions is done in the Fourier-domain (or, as it is often referred to, the spectral-
domain), the further development of the integral equation, together with its
discretization, is best done in the spatial domain. We will now give an example
of the transformation from the Fourier-domain to the spatial domain.

λ 2 e−λ0 |z−z |
Consider the zz-term of (2.30); the kernel of the integral operator, 02 ,
k0 2 λ 0

1 d 2 e−λ0 |z−z | δ (z − z )
is equivalent to the kernel 2 2 + , as simple differentiation
k0 dz 2λ0 k02
proves. That is, both kernels behave identically as integral operators. Hence,


  

 λ02 e−λ0 |z−z | δ (z − z ) ˜
Ẽz (kx , ky ; z) = − jω μ0 dz 1+ 2 − Jez (kx , ky ; z )
k0 2λ0 k02


  

 1 d 2 e−λ0 |z−z | ˜
= − j ω μ0 dz 1 + 2 2 Jez (kx , ky ; z )
k0 dz 2λ0

  
1 d2 e−λ0 |z−z | ˜
= − j ω μ0 1 + 2 2 dz Jez (kx , ky ; z ) . (3.1)
k0 dz 2λ0

Hence, the spatial-domain field is obtained by taking the inverse Fourier trans-
form of (3.1)
    ∞ −λ0 |z−z |
1 ∂2 1 e
Ez (r) = − jω μ0 1 + 2 2 dz 2 J˜ez (kx , ky ; z )e− j[kx x+ky y] dkx dky
k0 ∂ z 4π −∞ 2λ0

H.A. Sabbagh et al., Computational Electromagnetics and Model-Based Inversion, 35

Scientific Computation, DOI 10.1007/978-1-4419-8429-6 3,
© Springer Science+Business Media New York 2013
36 3 The Volume-Integral Equations for Plane-Layered Media

 
1 ∂2
= 1+ 2 2 dr Φ (e) (r − r )Jez (r ) , (3.2)
k0 ∂ z

where
  ∞ −λ0 |z−z |
− j ω μ0 e  
Φ (e) (r − r) = e− j[kx (x−x )+ky (y−y )] dkx dky . (3.3)
4π 2 −∞ 2λ0

The transform
function in (3.3) is spherically symmetric, because it is a function
of λ0 = kt2 − k02 . Hence, we can transform the two-dimensional integrals into one-
dimensional integrals in the following way: transform to cylindrical coordinates in
both physical and k-space

x − x = r cos φ ; kx = l cos α
y − y = r sin φ ; ky = l sin α . (3.4)

(We are replacing the variable, kt , by l.) Then the integral in (3.3) becomes
 2π  ∞ −λ0 |z−z |
− j ω μ0 e
Φ (e) 
(r − r ) = dα e− jlr cos(α −φ ) ldl . (3.5)
4π 2 0 0 2λ0

According to a well-known identity involving Bessel functions, we have

∞
e− jlr cos(α −φ ) = J0 (lr) + 2 ∑ (− j)n Jk (lr) cos n(α − φ ) . (3.6)
n=1

When (3.6) is substituted into (3.5), only the J0 term survives the integration over
2π radians, so that
 ∞ −λ0 |z−z |
(e)  − j ω μ0 e
Φ (r − r ) = J0 (rl)ldl
2π 0 2λ0

e− jk0 |r−r |
= − j ω μ0 . (3.7)
4π |r − r|

3.2 The Electric Differential Volume-Integral Equation

We decompose the total Green dyad into the “infinite” part, which is the field
produced by the source in infinite space, and the “layered” part, which is due to the
presence of the various layers of the workpiece. If we consider only nonmagnetic
problems (μ = μ0 ) right now, and let J(r ) be the unknown anomalous electric
current produced by the flaw, then the infinite part of the dyad produces the “infinite-
space” contribution to the total electric field:
3.2 The Electric Differential Volume-Integral Equation 37

(0) (e) 1 ∂ 2 (e) 1 ∂ 2 (e) 1 ∂ 2 (e)

Ex (r) = Ax + Ax + 2 Ay + 2 Az
k02 ∂ x2 k0 ∂ x∂ y k0 ∂ x∂ z

(0) 1 ∂ 2 (e) (e) 1 ∂ 2 (e) 1 ∂ 2 (e)

Ey (r) = A x + A y + Ay + 2 Az
k02 ∂ y∂ x k02 ∂ y2 k0 ∂ y∂ z

(0) 1 ∂ 2 (e) 1 ∂ 2 (e) (e) 1 ∂ 2 (e)

Ez (r) = Ax + 2 Ay + Az + 2 2 Az , (3.8)
k0 ∂ z∂ x
2 k0 ∂ z∂ y k0 ∂ z

where the vector potential is given by


A(e) (r) = Φ (e) (r − r)J(e) (r )dr , (3.9)

and

e− jk0 |r−r |
Φ (e) (r − r) = − jω μ0
4π |r − r|
 ∞ −λ0 |z−z |
− j ω μ0 e
= J0 (rl)ldl . (3.10)
2π 0 2λ0

In the Bessel transform, 3.10, r = [(x − x )2 + (y − y)2 ]1/2 .

The “layered-space” contribution to the total field is given by

(s) (1) 1 ∂2 1 ∂2 1 ∂2
Ex (r) = Fx + Fx + Fy + Fz
k02 ∂ x2 k02 ∂ x∂ y k02 ∂ x∂ z

(s) 1 ∂2 (1) 1 ∂2 1 ∂2
Ey (r) = Fx + Fy + 2 2 Fy + 2 Fz
k0 ∂ y∂ x
2 k0 ∂ y k0 ∂ y∂ z

(s) 1 ∂2 1 ∂2 1 ∂2
Ez (r) = Fzx + 2 Fzy + Fz + 2 2 Fz , (3.11)
k0 ∂ z∂ x
2 k0 ∂ z∂ y k0 ∂ z

where

(1) (1)
Fx (r) = dr Gxx (x − x , y − y ; z, z )Jx (r )

(2)
Fx (r) = dr Gxx (x − x , y − y ; z, z )Jx (r )

(2)
Fy (r) = dr Gxx (x − x , y − y ; z, z )Jy (r )

Fz (r) = dr Gxz (x − x , y − y ; z, z )Jz (r )

(1) (1)
Fy (r) = dr Gxx (x − x , y − y ; z, z )Jy (r )
38 3 The Volume-Integral Equations for Plane-Layered Media


(a) (a)
Fzx (r) = dr Gxz (x − x , y − y ; z, z )Jx (r )

(b) (b)
Fzx (r) = − dr Gxz (x − x , y − y ; z, z )Jx (r )

(a) (a)
Fzy (r) = dr Gxz (x − x , y − y ; z, z )Jy (r )

(b) (b)
Fzy (r) = − dr Gxz (x − x , y − y ; z, z )Jy (r ) . (3.12)

The unique kernels that appear in (3.12) are given by

 ∞
 
(1) − j ω μ0 c(z − z) −2λ0 T (E) eλ0 (z+z ) −2λ0 z0 (E)
Gxx = e G1,−1 + e G1
2π 0 2λ0 2λ0


e−λ0 (z+z ) 2λ0 z−1 (E)
+ e G−1 J0 (rl)ldl
2λ0
 
 2
(2) − jω μ0 ∞ c(z − z) −2λ0 T (E) k0
2
(M) λ0
Gxx = e G1,−1 2 + G1,−1 2
2π 0 2λ0 l l
  2
eλ0 (z+z ) −2λ0 z0 (E) k0
2
(M) λ0
+ e G1 2 + G1 2
2λ0 l l
 
 2
e−λ0(z+z ) 2λ0 z−1 (E) k0
2
(M) λ0
+ e G−1 2 + G−1 2 J0 (rl)ldl
2λ0 l l

 
− jω μ0 ∞ c(z − z) −2λ0 T (M) eλ0 (z+z ) −2λ0 z0 (M)
Gxz = e G1,−1 − e G1
2π 0 2λ0 2λ0


e−λ0 (z+z ) 2λ0 z−1 (M)
− e G−1 J0 (rl)ldl. (3.13)
2λ0

As before, r = [(x−x )2 +(y−y )2 ]1/2 in the Bessel transform. The results of (3.11)–
(3.13) follow from (2.54), upon replacing multiplication by − jkx , − jky , and λ0 by
∂ /∂ x, ∂ /∂ y, ∂ /∂ z, respectively. In the expression for G̃zz we use kt2 = λ02 + k02 .
The integro-differential equation to which we will apply the method of moments
is simply gotten by equating the total electric field, J(r)/σa , to the sum of the
incident field, due to the coil and the infinite-space and layered-space scattered
fields:

(i) Jx (r) (0) (s)

Ex (r) = − Ex (r)[J] − Ex (r)[J]
σa (r)
3.3 Vector Form of the Integral Equation 39

(i) Jy (r) (0) (s)

Ey (r) = − Ey (r)[J] − Ey (r)[J]
σa (r)
(i) Jz (r) (0) (s)
Ez (r) = − Ez (r)[J] − Ez (r)[J] , (3.14)
σa (r)

where σa (r) = jω (ε̂ (r) − ε̂h ) is the anomalous conductivity.

3.3 Vector Form of the Integral Equation

The equations of the model, (3.8)–(3.14), can be rewritten to bring out the vector
nature more clearly. In deriving the new equations we repeatedly make use of
integration-by-parts and then use the fact that the current distribution is limited in
(e)
space. For example, consider the term ∂ Ax (r)/∂ x:

(e) 
∂ Ax (r) ∂ (e)
= Φ (e) (r − r )Jx (r )dr
∂x ∂x

∂ (e) (e)
= Φ (r − r)Jx (r )dr
∂x

∂ (e)
= −  Φ (e) (r − r )Jx (r )dr
∂x
 (e)
∂ Jx (r ) 
= Φ (e) (r − r ) dr . (3.15)
∂ x

Following this pattern we derive two other important results for z-derivatives:
 
∂ (a) (a) ∂ Jz (r )
dr Gxz (x − x , y − y, z − z )Jz (r ) = dr Gxz (x − x , y − y , z − z )
∂z ∂ z
 
∂ (b) (b) ∂ Jz (r )
dr Gxz (x − x , y − y, z + z )Jz (r ) = − dr Gxz (x − x , y − y , z + z ) .
∂z ∂ z
(3.16)

Applying integration-by-parts to (3.8) allows us to rewrite the infinite-space

contribution as a single vector equation:
 
1
E(0) (r)[J(e) ] = Φ (e) (r − r )J(e) (r )dr + ∇ Φ (e) (r − r )∇ · J(e) (r )dr .
k02
(3.17)
40 3 The Volume-Integral Equations for Plane-Layered Media

In a similar manner, we can rewrite each of the component-equations in (3.11) to

get the following vector equation for the layered-space contribution:

(1)(a) (e)
E(s) (r)[J(e) ] = dr Gxx (x − x , y − y , z − z )Jt (r )

(a) (e)
+az dr Gxz (x − x , y − y , z − z )Jz (r )

(1)(b) (e)
+ dr Gxx (x − x , y − y , z + z )Jt (r )

(b) (e)
+az dr Gxz (x − x , y − y , z + z )Jz (r )

(2)(a) (e)
+∇t dr Gxx (x − x , y − y, z − z )∇t · Jt (r )

(2)(b) (e)
+∇t dr Gxx (x − x , y − y, z + z )∇t · Jt (r )

1 (a)
+ ∇ dr Gxz (x − x , y − y, z − z )∇ · J(e) (r )
k02

1 (b)
− ∇ dr Gxz (x − x , y − y, z + z )∇ · J(e) (r ) , (3.18)
k02

∂ ∂ ∂ Jx ∂ Jy
where ∇t = ax + ay is the transverse-gradient operator; ∇t · J = + is
∂x ∂y ∂x ∂y
the transverse divergence; Jt = ax Jx + ay Jy , and
⎡ ⎤
 ∞  − z) (E) (M)
G1,−1 − G1,−1
(2)(a) − jω μ0 ⎣ c(z λ ⎦ J0 (rl)ldl
Gxx (x − x , y − y , z − z ) = −2
e 0 T
2π 0 2λ0 l2

 ∞ λ0 (z+z ) (E) (M)
(2)(b) − jω μ0 e G − G1
Gxx (x − x , y − y , z + z ) = e−2λ0 z0 1
2π 0 2λ0 l2
 (E) (M)

e−λ0 (z+z ) 2λ0 z−1 G−1 − G−1
+ e J0 (rl)ldl.
2λ0 l2

(3.19)

The remaining expressions are given in (3.13).

3.4 The Volume-Integral Equations in Terms of Amperian Currents 41

3.4 The Volume-Integral Equations in Terms of Amperian

Currents

The development of the Green’s dyad in Chap. 2 assumed that the unknowns were
the dual currents, Jm and Je . By “dual” we mean that one appears in the Faraday–
Maxwell law (the first equation in (2.2)) and the other in the Ampere–Maxwell
law (the second equation in (2.2)). This works well in presenting the theory in a
coherent and consistent manner, but it also means that there will be “dual” Green’s
dyads for electric–electric and magnetic–magnetic interactions, as shown in Chap. 2.
It is advantageous from a computational perspective to be able to use as much
code as possible when computing the various dyadic components, and in order to
accomplish this we find that writing the anomalous magnetic currents as “Amperian
currents” in Ampere’s law is beneficial.
We start with Maxwell’s equations
∇ × E = − jω B
∇ × H = jω D + J(e) . (3.20)
Now H = B/μ (r) = B/ μh + B/μ (r) − B/μh = B/μh − Ma , where μh is the host
permeability and Ma is the anomalous magnetization vector. Thus the second of
Maxwell’s equations may be written
∇ × B/μh = jω D + J(e) + ∇ × Ma , (3.21)
which makes clear that the Amperian current, ∇ × Ma , is an equivalent anomalous
electric current that arises because of the departures of the magnetic permeability of
the workpiece from the host permeability, μh . J(e) , on the other hand, is an electric
current that includes the anomalous current that arises due to differences in electrical
conductivity; J(e) = σh E + (σ (r) − σh)E = σh E + Ja.
Even though the Amperian current is electrical, because it appears as a source
term in the second Maxwell equation (Ampere’s law), we will refer to it as J(m) , to
remind us that it is of magnetic origin, and to distinguish it from J(e) (which now
stands for the anomalous electric current, Ja ). The important point, however, is that
because the Amperian current behaves as an electrical current, we need only use
electric–electric Green functions in the new formulation of the problem, as we now
show.
The coupled volume-integral equations for J(e) and J(m) are
J(e) (r)        
E(i) (r) = − E(0) (r) J(e) − E(s) (r) J(e) − E(0) (r) J(m) − E(s) (r) J(m)
σa (r)

μ (r)μh 1   1  
B(i) (r) = Ma + ∇ × E(0) (r) J(e) + ∇ × E(s)(r) J(e)
μ (r) − μh jω jω

1   1  
+ ∇ × E(0)(r) J(m) + ∇ × E(s) (r) J(m) . (3.22)
jω jω
42 3 The Volume-Integral Equations for Plane-Layered Media

In arriving at the second equation, we have used the fact that B = −(1/ jω )∇ × E,
and Ma = ((μ (r) − μh )/ μ (r)μh )B.
The first part of the first equation in (3.22) is the usual electric–electric model,
(3.14). The remaining parts must be determined, and in order to do this we will use
the form of the equations given in Sect. 3.3. Because J(m) has zero divergence, we
can write

E(0) (r)[J(m) ] = Φ (e) (r − r )J(m) (r )dr

(1)(a) (m)
E(s) (r)[J(m) ] = dr Gxx (x − x , y − y , z − z )Jt (r )

(a) (m)
+az dr Gxz (x − x , y − y , z − z )Jz (r )

(1)(b) (m)
+ dr Gxx (x − x , y − y, z + z )Jt (r )

(b) (m)
+az dr Gxz (x − x , y − y , z + z )Jz (r )

(2)(a) (m)
+∇t dr Gxx (x − x , y − y, z − z )∇t · Jt (r )

(2)(b) (m)
+∇t dr Gxx (x − x , y − y, z + z )∇t · Jt (r ) , (3.23)

∂ ∂ ∂ Jx ∂ Jy
where ∇t = ax + ay ; ∇t · J = + ; Jt = ax Jx + ay Jy , and
∂x ∂y ∂x ∂y
 ∞  
(1)(a) − jω μ0 c(z − z) −2λ0 T (E)
Gxx (x − x , y − y , z − z ) = e G1,−1 J0 (rl)ldl
2π 0 2λ0
 

(1)(b) − jω μ0 ∞ eλ0 (z+z ) −2λ0 z0 (E)
Gxx (x − x , y − y , z + z ) = e G1
2π 0 2λ0


e−λ0 (z+z ) 2λ0 z−1 (E)
+ e G−1 J0 (rl)ldl
2λ0
⎡ ⎤
 ∞  − z) (E)
−
(M)
(2)(a) − j ω μ c(z G G
Gxx (x − x , y − y , z − z ) = ⎣ e−2λ0 T ⎦ J0 (rl)ldl
0 1,−1 1,−1
2π 0 2λ0 l2

 ∞ λ0 (z+z ) (E) (M)
(2)(b) − jω μ0 e G − G1
Gxx (x − x , y − y , z + z ) = e−2λ0 z0 1
2π 0 2λ0 l2
 (E) (M)

e−λ0 (z+z ) 2λ0 z−1 G−1 − G−1
+ e J0 (rl)ldl
2λ0 l2
3.4 The Volume-Integral Equations in Terms of Amperian Currents 43

 ∞  
(a) − jω μ0 c(z − z) −2λ0 T (M)
Gxz (x − x , y − y , z − z ) = e G1,−1 J0 (rl)ldl
2π 0 2λ0
 ∞
 
(b) − jω μ0 eλ0 (z+z ) −2λ0 z0 (M)
Gxz (x − x , y − y , z + z ) = − e G1
2π 0 2λ0


e−λ0 (z+z ) 2λ0 z−1 (M)
− e G−1 J0 (rl)ldl . (3.24)
2λ0

Furthermore,
  
∇ × E(0) (r) J(e) = ∇ × Φ (e) (r − r)J(e) (r )dr
  
(1)(a) (e)
∇ × E(s) (r) J(e) = ∇ × dr Gxx (x − x , y − y, z − z )Jt (r )

(a) (e)
+∇ × az dr Gxz (x − x , y − y, z − z )Jz (r )

(1)(b) (e)
+∇ × dr Gxx (x − x , y − y , z + z )Jt (r )

(b) (e)
+∇ × az dr Gxz (x − x , y − y, z + z )Jz (r )

(2)(a) (e)
+∇ × ∇t dr Gxx (x − x , y − y , z − z )∇t · Jt (r )

(2)(b) (e)
+∇×∇t dr Gxx (x−x , y−y , z+z )∇t · Jt (r ) , (3.25)

   
with the same expressions holding for ∇ × E(0)(r) J(m) , ∇ × E(s) (r) J(m) .
Chapter 4
Discretization via the Galerkin Method
of Moments

We will discretize (3.22) by employing Galerkin’s method, which uses the same
vector functions for expansion and testing. The spatial derivatives that cause
problems will be removed by the testing process. In order to test these derivatives,
we introduce special vector expansion functions, called “facet elements” and “edge
elements,” that comprise products of pulse and tent functions.
Facet elements have been called “volumetric rooftop” functions by Catedra
et al. [36]. Volumetric rooftop functions have also been used in [37, 38]. These
functions are a generalization of two-dimensional rooftop functions that have been
used in problems involving scattering from two-dimensional structures and three-
dimensional surfaces [30, 32, 33].
Facet elements and edge elements are a subset of a more general class of spline-
generated basis-functions that are based upon higher-order convolutions of the unit
pulse

1, if 0 ≤ x < 1
π (x) = (4.1)
0, otherwise.

The reader is invited to study [31, 39] for a more complete development of the
subject.

4.1 Expansion of the Anomalous Currents

We introduce facet elements oriented in the x-, y-, and z-directions, such that the
derivative with respect to x, y, and z is bounded. This ensures that the divergence of
the vector field is bounded, and for this reason facet elements are often referred to

H.A. Sabbagh et al., Computational Electromagnetics and Model-Based Inversion, 45

Scientific Computation, DOI 10.1007/978-1-4419-8429-6 4,
© Springer Science+Business Media New York 2013
46 4 Discretization via the Galerkin Method of Moments

Fig. 4.1 Showing the 1

location of the tent and pulse
functions for the facet
(x)
element Tklm (x, y, z) =
π2k (x/δ x)π1l (y/δ y)π1m (z/δ z)
X
kδx (k+1) δx (k+2) δx

Y
l δy (l+1) δy

Z
m δz (m+1) δz

(q)(e)
as “divergence-conforming.” We write Tklm (x, y, z) for the facet element oriented
(q)(e)
in the qth direction. The expressions for Tklm are:

(x)(e)
Tklm (r) = π2k (x/δ x)π1l (y/δ y)π1m (z/δ z)
(y)(e)
Tklm (r) = π1k (x/δ x)π2l (y/δ y)π1m (z/δ z)
(z)(e)
Tklm (r) = π1k (x/δ x)π1l (y/δ y)π2m (z/δ z) (k, l, m) = (0, 0, 0), . . . , (Nx , Ny , Nz ) ,
(4.2)

where π1m (y/δ y) is the mth unit pulse function and π2k (x/δ x) is the kth tent
function, which is the convolution of π1k (x/δ x) with itself. The T (q)(e) (r)klm are
called “facet elements,” because the qth element is constant over the qth facet of the
klmth cell.
Figure 4.1 shows the position of the tent and pulse functions for the facet element
(x)(e)
Tklm (x, y, z).
(x)(e)
The support of Tklm (x, y, z) is shown in Fig. 4.2. We assume that the conductiv-
ity is constant, with the value

σcell = σmax + Vc (σmin − σmax ), (4.3)

4.1 Expansion of the Anomalous Currents 47

(k+2,l+1,m)
(k+1,l+1,m)

(k,l+1,m)

(klm) δy

(k+1,l,m+1)
δz
δx

(x)(e)
Fig. 4.2 Support of Tklm (x, y, z). The conductivity is assumed to be constant within each cell of
dimension δ x × δ y × δ z

within each cell of dimension δ x × δ y × δ z. σmax and σmin are, respectively, the
maximum and minimum conductivities in the problem and Vc is the conductivity
volume-fraction.
We expand the anomalous electric current vector in terms of the facet elements as

∑ JKLM TKLM (r)

(e) (x) (x)
Jx (r) =
KLM

∑ JKLM TKLM (r)

(e) (y) (y)
Jy (r) =
KLM

∑ JKLM TKLM (r),

(e) (z) (z)
Jz (r) = (4.4)
KLM

and will then use these same basis-functions for testing the integral equations.
Because J(m) (r) = ∇ × Ma (r), we expand Ma (r) in “curl-conforming” edge
elements, which have the required differentiability of the curl operation

∑ MKLM TKLM
(x) (x)(m)
Mx (r) = (r)
KLM

∑ MKLM TKLM
(y) (y)(m)
My (r) = (r)
KLM

∑ MKLM TKLM
(z) (z)(m)
Mz (r) = (r), (4.5)
KLM
48 4 Discretization via the Galerkin Method of Moments

(l,m+2) (l+1,m+2) (l+2,m+2)

(l,m+1) (l+1,m+1)

δz

(l,m) (l+1,m)
Y
δy
(x)(m)
Fig. 4.3 Support of Tklm (x, y, z) in (y, z)-space. The support extends one cell in the x-direction,
normal to the page. The permeability is assumed to be constant within each cell of dimension
δx×δy×δz

where
(x)(m)
TKLM (r) = π1K (x)π2L (y)π2M (z)
(y)(m)
TKLM (r) = π2K (x)π1L (y)π2M (z)
(z)(m)
TKLM (r) = π2K (x)π2L (y)π1M (z). (4.6)
These functions are called edge elements because the expansion coefficient,
(x)
MKLM , is the (constant) value of Mx along the x-directed edge, (y = (L + 1)δ y, z =
(y) (z)
(M + 1)δ z). There are similar interpretations for MKLM and MKLM .
(x)(m)
The support of Tklm (x, y, z) is shown in Fig. 4.3. We assume that the magnetic
permeability is constant, with the value

μcell = μmax + Vp(μmin − μmax), (4.7)

within each cell of dimension δ x × δ y × δ z. μmax and μmin are, respectively, the
maximum and minimum permeabilities in the problem and Vp is the permeability
volume-fraction.
The components of the magnetic current vector are given by

(m) ∂ Mz ∂ My
Jx = −
∂y ∂z
 
= ∑ MKLM π2K (x)π2L
(z)  (y) 
(y)π1M (z) − MKLM π2K (x)π1L (y)π2M (z)
KLM
4.2 Testing the Integral Equations 49

(m) ∂ Mx ∂ Mz
Jy = −
∂z ∂x
 
= ∑ MKLM π1K (x)π2L (y)π2M
(x)  (z) 
(z) − MKLM π2K (x)π2L (y)π1M (z)
KLM

(m) ∂ My ∂ Mx
Jz = −
∂x ∂y
 
= ∑ MKLM π2K
(y)  (x) 
(x)π1L (y)π2M (z) − MKLM π1K (x)π2L (y)π2M (z)
KLM

(m) ∂ 2 Mx ∂ 2 My
∇t · Jt = −
∂ y∂ z ∂ x∂ z
 
= ∑ MKLM π1K (x)π2L
(x)   (y)  
(y)π2M (z) − MKLM π2K (x)π1L (y)π2M (z) , (4.8)
KLM

where the last term is the transverse divergence of the Amperian current.

4.2 Testing the Integral Equations

The first step in discretizing the integral equations, (3.22), is to substitute the
expansions, (4.4) and (4.8), for the anomalous currents into the equations and then
to “test” the resulting equation. By “testing” we mean taking moments, which is
done by multiplying a functional equation by a test function and then integrating
over space. When the test function is the same as the expansion function for the
unknowns, the method is called the Galerkin variant of the method of moments. We
will test each component of the electric equations of (3.22) by the corresponding
(q)(e)
facet function, Tklm (r), and each component of the magnetic equations by the
(q)(m)
corresponding edge function, Tklm (r). The procedure is straightforward, but quite
lengthy, so we will show only the results. The terms of (3.22) that involve spatial
derivatives are “mollified” by using the two vector identities that are defined in
Appendix A.1.
The discretized electric equation is:
⎡ ⎤ ⎡ (x) ⎤(ee) ⎡ (x) ⎤
E(ix) Q 0 0 J
⎣ E(iy) ⎦ = ⎣ 0 Q(y) 0 ⎦ ⎣ J(y) ⎦
E(iz) 0 0 Q(z) J(z)
⎡ (xx) (xy) (xz) ⎤(ee)
G G(0) G(0) ⎡ (x) ⎤
⎢ (0) ⎥ J
⎢ (yx) (yy) (yz) ⎥ ⎣
− ⎣ G(0) G(0) G(0) ⎦ J(y) ⎦
(zx) (zy)
G(0) G(0) G(0)
(zz) J(z)
50 4 Discretization via the Galerkin Method of Moments

⎡ ⎤(ee)
(xx) (xy) (xz) ⎡ ⎤
G(a) G(a) G(a)
⎢ (yx) (yy) (yz) ⎥ J(x)
−⎢
⎣ G(a) G(a) G(a) ⎦
⎥ ⎣ J(y) ⎦
(zx) (zy) (zz)
G(a) G(a) G(a) J(z)
⎡ ⎤(ee)
(xx) (xy) (xz) ⎡ ⎤
G(b) G(b) G(b)
⎢ (yx) (yy) (yz) ⎥ J(x)
−⎢
⎣ G(b) G(b) G(b) ⎦
⎥ ⎣ J(y) ⎦
(zx) (zy) (zz)
G(b) G(b) G(b) J(z)
⎡ (x) ⎤
 (em) M
− G(0) G(a) G(b) ⎣ M(y) ⎦ , (4.9)
M(z)

where the Q’s are tri-diagonal matrices, the G(0) ’s the infinite-space matrices, the
G(a) ’s the convolutional layered-space matrices, and the G(b) ’s the correlational
layered-space matrices. The infinite-space matrices are convolutional, also. The
superscript (ee) denotes electric–electric matrices and (em) denotes electric–
magnetic matrices. The J’s are the unknown electric currents and the M’s are
the unknown magnetic polarization vectors. The last block in (4.9) is simply a
shorthand representation of the three blocks above it, except that it represents
electric–magnetic interactions.
The discretized magnetic equation is similar to (4.9) and is given by
⎡ ⎤ ⎡ (x) ⎤(mm) ⎡ (x) ⎤
B(ix) Q 0 0 M
⎣ B(iy) ⎦ = ⎣ 0 Q(y) 0 ⎦ ⎣ M(y) ⎦
B(iz) 0 0 Q(z) M(z)
⎡ (xx) (xy) (xz) ⎤(mm)
G G(0) G(0) ⎡ (x) ⎤
⎢ (0) ⎥ M
(yx) (yy) (yz) ⎥
+⎢ G G
⎣ (0) (0) (0) ⎦ G ⎣ M(y) ⎦
(zx) (zy)
G(0) G(0) G(0)
(zz) M(z)
⎡ ⎤(mm)
(xx) (xy) (xz) ⎡ ⎤
G(a) G(a) G(a)
⎢ (yx) (yy) (yz) ⎥ M(x)
+⎢
⎣ G(a) G(a) G(a) ⎦
⎥ ⎣ M(y) ⎦
(zx) (zy) (zz)
G(a) G(a) G(a) M(z)
⎡ ⎤(mm)
(xx) (xy) (xz) ⎡ ⎤
G(b) G(b) G(b)
⎢ (yx) (yy) (yz) ⎥ M(x)
+⎢
⎣ G(b) G(b) G(b) ⎦
⎥ ⎣ M(y) ⎦
(zx) (zy) (zz)
G(b) G(b) G(b) M(z)
⎡ (x) ⎤
 (me) J
+ G(0) G(a) G(b) ⎣ J(y) ⎦ , (4.10)
J(z)
4.3 Solution via Conjugate Gradients 51

where B is the incident magnetic flux density due to the coil, the superscript (mm)
stands for magnetic–magnetic interactions, and (me) stands for magnetic–electric
interactions. The magnetic–magnetic Q matrices are a little more complicated than
the electric–electric ones.

4.3 Solution via Conjugate Gradients

The system of equations, (4.9) and (4.10), that is produced by the method of
moments contains a dense matrix, as opposed to the sparse matrices that are
produced by finite-element or finite-difference techniques. Hence, it is necessary
to develop efficient means of solving the system if the volume-integral method is
to be viable in solving realistic three-dimensional problems. When the number of
unknowns is  3,000, we can use a direct matrix-decomposition solver, such as
the LU-decomposition [41], but for large problems we use the conjugate-gradient
algorithm [42].
The conjugate-gradient algorithm, being an iterative method, requires many
matrix-vector multiplications. In order for this method to be useful, therefore, there
must be an efficient method for computing these products. Fortunately, because we
have formulated the volume-integral equations on a regular grid, there exists a very
efficient numerical scheme for computing vector-matrix products.

4.3.1 Efficient Computation of Convolutions and Correlations

The triple sums that appear in (4.9) and (4.10) consist of three-dimensional
convolutions, or two-dimensional convolutions and one-dimensional correlations,
which means that we can use three-dimensional FFTs to compute them.
The appropriate theorems (in one dimension) that relate discrete Fourier
transforms and convolutions and correlations are ( ⇐⇒ denotes a discrete transform-
pair):

If g( j) ⇐⇒ G(n)
h( j) ⇐⇒ H(n)
k=0 g( j + k)h(k) = N ∑k=0 g(k)h(k − j)
Then N1 ∑N−1 1 N−1

⇐⇒ G(n)H(−n)
= G(n)H(N − n)
∑
1 N−1
N k=0 g(k)h( j + k) = N1 ∑N−1
k=0 g(k − j)h(k)
⇐⇒ G(−n)H(n)
= G(N − n)H(n)
∑
1 N−1
N k=0 g(k)h( j − k) = N1 ∑N−1
k=0 g( j − k)h(k)
⇐⇒ G(n)H(n), (4.11)
52 4 Discretization via the Galerkin Method of Moments

where j = 0, . . . , N − 1, n = 0, . . . , N − 1 in all of these. Several points should be

made: first note that correlation summing is not commutable and that one must use
negative frequencies in the discrete Fourier transform (which, of course, introduces
the term N − n).
Let’s look at the matrix structure of convolution and correlation sums and see
how we can use FFT techniques to compute them. We’ll work in one dimension.
Consider the following convolution, which is written as a vector-matrix equation:

⎡ ⎤ ⎡ ⎤⎡ ⎤
y0 m0 m−1 m−2 m−3 x0
⎢ y 1 ⎥ ⎢ m1 m0 m−1 m−2 ⎥ ⎢ x1 ⎥
⎢ ⎥=⎢ ⎥⎢ ⎥. (4.12)
⎣ y 2 ⎦ ⎣ m2 m1 m0 m−1 ⎦ ⎣ x2 ⎦
y3 m3 m2 m1 m0 x3

Rewrite this in the expanded form (padding with zeros to get a power of two) in
order to achieve a circulant-matrix:
⎡ ⎤ ⎡ ⎤⎡ ⎤
y0 m0 m−1 m−2 m−3 0 m3 m2 m1 x0
⎢ y 1 ⎥ ⎢ m1 m0 m−1 m−2 m−3 0 m3 ⎥ ⎢
m2 ⎥ ⎢ x 1 ⎥
⎢ ⎥ ⎢ ⎥
⎢y ⎥ ⎢m m3 ⎥ ⎢ ⎥
⎢ 2⎥ ⎢ 2 m1 m0 m−1 m−2 m−3 0 ⎥ ⎢ x2 ⎥
⎢ ⎥ ⎢ ⎥⎢ ⎥
⎢ y 3 ⎥ ⎢ m3 m2 m1 m0 m−1 m−2 m−3 0 ⎥ ⎢ x3 ⎥
⎢ ⎥=⎢ ⎥⎢ ⎥, (4.13)
⎢ ∗ ⎥ ⎢0 m3 m2 m1 m0 m−1 m−2 m−3 ⎥ ⎢ 0 ⎥
⎢ ⎥ ⎢ ⎥⎢ ⎥
⎢ ∗ ⎥ ⎢ m−3 0 m3 m2 m1 m0 m−1 m−2 ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥⎢ 0 ⎥
⎣ ∗ ⎦ ⎣ m−2 m−3 0 m3 m2 m1 m0 m−1 ⎣ 0 ⎦
⎦
∗ m−1 m−2 m−3 0 m3 m2 m1 m0 0

where the ∗ denotes a discarded entry. Hence, the sequences to be FFT’d are:
(m0 , m1 , m2 , m3 , 0, m−3 , m−2 , m−1 ) and (x0 , x1 , x2 , x3 , 0, 0, 0, 0), and the output se-
quence is (y0 , y1 , y2 , y3 , ∗, ∗, ∗, ∗). The order of the entries in the sequences is very
important.
Now, for correlations:
⎡ ⎤ ⎡ ⎤⎡ ⎤
y0 m0 m1 m2 m3 x0
⎢ y 1 ⎥ ⎢ m1 m2 m3 m4 ⎥ ⎢ x 1 ⎥
⎥ ⎢
⎢ ⎥=⎢ ⎥. (4.14)
⎣ y 2 ⎦ ⎣ m2 m3 m4 m5 ⎦ ⎣ x 2 ⎦
y3 m3 m4 m5 m6 x3

Rewrite this in the expanded form (padding with zeros to get a power of two) in
order to achieve a circulant-matrix:
4.3 Solution via Conjugate Gradients 53

⎡ ⎤ ⎡ ⎤⎡ ⎤
y0 m0 m1 m2 m3 m4 m5 m6 0 x0
⎢ y 1 ⎥ ⎢ m1 m2 m3 m4 m5 m6 0 ⎥ ⎢
m0 ⎥ ⎢ x 1 ⎥
⎢ ⎥ ⎢ ⎥
⎢y ⎥ ⎢m m1 ⎥ ⎢ ⎥
⎢ 2⎥ ⎢ 2 m3 m4 m5 m6 0 m0 ⎥ ⎢ x2 ⎥
⎢ ⎥ ⎢ ⎥⎢ ⎥
⎢ y 3 ⎥ ⎢ m3 m4 m5 m6 0 m0 m1 m2 ⎥ ⎢ x 3 ⎥
⎢ ⎥=⎢ ⎥⎢ ⎥, (4.15)
⎢ ∗ ⎥ ⎢ m4 m5 m6 0 m0 m1 m2 m3 ⎥ ⎢ 0 ⎥
⎢ ⎥ ⎢ ⎥⎢ ⎥
⎢ ∗ ⎥ ⎢ m5 m6 0 m0 m1 m2 m3 m4 ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥⎢ 0 ⎥
⎣ ∗ ⎦ ⎣ m6 0 m0 m1 m2 m3 m4 m5 ⎣ 0 ⎦
⎦
∗ 0 m0 m1 m2 m3 m4 m5 m6 0

where the ∗ denotes a discarded entry. Hence, the sequences to be FFT’d are:
(m0 , m1 , m2 , m3 , m4 , m5 , m6 , 0) and (x0 , x1 , x2 , x3 , 0, 0, 0, 0), and the output sequence
is (y0 , y1 , y2 , y3 , ∗, ∗, ∗, ∗). The order of the entries in the sequences is very important,
and also don’t forget to negate the frequencies in the transform of the x-sequence.
To summarize: we expand the original data, padding with zeros as necessary to
get a circulant matrix, and then take FFTs.

4.3.2 The Conjugate-Gradient Algorithm

Let us write (4.9) and (4.10) as the operator equation

Y = A X, (4.16)

where Y denotes the known left-hand side and X denotes the vector of unknown
currents. The conjugate gradient (CG) algorithm starts with an initial guess, X0 ,
from which we compute R0 = Y − A X0 , P1 = Q0 = A ∗ R0 , where A ∗ is the adjoint
operator that corresponds to the conjugate-transpose of the matrix blocks in (4.9)
and (4.10). In addition, we have a convergence parameter, ε . Then for k = 1, . . ., if
Test = Rk /Y  < ε , stop; Xk is the optimal solution of (4.16). Otherwise, update
Xk by the following steps:

Sk = A Pk
Qk−1 2
ak =
Sk 2
Xk = Xk−1 + ak Pk
Rk = Rk−1 − ak Sk
Qk = A ∗ R k
Qk 2
bk =
Qk−1 2
Pk+1 = Qk + bk Pk . (4.17)
54 4 Discretization via the Galerkin Method of Moments

The convolution and correlation operations that are a part of A and A ∗ are
evaluated by using the FFT, as described in the preceding section. This, together
with the fact that the storage requirements are reasonably modest, is the reason why
the conjugate gradient algorithm becomes attractive for solving potentially large
problems in our model.

4.4 Comments and Conclusions

We have formulated the volume-integral approach in terms of the Galerkin variant of

the method of moments, in which the unknown anomalous currents and the testing
functions are expressed in terms of basis functions that are defined on a regular
grid. This results in operators that have very special structures; they are either three-
dimensional convolutions or two-dimensional convolutions and one-dimensional
correlations, which means that we can use three-dimensional FFTs to accelerate the
matrix-vector operations occurring within a conjugate-gradient search algorithm.
The use of a highly irregular mesh in the finite-element technique does not allow a
similar advantage in the solution process. In the remainder of this text, we will show
how this formulation produces extremely efficient solutions of complex problems.
We have not gone into certain technical details, such as comparing operation
counts for a direct matrix-vector multiply versus an FFT-assisted operation. Recent
texts, such as [43–45], deal with these issues in more generality, while [46] deals
with other “fast” algorithms for computational electromagnetics.

Appendix

A.1 Two Theorems of Vector Analysis

Let the testing vector function, B(r), have a finite support, then the curl-operator
can be transferred from one vector function to another within an integral:
  
B · ∇ × Adr = A · ∇ × Bdr + ∇ · (A × B)dr

= A · ∇ × Bdr + (A × B) · dS

= A · ∇ × Bdr, (4.18)

where the surface-integral in the second equality appears as a result of Gauss’

divergence theorem. The surface over which this integral is evaluated extends
beyond the support of B(r).
A.1 Two Theorems of Vector Analysis 55

Similarly, we have a transformation of a gradient operator to a divergence

operator:
  
B · ∇V dr = (∇ · BV )dr − V ∇ · Bdr

=− V ∇ · Bdr.
Chapter 5
Computing Network Immittance Functions
from Field Calculations

VIC-3D R
computes impedances from field calculations by using the reaction
concept. We will follow Harrington [47, pp. 116–120], in developing this concept.
Harrington defines reaction as
  
[a, b] = E(a) · J(b) dV , (5.1)

where E(a) is the field due to source a and J(b) is source b. The reciprocity theorem
states that

[a, b] = [b, a] ; (5.2)

that is, the reaction of field a on source b is equal to the reaction of field b on
source a.
Harrington also shows that the driving-point and transfer impedances of a linear
network can be defined in terms of reactions by

[ j, i] [i, j]
zi j = − =− . (5.3)
Ii I j Ii I j

This is a way of relating circuit-theoretic ideas, such as zi j , Ii , I j , with field-theoretic

concepts, such as fields, distributed currents, and reaction.
In order to gain further insight into the relationship between field-theoretic
and circuit models for eddy-current probes, we employ inductively coupled-circuit
theory.

H.A. Sabbagh et al., Computational Electromagnetics and Model-Based Inversion, 57

Scientific Computation, DOI 10.1007/978-1-4419-8429-6 5,
© Springer Science+Business Media New York 2013
58 5 Computing Network Immittance Functions from Field Calculations

5.1 The Classical Bistatic Arrangement

In Fig. 5.1 we show the classical bistatic arrangement. This arrangement subsumes
the usual “driver-pickup” and “remote-field” configurations. The transmitter in
Fig. 5.1 is excited, and the receiver drives an infinite-impedance load. Hence, the
equivalent circuit is as shown in the lower part of Fig. 5.1.
The circuit equations are

V1 = jω L1 I1 + jω M13 I3
0 = jω M13 I1 + (R3 + jω L3 )I3 . (5.4)

I1 is the actual current flowing in the exciting (transmitter) coil and I3 is the
distributed anomalous current due to the flaw. I1 is a circuit current and I3 a
distributed current.
From the second equation in (5.4) we have

Transmit Receive

Flaw

A Classical Bistatic Arrangement

I1

V1 L1 L2 V2

Transmit Receive

L3

Flaw I3

Fig. 5.1 A classical bistatic R3

sensor arrangement, upper,
and its equivalent circuit, Equivalent Circuit of the
lower Classical Bistatic Arrangement
5.1 The Classical Bistatic Arrangement 59

jω M13
I3 = − I1 . (5.5)
R3 + jω L3

Thus, the distributed current is due solely to the exciting current, I1 .

The output voltage, V0 , is given by

V0 = jω M12 I1 + jω M23 I3
[2, 3]
= jω M12 I1 − , (5.6)
I2

where we have once again used (5.3) in writing the second term.
The transfer impedance, Z01 , is defined to be

V0
Z01 =
I1
[2, 3]
= jω M12 −
I1 I2
[2, 3]
= z12 − . (5.7)
I1 I2

Now, z12 represents the direct coupling between the transmitter and receiver; as
such it does not take into account the presence of the flaw, which is represented by
the distributed current, I3 , but, rather, represents the background against which the
flaw must be detected. In remote-field inspection, the transmitter and receiver are
sufficiently far apart (remote) that z12 ≈ 0; that is, the direct coupling is practically
zero.
We can look at this another way, as well. Because z12 has nothing to do with
the flaw, we can subtract its effect initially, even if the bistatic arrangement is not
remote-field. Hence, we get for the impedance change that is due solely to the flaw:

[2, 3]
ΔZ = − , (5.8)
I1 I2

where
  
[2, 3] = E(2) · J(3)dV . (5.9)
flaw

Note that I1 is the current that actually drives the transmitter, but I2 is an apparent
circuit current that drives the receiver, when the receiver is treated as a transmitter,
as well. (It is a feature of reciprocity that it transforms transmitters into receivers
and receivers into transmitters.)
Because I1 and I2 are circuit currents, they are scalars that can be set to unity.
If we normalize I1 and I2 to unity, then E(2) in (5.9) is the incident field within the
60 5 Computing Network Immittance Functions from Field Calculations

flaw, due to one ampere in the receiver coil, when the receiver acts like a transmitter.
Note that J(3) is independent of E(2) , because there is only one source for J(3) ,
namely the one-ampere current in the true transmitter, coil 1.
VIC-3D R
computes the reaction [2, 3] by means of a scalar product of two
discrete vectors. This can be seen by substituting the current expansion, (4.4),
into (5.9):
  
[2, 3] = E(2) · J(3)dV
flaw
   
∑
(3)(x)
= JKLM E (2)(x) (r)T (x)(e) (r)dV
KLM flaw
  
(3)(y)
+ JKLM E (2)(y) (r)T (y)(e) (r)dV
flaw
   
(3)(z) (2)(z) (z)(e)
+ JKLM E (r)T (r)dV
flaw
 
∑
(3)(x) (2)(x) (3)(y) (2)(y) (3)(z) (2)(z)
= JKLM EKLM + JKLM EKLM + JKLM EKLM , (5.10)
KLM

(2)
where EKLM is the KLMth moment of E(2) . This expansion is extremely important
and will appear throughout this text.

5.2 The Differential Probe

The equivalent circuit for a differential-probe system is shown in Fig. 5.2. The
exciting coil is driven by the ac source, while the two sense coils are connected

Sense
I1 Coils +

L2 V2
L1
V1
Op-Amp
Exciting
Coil L3 V3
−

L4
Flaw I4

Fig. 5.2 Equivalent circuit

R4
for the differential probe
5.2 The Differential Probe 61

to an infinite-impedance operational amplifier. This means that neither sense coil

carries a current. The system equations are

V1 = jω L1 I1 + jω M14 I4
0 = jω M14 I1 + (R4 + jω L4 )I4 , (5.11)

where I1 is the actual current flowing in the exciting coil and I4 represents the effects
of the distributed anomalous electric current that flows within the flaw. As before,
I1 is a circuit-current and I4 is a distributed current. The second equation in (5.11)
shows that the distributed current is due only to the exciting current:

jω M14
I4 = − I1 . (5.12)
R4 + jω L4

The output voltage, V0 = V2 − V3, is given by

V0 = V2 − V3 = jω (M12 − M13 )I1 + jω (M24 − M34)I4 . (5.13)

Now, because of the symmetrical placement of the two sense-coils with respect
to the exciting coil, we have M12 = M13 . This holds regardless of the presence of
the flaw, because the flaw effects are included in the second term in (5.13). Hence,
(5.13) becomes

V0 = jω (M24 − M34)I4
[3, 4] [2, 4]
= − . (5.14)
I3 I2

We call the ratio, V0 /I1 = Z01 , the open-circuit transfer impedance of the linear
network consisting of the coupled circuits of Fig. 5.2. In terms of the reactions, then,
we have

[3, 4] [2, 4]
Z01 = − . (5.15)
I1 I3 I1 I2

I1 is the actual current in the exciting-coil, whereas I2 and I3 are fictitious currents
in the two sense-coils, when these coils are treated as transmitters when applying
reciprocity theory. If I1 = I2 = I3 = 1, then (5.15) is expressed solely in terms of
reactions

Z01 = ([3, 4] − [2, 4]) . (5.16)

Each of these reaction terms is interpreted as in the classical bistatic arrangement

discussed earlier.
62 5 Computing Network Immittance Functions from Field Calculations

5.3 Impedance of Ferrite-Core Probes

We calculate the driving-point impedance first. This is the impedance seen at the
terminals of the probe in the absence of the flaw; that is, I4 = 0 in (5.11). This
impedance consists of two parts, the first being the contribution of the coil and
the second the contribution of the ferrite core. We will compute each of these
contributions by using the reaction principle.
(1)
The reaction of field E(1) on source Je is

(1)
[1, 1] = Je · E(1) dV . (5.17)

The source with superscript 1 is the primary source due to the exciting coil. If Ic is
the current in the exciting coil, then the driving-point impedance (or self-impedance)
seen by the coil is
[1, 1]
Z=− . (5.18)
Ic2
If we normalize the excitation to be Ic = 1, then

Z=− Je · E(i) dV

=− (Jc + ∇ × MC ) · E(i) dV
 
=− Jc · E(i) dV − MC · ∇ × E(i) dV
 
=− Jc · E(i) dV + MC · jω B(i) dV , (5.19)

where we have replaced the superscript 1 by (i) to denote incident fields. The
transference of the curl operator in going from the second to the third equation is
valid for M with finite support (see Chap. 4, Appendix A.1). Jc is the current density
in the coil and MC is the magnetization of the core.
Upon substituting the expansions for the magnetic solution vectors, (4.5), into
(5.19), we get

Z=− Jc · E(i)
 
+ jω ∑ MKLM BKLM + MKLM BKLM + MKLM BKLM ,
(x) (ix) (y) (iy) (z) (iz)
(5.20)
KLM

(i)
where BKLM is the KLMth moment of B(i) . The first term is the contribution of the
coil to the driving-point impedance and is computed in Appendix A.1; the second is
the contribution of the core. The scalar product in this term is reminiscent of (5.10).
5.4 Computation of Impedance Changes due to the Presence of a Flaw 63

5.4 Computation of Impedance Changes due to the Presence

of a Flaw

Here, the crucial thing that must be computed is the reaction between the anomalous
current produced by the flaw and the incident field of a coil. In the case of the
differential ferrite-core probe, we want to compute [2, 4] and [3, 4] of (5.16).
(1)
The reaction of field, E(2) , on source, Je , is

(1)
[1, 2] = Je · E(2) dV . (5.21)

The source with superscript 1 is the primary source due to the exciting coil and
superscript 2 denotes scattered fields (and their sources) due to the flaw. If Ic is the
current in the exciting coil, then the change in impedance due to the flaw, as seen by
the coil is

[1, 2] [2, 1]
ΔZ = − =− 2 , (5.22)
Ic2 Ic

where we have used the reciprocity theorem. If we normalize the excitation to be

Ic = 1, then

ΔZ = − Je · E(i) dV

=− (J + ∇ × M) · E(i) dV
 
=− J · E(i) dV − M · ∇ × E(i) dV
 
=− J · E(i) dV + M · jω B(i) dV , (5.23)

where we have dropped the superscript 2 and replaced the superscript 1 by (i) to
denote incident fields. The transference of the curl operator in going from the second
to the third equation is valid for M with finite support (see Chap. 4, Appendix A.1).
Upon substituting the expansion for the magnetic solution vectors, (4.5), and the
corresponding one for the electric current, (4.4), into (5.23), we get

∑
(x) (ix) (y) (iy) (z) (iz)
ΔZ = − JKLM EKLM + JKLM EKLM + JKLM EKLM
KLM
 
(x) (ix) (y) (iy) (z) (iz)
− jω MKLM BKLM + MKLM BKLM + MKLM BKLM . (5.24)

This is a sum of scalar products of the electric current and magnetic polarization
solution vectors with the incident electric field and magnetic flux-density moment
vectors.
64 5 Computing Network Immittance Functions from Field Calculations

Appendix

A.1 Calculation of Circular Coil Impedance

The incident field due to a circular coil lying in region 0 is given by

 
(i)
E0 (r, z) = − jω μ0 2π aφ G00 (r, z; r , z ) · Jc (r , z )r dr dz , (5.25)
coil

where
 ∞ −λ0 |z−z | 
1 e + R0 e−λ0 (z+z )
G00 (r, z; r , z ) = aφ J1 (rl)J1 (r l)ldl . (5.26)
2π 0 2λ0
R0 is the reflection coefficient from the top surface of the workpiece and is calculated
using the ideas presented in Chap. 2.
We assume that the coil has a rectangular cross-section and carries a current,
Ic = 1A, that is uniformly distributed over this cross-section with Nc turns per square
meter (Jc = Nc aφ ). The integrals with respect to (r , z ) over the coil are evaluated
first:

G00 (r, z; r , z ) · Jc (r , z )r dr dz =
coil
 ∞
 
Nc I (r1 l, r2 l) F2 (z1 , z2 , z) + R0 e−λ0 z F1 (z1 , z2 )
J1 (rl)dl , (5.27)
2π 0 l 2λ0

where (r1 , r2 ) are the inner and outer radii of the coil and (z1 , z2 ) the bottom and top
coordinates of the coil.
Then

Zc = − Jc · E(i) dV
 
= −2π Nc E (i) (r, z)rdrdz
 
= jω μ0 (2π )2 Nc rdrdz G00 (r, z; r , z ) · Jc (r , z )r dr dz
coil coil


I (r1 l, r2 l)∞
= jω μ0 2π Nc2 rdrdz
coil 0 l
  
F2 (z1 , z2 , z) + R0 e−λ0 z F1 (z1 , z2 )
× J1 (rl)dl
2λ0
 ∞ 2  
I (r1 l, r2 l) F3 (z1 , z2 ; z1 , z2 ) + R0F12 (z1 , z2 )
= jω μ0 2π Nc 2
dl .(5.28)
0 l3 2λ0
A.1 Calculation of Circular Coil Impedance 65

The functions F1 (z1 , z2 ), F2 (z1 , z2 , z), F3 (z1 , z2 ; z1 , z2 ) and I (r1 l, r2 l) are defined
by [27]
 b
F1 (a, b) = e−λ0 z dz
a

e− λ 0 a − e− λ 0 b
=
λ0
 b
F2 (a, b, c) = e−λ0 |c−z| dz
a

e−λ0 (a−c) − e−λ0 (b−c)

= , c≤a
λ0
2 − e−λ0(c−a) − e−λ0(b−c)
= , a≤c≤b
λ0
e−λ0 (c−b) − e−λ0 (c−a)
= , b≤c
λ0
 z2  z(+)
(−) (+) 
dze−λ0 |z−z |
i
F3 (z1 , z2 , zi , zi ) = dz (−)
z1 zi

= I1 + I2 + I3 + I4
(+)
I1 = −2x1 /λ0 − e−λ0 x1 /λ02 , x1 > 0, x1 = zi − z2
= −eλ0 x1 /λ02 , x1 < 0
(−)
I2 = −2x2 /λ0 − e−λ0 x2 /λ02 , x2 > 0, x2 = zi − z1
= −eλ0 x2 /λ02 , x2 < 0
(−)
I3 = 2x3 /λ0 + e−λ0x3 /λ02 , x3 > 0, x3 = zi − z2
= eλ0 x3 /λ02 , x3 < 0,
(+)
I4 = 2x4 /λ0 + e−λ0x4 /λ02 , x4 > 0, x4 = zi − z1
= eλ0 x4 /λ02 , x4 < 0
 b
I (a, b) = zJ1 (z)dz , (5.29)
a

and J1 (r) is the Bessel function of the first kind, order 1.

66 5 Computing Network Immittance Functions from Field Calculations

A.2 A Coupled-Circuit Model of the Volume-Integral

Equation

Maxwell’s second equation defines the three current systems that we are inter-
ested in:

∇ × H = jω D + Jc + σ (r)E
= jω D + Jc + σh E + (σ (r) − σh)E
= j ω D + Jc + Jh + Ja , (5.30)

where jω D is the displacement current, Jc the coil current, Jh the host current, and
Ja the anomalous current due to the anomalous conductivity, σ (r) − σh .
Setting aside consideration of the displacement current, the three remaining
current systems are going to be modeled by the three coupled circuits of Fig. 5.3 in
order to get a simpler interpretation of the volume-integral equation. Mch = Mhc is
the mutual inductance between the coil and host, Mca = Mac is the mutual inductance
between the coil and anomaly, and Mha = Mah is the mutual inductance between the
host and anomaly.
The loop equations (Kirchoff’s Voltage Law) for this system are

V0 = Ic (Rc + jω Lc ) + jω Mch Ih + jω Mca Ia

0 = jω Mch Ic + (Rh + jω Lh )Ih + jω Mha Ia
0 = jω Mca Ic + jω Mha Ih + (Ra + jω La )Ia . (5.31)

The circuit elements, R and L, should not be taken literally. They represent such
things as energy loss and storage, which occur in both lumped circuits (such as this
one) and distributed fields.

Ic Ih

Rc
+
Mch
Coil V0 Lc Lh Rh Host
−

Mca Mha
Ia
La

Fig. 5.3 A coupled-circuit Ra

model showing the three
current systems, Ic , Ih , and Ia Anomaly
A.2 A Coupled-Circuit Model of the Volume-Integral Equation 67

If we eliminate Ih from the bottom two equations in (5.31), we get an equation

between the coil current and the anomalous current:
ω 2 Mch Mha Ic ω 2 Mha
2 I
a
0 = jω Mca Ic + + + (Ra + jω La )Ia , (5.32)
Rh + jω Lh Rh + jω Lh
which can be rewritten as

ω 2 Mch Mha Ic ω 2 Mha

2 I
a
− jω Mca Ic − = (Ra + jω La )Ia +
Rh + jω Lh Rh + jω Lh
or
Ja
Einc = + E(r) [Ja ] . (5.33)
σa
From this equation it is clear that the incident field and the “integral operator
functional” depend upon the host parameters, as we knew. If the host is freespace,
we are tempted to let Rh → ∞, which would make the host conduction current,
Ih , vanish. We must replace Rh in this case by a capacitor to simulate the host
displacement current, thereby maintaining the presence of the integral operator
functional in the last equation of (5.33). Equation (5.33) illustrates the role played
by the mutual inductances in modeling “action at a distance” terms, such as the
incident field and integral operator. We also see that if Ra → ∞ (which corresponds
to σa → 0), then the anomalous current also vanishes, Ia → 0. The second of (5.31)
also makes clear that the host current, Ih , depends upon both the coil current and
the anomalous current, Ia . The host current is not simply due to Ic alone, unless the
anomaly is not excited.
We can go further and compute the driving-point impedance of the coil. It is a
straightforward procedure to eliminate Ih and Ia in favor of Ic in the bottom two
equations of (5.31) and then substitute this result into the first equation to get a
relation between the driving voltage, V0 , and the loaded coil current, Ic . From this
we obtain the driving-point impedance as
V0 ω 2 Mch
2 (R + j ω L ) − j ω 3 M M M
a a ch ha ca
Zin = = Rc + jω Lc +
Ic (Rh + jω Lh )(Ra + jω La ) + ω 2 Mha
2

ω 2 Mca
2 (R + j ω L ) − j ω 3 M M M
h h ch ha ca
+
(Rh + jω Lh )(Ra + jω La ) + ω 2 Mha
2

ω 2 Mch
2
= Rc + jω Lc +
Rh + jω Lh
ω 2 Mca
2 (R + j ω L )2 − j2ω 3 M M M (R + j ω L ) − ω 4 M 2 M 2
h h ch ca ha h h
+ ! " ch ha .
(Rh + jω Lh ) (Rh + jω Lh )(Ra + jω La ) + ω 2Mha
2

(5.34)

The first two terms in (5.34) are the free-space coil driving-point impedance
and the third is the additional term due to the host in the absence of an anomaly
(the “unflawed host”). This term vanishes when the coil is well removed from the
workpiece, so that Mch → 0.
68 5 Computing Network Immittance Functions from Field Calculations

The final term is the change caused by the presence of an anomaly. It vanishes
under two distinct conditions: Ra → ∞, which, as we have seen before, corresponds
to the absence of the flaw, and Mca , Mha → 0. This condition corresponds to the
vanishing of Einc and the integral operator in (5.33), thereby forcing the anomalous
current, Ia , to vanish. This occurs when the coil is well removed from the anomaly.
Let’s take another look at (5.34). Each of the last two terms can be interpreted
as an “Einc · J” with an appropriate Einc and J. If Ia = 0, then the solution of the
second equation in (5.31) for Ih is Ih = − jω Mch Ic /(Rh + jω Lh ), and when this is
multiplied by Einc = − jω Mch Ic we get the second term in (5.34), after negation and
setting Ic = 1. Hence, we interpret Ih as being the anomalous or “scattering” current
when Ia = 0, and − jω Mch Ic as the incident field acting in free space.
The more usual case in the hierarchy is Ia = 0, which means that (5.33) applies.
If we solve this equation for Ia , and then multiply that result by the left-hand side,
we get the third term in (5.34), after negation and setting Ic = 1. Again, we have
consistency with the theory of calculating changes in impedance, but now Einc must
account for the presence of the host current, Ih , as well as the coil current, Ic . The
environment of each of the scattering currents is different, and that determines how
Einc is to be computed. Note that Einc in (5.33) contains two terms, the first being
the direct contribution of the coil acting in free-space and the second the effect of
the coil acting through the host current, Ih .

A.2.1 Frequency-Response Loci

We can now derive some very useful results for analyzing and interpreting model
results based on the frequency response of the measured impedance. Return to (5.34)
and set Mca = Mha = 0, which corresponds to the unflawed host. Throughout this
discussion we will work with Z  = Zin − Rc , which is done because Rc has nothing
to do with magnetic coupling of the coil to the host. Furthermore, we will work with
Z  /ω Lc , which is the impedance normalized to the free-space reactance of the coil.
It is, of course, a dimensionless quantity.
We have, after rationalizing the third (coupling) term in (5.34),

 2 L /L 
Z ω Mch2 R
h ω 2 Mch h c
= + j 1 −
ω Lc Lc (R2h + ω 2 L2h ) R2h + ω 2 L2h
 2 
ω k2 Lh Rh Rh + ω 2 L2h − ω 2k2 L2h
= 2 +j , (5.35)
Rh + ω 2 L2h R2h + ω 2 L2h

where we have introduced the coupling-coefficient, k2 = Mch /Lc Lh [108, page 399].
A.2 A Coupled-Circuit Model of the Volume-Integral Equation 69

Fig. 5.4 Illustrating the X’

semicircle frequency locus of ω Lc
the normalized impedance
when the circuit elements, Rh , 1
Lh , and k are frequency
independent. This is the
“ideal” Förster plot
k2
2 ω

k2
1−
2

1−k2
R’
ω Lc

It will be advantageous to shift the origin of the coordinate system for the
normalized impedance to j(1 − k2 /2):
   2  
Z k2 ω k2 Lh Rh Rh + ω 2 L2h − ω 2 k2 L2h k2
− j 1− = 2 +j − 1−
ω Lc 2 Rh + ω 2 L2h R2h + ω 2L2h 2
 
k2 2ω Lh Rh R2h − ω 2 L2h
= + j , (5.36)
2 R2h + ω 2 L2h R2h + ω 2 L2h

and from this it is easily shown that

#   #2
# Z k2 ## k4
#
# ω Lc − j 1 − 2 # = 4 . (5.37)

This is an extremely important result and is the basis for interpreting both model
and experimental results. If the circuit parameters, Rh , Lh , and k are all frequency
independent, as in an “ideal” lumped circuit, then the locus of impedance vs.
frequency is a semicircle, the “ideal” Förster plot shown in Fig. 5.4. The diameter
of this semicircle directly measures the coupling coefficient, k2 . The terminal points
of the diameter correspond to ω = 0 and ω → ∞, where R /ω Lc vanishes.
Of course, field calculations of eddy-current phenomena do not lead to “ideal” √
lumped circuit elements. For example, we know that at high frequencies Rh ∼ ω
because of the skin effect.1 This means that R /ω Lc in (5.35) dies out more slowly

1 See [109, pp. 286–303] for a thorough discussion of skin effect on circuit elements.
70 5 Computing Network Immittance Functions from Field Calculations

0 0

−0.01
−0.05 −0.02

−0.03
Normalized X

Normalized X
−0.1
−0.04

−0.05
−0.15
−0.06

−0.2 −0.07

−0.08

−0.25 −0.09
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0 0.01 0.02 0.03 0.04 0.05 0.06
Normalized R Normalized R

Fig. 5.5 Normalized impedance responses over the frequency range of 100 Hz to 1 MHz, for a
given coil and lift-off, but for two half-spaces with different conductivities. The half-space on the
left has σ = 3 × 107 S/m, and that on the right σ = 3 × 105 S/m

with ω , thereby effectively increasing k2 , where we now interpret this parameter to

be the distance between the terminal points of the locus on the X  /ω Lc axis, as in the
ideal case, even though this locus is no longer semicircular. We can anticipate the
increase in k2 due to skin effect on physical grounds; at higher frequencies, the
current in the host material is forced closer and closer to the upper boundary
(the “skin”), thereby increasing the coupling between it and the coil current.
Figure 5.5 shows a model calculation of the normalized impedance function for a
coil over a half-space whose conductivity is σ = 3 × 107 S/m. The frequency range
is 100 Hz to 1 MHz, with the locus of normalized impedance going clockwise. We
clearly see that the normalized impedance response is strongly elongated in the
reactance component. We can take the chord length along this direction from ω = 0
to ω → ∞ and conclude that the “effective” k2 ≈ 0.23.
When we redo the calculation with the same probe over the same frequency
range, but with a half-space whose conductivity is σ = 3 × 105 , we get the result
shown in the right-hand side of Fig. 5.5. What is interesting about these two plots
is that the loci of normalized impedances lie on the same curve. S.N. Vernon [110]
was the first to observe that, given a probe, data collected from materials of very
different resistivities fall on the same curve that she called a “universal impedance
diagram.” She did this empirically; Bowler et al. [111] confirmed this result using
model calculations based on the precursor of VIC-3D R
[25].
It is important to realize, however, that despite the fact that the impedance curve
loci are identical over the entire frequency range of 0 < f < ∞, the responses for
each frequency are considerably different. Compare, for example, that at 1 MHz,
the left-hand curve has virtually closed on the imaginary axis, whereas the right-
hand curve has barely reached the point of maximum normalized R .
Chapter 6
Advanced Probe Models Based on Magnetic
Dipoles and Ferrite Cores

6.1 Modeling Nonstandard Probes

By a “nonstandard probe,” we mean one that is not the usual air-core circular coil
with a square cross-section, whose axis is normal to a plane surface, as in the
pancake coil of Appendix A.1 of Chap. 5, or with its axis coinciding with the axis
of a tube, as in the bobbin coils of Chap. 9. In this chapter, we develop a theory
that allows us to efficiently discretize the current density in such probes, thereby
allowing us to use transfer matrices (these will be discussed in Chap. 10) and similar
mathematical constructs for solving problems.
In particular, this chapter deals with probes that are characterized by their
magnetic-dipole moments, which means that all equivalent electric currents are
solenoidal (divergence-free). These models allow us to calculate inductive effects,
only, which, of course, is a main concern in eddy-current modeling. In the next
chapter, we analyze probes that are driven by electric dipoles, which allows us to
calculate capacitive effects as well as inductive effects. This allows the probe models
to be used over a greater frequency range.

6.2 The Incident Field Due to a Ferrite-Core Probe

We only need to compute the incident electric field at a flaw, because the incident
magnetic field can be expressed in terms of the electric field, as we know.
The Green function that we have worked with previously was that for which
the source and field points were in the same layer of the workpiece. Now, upon
referring to Fig. 2.1 of Chap. 2, we need the Green function, G(±q0) , in which the
source point is in region 0, as before, but the field point is in the qth layer above
region 0, or the qth layer below the region. The first situation carries the label +q0
and the second −q0.

H.A. Sabbagh et al., Computational Electromagnetics and Model-Based Inversion, 71

Scientific Computation, DOI 10.1007/978-1-4419-8429-6 6,
© Springer Science+Business Media New York 2013
72 6 Advanced Probe Models Based on Magnetic Dipoles and Ferrite Cores

We compute the Green functions using the same four-vector algebraic approach
that was used in Chap. 2, and then the resulting volume-integral differential equation
is derived exactly as is explained in Chap. 3. Because the Amperian currents,
which are of magnetic origin, are actually equivalent electric currents [because
they appear in the second of Maxwell’s equations (Ampere’s law)], we only need
to compute the electric–electric dyadic Green functions. The resulting differentio-
integral equations are

Ex (r) = G(q0)(1)(x − x , y − y ; z, z )Jx (r )dr

∂
+ G(q0)(2)(x − x , y − y; z, z )∇t · Jt (r )dr
∂x

1 ∂
+ 2 G(q0)(3) (x − x , y − y; z, z )∇ · J(r )dr
k0 ∂ x

Ey (r) = G(q0)(1)(x − x , y − y ; z, z )Jy (r )dr

∂
+ G(q0)(2)(x − x , y − y ; z, z )∇t · Jt (r )dr
∂y

1 ∂
+ G(q0)(3)(x − x , y − y ; z, z )∇ · J(r )dr
k02 ∂ y

Ez (r) = G(q0)(4)(x − x , y − y ; z, z )Jz (r )dr

1 ∂
+ G(q0)(5) (x − x , y − y; z, z )∇ · J(r )dr , (6.1)
k02 ∂ z

which holds for r in region q. If r is in region −q, then replace (q0) by (−q0).
The result of (6.1) is quite general and could be used, for example, in finding
the interaction between a flaw in region 0 and one in region ±q, as in the Spatial
Decomposition Algorithm of Chap. 10. The Amperian current produced by the core
of such a probe has zero-divergence, as does the current in the exciting coil of the
probe. Hence, the last term in each of the equations in (6.1) vanishes. The electric
currents that flow in the various coils of the probe lie in the transverse, (x, y), plane,
because of the orientation of the coils. This means that ∇t · J(r ) = 0 for these
currents.
Upon calling the coil current, J(c) (r), and the current of magnetic origin (the
Amperian current) J(m) (r), then (6.1) becomes

(c)
Ex (r) = G(q0)(1) (x − x , y − y ; z, z )Jx (r )dr

(m)
+ G(q0)(1) (x − x , y − y; z, z )Jx (r )dr

∂ (m)
+ G(q0)(2) (x − x , y − y ; z, z )∇t · Jt (r )dr
∂x
6.3 A Solenoidal Current Model 73


(c)
Ey (r) = G(q0)(1) (x − x , y − y ; z, z )Jy (r )dr

(m)
+ G(q0)(1) (x − x , y − y; z, z )Jy (r )dr

∂ (m)
+ G(q0)(2) (x − x , y − y; z, z )∇t · Jt (r )dr
∂y

(m)
Ez (r) = G(q0)(4) (x − x , y − y ; z, z )Jz (r )dr . (6.2)

The kernels in (6.1) and (6.2) are computed using the same four-vector approach of
Chap. 2, but because of their complexity will not be shown here.
Equation (6.2) is still quite rigorous, within the context of the nature of the
currents, J(c,m) (r). We shall see shortly that is possible for common eddy-current
problems to eliminate the Ez (r) term, also. This occurs when the coil region is air,
and the field point is in a highly-conducting region, which is the usual combination
in eddy-current problems. The contributions of the coil current in the x- and
y-equation of (6.2) are precisely the classical terms produced by circular coils of
rectangular cross-section that VIC-3D R
already computes. Of course, if the coils
are noncircular, but, say, are rectangular in the (x, y)-plane, then we must evaluate
the contributions in the same manner as in the case of the magnetic currents. See
[92] for a benchmark validation of this model.

6.3 A Solenoidal Current Model

At the typical frequencies of interest in eddy-current NDE, currents that flow in

probes can be assumed to be divergenceless: ∇ · Jc = 0, which implies that we can
write Jc = ∇ × M, with the magnetization vector, M, expanded in edge elements,
as in (4.5) and (4.6) of Chap. 4. We cannot determine M uniquely from the curl
equation without specifying another constraint condition. Helmholtz’ theorem [48]
uses knowledge of ∇ · M as this additional condition, but we will simply set Mx = 0.
This will limit slightly our flexibility in defining arbitrary current sources but will
suffice for our purposes.
Thus, we rewrite expansion (4.8) for the current density in the following form:
 
∑ MKLM
(z) (z)(m) (z)(m)
Jc (x, y, z) = TKLM,y (r)ax − TKLM,x (r)ay
KLM
 
∑ MKLM
(y) (y)(m) (y)(m)
+ TKLM,x (r)az − TKLM,z (r)ax , (6.3)
KLM

(p)(m)
where TKLM,q denotes the qth spatial derivative of the pth-oriented edge element.
The parameters, δx , δy , δz , refer to the source grid, not the flaw grid.
74 6 Advanced Probe Models Based on Magnetic Dipoles and Ferrite Cores

(l+1)*dy (m+1)*dz

l*dy m*dz
y

z
(l-1)*dy (m-1)*dz
(k-1)*dx k*dx (k+1)*dx (k-1)*dx k*dx (k+1)*dx
x x
(z)(m) (z)(m)
Fig. 6.1 Circulations. Left: Tklm,x (r)ay − Tklm,y (r)ax , in the (x, y)-plane, about the edge joining
(y)(m) (y)(m)
the points (xk , yl , zm−1 ) and (xk , yl , zm ); right: Tklm,x (r)az − Tklm,z (r)ax , in the (x, z)-plane, about
the edge joining the points (xk , yl−1 , zm ) and (xk , yl , zm )

(z)(m) (z)(m)
The physical interpretation of the basis vectors is: Tklm,x (r)ay − Tklm,y (r)ax is
a circulation in the x–y plane, about the edge joining the points (xk , yl , zm−1 ) and
(y)(m) (y)(m)
(xk , yl , zm ). Tklm,x (r)az − Tklm,z (r)ax is a circulation in the x–z plane, about the
edge joining the points (xk , yl−1 , zm ) and (xk , yl , zm ). The circulation is confined to
four adjacent cells sharing the edge on the axis of circulation. These circulations are
plotted in Fig. 6.1.
The solution of (6.3) for the expansion coefficients is given by
 yl  xk
(y) 1 (z)
Mklm = Jc (x, y, zm )dxdy
δy yl−1 x0

 zm  xk
(z) 1 (y)
Mklm = − Jc (x, yl , z)dxdz , (6.4)
δz zm−1 x0


(y)
where x0 lies outside the source region. Now that we have MKLM = 0, MKLM ,

(z)
MKLM , we can use the transfer matrices of (10.4) to calculate incident fields on
flaws.
6.4 Response of a Rectangular Coil When Rotated Relative to a Crack 75

6.4 Response of a Rectangular Coil When Rotated Relative

to a Crack

We will apply the theory of nonstandard probes to model an experimental bench-

mark test, thereby validating the modeling algorithm that was developed in the
last section. R.J. Ditchburn and S.K. Burke, writing in NDT&E International 38
(2005), pp. 690–700, show that the maximum response of an infinite crack to a
rectangular coil is maximum when the long axis of the coil is parallel to the crack.
They conjecture that the opposite orientation dependence holds for small cracks.
Here, we demonstrate that their conjecture holds for the case of the coil shown in
Fig. 6.2, and the half-penny model of the fatigue crack shown in Figs. 6.3 and 6.4.
The host material is Inconel 718, which is nonmagnetic.
In Fig. 6.5, we show the driving-point impedance of the coil as it is rotated
through ninety degrees while it remains centered over the crack. The zero angle
corresponds to the case in which the x-axis of the coil coincides with the x-axis of
the crack.

6.4.1 The Anomalous (Scattering) Currents

In order to get a good physical understanding of the response of eddy-current probes,

it is useful to compute and plot the currents within the anomaly, as well as fields
produced by those currents. We will consider here only the anomalous currents.

Coil

Coil
Y − length = 30mils
X − width = 70mils
thickness = 2mils
height = 0.2mils
turns = 1

Fig. 6.2 A rectangular coil

76 6 Advanced Probe Models Based on Magnetic Dipoles and Ferrite Cores

Workpiece : σ = 1.021 x 106 Y

30mils
0.3mil
X
Fatigue Crack

Fig. 6.3 Top view of a half-penny fatigue crack

Fig. 6.4 Side view of the Z

same half-penny fatigue crack X
15mils

Fig. 6.5 Driving-point Response of Rotated Rectangular Coil

impedance of the transmit 0.0008
coil as it is rotates while fixed R
X
over the center of the crack
Z
0.0006
Impedance (Ohms)

0.0004

0.0002

−0.0002

−0.0004
0 10 20 30 40 50 60 70 80 90
Probe Angle (degrees)
6.4 Response of a Rectangular Coil When Rotated Relative to a Crack 77

3
Y-index

0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32
X-index

Fig. 6.6 Projections onto the xy-plane of the currents in the top layer of the anomaly: real
component at 0◦ multiplied by 2.0 × 10−9

Figures 6.6–6.9 illustrate the projections onto the xy-plane of the anomalous
(scattering) currents in the top layer of the crack. The coil is oriented at either 0◦ or
90◦ relative to the crack, as explained above. In order to facilitate the interpretation
of these figures, we have converted Jx into units of cell width and Jy into units of
cell length. By doing this, a current that flows toward the corner of a cell will be
plotted as a vector that points toward the corner of the cell. It is clear, now, why the
response at 90◦ is much larger than at 0◦ ; the anomalous currents are an order of
magnitude larger with the 90◦ orientation than with the 0◦ . Furthermore, it should
be clear that the currents within the very narrow crack are non-solenoidal; i.e., they
have a nonzero divergence. In extreme cases of very thin, very long cracks this can
cause problems with the convergence of the conjugate-gradient solver unless special
preconditioners are used, and the development of such preconditioners for volume-
integral equations is an active research area. Finally, we note that the currents are
predominately oriented in the y-direction, as one might expect for a long, thin crack
with the shape shown in Fig. 6.3.
In order to eliminate the distortion due to the large cell aspect-ratio, we plot
the current in a single cell so that the plotted cell has its true aspect ratio.
Figures 6.10–6.13 are the result.
78 6 Advanced Probe Models Based on Magnetic Dipoles and Ferrite Cores

3
Y-index

0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32
X-index

Fig. 6.7 Projections onto the xy-plane of the currents in the top layer of the anomaly: imaginary
component at 0◦ multiplied by 2.0 × 10−8

3
Y-index

0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32
X-index

Fig. 6.8 Projections onto the xy-plane of the currents in the top layer of the anomaly: real
component at 90◦ multiplied by 2.0 × 10−10
6.4 Response of a Rectangular Coil When Rotated Relative to a Crack 79

3
Y-index

0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32
X-index

Fig. 6.9 Projections onto the xy-plane of the currents in the top layer of the anomaly: imaginary
component at 90◦ multiplied by 1.0 × 10−10

5
Y-index

0
4 5
X-index

Fig. 6.10 Projections onto the xy-plane of the currents in a single cell of the anomaly: real
component at 0◦ multiplied by 2.0 × 10−9
80 6 Advanced Probe Models Based on Magnetic Dipoles and Ferrite Cores

5
Y-index

0
1 2
X-index

Fig. 6.11 Projections onto the xy-plane of the currents in a single cell of the anomaly: imaginary
component at 0◦ multiplied by 2.0 × 10−8

5
Y-index

0
4 5
X-index

Fig. 6.12 Projections onto the xy-plane of the currents in a single cell of the anomaly: real
component at 90◦ multiplied by 2.0 × 10−10

6.5 Validation via Benchmark Experiments

We will use the theory of this chapter to model several benchmark experiments
that are described in the Ditchburn–Burke (D–B) paper referred to above. In
particular, we will show model results for several experiments that involve the 50-
turn rectangular spiral coil whose aspect ratio is 2:1. We use the same dimensions
as in the paper: width = 18.52 mm, length = 36.68 mm, thickness = 8.71 mm,
6.6 Modeling a Differential-Receive Ferrite-Core Probe 81

5
Y-index

0
4 5
X-index

Fig. 6.13 Projections onto the xy-plane of the currents in a single cell of the anomaly: imaginary
component at 90◦ multiplied by 1.0 × 10−10

height = 0.01 mm. This value of height is the same as the thickness of the individual
tracks of copper forming each spiral turn, ≈ 10 μm. With these values, we calculate
the freespace inductance of the coil to be L0 = 48.33 μH, whereas the measured
value is 50.25 μH, the error being less than 4 %.
Next, we model experimental results for the coil over defect-free aluminum and
steel plates. The conductivity of aluminum is 1.653 × 107 S/m, and for steel, 3.472 ×
106 S/m. The relative magnetic permeability of the steel plate is 85. In both cases,
we used the same “fitted lift-off” in the paper: for aluminum, 0.12 mm, and for steel,
0.05 mm.
The results for the aluminum plate are shown in Fig. 6.14 and for steel in
Fig. 6.15.

6.6 Modeling a Differential-Receive Ferrite-Core Probe

The differential-receive ferrite-core probe, often called a “split-core differential”

or “split-D” probe, is used to detect flaws in bolt-holes in aircraft structures or
jet engine disks. We demonstrate a model result for a split-core differential probe
interacting with a flaw in an aluminum substrate. Figure 6.16 shows a model split-
D probe. Note the presence of the “split” ferrite cores and D-shaped receive coils,
which give the probe its name. The response produced by this probe is the transfer
impedance between the transmit (or driver) coil and the series-connected receive
coils.
82 6 Advanced Probe Models Based on Magnetic Dipoles and Ferrite Cores

0.6
Ditchburn and Burke dX/X0

Normalized Impedance Change

0.4 Ditchburn and Burke dR/X0
VIC-3D dX/X0
VIC-3D dR/X0
0.2

−0.2

−0.4

−0.6

−0.8

−1
100 1000 10000 100000 1e+006 1e+007
Frequency (Hz)

Fig. 6.14 VIC-3D

R
model results for the D–B experiment of the coil over a defect-free host of
aluminum

1.2
Ditchburn and Burke dX/X0
Normalized Impedance Change

1 Ditchburn and Burke dR/X0

0.8 VIC-3D dX/X0
VIC-3D dR/X0
0.6
0.4
0.2
0
−0.2
−0.4
−0.6
−0.8
−1
100 1000 10000 100000 1e+006 1e+007
Frequency (Hz)

Fig. 6.15 VIC-3D

R
model results for the D–B experiment of the coil over a defect-free host of
steel

6.6.1 Sizing Surface Cracks: Experimental Test Cases

We’ll apply a model of the split-D probe to the problem of characterizing sur-
face cracks within bolt holes in a benchmark test case. Figure 6.17 illustrates
the test cases, and Fig. 6.18 shows a photomicrograph of an actual through-wall
fatigue crack of the type that we are going to size. The bottom part of Fig. 6.18 shows
the EDM notch that is used for calibrating the data. The eddy-current instrument that
6.6 Modeling a Differential-Receive Ferrite-Core Probe 83

2mm
Transmit Coil:
IR = 9.34mm
OR = 18.4mm
HT = 9mm Receive Coils:
Turns = 408 IR = 7.34mm
OR = 8.34mm
HT = 9mm
Turns = 100

Cores:
R = 7.34mm
HT=9mm
μ=2000

Fig. 6.16 The split-D coil configuration

Fig. 6.17 Illustrating experimental test cases for sizing surface fatigue cracks using a split-D probe
within a bolt hole in an aluminum substrate. The probe (small white circle) is embedded in a
mandrel, which is then rotated azimuthally about a vertical axis that is not shown. Simultaneously,
the mandrel is lifted vertically, thereby generating a two-dimensional raster scan. The top figure
(a) illustrates the cross-section of a through-wall crack in a 0.100 -thick host, the middle figure
(b) illustrates a mid-bore crack in a 0.250  -thick host, and the bottom figure (c) illustrates the
EDM notch in a 0.250  -thick host that is used to calibrate the test setup (see Fig. 6.18)
84 6 Advanced Probe Models Based on Magnetic Dipoles and Ferrite Cores

Fig. 6.18 Illustrating a photomicrograph of an actual fatigue crack extending from the surface
of a bolt hole within an aluminum substrate (top), together with an EDM notch that is used for
calibrating the entire test setup (bottom). The “nominal” width of the crack is 0.007 mm (0.28 mils)
and the width of the notch is 6.1 mils
6.6 Modeling a Differential-Receive Ferrite-Core Probe 85

is used in the tests measures “instrument volts,” which are then scaled to impedances
for further processing by VIC-3D R
using the β -transformation scheme described
in Sect. 15.2.

6.6.2 An Interpolation Algorithm and Inverse Problem

In order to reduce the burden of computing the response of very thin, long cracks, we
will resort to a “surrogate” interpolation algorithm. The theory behind the algorithm
is based on the following formula for the change in impedance of a probe due to
scattering from an anomaly, as determined by fundamental electromagnetic theory
(recall Chap. 5):
  
Za = E0 · Ja dV , (6.5)
Va

where we assume that the probe is excited with a current of one ampere. E0
is the incident field due to the probe in the absence of the anomaly, Ja is the
anomalous (scattering) current, and Va is the volume of the anomaly. E0 and Ja
are both bounded, with bounded derivatives, so that the integral in (6.5) vanishes
to polynomial order in Va as Va → 0. This certainly includes the case where the
cross-section of the anomaly remains fixed and the width, W → 0.
Thus, we can derive an interpolation scheme in W that takes advantage of
this fact. Assuming that W is “small enough,” we can postulate the second-order
approximation, Za = aW + bW 2 . If we compute values of Za at W = 0.125 and
0.25 mil, calling them Z0.125 and Z0.25 , respectively, then it is easy to show that
a = 16Z0.125 − 4Z0.25 , b = −64Z0.125 + 32Z0.25, which yields the final result

Za = (16W − 64W 2 )Z0.125 − (4W − 32W 2 )Z0.25 . (6.6)

NLSE uses higher-order spline interpolations to perform similar surrogate duties

when solving inverse problems.1 What is novel in our next example is to use
interpolation to solve the forward problem, using the value of W that NLSE gives us.
We will solve an inverse problem to reconstruct the width of an 0.018-deep
through-wall crack. We assume the length to be 0.100, as in Fig. 6.17a, and we will
use a block model of the crack. In order to simplify the calculations, however, we
will assume that the crack lies in a half-space, as in Figs. 6.3 and 6.4. Figure 6.19
shows the blending functions, Z0.125 and Z0.25 , for the interpolating grid with nodes
at W = 0, 0.125, 0.250 mil. This allows a quadratic interpolation in NLSE. The
scaled input data for this crack are shown in Fig. 6.20.2

1 This will be fully explained in Chap. 12.

2 These data were supplied by David Forsyth and Mark Keiser of Texas Research Institute Austin.
86 6 Advanced Probe Models Based on Magnetic Dipoles and Ferrite Cores

Blending Functions for Three Values of W Blending Functions for Three Values of W
0.004 0.03
0.25
0.125 0.25
0.003 0 0.02 0.125
0
0.002
Resistance (Ohms)

Reactance (Ohms)
0.01
0.001

0 0

−0.001
−0.01
−0.002
−0.02
−0.003

−0.004 −0.03
−60 −40 −20 0 20 40 60 −60 −40 −20 0 20 40 60
Probe Position (mils) Probe Position (mils)

Fig. 6.19 Blending functions for the interpolation table with nodes at W = 0, 0.125 and 0.25 mil.
The functions associated with the latter two nodes are the Z0.125 and Z0.25 of (6.6), respectively

Scaled Measured Data

0.02

0.015
R
Resistance/Reactance (Ohms)

X
0.01

0.005

−0.005

−0.01

−0.015

−0.02
−60 −40 −20 0 20 40 60
Probe Position (mils)

Fig. 6.20 Showing the scaled measured data taken by the split-D probe. These data will be
submitted to NLSE for inversion to determine W

The result of the inversion is W = 0.0793, and the model results that are obtained
when using the quadratic interpolation formula derived in (6.6) with W = 0.0793
are shown in Fig. 6.21, along with the scaled input data. Clearly, there is a good fit
in reactance. The model resistance shows a higher spatial frequency content (more
wiggles) than the data, but this is probably due to uncertainties in measuring the
resistance. In addition, there remains a slight “shoulder” in the scaled reactance data
that the model does not reproduce, and this may be due in part to the measurements,
6.7 Validation Test: Tangent Coil Over a Crack in a Thin Plate 87

Interpolated Results With W=0.0793 vs Scaled Measured Data

0.02

0.015
R(Data)
Resistance/Reactance (Ohms)

X(Data)
0.01 R(Interp)
X(Interp)
0.005

−0.005

−0.01

−0.015

−0.02
−60 −40 −20 0 20 40 60
Probe Position (mils)

Fig. 6.21 Comparing the model result with W = 0.0793 mil and the scaled data input to NLSE

as well as to an imperfect model that does not include the effects of the finite edges
of the workpiece.
The ability to detect and size a fatigue crack whose width is of the order of
0.05–0.5 mil is enhanced through the use of a split-D probe, rather than a single
absolute coil. The split-D probe is essentially a mathematical differentiator, and
taking the “derivative” of the response past a crack will give a signal that is
sensitive to the width of the crack. More complex applications of surrogate models
in computational electromagnetics are given in [9].

6.7 Validation Test: Tangent Coil Over a Crack

in a Thin Plate

One of the important new features in modeling probes is the ability to rotate coils
in a variety of directions, allowing us, for example, to model circular tangent coils.
Figure 6.22 illustrates a tangent coil over a flawed workpiece, comprising a through-
wall crack in a thin plate. Burke and Rose have used this arrangement to produce
benchmark data which we have used to validate the rotation feature (see [49]).
Figure 6.23 shows a comparison of the benchmark data and a model calculation,
when the tangent coil is scanned over the slot at 2 kHz, and Fig. 6.24 compares the
benchmark data and a calculation of a frequency scan of the tangent coil, when it
is located over the center of the slot. The frequency range extends from 251 Hz to
10 kHz.
88 6 Advanced Probe Models Based on Magnetic Dipoles and Ferrite Cores

COIL

0.90 FLAW
101.0
WIDTH OF SLOT = 0.30mm

Fig. 6.22 Illustrating a tangent coil over a flawed workpiece, corresponding to the benchmark test
of [49]

RESULTS FOR TANGENT COIL OVER SLOT

14
X(data)
X(Vic3D)
12 R(data)
Resistance (R) or Reactance (X) in Ohms

R(Vic3D)
10

−2

−4

−6
0 10 20 30 40 50 60 70 80 90
Scan position (mm)

Fig. 6.23 Comparison of calculations with the position-scan benchmark data of [49]

Clearly, the magnetic-dipole model that we have developed in this chapter is

able to model the interaction of a tangent coil with a through-wall slot over a broad
frequency range, even for a slot as large as 101 mm. The results shown here were
obtained using a modest flaw grid of 128 × 2 × 4 cells, and a very coarse grid of
2 × 4 × 4 cells for the probe.
A.1 Analysis of a Bridge Circuit for a Split-D Probe 89

RESULTS FOR TANGENT COIL OVER SLOT

12
L(data)
L(Vic3D)
10
R(data)
R(Vic3D)
Inductance (mH) or Resistance (Ohms)

−2

−4

−6
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
f (Hz)

Fig. 6.24 Comparison of calculations with the frequency-scan benchmark data of [49]

Appendix

A.1 Analysis of a Bridge Circuit for a Split-D Probe

The split-D coil configuration shown in Fig. 6.16 comprises one transmit and two
receive coils. Another common split-D probe contains no separate transmit coil but
has two identical coils connected in a bridge circuit that serves as the driver of the
two coils.
Figure 6.25 shows a standard bridge circuit that drives the split-D coils, labeled
L1 and L2 , in the presence of a flaw, shown as an inductor, L3 , and resistor, R3 . The
inductance of the two coils is L; M is the mutual inductance between these two coils,
and M13 , M23 , are, respectively, the mutual inductance between L3 and L1 and L3
and L2 . The bridge is open-circuited, and its output voltage is V0 .
The equations for the circuit of Fig. 6.25 are

Vin = I1 (R0 + jω L) + jω MI2 + jω M13 I3

90 6 Advanced Probe Models Based on Magnetic Dipoles and Ferrite Cores

Fig. 6.25 Bridge circuit for a

split-D probe with two
transmit coils R0 R0

Vin I1 + − I2

V0

L 1 L 2

M13 L3 M23

R3
I3

Vin = jω MI1 + I2(R0 + jω L) + jω M23I3

0 = jω M13 I1 + jω M23 I2 + (R3 + jω L3 )I3 , (6.7)

from which we immediately get

jω M13 I1 + jω M23 I2
I3 = − .
R3 + jω L3

When this is substituted into (6.7), we have

   
ω 2 M13
2
ω 2 M13 M23
Vin = I1 R0 + jω L + + I2 jω M +
R3 + jω L3 R3 + jω L3
   
ω 2 M13 M23 ω 2 M23
2
Vin = I1 jω M + + I2 R0 + jω L +
R3 + jω L3 R3 + jω L3
 
ω 2 M13
2
ω 2 M13 M23
V0 = I1 jω L + − − jω M
R3 + jω L3 R3 + jω L3
 
ω 2 M13 M23 ω 2 M23
2
+I2 jω M − jω L + − . (6.8)
R3 + jω L3 R3 + jω L3

Now, if coil 1 and coil 2 are both considered to be transmit coils, then VIC-3D
R

considers them to be in series carrying one ampere. Hence, I1 = I2 = 1 in the

expression for V0 in (6.8), from which follows the result
A.1 Analysis of a Bridge Circuit for a Split-D Probe 91

ω 2 (M13
2 − M2 )
V0 = Zout = 23
. (6.9)
R2 + jω L3

One way to guarantee that the circuit of Fig. 6.25 will behave as a current source
such that I1 ≈ I2 is to make the balance resistance, R0 , much greater than any of
the other impedance elements over the frequency range of interest. In this case,
I1 ≈ I2 ≈ Vin /R0 .
M13 and M23 depend upon the scan position of the probe relative to the flaw. All
other parameters are independent of the probe position. It is clear from (6.9) that if
the flaw is symmetrically placed with respect to the probe, then M13 = M23 , which
means that V0 = 0. Furthermore, because the probe is symmetrical, it follows that V0
will be an antisymmetrical function of the probe position if the flaw is symmetrical
with respect to the probe position. This gives the usual antisymmetrical response
that we expect of a differential probe.
Chapter 7
Advanced Probe Models Based
on Electric Dipoles

7.1 Introduction

Eddy-current measurements of conductivity spectra are needed to gauge the

near-surface residual stress in surface-treated turbine engine components for life
extension, where conductivity and stress are related via the piezoresistive effect.
Broadband (0.1–100 MHz) planar spiral-coil eddy-current probes are currently
being developed using photolithography for the purpose of efficiently measuring the
conductivity spectra for subsequent inversion to residual stress versus depth profiles.
The key problem is that the probes do not perform well at high frequencies
(>30 MHz) due to stray capacitive effects, which begin to dominate as the resonant
frequency of the probe is approached. This parasitic effect is mainly due to
capacitive coupling between adjacent traces of the coil, which makes it especially
difficult to predict in terms of sensitivity. At high frequencies, the lift-off sensitivity
of current spiral coils becomes unstable above about 25 MHz. There is a need,
therefore, to develop models to account for these problematic parasitic effects and
allow for probe optimization in the design stage, rather than through the tedious
and inefficient trial-and-error approach. The key parameters of interest to the probe
designer are coil diameter, trace width, and trace spacing. The probes could be
single-layered or multilayered, circular or rectangular, shielded or unshielded.
In this chapter, we will apply the volume-integral algorithm to develop a
computer code to design spiral-coil eddy-current probes for optimal broadband
performance. This code will allow the designer to predict probe performance as
a function of frequency.

7.2 Development of the Electric-Dipole Mathematical Model

The current version of VIC-3D R

assumes that all applied current sources can be
modeled as magnetic dipoles. This is reasonable because the currents are assumed
to be solenoidal (zero divergence) and flow in parallel planes. Thus, the sources can
H.A. Sabbagh et al., Computational Electromagnetics and Model-Based Inversion, 93
Scientific Computation, DOI 10.1007/978-1-4419-8429-6 7,
© Springer Science+Business Media New York 2013
94 7 Advanced Probe Models Based on Electric Dipoles

Boundary of Anomalous Region

δx
Applied
current
facet
element

I = 1A
x
I = 1A

Fig. 7.1 A printed-circuit rectangular-spiral coil lying within an “anomalous region.” The electric
dipole produced by the driving-point current excites the metallic traces of the coil. The convention
in electric-circuit theory is that current enters the positive terminal of a passive load and leaves
through the negative terminal

be defined by a single magnetic dipole that is oriented normal to the planes. The
field produced by these sources, then, are incident upon the anomalous region of the
problem. This model works well for typical eddy-current problems but allows only
the computation of the inductance of the source. The solenoidal nature of the current
precludes any possibility of computing the driving-point capacitance of the coil. In
order to compute capacitance, and therefore the resonant frequency of the coil, we
must relax the assumption of solenoidal current-flow and assume that the excitation
of the coil is an electric dipole.
We are going to apply our volume-integral algorithm to the metallic traces
contained within the anomalous region shown in Fig. 7.1. These traces are the
“anomaly” of the problem (much like a ferrite core in a conventional eddy-current
probe), and the driving-point terminals form an electric dipole that will drive
the coil.
7.2 Development of the Electric-Dipole Mathematical Model 95

The discretized electric equation is:

⎡ ⎤ ⎡ (x) ⎤(ee) ⎡ (x) ⎤
E(ix) Q 0 0 J
⎣ E(iy) ⎦ = ⎣ 0 Q(y) 0 ⎦ ⎣ J(y) ⎦
E(iz) 0 0 Q(z) J(z)
⎡ (xx) (xy) (xz) ⎤(ee)
G G(0) G(0) ⎡ (x) ⎤
⎢ (0)
(yx) (yy) ⎥
(yz) ⎥
J
−⎢ G G
⎣ (0) (0) (0) ⎦ G ⎣ J(y) ⎦
G
(zx)
G
(zy)
G
(zz) J(z)
(0) (0) (0)
⎡ ⎤(ee)
(xx) (xy) (xz) ⎡ ⎤
G(a) G(a) G(a)
⎢ (yx) (yy) (yz) ⎥ J(x)
−⎢
⎣ G(a) G(a) G(a) ⎦
⎥ ⎣ J(y) ⎦
(zx) (zy) (zz)
G(a) G(a) G(a) J(z)
⎡ ⎤(ee)
(xx) (xy) (xz) ⎡ ⎤
G(b) G(b) G(b)
⎢ (yx) (yy) (yz) ⎥ J(x)
−⎢
⎣ G(b) G(b) G(b) ⎦
⎥ ⎣ J(y) ⎦
(zx) (zy) (zz)
G(b) G(b) G(b) J(z)
⎡ (x) ⎤
 (em) M
− G(0) G(a) G(b) ⎣ M(y) ⎦, (7.1)
M(z)

where the Q’s are tri-diagonal matrices, the G(0) ’s the infinite-space matrices, the
G(a) ’s the convolutional layered-space matrices, and the G(b) ’s the correlational
layered-space matrices. The infinite-space matrices are convolutional, also. The
superscript (ee) denotes electric–electric matrices and (em) denotes electric–
magnetic matrices. The J’s are the unknown electric currents and the M’s are
the unknown magnetic polarization vectors. The last block in (7.1) is simply a
shorthand representation of the three blocks above it, except that it represents
electric–magnetic interactions.
From here on we will ignore magnetic interactions and will combine the various
G-matrices into one. The incident electric field moments on the left-hand are due,
of course, to any independent sources, currents, or charges, as in the usual case of a
separate probe coil. We compute an effective incident electric field-moment due to
the dipole in the following way. Assume that the positive terminal in Fig. 7.1 carries
1A of current at cell k+ l + m+ and that the negative terminal carries 1A of current
(x)
at cell k− l − m− . Then in the configuration shown in Fig. 7.1, we have Jk+ l + m+ =
(y) (z) (x) (y) (z)
−1/δ yδ z, Jk+ l + m+ = Jk+ l + m+ = 0, and Jk− l − m− = 1/δ yδ z, Jk− l − m− = Jk− l − m− = 0,
where δ x, δ y, δ z are the dimensions of a cell in the anomalous region. Then, from
(7.1) we have for the incident field moments
96 7 Advanced Probe Models Based on Electric Dipoles

  1
(ix) (xx) (xx)
Eklm = −Gklm,k+ l + m+ + Gklm,k− l − m− ×
δ yδ z
  1
(iy) (yx) (yx)
Eklm = −Gklm,k+ l + m+ + Gklm,k− l − m− ×
δ yδ z
  1
(iz) (zx) (zx)
Eklm = −Gklm,k+ l + m+ + Gklm,k− l − m− × . (7.2)
δ yδ z
This electric-dipole moment vector can be thought of as the “constraint electric
field” that arises because of the unit-constrained current at k+ l + m+ and k− l − m− .
Now that we have transformed the known current-density vectors at k+ l + m+ and
k− l − m− into the incident electric-field moments (this is akin to using the G-matrices
as “transfer matrices”), we are left with an anomalous current of zero at these
cells. After all, a current at a location is either known or unknown (anomalous),
but not both. Thus, if the current is known, then the anomalous current must be
zero. The way to force the anomalous current to vanish is to make the cells at
k+ l + m+ , k− l − m− to contain only host material. Hence, with this proviso, the
problem can be solved completely as a standard problem for any anomalous region,
using (7.2) as the incident field. In this case, the anomalous region of Fig. 7.1 will
contain the source dipole.
The position of the applied current facet elements (plus and minus) are included
within the grid for the anomalous region. The magnitude of the current density at
the boundary of the cell for the + terminal is −1/δ yδ z, and the magnitude of the
current density for the − terminal is 1/δ yδ z.
Once we have the J’s, we can compute the change in driving-point impedance
seen at the terminals of the unit-current source in the usual way as the dot-product
of the J’s with the incident field moments:
 
(x) (ix) (y) (iy) (z) (iz)
Δ Z = −Σklm Jklm Eklm + Jklm Eklm + Jklm Eklm
1 $  
(x) (xx) (xx)
= Σklm Jklm −Gklm,k+ l + m+ + Gklm,k− l − m−
δ yδ z
 
(y) (yx) (yx)
+Jklm −Gklm,k+ l + m+ + Gklm,k− l − m−
 %
(z) (zx) (zx)
+Jklm −Gklm,k+ l + m+ + Gklm,k− l − m− (7.3)

The unloaded driving-point impedance of the dipole is given by taking the

“reaction” of the driving-point current with itself, i.e., by taking the dot product
of the driving-point current with the electric field moment produced by that current:
 2 
(0) 1 (xx) (xx) (xx)
Z =− Gk+ l + m+ ,k+ l + m+ + Gk− l − m− ,k− l − m− − Gk+ l + m+ ,k− l − m−
δ yδ z

(xx)
−Gk− l − m− ,k+ l + m+ . (7.4)

The driving-point impedance of the loaded dipole is the sum of (7.3) and (7.4).
7.3 An Example 97

7.3 An Example

Consider the 8 × 8 × 2 mm grid with 8 × 8 × 2 cells shown in Fig. 7.2. If the bottom
layer comprises only blue (host) cells, then the current is constrained to flow only
in the top layer, which means that there will be no z-directed currents at all within
the grid.
The current enters cell 3, which makes it the positive terminal, and leaves cell
1, making it the negative terminal. Thus, cells 1 and 3 make up an electric dipole
that will excite the remaining cells. In order to determine the excitation source, we
constrain the expansion coefficient for the x-directed current at the facet between
cells 3 and 4 to be −1 × 106 A/m2 and for the x-directed current at the facet between
cells 1 and 2 to be +1 × 106 A/m2 . Clearly, the current in cells 1 and 2 and 3 and 4
is constrained to flow only in the x-direction.
At frequencies that are low enough that this body is electrically large, we can
model this structure by the electrical equivalent circuit shown in Fig. 7.3. The
capacitor, C0 , is associated with the stored electric energy within the gap between
cells 1 and 3 of the grid, the resistor is associated with energy loss in the metal,
and the inductance is associated with the stored magnetic energy within the area
enclosed by the coil.

I I
X : σ = 0 S/m
(Host)
: σ = 1E7 S/m
1 3
(−) (+)

2 4

Fig. 7.2 Illustrating a square coil on an 8 × 8 × 2 grid

98 7 Advanced Probe Models Based on Electric Dipoles

Fig. 7.3 An equivalent +(3) R0

electric circuit for the I
structure shown in Fig. 7.2 at
low frequencies

C0 L0

−(1)

7.4 Another Example

Consider the same problem as before, but this time let the excitation current lie in
the y-direction, as shown in Fig. 7.4. In this case, the excitation currents lie in the
y-direction, and we show that the sources are limited to a single cell. Hence, for the
(−) terminal, the current is a “ramp” function, and in the (+) terminal, it is a “slide”
function. Thus, in computing the incident field vectors in (7.2), we would use only
the appropriate “ramp” or “slide” portions of the matrix elements.
The electric-dipole vector oriented in the y-direction is a little clearer now,
because of the continuity equation for electric charge:

∂ρ
+∇·J = 0 , (7.5)
∂t
from which we deduce that
1
ρ=− ∇·J
jω
109
=− Q/m3 (−) terminal
jω
109
=+ Q/m3 (+) terminal. (7.6)
jω

7.5 Initial Results

Figures 7.5–7.11 show the results of the dipole calculation for the example of
Fig. 7.2, but using ramp functions at the terminals (cells 1 and 3) instead of tent
functions. The tent function that spans cells 1 and 2, as well as the tent function that
7.5 Initial Results 99

To (−) terminal X From (+) terminal : σ = 0 S/m

of voltage source of voltage source (Host)
(−) (+)
: σ =1E7 S/m

Applied Currents

1 x 106 A/m2
Jy in cell(−) Jy in cell(+)

Fig. 7.4 Illustrating a square coil on an 8 × 8 × 2 grid. The excitation current lies in the y-direction
this time, and the nature of the excitation currents within their cells is shown

spans cells 3 and 4 are treated as unknowns to be computed. These are “freespace”
results—there is no workpiece, and the grid has 8 × 8 × 2 cells. The frequency of the
broad resonance increases as the gap increases, which suggests that the associated
capacitance decreases (see the discussion of Fig. 7.12). The sharp resonances, on the
other hand, do not change frequency with the gap, which suggests that they may be
transmission-line effects associated with the length of the loop.
Figure 7.12 is a plot of the terminal capacitance as a function of the distance
between the terminals, as deduced from the frequency of the zero-crossing of the
broad resonance in the reactance functions. This result is reasonable, because we
expect this capacitance to decrease with an increased separation of the terminals.
In addition, this figure includes the capacitance of the “naked terminals” as a
function of the gap between the terminals. These capacitances are computed from
the reactances (not shown) of the naked terminals at 108 Hz. The latter capacitances
are smaller than the former since they include only the ramp functions of the
terminal and not the tent functions that complete the terminal posts.
We can do a simple analysis of the circuit of Fig. 7.3 and estimate the parameters
from the frequency responses of the driving-point impedances that we have just
computed. The driving-point impedance is given by

1
Zin =
1
jω C0 +
R0 + jω L0
100 7 Advanced Probe Models Based on Electric Dipoles

Resistance vs Frequency
1
Gap: 6 cells
5 cells
4 cells
3 cells
Resistance (Ohm) 2 cells
0.1 1 cells
Closed loop

0.01

0.001
1e+007 1e+008 1e+009
Frequency (Hz)

Reactance vs Frequency
100
Gap: 1 cell
2 cell
3 cell
4 cell
Reactance (Ohm)

5 cell
6 cell
10 Closed loop
aircore

1e+007 1e+008 1e+009

Frequency (Hz)

Fig. 7.5 Results of the dipole model calculation in freespace at frequencies below the first
resonance. The curve labeled “aircore” is the result of VIC-3D R
’s standard aircore calculation
for a single-loop coil enclosing the same area. Top: resistance. Bottom: reactance

R0 + jω L0
=
1 − ω 2L0C0 + jω R0C0
R0 + jω (L0 [1 − ω 2L0C0 ] − R20C0 )
= . (7.7)
(1 − ω 2L0C0 )2 + (ω R0C0 )2
7.5 Initial Results 101

Resistance vs Frequency
14000
Gap: 1 cell
12000
Resistance (Ohm)
10000

8000

6000

4000

2000

5e+009 6e+009 7e+009 8e+009

Frequency (Hz)

Reactance vs Frequency
8000

6000
Gap: 1 cell
4000
Reactance (Ohm)

2000

−2000

−4000

−6000

−8000
5e+009 6e+009 7e+009 8e+009
Frequency (Hz)

Fig. 7.6 Results for the 8 × 8 × 2 grid, one-cell-gap (1 mm) model. Top: resistance. Bottom:
reactance

Resonance occurs at that frequency for which the total reactance vanishes, or L0 (1 −
ωR2 L0C0 ) = R20C0 , from which

&
1 L0 − C0 R20
ωR =
L0C0 L0

1
≈ , (7.8)
L0C0

for R20C0  L0 .
102 7 Advanced Probe Models Based on Electric Dipoles

Resistance vs Frequency
10000

8000

Resistance (Ohm)
6000
Gap: 2 cells
4000

2000

5e+009 6e+009 7e+009 8e+009

Frequency (Hz)

Reactance vs Frequency
8000

6000
Gap: 2 cells
4000
Reactance (Ohm)

2000

−2000

−4000

−6000

−8000
5e+009 6e+009 7e+009 8e+009
Frequency (Hz)

Fig. 7.7 Results for the 8 × 8 × 2 grid, two-cell-gap (2 mm) model. Top: resistance. Bottom:
reactance

We can estimate R0 as follows: the coil conductivity is 1 × 107 S/m, and the cells
are cubes of 1 mm on a side. Thus, a path of N cells will have a resistance of

R0 = ρ × l/A
= 1 × 10−7 × N × 1 × 10−3/(1 × 10−6)
= N × 1 × 10−4 Ω , (7.9)

from which we develop the following table (Table 7.1).

7.5 Initial Results 103

Resistance vs Frequency
10000
Gap: 3 cells
8000
Resistance (Ohm)
6000

4000

2000

5e+009 6e+009 7e+009 8e+009

Frequency (Hz)

Reactance vs Frequency
8000

6000
Gap: 3 cells
4000
Reactance (Ohm)

2000

−2000

−4000

−6000

−8000
5e+009 6e+009 7e+009 8e+009
Frequency (Hz)

Fig. 7.8 Results for the 8 × 8 × 2 grid, three-cell-gap (3 mm) model. Top: resistance. Bottom:
reactance

These results are in good agreement with the low-frequency resistances plotted
in Figs. 7.5 and 7.16.
We can estimate L0 from the reactance at 1 × 108 Hz in Fig. 7.5: L0 = XL /
(2π × 1 × 108) (Table 7.2).
Finally, we estimate the capacitance from the resonant frequencies as shown in
(7.8). These are the “Full Circuit” capacitance values that are plotted in Fig. 7.12
(Table 7.3).
Figure 7.13 shows the change in impedance when a 1 mm plate is placed below
the loop, and Figs. 7.14 and 7.15 show the results of a lift-off scan at 10 MHz. The
plotted results in the latter two figures are the change due to the plate.
104 7 Advanced Probe Models Based on Electric Dipoles

Resistance vs Frequency
10000
Gap: 4 cells
8000
Resistance (Ohm)
6000

4000

2000

5e+009 6e+009 7e+009 8e+009

Frequency (Hz)

Reactance vs Frequency
8000

6000
Gap: 4 cells
4000
Reactance (Ohm)

2000

−2000

−4000

−6000

−8000
5e+009 6e+009 7e+009 8e+009
Frequency (Hz)

Fig. 7.9 Results for the 8 × 8 × 2 grid, four-cell-gap (4 mm) model. Top: resistance. Bottom:
reactance

7.6 Results With a Grid Of 16×16×4 Cells

Figure 7.16 shows the low-frequency response of the refined grid, which should be
compared with the corresponding results of Fig. 7.5. Finally, we show in Fig. 7.17
the impedance response of the “naked” terminal-pair.
The result for the “naked” terminal pair is interesting. It represents a resistor,
whose resistance varies with frequency, in series with an object whose log-log
plot varies linearly (for the most part), with a negative slope. Thus, this object’s
7.6 Results With a Grid Of 16×16×4 Cells 105

Resistance vs Frequency
10000
Gap: 5 cells
8000

Resistance (Ohm)
6000

4000

2000

5e+009 6e+009 7e+009 8e+009

Frequency (Hz)

Reactance vs Frequency
8000

6000
Gap: 5 cells
Reactance (Ohm)

4000

2000

−2000

−4000

−6000

−8000
5e+009 6e+009 7e+009 8e+009
Frequency (Hz)

Fig. 7.10 Results for the 8 × 8 × 2 grid, five-cell-gap (5 mm) model. Top: resistance. Bottom:
reactance

frequency response is a simple negative-power law that obviously represents a

capacitor. In order to compute the capacitance of this object, we take the logarithm
of the usual expression, XC = 1/(2π × f × C), and get log XC = − log 2π − log f −
logC. Now, the reactance (not shown) of the naked terminals is XC = 5.06 × 104 Ω
when f = 1 × 108 Hz, so that logC = −13.489, or C = 3.24 × 10−14 which is in
good agreement with the corresponding one-cell results of Fig. 7.12. The lack of
smoothness in the resistance plot of Fig. 7.17 is due to lack of precision in the
calculations.
106 7 Advanced Probe Models Based on Electric Dipoles

Resistance vs Frequency
10000
Gap: 6 cells
8000
Resistance (Ohm)
6000

4000

2000

5e+009 6e+009 7e+009 8e+009

Frequency (Hz)
Reactance vs Frequency
8000

6000
Gap: 6 cells
4000
Reactance (Ohm)

2000

−2000

−4000

−6000

−8000
5e+009 6e+009 7e+009 8e+009
Frequency (Hz)

Fig. 7.11 Results for the 8 × 8 × 2 grid, six-cell-gap (6 mm) model. Top: resistance. Bottom:
reactance

7.7 Computation of the Divergence of the Currents

Figure 7.18 shows the divergence of the current in the loop with a 1-cell (1 mm)
gap at 10 MHz, and Fig. 7.19 shows the same thing at 1 GHZ. In both of these
figures, Jx flows out of the terminal at 0 < x < 0.001, 0.002 < y < 0.003 and into
the terminal at 0 < x < 0.001, 0.004 < y < 0.005. The units are A/m3 . Figure 7.20
shows the divergence of the current in the closed-loop at 10 MHz. The two driving
tent functions span the cells in the region 0.002 < x < 0.005, 0 < y < 0.001 and
flow in the positive x-direction. The units are A/m3 .
7.7 Computation of the Divergence of the Currents 107

Capacitance vs Terminal Separation

1e-013
Full Circuit
9e-014 Naked terminal

Capacitance (F) 8e-014

7e-014

6e-014

5e-014

4e-014

3e-014

2e-014
0 1 2 3 4 5 6 7
Terminal Gap Size (mm)

Fig. 7.12 Showing the “loaded” and “naked” terminal capacitance as a function of the distance
between the terminals

Table 7.1 Estimating R0 Test N (cells) R0 (Ω)

1 27 2.7 × 10−3
2 26 2.6 × 10−3
3 25 2.5 × 10−3
4 24 2.4 × 10−3
5 23 2.3 × 10−3
6 22 2.2 × 10−3

Table 7.2 Estimating L0 Test XL (Ω) L0 (H)

1 6.078 9.678 × 10−9
2 5.962 9.494 × 10−9
3 5.918 9.424 × 10−9
4 5.897 9.390 × 10−9
5 5.990 9.538 × 10−9
6 6.100 9.713 × 10−9

Table 7.3 Estimating C0 Test f R (Hz) L0 (H) C0 (F)

1 5.46 × 109 9.678 × 10−9 8.79 × 10−14
2 5.92 × 109 9.494 × 10−9 7.62 × 10−14
3 6.19 × 109 9.424 × 10−9 7.02 × 10−14
4 6.50 × 109 9.390 × 10−9 6.39 × 10−14
5 6.66 × 109 9.538 × 10−9 5.99 × 10−14
6 6.78 × 109 9.713 × 10−9 5.68 × 10−14
108 7 Advanced Probe Models Based on Electric Dipoles

Fig. 7.13 Change in Resistance Change vs Frequency

impedance of the dipole 0.2
model calculation for the loop
over a 1 mm plate. Top: Gap: 1 cell

Resistance Change (Ohm)

resistance. Bottom: reactance Closed loop
0.15 aircore

0.1

0.05

0
1e+007 1e+008 1e+009
Frequency (Hz)

Resistance Change vs Frequency

200
Gap: 1 cell
Closed loop
150
Reactance Change (Ohm)

aircore

100

50

−50

−100
1e+007 1e+008 1e+009
Frequency (Hz)

Note that the 1-mm gap coil and the closed-loop coil have comparable divergence
values at 10 MHz, but the response of the 1-mm gap coil at 1 GHz is orders-
of-magnitude larger, as we would expect. At this frequency resonance effects, as
manifested by the increased capacitive reactance, begin to assert themselves, and
we always associate capacitive effects with charge buildup, which is the same as
divergence of the current.
7.8 Three-Turn Spiral Coil 109

Resistance Change vs Liftoff

0.014

Resistance Change (Ohm)

0.012 Gap: 1 cell

0.01

0.008

0.006

0.004

0.002

0
0.1 1 10 100 1000
Liftoff (mm)

Reactance Change vs Liftoff

0

−0.05 Gap: 1 cell

Reactance Change (Ohm)

−0.1

−0.15

−0.2

−0.25

−0.3

−0.35

−0.4
0.1 1 10 100 1000
Liftoff (mm)

Fig. 7.14 Change in impedance of the dipole model calculation over a 1 mm plate as a function of
lift-off at 10 MHz. Top: resistance. Bottom: reactance

7.8 Three-Turn Spiral Coil

Figure 7.21 shows a three-turn spiral coil, whose results are displayed in the next few
figures. Figure 7.22 shows the frequency scan of this coil in freespace, and Fig. 7.23
shows these same results below the resonance around 2 GHz. Finally, Figs. 7.24 and
7.25 illustrate the results of the dipole model of the three-turn spiral coil when a
1 mm plate is placed below the coil.
110 7 Advanced Probe Models Based on Electric Dipoles

dX vs dR
0

−0.05 Gap: 1 cell

Reactance Change (Ohm)
−0.1

−0.15

−0.2

−0.25

−0.3

0 0.002 0.004 0.006 0.008 0.01

Resistance Change (Ohm)

Fig. 7.15 Impedance-plane plot of the change in impedance of the dipole model calculation over
a 1 mm plate as a function of lift-off at 10 MHz

7.9 A Two-Layered Spiral Coil

Figure 7.26 shows a two-layered spiral coil whose dimensions are 0.8 mm ×
0.8 mm × 0.15 mm. The low-frequency response of this coil is shown in Fig. 7.27,
and the response through resonance is shown in Fig. 7.28. Figure 7.29 shows the
high-frequency response of the two-layered coil, in which two weaker resonances
appear.
The conductivity of the metallic traces is σ = 1 × 107 S/m, as before. Given the
dimensions that were stated above, we compute the resistance of each voxel to be
Rvoxel = 0.002 Ω, and because there are 260 voxels connected in series in Fig. 7.26,
we compute the dc resistance of the coil to be R0 = 0.520 Ω, which is in excellent
agreement with the low-frequency resistance shown in Fig. 7.27.
Using the results of Figs. 7.27 and 7.28, we can estimate the other elements of the
equivalent circuit of Fig. 7.3. From Fig. 7.27 we estimate the inductance to be L0 =
1.59 × 10−8H, and from Fig. 7.28 we estimate the resonant frequency to be 6.4 ×
109 Hz, or ω0 = 40.2 × 109rad/sec. Using these results we compute C0 = 1/(L0 ×
ω02 ) = 3.89 × 10−14Fd. For a high-Q parallel resonant circuit of this type, the peak
resistance at resonance is given by Rres = L0 /(C0 R0 ) = 7.86 × 105 Ω, which is in
excellent agreement with the result shown in Fig. 7.28.
7.10 A Model of a THz Transmitter on GaAs 111

Resistance vs Frequency
0.1

Gap: 1 mm

Resistance (Ohm)

0.01

0.001
1e+007 1e+008 1e+009
Frequency (Hz)

Reactance vs Frequency
100

Gap: 1 mm
Reactance (Ohm)

10

1e+007 1e+008 1e+009

Frequency (Hz)

Fig. 7.16 Results of the 16 × 16 × 4-cell 1-mm gap model at low frequencies. Top: resistance.
Bottom: reactance

7.10 A Model of a THz Transmitter on GaAs

In [93], the authors describe a high-efficiency transmitter structure for terahertz

(THz) frequencies. The antenna consists of two 10 μm-wide metal lines deposited
on semi-insulating GaAs, with a separation of 100 μm. Two metal tabs extend out
from these lines toward each other. Figure 7.30 illustrates the structure. We have
modeled this system by using a 16 × 16 × 2-grid (160 μm × 160 μm × 20 μm) placed
on a 1 mm thick plate having the nominal parameters for GaAs, σ = 100 S/m and
εrel = 10.9 [94].
112 7 Advanced Probe Models Based on Electric Dipoles

(-)Terminal Resistance vs Frequency

100

10 Gap: 1 mm

Resistance (Ohm) 1

0.1

0.01

0.001

0.0001

1e-005

1e-006
2e+008
Frequency (Hz)

(-)Terminal Reactance vs Frequency

100000 Gap: 1 mm
Reactance (Ohm)

10000

1000

100

10

1
2e+008
Frequency (Hz)

Fig. 7.17 Results for the “naked” driving-point terminal response for the 16 × 16 × 4-cell 1-mm
gap model. Top: resistance. Bottom: reactance

The results of the impedance calculation are shown in Figs. 7.31 and 7.32. It
is clear from the high-frequency results shown in the bottom of Fig. 7.32 that the
impedance response can be modeled by the series-parallel circuit of Fig. 7.33 up to
a frequency of about 8 × 1011 Hz. We can estimate the parameters of Fig. 7.33 by
using the data shown in Figs. 7.31 and 7.32.
7.10 A Model of a THz Transmitter on GaAs 113

’divJsurf_10MHz.txt’ index 0 using 1:2:4

div(J) (A/m**3)

500

−500

0.008
0.007
0.006
0.005
0.004
0.003
0.002 x(m)
0.001
0.008 0.007 0.006 0.005 0
0.004 0.003 0.002 0.001 0
y(m)

’divJsurf_10MHz.txt’ index 0 using 1:2:5

div(J) (A/m**3)

−2

0.008
0.007
0.006
0.005
0.004
0.003 x(m)
0.002
0.001
0.008 0.007 0.006 0
0.005 0.004 0.003 0.002 0.001 0
y(m)

Fig. 7.18 Divergence of the current of the 1-mm gap model at 10 MHz. Top: real. Bottom:
imaginary
114 7 Advanced Probe Models Based on Electric Dipoles

’divJsurf_1GHz.txt’ index 0 using 1:2:4

div(J) (A/m**3)

5e+006

−5e+006

0.008
0.007
0.006
0.005
0.004
0.003
0.002 x(m)
0.008 0.001
0.007 0.006 0.005 0
0.004 0.003 0.002 0.001 0
y(m)

div(J) (A/m**3) ’divJsurf_1GHz.txt’ index 0 using 1:2:5

10000

5000

−5000

−10000

0.008
0.007
0.006
0.005
0.004
0.003 x(m)
0.002
0.001
0.008 0.007 0.006 0.005 0
0.004 0.003 0.002 0.001 0
y(m)

Fig. 7.19 Divergence of the current of the 1-mm gap model at 1 GHz. Top: real. Bottom: imaginary
7.10 A Model of a THz Transmitter on GaAs 115

’divJsurf_10MHz.txt’ index 0 using 1:2:4

div(J) (A/m**3)

500

−500

0.008
0.007
0.006
0.005
0.004 y(m)
0.003
0.002
0.001
0 0.001 0.002 0
0.003 0.004 0.005 0.006 0.007 0.008
x(m)

’divJsurf_10MHz.txt’ index 0 using 1:2:5

div(J) (A/m**3)

−2

0.008
0.007
0.006
0.005
0.004 y(m)
0.003
0.002
0.001
0 0.001 0.002 0
0.003 0.004 0.005 0.006 0.007 0.008
x(m)

Fig. 7.20 Divergence of the current of the closed-loop model at 10 MHz. Top: real. Bottom:
imaginary
116 7 Advanced Probe Models Based on Electric Dipoles

Fig. 7.21 A three-turn spiral coil. The bottom layer of the three-layer system contains the spiral
coil (red). The middle layer (upper right) contains a single via (red) that connects the bottom to
the top layer (upper left), which contains the lead (red) to the negative terminal on the left edge.
The positive terminal is at the upper left corner of the bottom-most layer. This makes the input gap
one-layer (the middle layer) high in the z-direction, and five cells wide along the left edge

1
C0 ≈
2π × 109 × 24,000
= 6.63 × 10−15 Fd
R ≈ 70 Ω
1
L≈
(2π × 3 × 1011)2 × 6.63 × 10−15
= 4.25 × 10−11 H
1
C1 ≈
(2π × 5 × 1011)2 × 4.25 × 10−11
= 2.38 × 10−15 Fd . (7.10)
7.10 A Model of a THz Transmitter on GaAs 117

Resistance vs Frequency
100

80 spiral coil
Resistance (Ohm)
60

40

20

−20
1e+007 1e+008 1e+009 1e+010
Frequency (Hz)

Reactance vs Frequency

0
Reactance (Ohm)

−2000 spiral coil

−4000

−6000

−8000

1e+007 1e+008 1e+009 1e+010

Frequency (Hz)

Fig. 7.22 Frequency scan of the three-turn spiral coil of Fig. 7.21 in freespace. Top: resistance.
Bottom: reactance

We argue that because C1  C0 , the shunt capacitor is an open circuit at the

lower frequencies, which means that the overall impedance response will be that
of a series RLC resonant circuit. That is the condition shown with the resonance at
about 3 × 1011 Hz. Immediately beyond the series resonance, the series branch is
inductive, and resonates with the shunt capacitor, C1 . The typical characteristic of a
parallel resonant circuit, in which the resistance rises significantly in the vicinity of
the reactance cross-over point, is clearly shown, with the parallel resonant frequency
118 7 Advanced Probe Models Based on Electric Dipoles

Resistance vs Frequency
0.3
spiral coil
0.25

Resistance (Ohm) 0.2

0.15

0.1

0.05

0
1e+007 1e+008 1e+009
Frequency (Hz)

Reactance vs Frequency
600

500
spiral coil
Reactance (Ohm)

400

300

200

100

−100
1e+007 1e+008 1e+009
Frequency (Hz)

Fig. 7.23 Results of the dipole model calculation of the three-turn spiral coil in freespace below
the resonance around 2 GHz. Top: resistance. Bottom: reactance

being approximately 5 × 1011 Hz. The wiggles beyond 8 × 1011 are attributed to
parasitic elements that are not easily represented in a simple lumped equivalent
circuit. This response is contrary to that shown in the earlier models, in which the
parallel resonance occurs first. The reason for that is that the simple closed-loop
circuit of the earlier models has a significant inductance, which dominates the low-
frequency response, contrary to the THz transmitter, which is simply a pair of open
terminals.
7.10 A Model of a THz Transmitter on GaAs 119

Resistance With and Without Plate

100

80
Resistance (Ohm) Without Plate
With Plate
60

40

20

−20
1e+007 1e+008 1e+009 1e+010
Frequency (Hz)

Reactance With and Without Plate

Without Plate
Reactance (Ohm)

−2000 With Plate

−4000

−6000

−8000

1e+007 1e+008 1e+009 1e+010

Frequency (Hz)

Fig. 7.24 Impedance of the dipole model calculation of the three-turn spiral coil in freespace and
over a 1 mm plate. Top: resistance. Bottom: reactance
120 7 Advanced Probe Models Based on Electric Dipoles

Resistance Change vs Frequency

1.5

Resistance Change (Ohm)

spiral coil
1

0.5

−0.5
1e+007 1e+008 1e+009
Frequency (Hz)

Reactance Change vs Frequency

100
Reactance Change (Ohm)

−100
spiral coil
−200

−300

−400

−500
1e+007 1e+008 1e+009
Frequency (Hz)

Fig. 7.25 Change in impedance of the dipole model calculation of the three-turn spiral coil over a
1 mm plate at low frequencies. Top: resistance. Bottom: reactance
7.10 A Model of a THz Transmitter on GaAs 121

Fig. 7.26 A two-layered spiral coil

122 7 Advanced Probe Models Based on Electric Dipoles

Resistance vs Frequency
100

spiral coil
Resistance (Ohm)
10

0.1
1e+007 1e+008 1e+009
Frequency (Hz)

Reactance vs Frequency
10000

spiral coil
1000
Reactance (Ohm)

100

10

1
1e+007 1e+008 1e+009
Frequency (Hz)

Fig. 7.27 Results of the dipole model calculation of the two-layered spiral coil in freespace below
resonance. Top: resistance. Bottom: reactance
7.10 A Model of a THz Transmitter on GaAs 123

Resistance vs Frequency
900000
800000
spiral coil
700000
Resistance (Ohm)
600000
500000
400000
300000
200000
100000
0
5e+009 6e+009 7e+009 8e+009 9e+009 1e+010
Frequency (Hz)

Reactance vs Frequency

400000
spiral coil
Reactance (Ohm)

200000

−200000

−400000

5e+009 6e+009 7e+009 8e+009 9e+009 1e+010

Frequency (Hz)

Fig. 7.28 Results of the dipole model calculation of the two-layered spiral coil in freespace
through resonance. Top: resistance. Bottom: reactance
124 7 Advanced Probe Models Based on Electric Dipoles

Resistance vs Frequency
80000

70000
spiral coil
Resistance (Ohm) 60000

50000

40000

30000

20000

10000

0
2e+010 4e+010 6e+010 8e+010 1e+011
Frequency (Hz)

Reactance vs Frequency
40000

30000
spiral coil
20000
Reactance (Ohm)

10000

−10000

−20000

−30000

−40000
2e+010 4e+010 6e+010 8e+010 1e+011
Frequency (Hz)

Fig. 7.29 Results of the dipole model calculation of the two-layered spiral coil in freespace at
higher frequencies; two weaker resonances appear. Top: resistance. Bottom: reactance
7.10 A Model of a THz Transmitter on GaAs 125

Fig. 7.30 Sketch of a THz

− Metal
transmitter (after Fig. 3 of
[93])

GaAs

+ Metal
126 7 Advanced Probe Models Based on Electric Dipoles

Resistance vs Frequency
400

350

Resistance (Ohm) 300

250

200

150

100

50

0
1e+009 1e+010 1e+011 1e+012
Frequency (Hz)

Reactance vs Frequency
5000

−5000
Reactance (Ohm)

−10000

−15000

−20000

−25000

−30000
1e+009 1e+010 1e+011 1e+012
Frequency (Hz)

Fig. 7.31 Results of the dipole-model calculation for the THz transmitter of Fig. 3 of [93]. Top:
resistance. Bottom: reactance
7.10 A Model of a THz Transmitter on GaAs 127

Reactance vs Frequency

10000
(-)Reactance (Ohm)

1000

100
1e+009 1e+010 1e+011 1e+012
Frequency (Hz)

Impedance vs Frequency
400

R
300 X
Impedance (Ohm)

200

100

−100

−200

4e+011 8e+011 1.2e+012

Frequency (Hz)

Fig. 7.32 Results of the dipole-model calculation for the THz transmitter of Fig. 3 of [93]. Top:
low-frequency reactance. Bottom: high-frequency impedance

Fig. 7.33 A series-parallel C1

circuit whose impedance is
equivalent to that of Fig. 7.32
up to about 8 × 1011 Hz

R L C0

Zin
Chapter 8
Planar and Conforming Arrays of Probes

In the last chapter, we developed a volume-integral approach to the analysis

of printed-circuit probes. Such probes are finding increased use in planar and
conformable arrays, because of their small size and the ability to be manufactured
in various shapes. Arrays are typically composed of a set of transmitter coils
and receiver coils, which allows a variety of multiplexing schemes to generate
responses with given features (see [96]). For example, one can generate a two-
dimensional “scan” with a suitable array, without moving it. This can be very
effective when doing voxel-based inversions, as will be discussed in a later volume
of this work.
In this chapter we will discuss a model for a conforming array that consists of
a periodic arrangement of transmit–receive coils in the azimuthal direction around
a center axis (a “necklace”), which is used for the inspection of steam generator
and heat-exchanger tubing in nuclear power plants. In the second part of the chapter
we will describe a planar array of printed-circuit coils that has been developed to
acquire data for a voxel-based inversion algorithm.
Throughout this chapter we implicitly make use of the reciprocity theorem of
Chap. 5 as it is applied to bistatic arrays comprising a transmit and receive coil.
Thus the impedances that we compute are transfer impedances.

8.1 Modeling a Circular Array

The ZETEC X-Probe is used for the rapid inspection of steam generator tubes in
nuclear power plants.1 It comprises a periodic azimuthal array of sensors, whose
fundamental basis cell is shown in Fig. 8.1 (we have chosen the dimensions shown

1 We are indebted to Prof. Sung-Jin Song of Sungkyunkwan University for bringing this probe
design to our attention.

H.A. Sabbagh et al., Computational Electromagnetics and Model-Based Inversion, 129

Scientific Computation, DOI 10.1007/978-1-4419-8429-6 8,
© Springer Science+Business Media New York 2013
130 8 Planar and Conforming Arrays of Probes

Axial Flaw
Y (Tube Axis)

R1 R2

Circumferential Flaw
R3
Scan Direction T2

R4
X (Circumferential Direction)
T1

Coordinates(mm): T1(0,0)
T2(−2.5,5)
R1(−2.5,10)
R2(2.5,10)
R3(7.5,5)
R4(10,0)

Fig. 8.1 Showing the basis cell of the X-Probe coil arrangement. “T” stands for a transmit coil
and “R” for a receive coil

in this figure). The array is designed for the inspection of steam generator tubes, and
operates by being pulled through the tube in the axial direction. The configuration
comprising T1R1 and T1R2 is sensitive to axial flaws, whereas T2R3 and T1R4
are sensitive to circumferential flaws. “T” stands for a transmit coil, and “R” for a
receive coil.
We have created a model for the interaction of the basis cell with a flaw of dimen-
sions 0.5 mm × 3 mm × 1.5 mm (width × length × height), located in an aluminum
half-space with conductivity σ = 2.801 × 107 S/m and excitation frequency of
100 kHz. The flaw is oriented in either the axial direction (the Y-axis in Fig. 8.1)
or the circumferential direction (the X-axis in Fig. 8.1). The long dimension of the
flaw determines its orientation.
We show the responses of T1R1 in the axial and circumferential directions
and T1R4 in the circumferential direction in Fig. 8.2. The response of T1R2 is
identical with that of T1R1 when the flaw is symmetrically placed with respect to
the Y -axis, so that we show only the T1R1 response. Clearly, T1R1 is sensitive to
axial flaws, while being relatively insensitive to circumferential flaws, while T1R4 is
sensitive to circumferential flaws, exactly as advertised. It is interesting to note that
the configuration T1R4 relative to the circumferential flaw is precisely the “parallel
scan” configuration discussed in [95].
8.1 Modeling a Circular Array 131

Response of T1R1 and T1R4

to axial and circumferential flaws
0.006

0.005
T1R1:axial
T1R1:circum
T1R4:circum
0.004
|Z|(Ohms)

0.003

0.002

0.001

0
−10 −5 0 5 10
Scan Position (mm)

Fig. 8.2 Comparing the responses of T1R1 and T1R4 to axial and circumferential cracks

Table 8.1 Parameters for

solid models Solid model A (mm) B (mm) Depth (mm)
Block 0.6125 0.6125 1.5
Hole 0.69 0.69 1.5
Cone 1.2 1.2 1.5
Semi-ellipsoid 0.846 0.846 1.5
A and B are transverse parameters

8.1.1 Modeling With the T1R4 Configuration

We’ll do some modeling with the T1R4 configuration using various solid models
shown in Table 8.1. These solids all have equal volumes, are surface-breaking, and
have the same maximum depth, 1.5 mm, into the host. The A and B parameters
determine the extent of the solid in the transverse (X,Y ) plane. Our first exercise
is to determine the sensitivity of the T1R4 configuration to the X-coordinate of the
hole as the probe is scanned in the Y -direction. The X-coordinates of the hole are
1, 2, 3, 4, and 5 mm. Recall that the center of T1 is X = 0 mm and the center of
R4 is X = 10 mm, which means that a hole location of 5 mm is at the center of the
configuration. We do not need to go beyond 5 mm because of the obvious mirror
symmetry about this point.
We contrast two conditions for the experiment; the first is with a host of
aluminum, σal = 2.801 × 107 S/m, and the second with a host of stainless steel,
σss = 1 × 106 S/m. Both problems are run at a frequency of 100 kHz. The results
are shown in Fig. 8.3. Two things, in particular, stand out in these results. In the
132 8 Planar and Conforming Arrays of Probes

Response of T1R4 to Hole Placement in AL

0.0025

0.002 1
2
3
0.0015 4
5
|Z|(ohms)

0.001

0.0005

−0.0005
−10 −5 0 5 10
Scan Position (mm)

Response of T1R4 to Hole Placement in SS

0.007

0.006
1
2
0.005 3
4
5
0.004
|Z|(ohms)

0.003

0.002

0.001

−0.001
−10 −5 0 5 10
Scan Position (mm)

Fig. 8.3 Response of the T1R4 configuration to the X-position of a hole as the probe is scanned in
the Y -direction. The coordinates of the hole are 1, 2, 3, 4, and 5 mm, and the frequency is 100 kHz.
Top: aluminum substrate. Bottom: stainless-steel substrate

first place, the response in the aluminum substrate is not a monotonic function of
the location of the hole, whereas it is in the stainless-steel substrate (locations 4
and 5 are virtually identical). This means that the parameters that define the mutual
inductive coupling between the hole and either coil depend upon the electromagnetic
properties of the structure as well as the frequency.
The second point is that the response is much larger in the stainless-steel substrate
than in the aluminum. The reason for this is that the skin depth, which defines the
8.1 Modeling a Circular Array 133

Response of T1R4 to Response of T1R4 to

Flaw Models in Stainless Steel Flaw Models in Stainless Steel
0.0002 0.001
0 0
−0.0002 −0.001
−0.0004 Block −0.002
−0.0006 Hole
SemiEllip −0.003
−0.0008 Cone
R

X
−0.004
−0.001
−0.005
−0.0012 Block
Hole
−0.0014 −0.006 SemiEllip
Cone
−0.0016 −0.007
−0.0018 −0.008
−10 −5 0 5 10 −10 −5 0 5 10
Scan Position (mm) Scan Position (mm)

Response of T1R4 to
Flaw Models in Stainless Steel
0.008

0.007

0.006 Block
Hole
0.005 SemiEllip
Cone
|Z|

0.004

0.003

0.002

0.001

0
−10 −5 0 5 10
Scan Position (mm)

Fig. 8.4 Response of each of the four solid models to the T1R4 configuration, when each is
centered at X = 5,Y = 0

penetration extent of the coil fields in any direction, is much smaller in aluminum
than in stainless steel. This means that more of the field of either the transmit or
receive coil in stainless steel interacts with the hole than in aluminum.
Models of the type that we have just computed are useful in making probability-
of-detection estimates. For example, by choosing frequencies correctly for a given
host material, so that the response with respect to flaw position is monotonic, as
in the bottom part of Fig. 8.3, we can choose the spacing between the transmit and
receive coils so that the response will have a given minimum value with respect to
flaw position. This allows one to estimate the probability of detection for the given
transmit–receive configuration.
The second exercise is to compute the response of the T1R4 configuration to each
of the solid models of Table 8.1, when each is located midway between T1 and R4,
i.e., at X = 5, Y = 0, in stainless steel, and the excitation frequency is 100 kHz. The
responses are plotted in Fig. 8.4.
134 8 Planar and Conforming Arrays of Probes

Response of T1R4 and T2R3 to Response of T1R4 and T2R3 to

the Hole in Stainless Steel the Hole in Stainless Steel
0.0002 0.001

0 0
T2R3
−0.0002 −0.001 T1R4
T2R3
T1R4
−0.0004
−0.002
−0.0006
−0.003
R

X
−0.0008
−0.004
−0.001
−0.005
−0.0012

−0.0014 −0.006

−0.0016 −0.007
−10 −5 0 5 10 −10 −5 0 5 10
Scan Position (mm) Scan Position (mm)

Fig. 8.5 Response of the T1R4–T2R3 configuration as the probe is scanned past the hole in
stainless steel

The significance of these results is that, because the four solids have the same
volume, their reactance and magnitude responses are quantitatively and qualitatively
similar, the major differences being due to the different lateral extents of the four
flaws. It is the resistance function, however, that gives insight into the depth-features
of each of the flaws. This function is completely different in magnitude and shape,
depending upon the nature of the profile-in-depth of the flaw. This is a typical
result in eddy-current NDE; flaws of equal volume have similar reactance functions,
whereas the resistance functions almost uniquely determine the shape of the flaw
in depth. (This is not a theorem, but a reasonable rule of thumb!) This is why it
is absolutely crucial to record both R and X of a response if one is to accurately
reconstruct the flaw from impedance data. This holds true for any measurement
technique; two “channels” are required, not just magnitude, say, or reactance.
We can think of the cavity of the flaw as storing magnetic energy, so that, if
all conditions are equal, flaws of equal volume will store equal energy, and since
the time-rate-of-change of magnetic energy determines the inductive reactance of a
system, we can understand that the reactance function of equal-volume flaws will
be equal, or nearly so.

8.1.2 Modeling with the T1R4–T2R3 Configuration

Consider the response of the combined T1R4–T2R3 configuration as the probe is

scanned past the hole of Table 8.1, which is located at (X = 1.6 mm,Y = 0 mm) in
stainless steel. The results are shown in Fig. 8.5. Because the T2R3 dipole is located
at Y = 5, its response occurs before the response of the T1R4 dipole as the combined
probe is scanned from Y = −10 to 10 mm. The advantage of the joint configuration
is that one dipole, in this case, T2R3, captures more of the hole than the other,
thereby giving an improved signal response.
8.3 N-Port Theory of the T/R Array 135

Fig. 8.6 Sketch of the

T/R-array with two large
drive-coils and an 8 × 8 array
of receive coils in the shaded
rectangle

8.2 Modeling a Planar T/R Array

Figure 8.6 illustrates a planar T/R array that is designed to produce benchmark data
to validate voxel-based inversion algorithms. The algorithms are not of importance
here; we intend to develop an approach to characterizing such arrays through
techniques that have been developed for analyzing electrical (and microwave)
N-port networks. The 8 × 8 array of printed-circuit receive-sensors is shown
in Fig. 8.7.
The freespace response (reactance) of the sensor when drive 1 is excited at 1 MHz
is shown in Table 8.2 and the response when drive 2 is excited at 1 MHz is shown
in Table 8.3. The response at any other frequency is directly proportional to the
frequency. In particular note the symmetries of the responses.

8.3 N-Port Theory of the T/R Array

The transmit coils are driven by the A-channel of an HP3577A network analyzer,
and the receive coils are connected to the B-channel of the analyzer, as shown in
Fig. 8.8. The electronic commutator at the transmit ports sequentially connects the
two transmit coils, represented as Ports A and B, to the A-channel of the network
analyzer. While connected to either of the transmit ports, the electronic commutator
will sequentially connect each of the receive coils, represented as Ports 1 through
16 (for the 4 × 4 array), to the B-channel of the analyzer. Thus, the operation
reduces to measuring a series of scattering coefficients for a simple 2-port network.
136 8 Planar and Conforming Arrays of Probes

Fig. 8.7 Illustrating the 8 × 8 array of receive-sensors

For example, when transmit coil A is connected to the analyzer, we will measure the
sequence of scattering coefficients, SAA, SAi , SiA , Sii , for i = 1, . . . , 16. With these
results, we can calculate the set of open-circuit transfer impedances from port A to
port i via

2Z0 SAi
ZAi = , i = 1, . . . , 16, (8.1)
1 − SAA − Sii + (SAASii − SAiSiA )

where Z0 is the characteristic impedance, 50 Ω, of the transmission lines that

connect the coils to the network analyzer channels. This relation is derived from
the application of scattering theory to electrical networks. See, e.g., N. Balabanian
and T. Bickart, Electrical Network Theory, J. Wiley & Sons, Inc., New York, 1969,
Chap. 8, or H.J. Carlin and A.B. Giordano, Network Theory, Prentice-Hall, Inc.,
8.3 N-Port Theory of the T/R Array 137

Table 8.2 Freespace response (reactance) of the sensor when drive 1 is excited at 1MHz

Row Column index

index 1 2 3 4 5 6 7 8
1 0.086667 0.07897 0.067643 0.054604 0.035038 0.016082 −0.011478 −0.049784
2 0.11716 0.10924 0.097852 0.082588 0.063816 0.042137 0.019594 −0.011792
3 0.14929 0.13967 0.12581 0.10896 0.089246 0.066876 0.041743 0.015422
4 0.18371 0.17076 0.15412 0.13412 0.11258 0.089016 0.063428 0.034018
5 0.22034 0.20369 0.18303 0.15891 0.13368 0.10812 0.081338 0.053026
6 0.26323 0.23871 0.21134 0.1824 0.15324 0.12421 0.09614 0.066032
7 0.31381 0.275 0.23754 0.20234 0.16917 0.13792 0.10758 0.076681
8 0.37296 0.31155 0.2607 0.21799 0.18134 0.14681 0.11491 0.084233

Table 8.3 Freespace response (reactance) of the sensor when drive 2 is excited at 1 MHz

Row Column index

index 1 2 3 4 5 6 7 8
1 −0.048081 −0.011022 0.015848 0.035748 0.054344 0.067393 0.078287 0.085688
2 −0.01108 0.019723 0.042402 0.064251 0.082398 0.096835 0.109 0.11698
3 0.015891 0.042336 0.0674 0.089226 0.10909 0.12534 0.13897 0.14869
4 0.035562 0.064129 0.08916 0.11344 0.13414 0.15422 0.17029 0.18254
5 0.053813 0.082191 0.10893 0.13406 0.15904 0.18267 0.20341 0.21926
6 0.067035 0.096412 0.12518 0.15409 0.18256 0.21131 0.23876 0.26188
7 0.078053 0.10861 0.13872 0.16998 0.20314 0.23853 0.27545 0.31296
8 0.085347 0.11685 0.14826 0.1822 0.21882 0.26143 0.31258 0.37377

2
A
3
Channel A
HP3577A Transmit Ports Receive Ports Channel B
HP3577A
B
62

63

64

Fig. 8.8 Illustrating an N-port system representation of the T/R array and its connection to the
HP3577A network analyzer

Englewood Cliffs, 1964, Chap. 4. For a reciprocal network of the type that we
have, SAi = SiA . Open-circuit transfer impedances are the input data for voxel-based
inversion algorithms.
138 8 Planar and Conforming Arrays of Probes

Table 8.4 Relationship between two-port parameters

Open-circuit Short-circuit
impedance admittance Chain
parameters parameters parameters
y22 −y12 A AD − BC
z11 z12
|y| |y| C C
−y21 y11 1 D
z21 z22
|y| |y| C C
z22 −z12 D −(AD − BC)
y11 y12
|z| |z| B B
−z21 z11 −1 A
y21 y22
|z| |z| B B
z11 |z| −y22 −1
A B
z21 z21 y21 y21
1 z22 −|y| −y11
C D
z21 z21 y21 y21

For completeness, we’ll write the open-circuit driving-point impedances at ports

A and i that correspond to (8.1):
Z0 (1 + SAA)(1 − Sii ) + SAiSiA
ZAA =
1 − SAA − Sii + (SAASii − SAiSiA )
Z0 (1 − SAA)(1 + Sii ) + SAiSiA
Zii = . (8.2)
1 − SAA − Sii + (SAASii − SAiSiA )
Note that these expressions reduce to the usual result for a one-port structure if
SAi = SiA are very small.

8.3.1 Two-Port Parameter Relations

In addition to the usual open-circuit impedance parameters, z11 , z12 , z21 , z22 ,
and short-circuit admittance parameters, y11 , y12 , y21 , y22 , there exists a set of
parameters, called the chain parameters, that specify the current and voltage at one
port in terms of those at the other. In this way, one can chain networks together very
conveniently. These parameters are labeled A, B, C, and D, where
V1 = AV2 − BI2
I1 = CV2 − DI2 . (8.3)
Table 8.4 gives the relationship between the three sets of two-port parameters.2
The vertical bars denote the determinant of the corresponding 2 × 2 matrix. It is a

2 Takenfrom N. Balabanian and T. Bickart, Electrical Network Theory, J. Wiley & Sons, Inc., New
York, 1969, Chap. 3.
8.3 N-Port Theory of the T/R Array 139

z12
simple matter to verify that the determinant of the chain matrix, AD − BC = =
z21
y12
= 1, for reciprocal two-ports. Hence, for such networks the inverse chain matrix
y21
always exists.
Two independent two-ports can be chained together to produce a composite third
two-port simply by multiplying their corresponding chain matrices:
     
A3 B3 A1 B1 A2 B2
= × . (8.4)
C3 D3 C1 D1 C2 D2

Let the matrix, M3 = M1 × M2 , represent the composite effect of the “model”

driver-receive coil two-port, M2 , and the remaining “parasitic” effects of the con-
nections within the T/R array, M1 . Then after measuring the scattering coefficients
of the composite system in, say, freespace, by means of the HP3577A, and using
(8.1) and (8.2) to compute the open-circuit impedance parameters, followed by
application of the appropriate entries in Table 8.4, we have M3 . We compute M2
by means of the VIC-3D R
model, and when we right-multiply M3 by M−1 2 , which
is guaranteed to exist, we have the desired parasitic chain matrix:

M1 = M3 × M−1
2 . (8.5)

If, then, we assume that the parasitic effects are unchanged whether the T/R array
is in freespace or over a host material, we simply left-multiply the new measured
composite chain matrix by M−1 1 to get the new model chain matrix that corresponds
to the current test configuration. Using the top two lines of Table 8.4 allows us
to transform from the A, B, C, D parameters of the model to the open-circuit
impedance parameters that VIC-3D R
uses.
If, as is likely, the parasitic effects change when the T/R array is over a host, we
redo the procedure described above by replacing freespace by an unflawed version
of the host, or a metal whose conductivity is close to that of the host.

8.3.2 Calculation of Freespace Chain Matrix at 1 MHz

There are 128 chain matrices for the two drive coils and 64 receive coils. We’ll
compute a few of them here. First, we note that the freespace driving-point
impedance of either drive coil is ZD = j562.64 Ω at 1 MHz, and the driving-point
impedance of each receive coil is ZR = j5.7559 Ω in freespace at 1 MHz. These
are model values computed by VIC-3D R
. The entries in Tables 8.2 and 8.3 are
the open-circuit transfer impedances between the appropriate drive coil and each of
the receive coils, also calculated by VIC-3D R
.
We’ll label each receive coil by its row-column indices, and each drive coil by
either A or B, with A corresponding to drive 1, and B to drive 2. Thus, our notation
140 8 Planar and Conforming Arrays of Probes

Table 8.5 DC resistance (Ω) of several receive coils

Connector 7 7 7 7 9 9 9 9
Pins 5 4 3 2 5 4 3 2
2.7 2.9 3.2 3.4 2.7 3.0 3.2 3.4

Table 8.6 Measured freespace scattering parameters at 1MHz taken with the
HP3577A
Connector 7 7 7 7
Pins 5 4 3 2
S11 Mag dBm −0.133 −0.133 −0.133 −0.133
Phase degrees −3.957 −3.992 −4.016 −4.044
S21 Mag dBm −62.4 −62.6 −60.3 −59.4
Phase degrees 162 156 159 158
S12 Mag dBm −62.4 −62.3 −60.2 −59.5
Phase degrees 162 157 159 158
S22 Mag dBm −1.101 −1.199 −1.280 −1.370
Phase degrees 160.70 0160.70 0160.50 0160.44

yields ZAA = ZBB = j562.64, Zi j,i j = j5.7559, for all i j, i j, and, for example, ZA,45 =
Z45,A = j0.11258. With this notation, we’ll compute the chain matrix for the coupled
pair (A, 45) using the relationships of Table 8.4:
⎡ ⎤
562.64 −562.64 × 5.7559 + (0.11258)2
  ⎢ 0.11258 ⎥
AB ⎢ j0.11258 ⎥
⎢
=⎢ ⎥
CD ⎥
(A,45) ⎣ 1 5.7559 ⎦
j0.11258 0.11258
 
4997.69 j28766.09
= . (8.6)
− j8.88257 51.1272

8.3.3 Application to Measured Data at 1 MHz

The dc resistance of the two drive coils are 14.0 Ω (yellow–orange) and 15.3 Ω (red–
brown). The dc resistance for several receive coils are listed in Table 8.5.
Table 8.6 shows measured freespace scattering parameters taken at 1 MHz with
the HP3577A, and Table 8.7 gives the scattering parameters expressed as complex
numbers.
Using (8.1) and (8.2) and Table 8.7, we compute the open-circuit driving-point
and transfer impedances at 1 MHz and list the results in Table 8.8. It is clear from
these results that the driver coil is operating beyond resonance because of the large
negative (capacitive) reactance in S11 , as well as a large resistance component.
8.3 N-Port Theory of the T/R Array 141

Table 8.7 Measured freespace scattering parameters at 1 MHz expressed as complex numbers
Connector 7 7 7 7
Pins 5 4 3 2
S11 0.9848 0.9848 0.9848 0.9848
0.9976 − j0.069 0.9976 − 0.0696 0.9975 − j0.070 0.9975 − j0.0705
S21 7.5858 × 10−4 7.4131 × 10−4 9.661 × 10−4 10.715 × 10−4
−0.951 + j0.3090 −0.9135 + j0.4067 −0.9336 + j0.3584 −0.9272 + j0.3746
S12 7.5858 × 10−4 7.4131 × 10−4 9.661 × 10−4 10.715 × 10−4
−0.951 + j0.3090 −0.9135 + j0.4067 −0.9336 + j0.3584 −0.9272 + j0.3746
S22 0.8809 0.8711 0.8630 0.8541
−0.9438 + j0.3305 −0.9438 + j0.3305 −0.9426 + j0.3338 −0.9423 + j0.3348

Table 8.8 Computed Z-parameters at 1 MHz

Pin Z11 Z12 = Z21 Z22
5 0.306552D + 03, −0.561103D − 01, 0.325738D + 01,
−0.137948D + 04 0.580112D + 00 0.846631D + 01
4 0.300809D + 03, 0.642640D − 02, 0.354377D + 01,
−0.136906D + 04 0.567862D + 00 0.845987D + 01
3 0.298757D + 03, −0.308957D − 01, 0.378596D + 01,
−0.136126D + 04 0.738724D + 00 0.854385D + 01
2 0.294125D + 03, −0.182279D − 01, 0.405050D + 01,
−0.135279D + 04 0.818309D + 00 0.856359D + 01
The first component listed is the resistance, and the second the reactance (Smatrix.f90)

This has an effect on the other parameters. For example, from Tables 8.2 and 8.3,
we expect the reactance of Z12 to have much smaller values than those shown in
Table 8.8.
By using the data of Table 8.8 in the two-port relations of Table 8.4, we compute
the ABCD-matrices for the array at 1 MHz. The results are shown in Table 8.9. The
integer at the left is the pin number, and the complex number beneath each matrix is
the determinant of the matrix. Note that each determinant is equal to the theoretical
value of unity.

8.3.4 Application of Two-Port Theory to Measured Data

Now, we’ll apply the chain-matrix algorithm described in (8.4) to the measured
data. In this example, we’ll let M3 correspond to the pin 5 measured data and M2
correspond to the Drive 2(1,1) model data. When we apply (8.4) to these matrices,
we compute M1 to be as shown in Table 8.10. When we substitute the Z-parameter
data for M1 in Table 8.10 into the generic T-network 2-port of Fig. 8.9, we obtain the
equivalent 2-port T-network for the parasitic structure shown in Fig. 8.10. Though
we do not need to show the parasitic network, since we are not trying to synthesize
it (in fact, we’re trying to eliminate it!), it is interesting to note that it is physically
realizable with passive elements.
142 8 Planar and Conforming Arrays of Probes

Table 8.9 ABCD matrices for the array based upon the HP3577A measurements at 1 MHz
2 −0.166034D+04 −0.322446D+03 −0.396387D+04 −0.155253D+05
−0.272074D−01 −0.122143D+01 0.103496D+02 −0.518038D+01
0.10000D+01 0.11595D−11
3 −0.185639D+04 −0.326783D+03 −0.423618D+04 −0.170986D+05
−0.565164D−01 −0.135132D+01 0.113315D+02 −0.559892D+01
0.10000D+01 −0.15228D−11
4 −0.240460D+04 −0.556934D+03 −0.380976D+04 −0.223168D+05
0.199263D−01 −0.176077D+01 0.149665D+02 −0.607118D+01
0.10000D+01 −0.22027D−12
5 −0.240655D+04 −0.295666D+03 −0.533580D+04 −0.213383D+05
−0.165187D+00 −0.170783D+01 0.139209D+02 −0.696156D+01
0.10000D+01 0.18492D−11
The integer at the left denotes the pin, and the complex number beneath each array is the value of
the determinant of the matrix (ABCD matrix HP3577A.f90)

Table 8.10 Data associated with M1 (system matrix.f90)

ABCD matrix elements of M1
−0.155705D+06 0.146371D+06 0.823540D+08 0.876060D+08
−0.125013D+03 −0.850823D+02 −0.478708D+05 0.703375D+05
Z-parameter matrix elements of M1
0.306623D+03 −0.137952D+04 −0.546689D−02 0.372069D−02
0.419358D−03 −0.562640D+03
Determinant of M1
0.999999D+00 0.413507D−06
ABCD matrix elements of M1inv
−0.478708D+05 0.703376D+05 −0.823541D+08 −0.876060D+08
0.125014D+03 0.850824D+02 −0.155705D+06 0.146371D+06
Consistency check: M1inv x M3 = M2
−0.117019D+05 0.483884D−02 −0.278519D−01 −0.673551D+05
0.860027D−05 0.207983D+02 −0.119713D+03 0.495023D−04

Fig. 8.9 Showing a 2-port

Z11−Z12 Z22−Z12
T-network constructed from
knowledge of
Z11 , Z12 = Z21 , and Z22
Z12

Figure 8.11 shows the networks associated with M1 and M2 coupled together
between the new ports 1 and 2. The coupled network has the same Z-parameters
as originally tabulated for Pin 5 in Table 8.8. (We’ll leave this as an exercise for
the reader.) This shows the consistency between the algebra of ABCD-matrices and
simple network theory.
8.3 N-Port Theory of the T/R Array 143

306.628 −1379.52 0.00589 −562.65

−0.00547

j0.00372

Fig. 8.10 Showing the equivalent T-network 2-port for the parasitic structure

M1 M2

306.628 −j1379.52 0.00589 −j562.65 j562.69 j5.8040

−0.00547
Port1 −j0.04808 Port2
j0.00372

Fig. 8.11 Showing the coupled networks that correspond to M1 (the parasitic elements) and M2
(the model data for Drive 2(1,1)). The composite network extending from Port1 to Port2 has the
same Z-parameters as shown in Table 8.8 for Pin 5

8.3.5 Modeling N-Port Array on an Aluminum Host:

σ Al = 2.801×107 S/m

Let M3 = M1 × M2 , where

M1 = “parasitic” effects measured on unflawed Al workpiece

M2 = model results on unflawed Al workpiece computed by VIC3D(c)
M3 = measured data on unflawed Al workpiece (8.7)

and let

M1 = “parasitic” effects measured on flawed Al workpiece (same as above)

M2 = model results on flawed Al workpiece
M3 = measured data on flawed Al workpiece (8.8)

Compute M1 = M3 × M−1 2 from (8.7) and M2 = M−1 

1 × M3 from (8.8). The

anomalous data for inversion are contained in M2 − M2 . Actually, all we want is
1 1
δ Z12 =  − , where C2 and C2 are entries in the ABCD-matrices for M2 and
C2 C2
M2 , respectively.
Chapter 9
Multilayered Media with Cylindrical
Geometries

9.1 Introduction

In this chapter, we describe the mathematical development of a general axisym-

metric model for VIC-3D R
. This model is capable of analyzing tubes with tube
supports and roll-expanded transition zones. Features such as magnetite and sludge,
are included, and materials may be either ferromagnetic or nonmagnetic. The model
described in this chapter will include only differential (or absolute) bobbin coils.
Flaws, or anomalies, can be of three types: (1) axisymmetric (such as circumfer-
ential rings, tube supports, roll-expanded transition zones), (2) thin axially-oriented
cracks, and (3) user-defined flaws, such as intergranular attack (IGA) or corrosion
pits. The incident fields due to bobbin coils are computed using the axisymmetric
theory developed in this chapter.

9.2 Some Typical Problems in Steam Generator Tubing

There are a number of rather complicated geometries that appear in the inspection
of steam generator tubing by means of eddy currents. Figures 9.1–9.3 illustrate
several of them, and Fig. 9.4 illustrates a number of different flaws that must be
modeled. The model must not only contend with these geometries, but it must deal
with ferromagnetic bodies, as well.

9.3 Coupled Ferromagnetic Integral Equations

It is possible to account for ferromagnetic effects by introducing an anoma-

lous magnetic current, together with magnetic–magnetic, magnetic–electric, and
electric–magnetic Green functions. For this problem, which is axisymmetric, and

H.A. Sabbagh et al., Computational Electromagnetics and Model-Based Inversion, 145

Scientific Computation, DOI 10.1007/978-1-4419-8429-6 9,
© Springer Science+Business Media New York 2013
146 9 Multilayered Media with Cylindrical Geometries

Fig. 9.1 Illustrating a steam

generator tube with a tube
support Tube support

Fig. 9.2 Illustrating a

roll-expanded steam
generator tube
Tube sheet

Fig. 9.3 Illustrating a Sleeve

laser-welded sleeve in a
steam generator tube
Tube

Laser weld

therefore solenoidal in the electric variables, it will be easier to use an “all electric”
model for the coupled system, following the same ideas that we developed in the
context of Amperian currents in Chap. 3. The electric–electric Green function is
simple, and because all electric variables are solenoidal, the matrices should be well
conditioned for all frequencies.
9.3 Coupled Ferromagnetic Integral Equations 147

Axial Flaw Sludge, roll expansion

Circumferential Flaw Inter-Granular Attack

Fig. 9.4 Illustrating a variety of anomalies (“flaws”) in steam generator tubes

We start with Maxwell’s equations

∇ × E = − jω B
∇ × H = jω D + J(e) . (9.1)
Now H = B/μ (r) = B/ μh + B/μ (r) − B/μh = B/μh − Ma , where μh is the host
permeability and Ma is the anomalous magnetization vector. Thus the second of
Maxwell’s equations may be written

∇ × (B/ μh ) = jω D + J(e) + ∇ × Ma , (9.2)

which makes clear that the Amperian current, ∇ × Ma , is an equivalent anomalous
electric current that arises because of the departures of the magnetic permeability of
the workpiece from the host permeability, μh . J(e) , on the other hand, is an electric
current that includes the anomalous current that arises due to differences in electrical
conductivity; J(e) = σh E + (σ (r) − σh)E = σh E + Ja.
The problem is axisymmetric, which means that E, Ja , and ∇ × Ma have only
a nonzero φ -component, whereas Ma has z- and r-components. Hence, we have as
the solutions of (9.1) and (9.2)

(ee)
E(r, z) = E(i) (r, z) + aφ 2π Gφ φ (r, z; r , z )Ja (r , z )r dr dz
flaw

(ee)
+ aφ 2π Gφ φ (r, z; r , z )(∇ × Ma )φ r dr dz
flaw
148 9 Multilayered Media with Cylindrical Geometries

1
B(r, z) = − ∇×E
jω

1 (ee)
= B(i) (r, z) − ∇ × aφ 2π Gφ φ (r, z; r , z )Ja (r , z )r dr dz
jω flaw

1 (ee)
− ∇ × aφ 2π Gφ φ (r, z; r , z )(∇ × Ma )φ r dr dz . (9.3)
jω flaw

E(i) and B(i) are the incident fields produced by the exciting coil. The curl operation
divided by − jω in (9.3) defines a magnetic–electric operator.
Upon multiplying E by the anomalous conductivity, σa (r) = jω (ε̂ (r) − ε̂h ), we
μ (r)μh
get the anomalous current, Ja . Similarly, B(r) = Ma (r). Hence, when
μ (r) − μh
we make these substitutions in (9.3), we get the coupled integral equations for
axisymmetric, ferromagnetic problems:

Ja (r, z) (ee)
= E (i) (r, z) + 2π Gφ φ (r, z; r, z )Ja (r , z )r dr dz
σa (r, z) flaw

(ee)
+ 2π Gφ φ (r, z; r, z )(∇ × Ma )φ r dr dz
flaw

μ (r, z)μh 1 (ee)
Ma (r, z) = B(i) (r, z) − ∇ × aφ 2π Gφ φ (r, z; r, z )Ja (r , z )r dr dz
μ (r, z) − μh jω flaw

1 (ee)
− ∇ × aφ 2π Gφ φ (r, z; r, z )(∇ × Ma )φ r dr dz .
jω flaw

(9.4)

We have dropped the vector components from the electrical variables; the scalar
equation in (9.4) is the φ -component of the electric field integral equation.

9.4 Discretization: Method of Moments

Because M(r, z) = Mr (r, z)aρ + Mz (r, z)az , it follows that (∇ × M)φ = ∂ Mr /∂ z −

∂ Mz /∂ r. Therefore, in order for M to belong to H(curl), the space of vector
functions with bounded curls, it follows that Mr must be everywhere differentiable
with respect to z, and Mz must be everywhere differentiable with respect to r.
An expansion of the following form will satisfy these criteria:

Ma (r, z) = ∑ Mlm π1l (r)π2m (z)

(r) (r)

lm

Ma (r, z) = ∑ Mlm π2l (r)π1m (z) ,

(z) (z)
(9.5)
lm
9.4 Discretization: Method of Moments 149

where π1m is the unit pulse function, which starts at the mth node, and π2m is the
triangular function, which starts at the mth node. Ja (r, z), on the other hand, can be
expanded in pulse functions

Ja (r, z) = ∑ Jlm π1l (r)π1m (z) , (9.6)

lm

because there are no derivatives appearing in (9.4)(a), except for ∇ × Ma .

μ (r, z)μh
We proceed to discretize (9.4), assuming that ν (r, z) = and σa (r, z)
μ (r, z) − μh
are constant in each cell. The result, after taking moments of (9.4)(a) is
(+) (−)
rl + rl δ rδ z
Jlm − ∑ Glm,LM JLM − ∑ Glm,LM MLM + ∑ Glm,LM MLM
(i) (ee) (em)(r) (r) (em)(z) (z)
Elm =
2 σlm LM LM LM

(9.7)
where
 
(i)
Elm = E (i) (r, z)π1l (r)π1m (z)rdrdz
   
(ee)
Glm,LM = 2π rdrdzπ1l (r)π1m (z) G(r, z; r , z )π1L (r )π1M (z )r dr dz
   
(em)(r)
Glm,LM = 2π rdrdzπ1l (r)π1m (z) G(r, z; r , z )π1L (r )π2M

(z )r dr dz
(ee) (ee)
Glm,LM − Glm,LM+1
=
δz
   
(em)(z)
Glm,LM = 2π rdrdzπ1l (r)π1m (z) G(r, z; r , z )π2L

(r )π1M (z )r dr dz
(ee) (ee)
Glm,LM − Glm,L+1M
= . (9.8)
δr
The discretized version of (9.4)(b) is

(+) (−)
rl + rl
∑ QmM MlM + ∑ Glm,LM JLM + ∑ Glm,LM
(i)(r) (r) (r) (me)(r) (mm)(rr) (r)
Blm = δr MLM
2 M LM LM

−∑
(mm)(rz) (z)
Glm,LM MLM
LM

= δ z ∑ QlL MLm − ∑ Glm,LM JLM − ∑ Glm,LM MLM + ∑ Glm,LM MLM ,

(i)(z) (z) (z) (me)(z) (mm)(zr) (r) (mm)(zz) (z)
Blm
L LM LM LM

(9.9)
150 9 Multilayered Media with Cylindrical Geometries

where the tri-diagonal matrices are symmetric and have the following nonzero
entries:

(r) (−)
QmM = ν (rl , z)π2M (z)π2m (z)dz
⎧
νlm if M = m − 1
δz ⎨
= 2(νlm + νlm+1 ) if M = m
6 ⎩
νlm+1 if M = m + 1

(z)
QlL = ν (r, mδ z)π2L (r)π2l (r)rdr
⎧  
⎪ 1 l + r0 /δ r
⎪
⎪ νlm + if L = l − 1
⎪
⎪ 12 6
⎪
⎪
⎪
⎪
⎪
⎨    
1 l + r0 /δ r 10 l + r0 /δ r
= δr 2
νlm + + νl+1,m + if L = l
⎪
⎪ 4 3 24 3
⎪
⎪
⎪
⎪  
⎪
⎪
⎪
⎪ 1 l + r0 /δ r
⎩ νl+1,m + if L = l + 1
4 6
(9.10)

(−) (+)
where r0 is the radial starting point of the grid. rl = r0 + l δ r and rl = r0 +
(l + 1)δ r. The first entry in each of these matrices is the lower diagonal, the second
the main diagonal, and the third the upper diagonal.
The other matrices are given by

 
(i)(r)
Blm = B(i) (r, z) · aρ π1l (r)π2m (z)rdrdz

(i) (i)
1 Elm − Elm+1
=−
jω δz
 
(i)(z)
Blm = B(i) (r, z) · az π2l (r)π1m (z)rdrdz

(i) (i)
1 Elm − El+1m
=
jω δr
   
(me)(r) 2π 
Glm,LM = rdrdzπ1l (r)π2m (z) G(r, z; r , z )π1L (r )π1M (z )r dr dz
jω
(ee) (ee)
1 Glm,LM − Glm+1,LM
=
jω δz
9.5 Calculation of Matrices 151

   
(mm)(rr) 2π 
Glm,LM = rdrdzπ1l (r)π2m (z) G(r, z; r , z )π1L (r )π2M

(z )r dr dz
jω
(ee) (ee) (ee) (ee)
Glm,LM − Glm+1,LM − Glm,LM+1 + Glm+1,LM+1
=
jωδ z2
   
(mm)(rz) 2π 
Glm,LM = rdrdzπ1l (r)π2m (z) G(r, z; r , z )π2L

(r )π1M (z )r dr dz
jω
(ee) (ee) (ee) (ee)
Glm,LM − Glm+1,LM − Glm,L+1M + Glm+1,L+1M
=
jωδ rδ z
   
(me)(z) 2π
Glm,LM = rdrdzπ2l (r)π1m (z) G(r, z; r , z )π1L (r )π1M (z )r dr dz
jω
(ee) (ee)
1 Glm,LM − Gl+1m,LM
=
jω δr
   
(mm)(zr) 2π
Glm,LM = rdrdzπ2l (r)π1m (z) G(r, z; r , z )π1L (r )π2M

(z )r dr dz
jω
(ee) (ee) (ee) (ee)
Glm,LM − Gl+1m,LM − Glm,LM+1 + Gl+1m,LM+1 (mm)(rz)
= = GLM,lm
jωδ rδ z
   
(mm)(zz) 2π
Glm,LM = rdrdzπ2l (r)π1m (z) G(r, z; r , z )π2L

(r )π1M (z )r dr dz
jω
(ee) (ee) (ee) (ee)
Glm,LM − Gl+1m,LM − Glm,L+1M + Gl+1m,L+1M
= . (9.11)
jωδ r2

9.5 Calculation of Matrices

Keep in mind that the Green function that appears in the integrands of (9.8) and
(ee)
(9.11) is really Gφ φ (r, z; r , z ). We will derive expressions for this function later,
but for the present we will use the results.
Refer to Fig. 9.5, which shows an axisymmetric layered tube with concentric
layers. A coaxial filamentary current source of unit strength is located at r = r in the
layer bounded by rs−1 < r < rs −1 . The wave parameter in the ith layer is given by
αi = (ki2 − h2)1/2 , where h is the one-dimensional spatial Fourier transform variable,
and ki2 = ω 2 μi ε̂i = ω 2 μi εi − jω μi σi . The eigenvectors, V1m , V2m , V3m , V4m , shown
in Fig. 9.5 are defined in Appendix A.1.
152 9 Multilayered Media with Cylindrical Geometries

Axis of Tube
r=0
V(1)=a1V11+c1V31 Region 1 α=α1
r1
V(2)=a2V12+b2V22+c2V32+d2V42 Region 2 α=α2
r2
..
..
rs-1
Region s
V(s)=asV1s+bsV2s+csV3s+dsV4s α=αs
Source Point
r=r | αs=αs|
|
V(s’)=as’V1s’+bs’V2s’+cs’V3s’+ds’V4s’ Region s
α=αs|
rs| -1
..
..
r2 |
|
Region 2
V(2’)=a2’V12’+b2’V22’+c2’V32’+d2’V42’ α=α2|
r1|
V(1’)=b1’V21’+d1’V41’ |
Region 1
α=α1|

Fig. 9.5 An axisymmetric layered tube with concentric layers

We introduce the following two coefficients that define a φ -directed filamentary

electric current source:
(2)
παs U12 (s )e− jαs rs −1 J1 (αs r ) + U22 (s )e− jαs Δs e jαs rs−1 H1 (αs r )
c1 = ·
j2 e− j2αs Δs U22 (s )V11 (s) − U12 (s )V21 (s)

(2)
παs V11 (s)e− jαs Δs e− jαs rs −1 J1 (αs r ) + V21(s)e jαs rs−1 H1 (αs r )
d1  = · , (9.12)
j2 e− j2αs Δs U22 (s )V11 (s) − U12 (s )V21 (s)

where Δs = rs −1 − rs−1 . These coefficients are derived in Appendix A.1.

The electric field in the various regions is defined in terms of these parameters,
as follows:
In region m (rm−1 < r < rm ), above the source region:
μm   (2)

Ẽφ (r, r ) = − jω c1 V11 (m)e− jαm (rm −r) j1 (αm r) + V21(m)e jαm (rm −r) h1 (αm r) ,
αm
(9.13)
and in region m
(rm < r < rm −1 ), below the source region:
μ   (2)

Ẽφ (r, r ) = − jω m d1  U12 (m )e− jαm (rm −r) j1 (αm r)+U22 (m )e jαm (rm −r) h1 (αm r) .
αm
(9.14)
9.5 Calculation of Matrices 153

In the source region, (rs−1 < r, r < rs −1 ):

μs   (2)

Ẽφ (r, r ) = − jω d1 U12 (s )e− jαs (rs−1 −r) j1 (αs r) + U22 (s )e jαs (rs−1 −r) h1 (αs r)
αs
r > r
μs  (2)

= − jω c1 V11 (s)e− jαs (rs −1 −r) j1 (αs r) + V21(s)e jαs (rs −1 −r) h1 (αs r)
αs
r < r . (9.15)

As an interesting application of (9.15), if we assume that there is only one region

that occupies all space, namely the source region, then (9.15) yields

ω μs (2)
Ẽφ (r, r ) = −π H (αs r )J1 (αs r) , r < r
2 1
ω μs (2)
Ẽφ (r, r ) = −π H (αs r)J1 (αs r ) , r < r . (9.16)
2 1
Alternatively, we can treat (9.16) as an “incident Green function,” which is then
scattered from the various layers, producing the total function in (9.15).
The spatial-domain Green function, in which the field point, r, is in region
m, m , s, and the source point in region s is given by the inverse Fourier transform of
the corresponding expression in (9.13)–(9.15)
 ∞
1 
G(r, z; r , z ) = Ẽ(r, r )e jh(z−z ) dh . (9.17)
(2π )2 −∞

The G(r, z; r , z ) that appear in (9.8) and (9.11) have r, r in the same (source)
region, so it corresponds to (9.15) being substituted into (9.17).
The U’s and V ’s are transfer coefficients that will be defined recursively shortly.
Note that when s = 1, i.e., when there are no layers above the source point in Fig. 9.5,
then V11 (1) = 1, V21 (1) = 0; when s = 1 , i.e., when there are no layers below the
source point, then U22 (1 ) = 1, U12 (1 ) = 0.
It will be more convenient to compute the infinite-space Green function by using
 ∞ −αs |z−z |
j ω μs e
G(r, z; r , z ) = − J1 (r l)J1 (rl)ldl , (9.18)
2π 0 2αs

which is the same as the free-space Green function that is used in the ferrite-
core module. Equations (9.17) and (9.18) produce the same result for a filamentary
current loop in infinite space, even though they correspond to different coordinate
systems.
When computing matrix elements that involve G(ee) in the source region, it will
be advantageous to separate the incident, or infinite-space, term, (9.16), from the
layered-space terms, as in Chap. 2. Because (9.15) includes both terms, we subtract
(9.16) from (9.15) to get the layered-space terms.
154 9 Multilayered Media with Cylindrical Geometries

We start by rewriting (9.12) in a slightly more familiar form that involves

reflection coefficients:
   j αs Δ s 
U12 (s ) α Δ − jαs rs −1 e (2)
e j2 s s e 
J1 (αs r ) + e jαs rs−1 H1 (αs r )
πα U (s  )V (s) V (s)
c1 =
s 22 11 11
j2 1 − Rs ,s e j2αs Δs
 j αs Δ s   
e − jαs rs −1 V21 (s) (2)
e 
J1 (αs r ) + e j2αs Δs e jαs rs−1 H1 (αs r )
πα U (s ) U (s  )V (s)
d1  =
s 22 22 11
,
j2 1 − Rs ,s e j2αs Δs
(9.19)

where the double reflection coefficient is defined by

U12 (s )V21 (s)

Rs ,s = . (9.20)
U22 (s )V11 (s)

Next, substitute (9.19) into (9.15):

 
 π
Ẽφ (r, r ) = − jω μs ·
j2
⎡  j αs Δ s    ⎤
e − jαs rs −1 ) + V21 (s) j2αs Δs e jαs rs−1 H (2) (α r  )
⎢ U22 (s ) e J (
1 sα r e s ⎥
⎢ U22 (s )V11 (s) 1
⎥·
⎣ α
1 − Rs ,s e s s
j2 Δ ⎦

 
(2)
U12 (s )e− jαs (rs−1 −r) j1 (αs r) +U22 (s )e jαs (rs−1 −r) h1 (αs r)

r > r
 
π
= − jω μs ·
j2
⎡   j αs Δ s  ⎤
U12 (s ) j2αs Δs e− jαs rs −1 J (α r  ) + e jαs rs−1 H (2) (α r  )
⎢ U22 (s )V11 (s) e 1 s e s ⎥
⎢ V11 (s) 1
⎥·
⎣ 1 − Rs ,s e j2α s Δ s ⎦

 
(2)
V11 (s)e− jαs (rs −1 −r) j1 (αs r) +V21 (s)e jαs (rs −1 −r) h1 (αs r)

r < r . (9.21)

Upon expanding and simplifying we get

9.5 Calculation of Matrices 155

⎡ 
U12 (s ) − jαs (2rs−1 −(r+r ))
  e j1 (αs r ) j1 (αs r)
π ⎢ 
⎢ U22 (s )
Ẽφ (r, r ) = − jω μs
j2 ⎣ 1 − Rs,s e j2αs Δs

(2) (2)
H1 (αs r)J1 (αs r ) + Rs,s e j2αs Δs H1 (αs r )J1 (αs r)
+
1 − Rs,s e j2αs Δs
  ⎤
V21 (s)  (2) (2)
e jαs (2rs −1 −(r+r )) h1 (αs r )h1 (αs r) ⎥
V11 (s) ⎥
+ ⎦
1 − Rs,s e j2αs Δs

r > r

⎡ 
U12 (s ) − jαs (2rs−1 −(r+r ))
  e j1 (αs r ) j1 (αs r)
π ⎢ 
⎢ U22 (s )
= − j ω μs
j2 ⎣ 1 − Rs,s e j2αs Δs

(2) (2)
H1 (αs r )J1 (αs r) + Rs ,s e j2αs Δs H1 (αs r)J1 (αs r )
+
1 − Rs,s e j2αs Δs
  ⎤
V21 (s)  (2) (2)
e jαs (2rs −1 −(r+r )) h1 (αs r )h1 (αs r) ⎥
V11 (s) ⎥
+ ⎦
1 − Rs,s e j2αs Δs

r > r (9.22)

U12 (s ) V21 (s)

Call 
= ρs e− j2αs Δs , = ρs e− j2αs Δs . Then, clearly, Rs ,s = ρs ρs e− j4αs Δs .
U22 (s ) V11 (s)
ρs corresponds to reflections from the surface r = rs −1 , and ρs corresponds to
reflections from the surface r = rs−1 .
Finally, subtract (9.16) from (9.22) to get the layered-space fields in the region
rs−1 < r, r < rs −1 :
    (2) (2)
(ls) π ρs e− jαs (2rs −1 −(r+r )) j1 (αs r  ) j1 (αs r) + ρs e jαs (2rs−1 −(r+r )) h1 (αs r  )h1 (αs r)
Ẽφ = − jω μs
j2 1 − ρs ,s e− j2αs Δs
 ⎤
(2) (2)
ρs ,s e− j2αs Δs H1 (αs r  )J1 (αs r) + H1 (αs r)J1 (αs r  )
+ ⎦. (9.23)
1 − ρs ,s e− j2αs Δs
156 9 Multilayered Media with Cylindrical Geometries

The layered-space field is continuous, along with its derivatives at r = r . We can

also replace the Hankel and Bessel functions with their scaled versions in (9.23):

(2) (2)  (2)

H1 (αs r )J1 (αs r) + H1 (αs r)J1 (αs r ) = e− jαs (r −r) h1 (αs r ) j1 (αs r)
 (2)
+ e jαs (r −r) h1 (αs r) j1 (αs r ).

Note the similarity between these results and the (a) and (b) terms (convolution
and correlation) in Chap. 2. Of course, because the scaled Bessel and Hankel
functions are not independent of (r, r ), the results of (9.23) are not rigorously
convolutions and correlations in (r, r ). The expression (9.23), however, does make
clear the presence of incident and reflected fields at the layers, r and r , just as in
Chap. 2.
In a similar manner, we can derive expressions for the transmitted fields above
and below the source region. In region m (rm−1 < r < rm ) above the source region,
we have, upon substituting (9.12) into (9.13):
  
 αs π
Ẽφ (r, r ) = − jω μm
αm j2
   (2)

ρs e− jαs Δs e− jαs (rs −1 −r ) j1 (αs r ) + e jαs (rs−1 −r ) h1 (αs r )
×
1 − ρs ,s e− j2αs Δs
 
V21 (m) jαm (rm −r) (2) V11 (m) − jαm (rm −r)
× e h ( α r) + e j ( α r) ,
V11 (s)e− jαs Δs
m
V11 (s)e− jαs Δs
1 1 m

(9.24)

and in region m (rm < r < rm −1 ) below the source region:
  
αs π
Ẽφ (r, r ) = − jω μm
αm j2
  (2) 

ρs e− jαs Δs e jαs (rs−1 −r ) h1 (αs r ) + e− jαs(rs −1 −r ) j1 (αs r )
×
1 − ρs,s e− j2αs Δs

U12 (m )
× e− jαm (rm −r) j1 (αm r)
U22 (s )e− jαs Δs

U22 (m ) j αm (rm −r) (2)
+ e h (α  r) . (9.25)
U22 (s )e− jαs Δs 1 m

By writing

V21 (m) V11 (m) V21 (m)

= ·
V11 (s) V11 (s) V11 (m)
9.5 Calculation of Matrices 157

V11 (m)
= Rm
V11 (s)
U12 (m ) U22 (m ) U12 (m )

= ·
U22 (s ) U22 (s ) U22 (m )
U22 (m )
= R , (9.26)
U22 (s ) m

then we can rewrite (9.24) and (9.25) as

   
 αs π V11 (m)
Ẽφ (r, r ) = − jω μm
αm j2 V11 (s)e− jαs Δs
   (2)

ρs e− jαs Δs e− jαs (rs −1 −r ) j1 (αs r ) + e jαs (rs−1 −r ) h1 (αs r )
×
1 − ρs,s e− j2αs Δs
 
(2)
× Rm e jαm (rm −r) h1 (αm r) + e− jαm(rm −r) j1 (αm r) , (9.27)

in region m (rm−1 < r < rm ) above the source region, and

   
 αs π U22 (m )
Ẽφ (r, r ) = − jω μm
αm j2 U22 (s )e− jαs Δs
  (2) 

ρs e− jαs Δs e jαs (rs−1 −r ) h1 (αs r ) + e− jαs (rs −1 −r ) j1 (αs r )
×
1 − ρs,s e− j2αs Δs
 
(2)
× Rm e− jαm (rm −r) j1 (αm r) + e jαm (rm −r) h1 (αm r) , (9.28)

in region m (rm < r < rm −1 ) below the source region. We will return to a
discussion of the computation of the reflection coefficients (the R’s) shortly, but
right now we’ll get into details of the computation of the matrix elements.
Consider
   
(ee)
Glm,LM = 2π rdrdzπ1l (r)π1m (z) G(r, z; r , z )π1L (r )π1M (z )r dr dz

where
G(r, z; r , z ) = G(0) (r, z; r , z ) + G(s) (r, z; r , z ) . (9.29)
158 9 Multilayered Media with Cylindrical Geometries

G(0) (r, z; r , z ) is the infinite-space Green function and G(s) (r, z; r , z ) is the layered-
space Green function:
 ∞ −αs |z−z |
jω μs e
G(0) (r,z;r  ,z ) = − J1 (r  l)J1 (rl)ldl
2π 0 2αs

ω μs ∞ jh(z−z )
G(s) (r,z;r  ,z ) = − e
8π −∞
   (2) (2)
ρ  e− jαs (2rs −1 −(r+r )) j1 (αs r  ) j1 (αs r) + ρs e jαs (2rs−1 −(r+r )) h1 (αs r  )h1 (αs r)
× s
1 − ρs ,s e− j2αs Δs
 ⎤
 (2)  (2)
ρs ,s e− j2αs Δs e− jαs (r −r) h1 (αs r  ) j1 (αs r) + e jαs (r −r) h1 (αs r) j1 (αs r  )
+ ⎦ dh .
1 − ρs ,s e− j2αs Δs

(9.30)

Hence, the infinite-space matrix element is:

 ∞
(0)(ee) F(m−M) I (L(hδ r), (L+1)(hδ r)) · I (l(hδ r), (l+1)(hδ r))
Glm,LM = − jω μs dh ,
0 2αs h3
(9.31)

where
⎧ ' (2
⎪
⎪ 1 − e− α s δ z
⎪
⎪ e α s δ z(m−M+1) , if m − M ≤ −1
⎪
⎪ αs δ z
⎪
⎪ ' (
⎪
⎨ e−αs δ z − 1 + αsδ z
F(m − M) = δ z2 2 , if m − M = 0
⎪
⎪ (αs δ z)2
⎪
⎪ ' (2
⎪
⎪
⎪ −αs δ z(m−M−1) 1 − e−αsδ z
⎪
⎪
⎩e , if m − M ≥ 1
αs δ z
 b
I (a, b) = zJ1 (z)dz . (9.32)
a

Note that F(m − M) = F(M − m); i.e., it is symmetrical in (m, M), and corresponds
to the F3 function in (5.29).
The layered-space matrix element is given by:


' (
(s)(ee) ω μs 2 ∞ 2 − e− jhδ z − e jhδ z
Glm,LM =− δz e jh(m−M)δ z
4 −∞ h2 δ z2

I (l(αs δ r), (l + 1)(αs δ r)) · I (L(αs δ r), (L + 1)(αs δ r))
× R(1) (αs )
αs4
H (l(αs δ r), (l + 1)(αs δ r)) · H (L(αs δ r), (L + 1)(αs δ r))
+ R(2) (αs )
αs4
9.5 Calculation of Matrices 159


I (l(αs δ r), (l + 1)(αs δ r)) · H (L(αs δ r), (L + 1)(αs δ r))
(3)
+ R (αs )
αs4

H (l(αs δ r), (l + 1)(αs δ r)) · I (L(αs δ r), (L + 1)(αs δ r))
+ dh ,
αs4
(9.33)
where
ρs e− jαs 2rs −1
R(1) (αs ) =
1 − ρs,s e− j2αs Δs
ρs e jαs 2rs−1
R(2) (αs ) =
1 − ρs,s e− j2αs Δs
ρs ,s e− j2αs Δs
R(3) (αs ) =
1 − ρs,s e− j2αs Δs
 z2
(2)
H (z1 , z2 ) = H1 (z)zdz. (9.34)
z1

9.5.1 Bobbin Coil Incident Field and Moments

Now we calculate the field produced by a bobbin coil that extends from l0 ≤ z ≤
l1 , r0 ≤ r ≤ r1 , and is located in region c. First, we note that the current density
flowing within such a coil is Nc I, where Nc is the density of turns (i.e., turns per unit
area), and I is the current flowing within the coil. Hence, we have
 l1  r1
E (i) (r, z) = 2π Nc I dz r dr G(r, z; r , z )
l0 r0
 ∞  l1  r1 
Nc I − jhz    
= e jhz
dh e dz r dr Ẽ(r, r ) . (9.35)
2π −∞ l0 r0

Ẽ(r, r ) is given by (9.27) if the field point is above region c, and by (9.28) otherwise.
Hence,
 ∞ jh(z−l0 )  r 
(i) Nc I e − e jh(z−l1) 1
  
E = r dr Ẽ(r, r ) dh . (9.36)
2π −∞ jh r0

Upon substituting (9.27) and (9.28) into (9.36), we get

    jh(z−l )
0 − e jh(z−l1 )
(i) Nc I ∞ αc V11 (m) e
E (r, z) = jω μm
4 −∞ αm V11 (c)e− jαc Δc h
 
I (αc r0 , αc r1 ) H (αc r0 , αc r1 )
× R(4) (αc ) + R (5)
(αc )
αc2 αc2
160 9 Multilayered Media with Cylindrical Geometries

 
(2)
× Rm e jαm (rm −r) h1 (αm r) + e− jαm(rm −r) j1 (αm r) dh

rm−1 < r < rm (above coil)

    jh(z−l )
Nc I ∞ αc U22 (m ) e 0 − e jh(z−l1 )
= jω μm α Δ
4 −∞ αm  −
U22 (c )e c c j h
 
H (αc r0 , αc r1 ) I (αc r0 , αc r1 )
× R(6) (αc ) + R (7)
(α c )
αc2 αc2
 
(2)
× Rm e− jαm (rm −r) j1 (αm r) + e jαm (rm −r) h1 (αm r) dh

rm < r < rm −1 (below coil), (9.37)

where

ρc e− jαc Δc e− jαc rc −1

R(4) (αc ) =
1 − ρc,c e− j2αc Δc
e jαc rc−1
R(5) (αc ) =
1 − ρc,c e− j2αc Δc
ρc e− jαc Δc e jαc rc−1
R(6) (αc ) =
1 − ρc,c e− j2αc Δc
e− jαc rc −1
R(7) (αc ) =
1 − ρc,c e− j2αc Δc
Δc = rc −1 − rc−1 . (9.38)

The first result in (9.37) would be used for an encircling-coil, because in that case
the entire tube is “above” the coil. Similarly, the second result would be used for an
internal-coil, because the entire tube is “below” the coil.
Because the dependency on r and r , as well as z and z , is separated in (9.37), it
is easy to compute the moments of the incident field, as given in (9.8):
   
(i) Nc I ∞ αc V11 (m)
Elm = ω μm
4 −∞ αm V11 (c)e− jαc Δc
 
e− jh(l0 −(m+1)δ z) − e− jh(l0 −mδ z) − e− jh(l1 −(m+1)δ z) + e− jh(l1 −mδ z)
×
h2
 
I (αc r0 , αc r1 ) H (αc r0 , αc r1 )
× R(4) (αc ) + R (5)
( αc )
αc2 αc2
 
H (l αm δ r, (l + 1)αm δ r) − jαm rm I (l αm δ r, (l + 1)αm δ r)
× Rm e j αm rm + e dh
αm2 αm2
rm−1 < r < rm (above coil)
9.5 Calculation of Matrices 161

 ∞   
Nc I αc U22 (m )
= ω μm
4 −∞ αm U22 (c )e− jαc Δc
 
e− jh(l0 −(m+1)δ z) − e− jh(l0 −mδ z) − e− jh(l1 −(m+1)δ z) + e− jh(l1 −mδ z)
×
h2
 
H (αc r0 , αc r1 ) I (αc r0 , αc r1 )
× R(6) (αc ) + R (7)
( αc )
αc2 αc2
 
− j α  r m
I (l αm δ r, (l + 1)αm δ r) j α  r m
H (l αm δ r, (l + 1)αm δ r)
× R m e m +e m dh
αm2  αm2 

rm < r < rm −1 (below coil) . (9.39)

9.5.2 Driving-Point Impedance of a Bobbin Coil

The next thing that we must do with the coil is compute its driving-point impedance.
We will use the reaction-formula

E · J
Z=− , (9.40)
I2

where J is the current density in the coil (= Nc Iaφ ), and I is the total current. Thus,

 
2π Nc I
Z=− E(r, z)rdrdz
I2
    
(2π )2 (Nc I)2
=− rdr dz r dr dz G(0) (r, z; r , z )
I2
)
(s)  
+ G (r, z; r , z ) , (9.41)

where G(0) (r, z; r , z ) and G(s) (r, z; r , z ) are the infinite-space and layered-space
Green functions given in (9.30). Note that the integral over G(0) gives the free-
space impedance, Z (0) , of the coil, and the integral over G(s) gives the change in
impedance, ΔZ, due to the presence of the scattering body (the tube). Hence,
162 9 Multilayered Media with Cylindrical Geometries

Z = Z (0) + ΔZ

 ∞ (c)
F (l1 − l0 ) I 2 (r0 h, r1 h)
= jω μc 2π Nc2 · dh
0 2αc h3

 ∞
' (
π  2 − e− jh(l1−l0 ) − e jh(l1 −l0 )
+ ω μc Nc2
2 −∞ h2


I 2 (αc r0 , αc r1 ) H 2 (αc r0 , αc r1 )
× R(1) (αc ) + R (2)
(αc )
αc4 αc4

(3) I (αc r0 , αc r1 )H (αc r0 , αc r1 )
+2R (αc ) dh , (9.42)
αc4

 
−1 + e −(l1−l0 )αc + α (l − l )
where F (c) (l1 − l0 ) = 2
c 1 0
is the specialization to the
αc2
coil of the F-function of (9.32).

Appendix

A.1 Cylindrical Eigenvectors and the Green Function

We will develop the general theory of eigenvectors in cylindrical coordinates and

then complete the development of the Green functions for axisymmetric problems.
Maxwell’s equations are

∇ × E = − jω μ H − M(i)
∇ × H = jω ε̂ E + J(i) , (9.43)

where M(i) and J(i) are impressed current sources, M(i) being magnetic, and J(i)
electric.
In cylindrical coordinates, the component forms of (9.43) are
A.1 Cylindrical Eigenvectors and the Green Function 163

1 ∂ Ez ∂ Eφ (i)
− = − jω μ Hr − Mr
r ∂φ ∂z

∂ Er ∂ Ez (i)
− = − jω μ Hφ − Mφ
∂z ∂r

1 ∂ (rEφ ) 1 ∂ Er (i)
− = − jω μ Hz − Mz
r ∂r r ∂φ

1 ∂ Hz ∂ Hφ (i)
− = jω ε̂ Er + Jr
r ∂φ ∂z

∂ Hr ∂ Hz (i)
− = jω ε̂ Eφ + Jφ
∂z ∂r

1 ∂ (rHφ ) 1 ∂ Hr (i)
− = jω ε̂ Ez + Jz . (9.44)
r ∂r r ∂φ

We expand all field variables in a Fourier series in φ and a Fourier integral in z:

∞  ∞
F(r, φ , z) = ∑ e jmφ
−∞
F̃(r, m, h)e jhz dh
m=−∞

 2π  ∞
1  
F̃(r, m, h) = e− jmφ d φ  e− jhz F(r, φ  , z )dz , (9.45)
(2π )2 0 −∞

where the second equation is the direct transform from physical space to frequency
(wave-number) space.
Upon transforming (9.44), we get, after cancelling the common factor, e jnφ e jhz :

jn (i)
Ẽz − jhẼφ = − jω μ H̃r − M̃r
r
d Ẽz (i)
jhẼr − = − jω μ H̃φ − M̃φ
dr
1 d(rẼφ ) n (i)
− j Ẽr = − jω μ H̃z − M̃z
r dr r
jn (i)
H̃z − jhH̃φ = jω ε̂ Ẽr + J˜r
r
d H̃z (i)
jhH̃r − = jω ε̂ Ẽφ + J˜φ
dr
1 d(rH̃φ ) n (i)
− j H̃r = jω ε̂ Ẽz + J˜z . (9.46)
r dr r
164 9 Multilayered Media with Cylindrical Geometries

The first and fourth equations are free of derivatives. Hence, they allow us to
eliminate H̃r and Ẽr in favor of the transverse components, H̃z , H̃φ , Ẽz , Ẽφ :

(i)
n h M̃r
H̃r = − Ẽz + Ẽφ −
ωμr ωμ jω μ
(i)
n h J˜r
Ẽr = H̃z − H̃φ − . (9.47)
ω ε̂ r ω ε̂ jω ε̂

The remaining equations are

d Ẽz (i)
jhẼr − = − jω μ H̃φ − M̃φ
dr

1 d(rẼφ ) n (i)
− j Ẽr = − jω μ H̃z − M̃z
r dr r

d H̃z (i)
jhH̃r − = jω ε̂ Ẽφ + J˜φ
dr

1 d(rH̃φ ) n (i)
− j H̃r = jω ε̂ Ẽz − J˜z . (9.48)
r dr r

Upon eliminating Ẽr and H̃r , we get

 
d Ẽz jhn jh2 jω μ h ˜(i) (i)
= H̃z + − + (rH̃φ ) − Jr + M̃φ
dr ω ε̂ r ω ε̂ r r ω ε̂
 
d(rẼφ ) jn2 jnh n ˜(i) (i)
= − jω μ r H̃z − (rH̃φ ) − Jr − rM̃z
dr ω ε̂ r ω ε̂ r ω ε̂
 2 
d H̃z jhn jh jω ε̂ (i) h (i)
=− Ẽz + − (rẼφ ) − J˜φ − M̃r
dr ωμr ωμr r ωμ
 
d(rH̃φ ) jn2 jnh (i) n (i)
= − + jω ε̂ r Ẽz + (rẼφ ) + rJ˜z − M̃r . (9.49)
dr ωμr ωμr ωμ
A.1 Cylindrical Eigenvectors and the Green Function 165

This system can be written in the following vector-matrix form:

⎡ ⎤ ⎡ ⎤⎡ ⎤
Ẽz 0 0 a11 (r) a12 (r) Ẽz
d ⎢ ⎥
⎢ rẼφ ⎥ =
⎢0
⎢ 0 a21 (r) a22 (r) ⎥ ⎢
⎥ ⎢ rẼφ ⎥
⎥
dr ⎣ H̃z ⎦ ⎣ b11 (r) b12 (r) 0 0 ⎦ ⎣ H̃z ⎦
rH̃φ b21 (r) b22 (r) 0 0 rH̃φ
⎡ ⎤
J˜r
⎡ ⎤
u11 0 0 0 u15 0 ⎢ ⎢
⎥
J˜φ
⎥
⎢ u21 ⎢ ⎥
0 0 0 0 u26 ⎥ ⎢ J˜z
⎥
+⎢
⎣ 0
⎥⎢ ⎥, (9.50)
u32 0 u34 0 0 ⎦ ⎢ M̃r ⎥
⎢ ⎥
0 0 u43 u44 0 0 ⎣ M̃φ ⎦
M̃z

where

jhn − jh2 jω μ jn2 − jnh

a11 = a12 = + a21 = − jω μ r a22 =
ω ε̂ r ω ε̂ r r ω ε̂ r ω ε̂ r

− jhn jh2 jω ε̂ − jn2 jnh

b11 = b12 = − b21 = + jω ε̂ r b22 =
ωμr ωμr r ωμr ωμr
(9.51)
−h −n
u11 = u15 = 1 u21 = u26 = −r
ω ε̂ ω ε̂
−h −n
u32 = −1 u34 = u43 = r u44 = .
ωμ ωμ

In order to compute the eigenvectors of (9.50), we work with the homogeneous

system, which we rewrite as dV/dr = MV. It is straightforward to prove, by
appealing to (9.51), that M2 = (n2 /r2 + h2 − ω 2 μ ε̂ )I, where I is the identity matrix.
Furthermore, we have

⎡ ⎤
0 0 0 0
dM M ⎢ 0 0 − j2ω μ 0⎥
= − +⎢
⎣
⎥.
dr r 0 0 0 0⎦
j2ω ε̂ 0 0 0

Then
166 9 Multilayered Media with Cylindrical Geometries

d2V dM dV
2
= V+M
dr dr dr

dM
= V + M2 V
dr

⎤ ⎡
0
 
1 dV n2 ⎢ − j2ω μ V3 ⎥
=− + 2 + h2 − ω 2 μ ε̂ V + ⎢
⎣
⎥.
⎦ (9.52)
r dr r 0
j2ω ε̂ V1

The solutions for the eigenvector

 2  Ẽz and H̃z , are obtained imme-
components,
d 2 v 1 dv n
diately from 2 + − 2 + h2 − ω 2 μ ε̂ v = 0. This is satisfied by the two
dr r dr r
linearly-independent functions, Jn (α r), the Bessel function of the first kind, of
(2)
order n, and Hn (α r), the Hankel function of the second kind, of order n, where
α = (k2 − h2 )1/2 and k2 = ω 2 μ ε̂ = ω 2 μ (ε + σ / jω ).
In order to compute the components, rẼφ and rH̃φ , we do not use (9.52), but
return to the fundamental system (9.50). Using the first and third equations, we
have d Ẽz /dr = a11 (r)H̃z + a12(r)(rH̃φ ). Hence
 
1 d Ẽz
H̃φ = − a11(r)H̃z
ra12 (r) dr
 
1 d Ẽz hn
= − 2 jω ε̂ + H̃z .
α dr r
By duality, we have
 
1 d H̃z
Ẽφ = − b11(r)Ẽz
rb12 (r) dr
 
1 d H̃z hn
= j ω μ − Ẽ z .
α2 dr r
Hence, the four eigenvectors of (9.50) are
⎡ ⎤ ⎡ (2) ⎤ ⎡ ⎤⎡ ⎤
⎡ ⎤ Jn (α r) Hn (α r) 0 0
Ẽz ⎢ hn ⎥ ⎢ ⎥ ⎢ jω μ  ⎥ ⎢ j ω μ (2) ⎥
⎢ ⎥ ⎢ − Jn (α r) ⎥ ⎢ − hn Hn(2) (α r) ⎥ ⎢ ⎥ ⎢ ⎥
⎢ rẼφ ⎥ ⎢ α2 ⎥ ⎢ α2 ⎥ ⎢ α rJn (α r) ⎥ ⎢ α rHn (α r) ⎥
⎢ ⎥=⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ (2) ⎥ .
⎣ H̃z ⎦ ⎢ 0 ⎥ ⎢ 0 ⎥ ⎢ Jn (α r) ⎥ ⎢ Hn (α r) ⎥
⎣ ⎦ ⎣ ⎦ ⎣ hn ⎦ ⎣ ⎦
rH̃φ j ω ε̂  j ω ε̂ (2) hn (2)
− rJn (α r) − rHn (α r) − 2 Jn (α r) − 2 Hn (α r)
α α α α
(9.53)
A.1 Cylindrical Eigenvectors and the Green Function 167

We label the vectors, V1 , V2 , V3 , V4 , respectively, from left to right. It is

straightforward to show that these vectors satisfy the homogeneous form of (9.50).
The first two eigenmodes are TM (Transverse Magnetic) to z, because H̃z = 0,
and the last two are TE (Transverse Electric) to z, because Ẽz = 0. When we
specialize these equations to the axisymmetric case, in which the azimuthal Fourier
wavenumber, n, vanishes, we will get another form of TE and TM, namely in the
radial direction. This, as we will see, is of great advantage in solving problems with
radially layered media.
We now establish the conditions of continuity/discontinuity on V at singular
current sources. Before doing this, we note that the Fourier transform of a unit
point-current source located at φ = φ  , z = z is given by F {δ (φ − φ  )δ (z − z )} =
1 − jnφ  − jhz
e e .
(2π )2
Now, if we have a unit vector current source located at (r , φ  , z ), then the source
term in (9.50) becomes
 
δ (r − r ) e− jnφ e− jhz
UI , (9.54)
r (2π )2
where I is a unit vector. Then upon integrating (9.50) an epsilon distance with
respect to r across the cylinder r = r , we get
 
U e− jnφ e− jhz
V(+) − V(−) = I , (9.55)
r (2π )2
where the superscript (+) denotes the limit r → r from above and (−) denotes the
limit r → r from below.

A.1.1 Application to Axisymmetric Systems

The specialization of the preceding to axisymmetric problems requires only that we

set the azimuthal wave number, n, to zero. Then the eigenvectors become:
⎡ ⎤ ⎡ (2)
⎤
J0 (α r) H0 (α r)
⎢ ⎥ ⎢ ⎥
⎢ 0 ⎥ ⎢ 0 ⎥
⎢
V1 = ⎢ ⎥ ⎢ ⎥
0 ⎥ , V2 = ⎢ 0 ⎥
⎣ ω ε̂  ⎦ ⎣ ω ε̂ (2) ⎦
−j rJ0 (α r) −j rH0 (α r)
α α
⎡ ⎤ ⎡ ⎤
0 0
⎢ j ω μ rJ  (α r) ⎥ ⎢ ω μ (2) ⎥
⎢ ⎥ ⎢j rH0 (α r) ⎥
V3 = ⎢ α 0 ⎢
⎥ , V4 = ⎢ α ⎥. (9.56)
⎣ J0 (α r) ⎦ (2) ⎥
⎣ H (α r) 0
⎦
0 0
168 9 Multilayered Media with Cylindrical Geometries

When we recall (9.47), we see that H̃r = 0, Ẽr = 0, for axisymmetric problems,
which means that we have TE and TM fields with respect to the radial direction, as
well.
In order to compute the Green function, recall Fig. 9.5. We use eigenvectors V1
and V3 in the region next to r = 0, because these two vectors contain J0 (α1 r), which
(2)
is continuous at r = 0. Conversely, V2 and V4 contain H0 (α1 r), which, because it
represents an out-going cylindrical wave, satisfies the radiation condition at infinity,
which includes Region 1 .
Our strategy is to eliminate (a2 , b2 , c2 , d2 ), . . . , (ak , bk , ck , dk ) in favor of (a1 , c1 ),
and (a2 , b2 , c2 , d2 ), . . . , (ak , bk , ck , dk ) in favor of (b1 , d1 ), and then use jump
conditions at r = r to determine (a1 , c1 , b1 , d1 ). The intermediate steps are
accomplished by equating the four-vectors at each interface.
For example, at r = r1 , we have a1 V11 + c1 V31 = a2 V12 + b2 V22 + c2 V32 + V42 ,
which produces the two systems

(2)
a1 J0 (α1 r1 ) = a2 J0 (α2 r1 ) + b2H0 (α2 r1 )
  
α2 ε̂1 (2)
a1 J0 (α1 r1 ) = a2 J0 (α2 r1 ) + b2H0 (α2 r1 )
α1 ε̂2

(2)
c1 J0 (α1 r1 ) = c2 J0 (α2 r1 ) + d2H0 (α2 r1 )
  
α2 μ1  (2)
c1 J (α1 r1 ) = c2 J0 (α2 r1 ) + d2H0 (α2 r1 ) . (9.57)
α1 μ2 0

Similarly, at the bottom, r = r1 , we have

(2) (2)
b1 H0 (α1 r1 ) = a2 J0 (α2 r1 ) + b2 H0 (α2 r1 )
  
α2 ε̂1 (2) (2)
b 1 H0 (α1 r1 ) = a2 J0 (α2 r1 ) + b2 H0 (α2 r1 )
α1 ε̂2

(2) (2)
d1 H0 (α1 r1 ) = c2 J0 (α2 r1 ) + d2 H0 (α2 r1 )
  
α2 μ1  (2) (2)
d 1 H0 (α1 r1 ) = c2 J0 (α2 r1 ) + d2 H0 (α2 r1 ) . (9.58)
α1 μ2 

In general, at r = rM ,

aM V1M + bM V2M + cM V3M + dM V4M

= aM+1 V1(M+1) + bM+1 V2(M+1) + cM+1 V3(M+1) + dM+1 V4(M+1)
A.1 Cylindrical Eigenvectors and the Green Function 169

produces

(2) (2)
aM J0 (αM rM ) + bM H0 (αM rM ) = aM+1 J0 (αM+1 rM ) + bM+1 H0 (αM+1 rM )

     
αM+1 ε̂M αM+1 ε̂M (2)
aM J0 (αM rM ) + bM H0 (αM rM )
αM ε̂M+1 αM ε̂M+1

(2)
= aM+1 J0 (αM+1 rM ) + bM+1 H0 (αM+1 rM )

(2) (2)
cM J0 (αM rM ) + dM H0 (αM rM ) = cM+1 J0 (αM+1 rM ) + dM+1H0 (αM+1 rM )

     
αM+1 μM αM+1 μM (2)
cM J0 (αM rM ) + dM H0 (αM rM )
αM μM+1 αM μM+1

(2)
= cM+1 J0 (αM+1 rM ) + dM+1 H0 (αM+1 rM ) . (9.59)

This system holds also for r = rM .

What about the source interface, r = r ? We will use (9.55) (with n = 0 since this
is an axisymmetric problem). Furthermore, since we need only consider electric
sources (because magnetic sources can be derived from electric ones, by virtue of
Amperian currents), we have J˜r = 0, J˜φ = 1, J˜z = 0, M̃r = 0, M̃φ = 0, M̃z = 0.
Hence, upon invoking (9.50) and (9.51) we have
⎡
⎤ ⎡ ⎤
0 0
⎢ 0 ⎥ ⎢ 0 ⎥
UI = ⎢ ⎥ ⎢ ⎥
⎣ u32 ⎦ = ⎣ −1 ⎦ , (9.60)
0 0

so that (9.55) becomes

⎤ ⎡
0
1 ⎢ ⎥ − jhz

V(s ) (r ) − V(s) (r ) = ⎢ 0 ⎥e . (9.61)
(2π )2 ⎣ −1 ⎦ r
0

Upon referring, once again, to Fig. 9.5, (9.61) is seen to produce the system
170 9 Multilayered Media with Cylindrical Geometries

⎡ ⎤ ⎡ (2) ⎤
J0 (αs r ) H0 (αs r )
⎢ ⎥ ⎢ ⎥
⎢ 0 ⎥ ⎢ 0 ⎥
⎢
(as − as ) ⎢ ⎥ ⎢ ⎥
0 ⎥ + (bs − bs ) ⎢ 0 ⎥
⎣ ω ε̂s   ⎦ ⎣ ω ε̂  ⎦
(2)
−j r J0 (αs r ) −j
s
r H0 (αs r )
αs αs
⎡ ⎤ ⎡ ⎤
0 0
⎢ ω μs   ) ⎥ ⎢ ω μs  (2) ⎥
⎢j r J (α r ⎥ ⎢j r H0 (αs r ) ⎥
+(cs − cs ) ⎢ ⎥ ⎢ α ⎥
s
⎢ αs
0
⎥ + (d s  − d )
s ⎢ s ⎥
(2)
⎣ J0 (αs r ) ⎦ ⎣ H0 (αs r ) ⎦
0 0
⎡ ⎤
0
⎢ 0 ⎥ − jhz
⎢ ⎥e
=⎢
⎢−1
⎥
⎥ (2π )2 . (9.62)
⎣ r ⎦
0

The solution of (9.62) is obtained in a straightforward manner:

a s − a s = 0
b s − b s = 0
(2) 
H1 (αs r )παs e− jhz
c s − c s = −
j2 (2π )2


J1 (αs r )παs e− jhz
d s − d s = . (9.63)
j2 (2π )2

In arriving at the final form of (9.63), we made use of the Bessel-function

(2) (2)
identities, J0 (z) = −J1 (z), H0 (z) = −H1 (z), and also of the Wronskian identity,
(2) (2)  2
H0 (z)J0 (z) − J0 (z)H0 (z) = j .
πz
Our problem is considerably simplified by realizing that all a’s and b’s vanish.
This agrees with the fact that Ez = 0 and Hφ = 0 in this problem, and the only places
that Ez and Hφ appear are in the eigenvectors, V1 , V2 , whose excitation coefficients
are a and b. a = 0, b = 0 satisfies (9.59) and (9.63), of course. Hence, all we need
to work with are the last two equations in (9.59) and (9.63).
For computational convenience, we introduce the scaled Hankel and Bessel
functions, h0,1 (z), j0,1 (z), through the relations
A.1 Cylindrical Eigenvectors and the Green Function 171

(2)
H0,1 (z) = e− jz h0,1 (z)

J0,1 (z) = e jz j0,1 (z) . (9.64)

Wherever the coefficient c appears in (9.59), it is multiplied by J, and d is multiplied

by H (2) . Hence, we introduce the scaled coefficients, c , d  through cJ = c j, dH =
d  h, which yields

c = ce jz
d  = de− jz . (9.65)

Now we proceed to solve the scaled version of (9.59)(c,d)

(2)
cM j0 (αM rM ) + dM

h0 (αM rM ) = cM+1 e− jαM+1 ΔM+1 j0 (αM+1 rM )
(2)

+ dM+1 e jαM+1 ΔM+1 h0 (αM+1 rM )

     
αM+1 μM αM+1 μM (2)
cM 
j1 (αM rM ) +dM h1 (αM rM )
αM μM+1 αM μM+1

(2)
= cM+1 e− jαM+1 ΔM+1 j1 (αM+1 rM ) +dM+1

e jαM+1 ΔM+1 h1 (αM+1 rM ) , (9.66)

where we have used the derivative identities defined below (9.63). We have also
defined

ΔM+1 = rM+1 − rM
ΔM +1 = −(rM +1 − rM ) , (9.67)

where ΔM +1 is defined with a negative sign, because rM +1 < rM .
We express (9.66) in matrix notation
    
cM+1 T11 (M + 1, M) T12 (M + 1, M) cM
 =  , (9.68)
dM+1 T21 (M + 1, M) T22 (M + 1, M) dM

where
172 9 Multilayered Media with Cylindrical Geometries

παM+1 rM e jαM+1 ΔM+1  (2)

T11 (M + 1, M) = h1 (αM+1 rM ) j0 (αM rM )
j2
  
(2) αM+1 μM
− h0 (αM+1 rM ) j1 (αM rM )
αM μM+1

παM+1 rM e jαM+1 ΔM+1  (2) (2)

T12 (M + 1, M) = h1 (αM+1 rM )h0 (αM rM )
j2
  
(2) (2) αM+1 μM
− h0 (αM+1 rM )h1 (αM rM )
αM μM+1

  
παM+1 rM e− jαM+1 ΔM+1 αM+1 μM
T21 (M + 1, M) = j0 (αM+1 rM ) j1 (αM rM )
j2 αM μM+1

− j1 (αM+1 rM ) j0 (αM rM )]

  
παM+1 rM e− jαM+1 ΔM+1 αM+1 μM (2)
T22 (M + 1, M) = j0 (αM+1 rM )h1 (αM rM )
j2 αM μM+1

(2)
− j1 (αM+1 rM )h0 (αM rM ) . (9.69)

In arriving at the final version of (9.69), we used the fact that the Wronskian for
(2)
unscaled Bessel functions holds also for scaled Bessel functions; j0 (z)h1 (z) −
(2)
j1 (z)h0 (z) = j2/π z.
The T-matrix is called the transition matrix from region M to M + 1. The
boundary condition on (9.68) is d1 = 0, c1 = 0.
The corresponding transition matrix from M  to M  + 1 (i.e., below the source) is:
    
cM +1 T11 (M  + 1, M  ) T12 (M  + 1, M  ) cM
 = , (9.70)
dM  +1 T21 (M  + 1, M  ) T22 (M  + 1, M  ) dM


where
A.1 Cylindrical Eigenvectors and the Green Function 173

παM +1 rM  e− jαM +1 ΔM +1  (2)

T11 (M  + 1, M  ) = h1 (αM +1 rM  ) j0 (αM  rM  )
j2

  
(2) αM  +1 μM 
− h0 (αM  +1 rM  ) j1 (αM  rM  )
αM  μM +1

παM +1 rM  e− jαM +1 ΔM +1  (2) (2)

T12 (M  + 1, M  ) = h1 (αM +1 rM  )h0 (αM  rM  )
j2

  
(2) (2) αM  +1 μM 
− h0 (αM  +1 rM  )h1 (αM  rM  )
αM  μM  +1

  
παM +1 rM  e jαM +1 ΔM +1 αM  +1 μM 
T21 (M  + 1, M  ) = j0 (αM +1 rM  ) j1 (αM  rM  )
j2 αM  μM +1

− j1 (αM +1 rM  ) j0 (αM  rM  )]

  
παM +1 rM  e jαM +1 ΔM +1 αM  +1 μM  (2)
T22 (M  + 1, M  ) = j0 (αM +1 rM  )h1 (αM  rM  )
j2 αM  μM +1


(2)
− j1 (αM +1 rM  )h0 (αM  rM  ) . (9.71)

In the source region, define the scaled source variables as:

cs = e jαs rs −1 cs , cs = e jαs rs−1 cs

ds = e− jαs rs −1 ds , ds  = e− jαs rs−1 ds . (9.72)

In terms of these scaled coefficients, the source equations, (9.63), become


παs jαs rs −1 (2) e− jhz
e jαs (rs −1 −rs−1 ) cs − cs =− e H1 (αs r )
j2 (2π )2


παs − jαs rs−1 e− jhz
ds  − e jαs (rs −1 −rs−1 ) ds = e J1 (αs r ) . (9.73)
j2 (2π )2

Note that all exponentials have exponents that are proportional to the difference
of nearby radii. For example, rs −1 − rs−1 is the width of the source region,
(2) 
and e jαs rs−1 H0,1 (αs r ) → e− jαs (r −rs−1 ) , where r > rs−1 , and e− jαs rs −1 J0,1 (αs r ) →

e− jαs (rs −1 −r ) , where rs −1 > r . Hence, the right-hand sides of (9.73) vanish to
174 9 Multilayered Media with Cylindrical Geometries

exponential order for ℜαs → ∞, thereby ensuring proper behavior of the scaled
coefficients. Note that the unknowns for the entire system of equations are scaled
coefficients.
Now
   
cs c1
= T (s, s − 1) · · · T (2, 1)
ds 0
   
cs 0
= T (s , s − 1) · · · T (2 , 1 ) , (9.74)
ds  d1 

where T is a 2 × 2 matrix, with entries T11 , T12 , T21 , T22 , as defined in (9.69) and
(9.71).
Call

T (s, s − 1) · · · T (2, 1) = V (s)

T (s , s − 1) · · ·T (2 , 1 ) = U(s ) , (9.75)

then

cs = V11 (s)c1 , ds = V21 (s)c1

(s ) (s )
cs = U12 d1  , ds  = U22 d1  . (9.76)

In general

cM = V11 (M)c1 , dM


= V21 (M)c1
cM = U12 (M  )d1  , dM
  
 = U22 (M )d1 . (9.77)

Couple (9.76) with (9.73), and call rs −1 − rs−1 = Δs ; the results are then

παs jαs r  (2) e− jhz
e jαs Δs U12 (s )d1  − V11(s)c1 = − e s −1 H1 (αs r )
j2 (2π )2


παs − jαs rs−1 e− jhz
U22 (s
)d1  − e jαs Δs V21 (s)c1 = e J1 (αs r ) , (9.78)
j2 (2π )2
A.1 Cylindrical Eigenvectors and the Green Function 175

which has for its solution:

(2) 
παs U12 (s )e jαs Δs e− jαs rs−1 J1 (αs r ) + U22(s )e jαs rs −1 H1 (αs r ) e− jhz
c1 =
j2 U22 (s )V11 (s) − e j2αs Δs U12 (s )V21(s) (2π )2

(2) 
παs V11 (s)e− jαs rs−1 J1 (αs r )+V21 (s)e jαs rs −1 e jαs Δs H1 (αs r ) e− jhz
d1  = (9.79)
j2 U22 (s )V11 (s)−e j2αs Δs U12 (s )V21 (s) (2π )2

This is the same as (9.12), with which we started the calculation of matrices.
Chapter 10
Some Special Topics in Computational
Electromagnetics

10.1 Spatial Decomposition Algorithm

“Spatial decomposition” is a generic term for the first step in transforming a

functional equation into its discrete version prior to numerical computation.
Umashankar [51] introduced the term “spatial decomposition technique” for a
method of solving two-dimensional scattering problems from electrically large
bodies using boundary-integral equations. As far as we know, we are the first to
apply the “spatial decomposition algorithm” to the solution of three-dimensional
scattering problems when the scatterer (the “flaw”) occupies several regions with
different electrical properties using volume-integral equations.
If the flaw extends over two or more layers with different electrical constitutive
properties, as in Fig. 10.1, then (4.9) and (4.10) still hold in each layer (with different
matrices), but now the incident field-moment vectors, [E(ix) , E(iy) , E(iz) ], [B(ix) ,
B(iy) , B(iz) ], depend upon the coil current plus anomalous currents in other layers.
This means that we must introduce new Green dyadics; the Green dyadic that we
have worked with previously was that for which the source and field points were in
the same layer of the workpiece. Now, upon referring to Fig. 2.1 of Chap. 2, we need
the Green function, G(±q0) , in which the source point is in region 0, as before, but the
field point is in the qth layer above region 0, or the qth layer below the region.
The first situation carries the label +q0, and the second −q0.
We compute the Green functions using the same four-vector algebraic approach,
(2.15), that was used in Chap. 2, and then the resulting volume-integral differential
equation is derived exactly as explained in Chaps. 2 and 3. Because the Amperian
currents, which are of magnetic origin, are actually equivalent electric currents
[because they appear in the second of Maxwell’s equations (Ampere’s law)], we
only need to compute the electric-electric dyadic Green functions.
The resulting differential volume-integral relations for the electric anomalous
current, J(e) (r), and the electric current of magnetic origin (the Amperian current or
simply “magnetic current”), J(m) (r), are

H.A. Sabbagh et al., Computational Electromagnetics and Model-Based Inversion, 177

Scientific Computation, DOI 10.1007/978-1-4419-8429-6 10,
© Springer Science+Business Media New York 2013
178 10 Some Special Topics in Computational Electromagnetics

Region A Flaw

Region B

Fig. 10.1 A flaw in multiple layers


(e)
Ex (r) = G(q0)(1) (x − x , y − y; z, z )Jx (r )dr

(m)
+ G(q0)(1) (x − x , y − y; z, z )Jx (r )dr

∂ (e)
+ G(q0)(2) (x − x , y − y ; z, z )∇t · Jt (r )dr
∂x

∂ (m)
+ G(q0)(2) (x − x , y − y ; z, z )∇t · Jt (r )dr
∂x

1 ∂
+ 2 G(q0)(3) (x − x , y − y ; z, z )∇ · J(e) (r )dr
k0 ∂ x

(e)
Ey (r) = G(q0)(1) (x − x , y − y; z, z )Jy (r )dr

(m)
+ G(q0)(1) (x − x , y − y; z, z )Jy (r )dr

∂ (e)
+ G(q0)(2) (x − x , y − y; z, z )∇t · Jt (r )dr
∂y

∂ (m)
+ G(q0)(2) (x − x , y − y; z, z )∇t · Jt (r )dr
∂y

1 ∂
+ G(q0)(3) (x − x , y − y; z, z )∇ · J(e) (r )dr
k02 ∂ y

(e)
Ez (r) = G(q0)(4) (x − x , y − y; z, z )Jz (r )dr

(m)
+ G(q0)(4) (x − x , y − y; z, z )Jz (r )dr

1 ∂
+ G(q0)(5)(x − x , y − y ; z, z )∇ · J(e) (r )dr , (10.1)
k02 ∂ z
10.1 Spatial Decomposition Algorithm 179

which hold for r in region q and r in region 0. The notation for the Z-layering,
including the source region, is shown in Fig. 2.1 of Chap. 2. If r is in region −q,
then replace (q0) by (−q0). In deriving the final version of (10.1), we used the fact
that ∇ · J(m) (r) = 0.
The kernels that appear in (10.1) are given by
 ∞
(q0)(1) j ω μq B̃q0 (E)
G =− G̃q0 (z, z )J0 (rl)ldl
2π 0 2λ0
⎡ ⎤
(E)  μ0 λ q (M)
 ∞ ⎢ B̃q0 G̃q0 (z, z ) − D̃q0 G̃q0 (z, z )
j ω μq μq λ 0 ⎥
G(q0)(2) =− ⎢ ⎥ J0 (rl)ldl
2π 0 ⎣ 2λ0 l 2 ⎦

 ∞
j ω μq μ0 λq D̃q0 (M)
G(q0)(3) = − G̃q0 (z, z )J0 (rl)ldl
2π 0 μq λ 0 2 λ 0
 ∞
(q0)(4) j ω μq μ0 D̃q0 (M)
G =− G̃q0 (z, z )J0 (rl)ldl
2π 0 μq 2 λ 0
 ∞
j ω μq μ0 λ0 D̃q0 (M)
G(q0)(5) = − G̃q0 (z, z )J0 (rl)ldl
2π 0 μq λ q 2 λ 0
 ∞
jω μ−q Ã−q0 (E)
G(−q0)(1) = − G̃ (z, z )J0 (rl)ldl
2π 2λ0 −q00
⎡ ⎤
(E)
(z,  ) − μ0 λ−q C̃ (M)
(z, )
 Ã G̃
−q0 −q0 z G̃
−q0 −q0 z
jω μ−q ∞ ⎢
⎢ μ−q λ0 ⎥
⎥ J0 (rl)ldl
G(−q0)(2) =− ⎣ ⎦
2π 0 2 λ 0 l 2

 ∞
jω μ−q μ0 λ−q C̃−q0 (M)
G(−q0)(3) = − G̃−q0 (z, z )J0 (rl)ldl
2π 0 μ−q λ0 2λ0
 ∞
(−q0)(4) jω μ−q μ0 C̃−q0 (M)
G =− G̃−q0 (z, z )J0 (rl)ldl
2π 0 μ−q 2λ0
 ∞
jω μ−q μ0 λ0 C̃−q0 (M)
G(−q0)(5) = − G̃−q0 (z, z )J0 (rl)ldl , (10.2)
2π 0 μ−q λ−q 2λ0

where r = [(x − x )2 + (y − y)2 ]1/2 , and the source point, z , is in region 0, and the
field point, z, is in region ±q.
180 10 Some Special Topics in Computational Electromagnetics

The G̃’s that appear in (10.2) are given by the product of terms that involve the
source parameters only and those that involve the field parameters:
⎡ ⎤
e −λ0 (z0 −z ) + R(E) e−λ0 (z0 +z −2z−1 )  
(E) ⎦ e−λq (z−zq ) + R(E) eλq (z−zq )
G̃q0 (z, z ) = ⎣
0(−)
(E) (E) q(+)
1 − R0(+)R0(−) e−2λ0 Δ0
⎡ ⎤
 (M) 
(M)
e−λ0 (z0 −z ) + R0(−)e−λ0 (z0 +z −2z−1 )  
G̃q0 (z, z ) = ⎣ ⎦ e−λq (z−zq ) + R(M) eλq (z−zq )
(M) (M) q(+)
1 − R0(+)R0(−) e−2λ0 Δ0
⎡ ⎤
e −λ0 (z0 −z ) − R(M) e−λ0 (z0 +z −2z−1 )  
(M) 
G̃q0 (z, z ) = ⎣
0(−) ⎦ e−λq (z−zq ) − R(M) eλq (z−zq )
(M) (M) q(+)
1 − R0(+)R0(−) e−2λ0 Δ0
⎡ ⎤
−λ0 (z −z−1 ) (E) −λ0 (2z0 −z −z−1 )
(E)
e + R e
G̃−q0 (z, z ) = ⎣ ⎦
0(+)
(E) (E) −2λ Δ
1 − R0(+)R0(−) e 0 0
 
(E)
eλ−q (z−z−(q+1) ) + R−q(−)e−λ−q (z−z−(q+1) )
⎡ ⎤
e −λ0 (z −z−1 ) + R(M) e−λ0 (2z0 −z −z−1 )
(M)
G̃−q0 (z, z ) = ⎣ ⎦
0(+)
(M) (M) −2λ Δ
1 − R0(+)R0(−) e 0 0
 
(M)
eλ−q (z−z−(q+1) ) + R−q(−)e−λ−q (z−z−(q+1) )
⎡ ⎤
e −λ0 (z −z−1 ) − R(M) e−λ0 (2z0 −z −z−1 )
(M) 
G̃−q0 (z, z ) = ⎣ ⎦
0(+)
(M) (M)
1 − R0(+)R0(−) e−2λ0 Δ0
 
(M)
eλ−q (z−z−(q+1) ) − R−q(−)e−λ−q (z−z−(q+1) ) . (10.3)

(E) (E) (M) (M)

The reflection coefficients, R0(+) , R0(−) , R0(+) , R0(−) , correspond, respectively,
(E) (E) (M) (M)
to R1 , R−1 , R1 , R−1 , of Chap. 2, and Δ0 corresponds to T , of Chap. 2.
(E,M)
The reflection coefficients, R(±)q(±) are defined in Fig. 2.5 of Chap. 2, where we also
derived a recursion relation for computing these coefficients. The transfer functions,
Ã−q0 , B̃q0 , C̃−q0 , and D̃q0 , are defined by recursion relations that follow from the
four-vector approach, but will be omitted here.
We discretize the functional equations of (10.1) in the usual way, by first
substituting the electric- and magnetic-current expansions, (4.4) and (4.8) of
10.1 Spatial Decomposition Algorithm 181

Chap. 4, into (10.1), and then testing with the electric facet elements, (4.2) of
Chap. 4. The result is:

∑ Tklm,KLM ∑ Tklm,KLM ∑ Tklm,KLM

(x) (ee)(xx)(q0) (x)(e) (ee)(xy)(q0) (y)(e) (ee)(xz)(q0) (z)(e)
Eklm = JKLM + JKLM + JKLM
KLM KLM KLM

∑ Tklm,KLM ∑ Tklm,KLM ∑ Tklm,KLM

(em)(xx)(q0) (x) (em)(xy)(q0) (y) (em)(xz)(q0) (z)
+ MKLM + MKLM + MKLM
KLM KLM KLM

∑ Tklm,KLM ∑ Tklm,KLM ∑ Tklm,KLM

(y) (ee)(yx)(q0) (x)(e) (ee)(yy)(q0) (y)(e) (ee)(yz)(q0) (z)(e)
Eklm = JKLM + JKLM + JKLM
KLM KLM KLM

∑ Tklm,KLM ∑ Tklm,KLM ∑ Tklm,KLM

(em)(yx)(q0) (x) (em)(yy)(q0) (y) (em)(yz)(q0) (z)
+ MKLM + MKLM + MKLM
KLM KLM KLM

∑ Tklm,KLM ∑ Tklm,KLM ∑ Tklm,KLM

(z) (ee)(zx)(q0) (x)(e) (ee)(zy)(q0) (y)(e) (ee)(zz)(q0) (z)(e)
Eklm = JKLM + JKLM + JKLM
KLM KLM KLM

∑ Tklm,KLM ∑ Tklm,KLM ∑ Tklm,KLM

(em)(zx)(q0) (x) (em)(zy)(q0) (y) (em)(zz)(q0) (z)
+ MKLM + MKLM + MKLM
KLM KLM KLM

∑ Tklm,KLM ∑ Tklm,KLM ∑ Tklm,KLM

(x) (me)(xx)(q0) (x)(e) (me)(xy)(q0) (y)(e) (me)(xz)(q0) (z)(e)
Bklm = JKLM + JKLM + JKLM
KLM KLM KLM

∑ ∑ ∑ Tklm,KLM
(mm)(xx)(q0) (x) (mm)(xy)(q0) (y) (mm)(xz)(q0) (z)
+ Tklm,KLM MKLM + Tklm,KLM MKLM + MKLM
KLM KLM KLM

∑ ∑ ∑ Tklm,KLM
(y) (me)(yx)(q0) (x)(e) (me)(yy)(q0) (y)(e) (me)(yz)(q0) (z)(e)
Bklm = Tklm,KLM JKLM + Tklm,KLM JKLM + JKLM
KLM KLM KLM

∑ ∑ ∑ Tklm,KLM
(mm)(yx)(q0) (x) (mm)(yy)(q0) (y) (mm)(yz)(q0) (z)
+ Tklm,KLM MKLM + Tklm,KLM MKLM + MKLM
KLM KLM KLM

∑ ∑ ∑ Tklm,KLM
(z) (me)(zx)(q0) (x)(e) (me)(zy)(q0) (y)(e) (me)(zz)(q0) (z)(e)
Bklm = Tklm,KLM JKLM + Tklm,KLM JKLM + JKLM
KLM KLM KLM

∑ ∑ ∑ Tklm,KLM
(mm)(zx)(q0) (x) (mm)(zy)(q0) (y) (mm)(zz)(q0) (z)
+ Tklm,KLM MKLM + Tklm,KLM MKLM + MKLM,
KLM KLM KLM

(10.4)
(q0) (q0)
where the transfer matrices satisfy Tklm,KLM = Tk−K,l−L;m,M ; i.e., they are Töplitz
(2D-convolution) in (X,Y ). Thus, one can still use two-dimensional FFTs to
efficiently execute the computations. The superscript (ee) on the transfer ma-
trices denotes “electric–electric,” which means that these matrices “transfer” an
electric current in region 0 into an electric field in region q. Similar interpreta-
tions attach to “electric–magnetic,” “magnetic–electric,” and “magnetic–magnetic”
transfer matrices; all transfer a source in region 0 into a field in region q. The other
transfer matrices can be deduced from the electric–electric ones by virtue of
Maxwell’s equations.
182 10 Some Special Topics in Computational Electromagnetics

For example, the magnetic-field moments shown on the left-hand side of the
bottom three equations in (10.4) can be easily computed in terms of the electric-
field moments:

(i)(x) (x)(m)
Bklm = B(i) (r) · Tklm (r)ax dr

1 (x)(m)
=− ∇ × E(i) (r) · Tklm (r)ax dr
jω

1 (x)(m)
=− E(i) (r) · ∇ × Tklm (r)ax dr
jω
 !
1
=− E(i) (r) · π1k (x)π2l (y)π2m

(z)ay
jω
"
−π1k (x)π2l (y)π2m (z)az dr
 (y) (y) (z) (z)

1 Eklm − Eklm+1 Eklm − Ekl+1m
=− −
jω δz δy
 (z) (z) (x) (x)

(y) 1 Eklm − Ek+1lm Eklm − Eklm+1
Bklm =− −
jω δx δz
 (x) (x) (y) (y)

(z) 1 Eklm − Ekl+1m Eklm − Ek+1lm
Bklm =− − . (10.5)
jω δy δx

This is simply −(1/ jω ) × the discrete curl operator, as we expected.

10.1.1 Summary of Discrete Equations With Transfer Matrices

The master discrete equations, (4.9) and (4.10), are augmented by the transfer
matrices when there is a second grid containing anomalous currents, yielding the
following equations for a coupled problem between grids q and 0:
⎡ ⎤ ⎡
E(i)(q) Q(ee)(qq) − G(ee)(qq) −G(em)(qq)
⎢ B(i)(q) ⎥ ⎢ G(me)(qq) Q(mm)(qq) + G(mm)(qq)
⎢ ⎥ ⎢
⎣ E(i)(0) ⎦ = ⎣ −T(ee)(0q) −T(em)(0q)
B(i)(0) −T (me)(0q)
−T(mm)(0q)
⎤ ⎡ (e)(q) ⎤
−T(ee)(q0) −T(em)(q0) J
−T(me)(q0) −T(mm)(q0) ⎥ ⎢ M(q) ⎥
⎥⎢ ⎥
Q(ee)(q0) − G(ee)(00) −G(em)(00) ⎦ ⎣ J(e)(0) ⎦. (10.6)
G(me)(00) Q(mm)(00) + G(mm)(00) M(0)
10.1 Spatial Decomposition Algorithm 183

Field Grid

Region A J =0 Flaw Source

a
Grid
Ja
Region B

Fig. 10.2 Showing a flaw that spans two regions and a grid in the probe region

10.1.2 Application: Computing the Scattered Fields

in the Vicinity of a Probe.

Consider Fig. 10.2, which shows a flaw spanning two regions, together with its
“source grid,” and a “field grid” in the upper-region where the probe is located.
The objective is to calculate the scattered magnetic field in the probe region due to
the anomalous current, Ja , of the flaw, because this is the variable that a magnetic
sensor, such as a Hall-effect probe or a giant magnetoresistive (GMR) sensor, would
measure.
We assume that we have used (10.4) to compute the magnetic-field moments on
the “Field Grid” of Fig. 10.2, so now we want to compute the B-field, itself. To do
this, we expand the B-field in terms of the magnetic edge elements:

∑ bKLM TKLM
(x) (x)(m)
B(x) (r) = (r)
KLM

∑ bKLM TKLM
(y) (y)(m)
B(y) (r) = (r)
KLM

∑ bKLM TKLM
(z) (z)(m)
B(z) (r) = (r), (10.7)
KLM

(x) (y) (z)

where the expansion coefficients, bKLM , bKLM , bKLM , are to be determined from
knowledge of the field moments. This is done as follows:
 
∑
(x) (x)(m) (x) (x)(m) (x)(m)
B (r)Tklm (r)dr = bKLM TKLM (r)Tklm (r)dr
KLM

∑
(x)
= δx bKLM δkK π2L (y)π2M (z)π2l (y)π2m (z)dydz
KLM

= δ x ∑ Qlm,LM bkLM
(yz) (x)

LM
184 10 Some Special Topics in Computational Electromagnetics


(r)dr = δ y ∑ Qkm,KM bKlM
(y)(m) (xz) (y)
B(y) (r)Tklm
KM

= δ z ∑ Qkl,KL bKLm.
(z)(m) (xy) (z)
B(z) (r)Tklm (r)dr (10.8)
KL

(yz) (xz) (xy)

The Qlm,LM , Qlm,LM , and Qlm,LM are tri-diagonal matrices that can be readily
(x) (y)
computed. Equation (10.8) is easily solved, and the solution vectors bKLM , bKLM and
(z)
bKLM can then be inserted into (10.7) to compute the field at any point within the
“Field Grid” of Fig. 10.2.
We can ask “In what sense does (10.7) satisfy the condition, ∇ · B = 0?”
To answer this question, we observe from the preceding equations that the b’s satisfy

(x) (x) (y) (y) (z) (z)

Bk+1lm − Bklm Bkl+1m − Bklm Bklm+1 − Bklm
+ + =0
δx δy δz
   
= ∑ Qlm,LM bk+1LM − bkLM + ∑ Qkm,KM bKl+1M − bKlM
(yz) (x) (x) (xz) (y) (y)

LM KM
 
+ ∑ Qkl,KL bKLm+1 − bKLm ,
(xy) (z) (z)
(10.9)
KL

where the first equality comes from (10.5). Hence, the divergence condition is met
in the discrete sense involving moments.

10.1.3 Test Results of Field Calculation in Probe Region

In a test problem, in which a probe is centered on a cubical flaw within a workpiece,

(x) (y) (z)
we calculate the coefficients, bKLM , bKLM , bKLM , of the magnetic edge elements
of the scattered field in the probe region (recall (10.7)). The results, which were
computed at 100 Hz, are shown in Fig. 10.3. Note that B(x) is antisymmetric in x and
B(y) is antisymmetric in y.
We validate the computation of the z-component by applying Faraday’s law to
compute the impedance of the probe coil, and get a number very close to that
calculated using the incident electric field and anomalous currents within the flaw.
Figure 10.4 shows the z-component of the B-field. The four dimples occur at the
corners of the flaw.
The details of the test problem are as follows: the host is a half space with σ = 106
S/m, μ = μ0 , ε = ε0 , and the flaw is a surface-breaking 1-mm cube. The probe is
a conventional 100-turn pancake coil, with inner radius 1 mm, outer radius 1.1 mm,
and height of 1 mm. The probe is centered over the flaw, with a lift-off of zero.
The flaw grid is 16 × 16 × 8, and the field grid, whose bottom is at z = 0, spans
2 mm × 2 mm × 1 mm with a 32 × 32 × 4 grid.
10.1 Spatial Decomposition Algorithm 185

’sd18Bx.txt’ index 1 using 1:2:6

2e-007

1e-007

-1e-007

-2e-007
30
25
20
15 y index
10
5
0 5 10 0
15 20 25 30
x index

’sd18By.txt’ index 1 using 1:2:6

2e-007

1e-007

-1e-007

-2e-007

0
5
10
15
x index 20
25
30 20 25 30
5 10 15
0
y index

Fig. 10.3 Plot of imaginary part of B(x) (top) and imaginary part of B(y) (bottom)
186 10 Some Special Topics in Computational Electromagnetics

’sd18Bz.txt’ index 1 using 1:2:6

4e-007

2e-007

0
30
25
20
15 y index
10
5
0 5 10 15 0
20 25 30
x index

Fig. 10.4 Plot of imaginary part of B(z) . The four dimples occur at the corners of the flaw

The change in driving-point impedance of the probe in this location is (−1.407 ×

10−8, 8.586 × 10−12). In applying Faraday’s law to calculate the impedance (thereby
(z)
validating the computation of bKLM ), we first average B(z) over the coil axis,
excluding edge elements whose centers lie outside the coil outer radius:
1
∑
(z) (z)
Bavg = bklm.
Ninside coilinterior

Then we have
(z)
Zfaraday = 100 × j2π × 100 × Bavg × π ((1 × 10−3)2 + (1.1 × 10−3)2 )/2
= (−1.329 × 10−8, 8.272 × 10−12),

which is in good agreement with the actual value above.

(x) (y)
The results plotted and described above are for the imaginary parts of bKLM , bKLM ,
(z)
and, bKLM . The real parts, which are several orders of magnitude smaller, are shown
in Fig. 10.5.
As a final test, we compute the divergence expressions of (10.9). In Fig. 10.6,
we plot the sum of the three partial derivatives (i.e., the divergence) divided by the
square root of the sum of the squares of their absolute values. These are computed
from the expansion coefficients written to an external file and, therefore, are limited
by having only six digits. When double precision values are used, both expressions
of (10.9) give values several orders of magnitude smaller, 10−10 to 10−13. Clearly,
the zero-divergence condition holds!
10.1 Spatial Decomposition Algorithm 187

’sd18Bx.txt’ index 1 using 1:2:5

5e-011

-5e-011
30
25
20
15 y index
10
5
0 5 10 0
15 20 25 30
x index
’sd18By.txt’ index 1 using 1:2:5

5e-011

-5e-011

0
5
10
15
x index 20
25
30 20 25 30
5 10 15
0
y index

It may be thought that the expansion coefficients would satisfy a zero-divergence

expression of the type 0 = (bk+1lm − bklm )/dx + (bkl+1m − bklm )/dy + (bklm+1 −
bklm )/dz, but that clearly doesn’t hold, as Fig. 10.7 illustrates. It is necessary to
use the Q-matrices in (10.9) in order to change bases.
188 10 Some Special Topics in Computational Electromagnetics

’sd18Bz.txt’ index 1 using 1:2:5

2e-010

1.5e-010

1e-010

30
25
20
15 y index
10
5
0 5 10 15 0
20 25 30
x index

Fig. 10.5 Real part of Bx (top: preceding page), B(y) (bottom: preceding page), and B(z) (this page)

’sd18divB2.txt’ index 0 using 1:2:4

2e-008

-2e-008

25
20
15 y index
10
5
0 5 10 0
15 20 25 30
x index

Fig. 10.6 Plot of the “normalized divergence” expression on the left of (10.9). Top: real part;
bottom: imaginary part
10.1 Spatial Decomposition Algorithm 189

’sd18divB.txt’ index 1 using 1:2:4

0.1

0.05

25
20
15 y index
10
5
0 5 10 0
15 20 25 30
x index

’sd18divB.txt’ index 1 using 1:2:5

0.5

-0.5

-1
25
20
15 y index
10
5
0 5 10 0
15 20 25 30
x index

Fig. 10.7 A “pseudo expression” for the divergence in terms of the expansion coefficients, only
190 10 Some Special Topics in Computational Electromagnetics

10.2 A Theory of Error Extrapolation

In this section, we develop a theory of error estimation that allows us to extrapolate

the error in a VIC-3D R
calculation with a given grid size to a grid with zero cell
size. That is, we estimate the solution of the continuous problem by extrapolating
the solutions of a sequence of discrete problems.
Let h be a measure of the fineness of a mesh. It could, for example, be
proportional to the volume of the largest cell of the mesh. For a regular mesh, such
as those that VIC-3D R
uses, h measures the volume of each cell. The accuracy of
the solution of a discrete problem depends upon h, and we would like to know what
the solution of the continuous problem that is approximated by a discrete problem is.
This can be gotten by extrapolating to the limit of h → 0 the solution of a sequence
of discrete problems, with h varying.
Let f (0) be the “exact” solution to the continuum problem, which is to be
achieved by passing to the limit h → 0 in a sequence of discrete problems. Further,
let A(h) be the solution to the discrete problem with mesh fineness, h, and e(h) be
the error with this value of h. Then we have

f (0) = A(h) + e(h)

= A(4h) + e(4h)
= A(16h) + e(16h)
= A(64h) + e(64h). (10.10)

We expand e(h) as a power series, stopping with the cubic term:

e(h) = C1 h + C2h2 + C3 h3 , (10.11)

where the Ci are independent of h.

In order to determine these constants, we form three equations out of the four in
(10.10) by taking differences;

Δh A = e(4h)−e(h) = C1 4h+C2 16h2+C3 64h3−C1 h−C2 h2 −C3 h3

= 3C1 h+15C2 h2 +63C3 h3
Δ4h A = e(16h)−e(4h) = C1 16h+C2 256h2+C3 4096h3−C1 4h−C2 16h2 −C3 64h3
= 12C1 h+240C2 h2 +4032C3h3
Δ16h A = e(64h)−e(16h) = C1 64h+C2 4096h2+C3 262144h3
− C1 16h−C2 256h2−C3 4096h3
= 48C1 h+3840C2 h2 +258048C3h3, (10.12)

where Δh A = A(h)− A(4h), Δ4h A = A(4h)− A(16h), and Δ16h A = A(16h)− A(64h).
10.2 A Theory of Error Extrapolation 191

The solution of (10.12) is easily obtained:

1024ΔhA − 80Δ4hA + Δ16hA

C1 h =
2160
−256 Δ h A + 68Δ4hA − Δ16hA
C2 h2 =
8640
64 Δ h A − 20 Δ4hA + Δ16hA
C3 h3 = . (10.13)
181, 440

Upon substituting this result into (10.11), we get an expression for the error, e(h),
which, when substituted into (10.10) gives the final result
1261 83 1
f (0) = A(h) + ΔhA − Δ4h A + Δ16hA
2835 2835 2835

4096A(h) − 1344A(4h) + 84A(16h) − A(64h)

= . (10.14)
2835
We apply this theory to model-data that correspond to an experiment in which
a bobbin coil is scanned past a through-wall notch drilled into an aluminum tube.1
The notch is 1.420 inches long and 0.017 inches wide. The wall thickness is 0.220
inch. The conductivity of the tube is 2.577 × 107 S/m, and the frequency of operation
is 500 Hz. The grids have cell counts of 2 × 16 × 4, 2 × 32 × 8, 2 × 64 × 16, and 2 ×
128 × 32, where the width of the flaw contains two cells in all the grids. The results
are shown in Fig. 10.8.
Next, we do a linear interpolation
* on the two+coarsest grids, using the two-point
extrapolation formula, f (0) = 4A(h) − A(4h) /3, which can be easily deduced
using the ideas set out above, and get the results shown in Fig. 10.9. In this example
it is clear that a simple linear extrapolation based on two relatively coarse grids will
suffice to give a very good response. This could represent a significant cost savings
in computer time.
The large discrepancy in resistance in both figures is due to the manner in which
the data were taken. Four slots were cut into the aluminum cylinder, equally spaced
in azimuth by 90◦ , and data were recorded for this composite anomaly. This was
done to get a sufficiently large signal for the sensitivity of the instrument. We then
divided this response by four and compared that with VIC-3D R
’s response to a
single slot. Because of the size of the slots, compared to the inner radius of the
aluminum tube, 1.53 inch, there is considerable interaction between the slots, which
means that simple division of the measurements by four will produce an error. This
error is manifested mostly in the resistance values.
Finally, we apply extrapolation theory to the proximity sensor problem that is
shown in Fig. 10.10. The sensor problem is characterized by, among other things,

1 Caius V. Dodd, Oak Ridge National Laboratory, private communication.

192 10 Some Special Topics in Computational Electromagnetics

Fig. 10.8 Comparing the Extrapolation Results

convergence of model 10
2x128x32
impedances with mesh 9 2x64x16
refinement to measured data. 2x32x8
8 2x16x4
The extrapolated results, Extrapolated

Resistance (Ohms)
based on equation (10.14), 7 Data
are also shown. Top: 6
resistance; bottom: reactance 5
4
3
2
1
0
0 0.2 0.4 0.6 0.8 1
Probe position (inch)
Extrapolation Results
25
2x128x32
2x64x16
2x32x8
20 2x16x4
Extrapolated
Reactance (Ohms)

Data
15

10

-5
0 0.2 0.4 0.6 0.8 1
Probe position (inch)

the need to apply the magnetic-dipole model of nonstandard probes to the racetrack
transmit and receive coils, as well as model the nickel mesa anomalous region.
This can result in a problem with a large number of unknowns. To help alleviate
this problem, we use extrapolation techniques to generate increasingly accurate
solutions, while simultaneously conserving computer resources and time. We
present the results of some numerical experiments that demonstrate the application
of extrapolation techniques to this particular sensor problem.
Table 10.1 lists the convergence results for the racetrack T/R coils in freespace.
Note that the results converge nicely with Nz , for a fixed Nx × Ny . In each of the first
three cases, the change in going from Nz = 8 to Nz = 16 is 0.51 %, so this suggests
that Nz = 16 is sufficient to model this sensor.
As for the extrapolation analysis, if we use the three-point formula,

768 × A(h) − 240 × A(4h) + 12 × A(16h)

f (0) = , (10.15)
540
10.2 A Theory of Error Extrapolation 193

Fig. 10.9 Comparing the Linear Extrapolation Results

convergence of model 9
impedances with mesh Extrapolated
8 Data
refinement to measured data.
The linearly extrapolated 7

Resistance (Ohms)
results, based on the two 6
coarsest grids,
5
2 × 16 × 4 and 2 × 32 × 8, are
also shown. Top: resistance; 4
bottom: reactance 3
2
1
0
0 0.2 0.4 0.6 0.8 1
Probe position (inch)
Linear Extrapolation Results
25
Extrapolated
Data
20
Reactance (Ohms)

15

10

-5
0 0.2 0.4 0.6 0.8 1
Probe position (inch)

on the first three bottom entries (with Nz = 16), we get an “ideal” solution of
0.107308, which is higher by less than 0.2 % than the computed solution for
32×256×16. If we apply the four-point formula, (10.14), to the bottom four entries,
we get 0.108 for the new “ideal” solution. This result differs from 0.10711 by 0.8 %.
In order to determine the convergence profile for modeling the anomalous region
within the host material, we ran a number of different models, whose results are
shown in Table 10.2. The first four entries demonstrate convergence under the
fixed conditions of frequency and host conductivity, while varying the cell-density
count, Ny × Nz , keeping Nx , the density across the width of the mesa, fixed. If
we use the three-point extrapolation over the first three entries, we get an ideal
result of f (0) = (−8.0617 × 10−7 , 3.3687 × 10−6), which should be compared
to (−8.197 × 10−7, 3.3913 × 10−6 ), the value corresponding to the finest grid
with 8 × 64 × 32 cells. The error in the real part is 1.68 %, and in the imaginary
0.67 %. If we use the four-point formula over the first four entries, we get f (0) =
(−8.3576 × 10−7, 3.4295 × 10−6).
194 10 Some Special Topics in Computational Electromagnetics

2in
Receive Coil 0.062 in
Circuit Board
Transmit Coil
Air 0.25 in

Poor Conductor : σ=10 E4 S/m 0.2 in

1.5 in
Air

Nickel Mesa
1 in
Host σ=1.15E7 S/m

2 in
Air

1.625 in

0.1875 in

Racetrack Transmit and Receive Coils

Fig. 10.10 A proximity sensor that utilizes racetrack transmit and receive coils

Table 10.1 Convergence results for racetrack coils in freespace

4 × 32 × 2 0.10376
4 × 32 × 4 0.077742 8 × 64 × 4 0.10123
4 × 32 × 8 0.074941 8 × 64 × 8 0.097598 16 × 128 × 8 0.10517
4 × 32 × 16 0.074559 8 × 64 × 16 0.0971 16 × 128 × 16 0.10463 32 × 256 × 16 0.10711
Cells Reactance Cells Reactance Cells Reactance Cells Reactance

Table 10.2 Convergence profile (iterations/% error in residuals) for iterative solver
Cells No. unknowns Frequency σhost Profile Solution
8×8×4 768 1 MHz 10 1389/1.00 % (−6.0793E−7,2.8248E−6)
8 × 16 × 8 3072 1 MHz 10 1643/1.0 0 % (−7.0964E−7,3.0987E−6)
8 × 32 × 16 12288 1 MHz 10 2016/1.00 % (−7.791E−7,3.2928E−6)
8 × 64 × 32 49152 1 MHz 10 3756/1.00 % (−8.197E−7,3.3913E−6)
8 × 64 × 32 49152 100 kHz 10 500/8.62 % (−1.6599E−6,−1.9207E−6)
8 × 64 × 32 49152 100 kHz 100 500/5.96 % (−1.7655E−6,−1.9889E−6)
8 × 64 × 32 49152 1 MHz 10 500/6.41 % (−7.1739E−7,3.0734E−6)
8 × 64 × 32 49152 1 MHz 100 500/4.36 % (−3.7532E−7,3.1366E−6)
10.2 A Theory of Error Extrapolation 195

Fig. 10.11 A ferrite core

probe
Y

TOP VIEW X

μ, σ

μ, σ

FRONT VIEW

Table 10.3 Results, using a cubic error term, for grids of increasing
refinement in a nonconducting ferrite core
Grid(Nx × Ny × Nz ) Z Grid(Nx × Ny × Nz ) Z
4×4×8 −j1.5078 4 × 4 × 16 −j1.7702
8×8×8 −j1.6049 8 × 8 × 16 −j1.9335
16 × 16 × 8 −j0.92719 16 × 16 × 16 −j0.98303
32 × 32 × 8 −j1.1119 32 × 32 × 16 −j1.0733
Extrapolated Value −j1.2139 Extrapolated Value −j1.1413

The last four entries in the table indicate convergence profiles as a function of
either frequency or host conductivity. Clearly, for a given frequency, convergence
is better at the higher host conductivity, whereas for a given host conductivity,
convergence is better at the higher frequency.

10.2.1 Application to a Ferrite Core Probe

A ferrite core probe is shown in Fig. 10.11. Consider the ferrite core problem with
σ = 0, μ = 4300, ε = 1, and grids of increasing refinement, with a cubic error term
as shown in Table 10.3. Since the core is nonconducting, the impedances are purely
imaginary. The extrapolated values differ by 6.4 %, which is reasonable for this
computation. The value for the 32×32×8 grid differs by 9.2 % from its extrapolated
value, whereas that for 32 × 32 × 16 differs by 6.3 %, which makes sense since the
finer mesh should produce a better estimate.
196 10 Some Special Topics in Computational Electromagnetics

Fig. 10.12 Showing the Reactance as a Function of Grid Refinement

reactance as a function of grid -0.8
fineness. The value at 0 is the
extrapolated value, using the
cubic error analysis -1
8 cells in Z
16 cells in Z
-1.2

X
-1.4

-1.6

-1.8

-2
0 10 20 30 40 50 60 70
Relative Fineness of Grid

It is clear that the complex shape of the probe produces an oscillating result for
the impedance as the grid is refined (see Fig. 10.12), contrary to what we saw in
Fig. 10.8 for a regular slot, for which the convergence was monotonic. In the case of
the slot, the finer mesh produces a more accurate estimate of the anomalous currents,
while doing nothing to improve the accuracy of the representation of the geometry of
the anomaly. In the case of the probe, however, we need a refined grid to accurately
represent both the anomaly and the anomalous currents.
It may be argued that the data for the four coarsest grids in Table 10.3 unduly
skews the extrapolation, or, to put it another way, these data may be outside the range
of a cubic error analysis. This argument may justify a simple linear extrapolation
over the results for grids 32 × 32 × 16 and 16 × 16 × 8:
A(h) − A(8h)
f (0) = A(h) +
7
8A(h) − A(8h)
=
7
8(1.0733) − 0.92719
= −j
7
= − j1.0942. (10.16)

The result for the finer grid differs by less than 2 % from the linearly extrapolated
value. Clearly, there is a numerical advantage in using finer grids, but the disadvan-
tage is in computational time. In any case, this analysis gives us an estimate of the
discretization error when using any of these grids.
These results hold for computations in freespace and suggest that using a
relatively coarse grid in X,Y may not be serious from a numerical viewpoint. When
the probe is placed over a host that includes a flaw, it is important to have enough
10.2 A Theory of Error Extrapolation 197

Fig. 10.13 Schematic Coil

configuration for the tilted EDM notch
notch benchmark problem
with a 30◦ EDM notch 4.33mm
30 degrees
Al: σ=1.92 E7

Offset Coil

Y
14mm

12.6mm X

EDM Notch 0.28mm

30 degrees

Al: σ=1.92 E7

cells in the transverse plane to produce an accurate incident field on the flaw, since
it is this field that will not only couple the probe to the host, but will produce the
flaw “signature” that will be used in an inversion process.

10.2.2 Application to Tilted Notch Benchmark Problems

Figure 10.13 illustrates the configuration for the tilted notch benchmark problem
with a 30◦ EDM notch. Data are taken at 7 kHz, which, with a conductivity of
1.92 × 107 S/m for the host, yields a skin depth of 1.38 mm. This implies that the
problem is quite large electrically, given the dimensions and extent of the EDM
notch anomaly in Fig. 10.13. Thus, we will apply extrapolation theory with cell
sizes of 32 × 32 × 32, 32 × 32 × 64, 64 × 32 × 64, and 64 × 32 × 128. The model
results, using the previous notation, are, therefore, A(h), A(2h), A(4h), and A(8h),
and the extrapolation formula, using a cubic error term, becomes
64A(h) − 56A(2h) + 14A(4h) − 1A(8h)
f (0) = . (10.17)
21
We tabulate in Table 10.4 the results for |Z| at 7 kHz for scans over the 30◦ notch
with these four grids.
These results are plotted in Figs. 10.14 and 10.15. The convergence is monotonic
from below, in contrast to the previous results shown in Fig. 10.12, confirming,
once again, that the geometry is well defined with a rather coarse grid and that the
refinement of the grid goes into creating a better approximation of the anomalous
currents. The asymmetry in the extrapolated model response is clearly evident, with
198 10 Some Special Topics in Computational Electromagnetics

Table 10.4 Extrapolated results for |Z| at 7 kHz for the 30◦ notch using four grids
Grid Size
Scan Pt 32 × 32 × 32 32 × 32 × 64 64 × 32 × 64 64 × 32 × 128 Extrapolated Data
−28 0.000315 0.000408 0.00055 0.00069 0.000893 0.013287
−21 0.002506 0.003369 0.004699 0.006004 0.007894 0.016114
−14 0.021642 0.030513 0.044342 0.057872 0.077437 0.069345
−7 0.070115 0.105 0.16076 0.21561 0.29506 0.312066
0 0.092742 0.14237 0.22237 0.30166 0.41685 0.426732
7 0.069253 0.11037 0.17197 0.23366 0.32380 0.316074
14 0.023636 0.035744 0.053971 0.0719 0.09790 0.080692
21 0.002819 0.004053 0.005868 0.007689 0.01035 0.008785
28 0.000344 0.000472 0.000658 0.000845 0.001119 0.008045

Fig. 10.14 Convergence of Convergence of Model Data With Mesh Fineness

the model results of 0.45
Table 10.4 for the 30◦ slot
0.4
with mesh fineness at 7 kHz Data
Extrapo
0.35 64x32x128
64x32x64
0.3 32x32x64
32x32x32
0.25
Z

0.2

0.15

0.1

0.05

0
-30 -20 -10 0 10 20 30
Scan Point (mm)

the larger signal occurring for positive values of x (the slot is located at x = 0),
where the slot is tilted towards the surface. The symmetry is less noticeable in the
actual data, probably because of a slightly non-flat surface. At lower frequencies,
this effect probably will be less of a problem.
Figure 10.16 illustrates the measured and extrapolated-model responses in the
impedance plane of the measured data for the 30◦ notch. The enclosed areas are
roughly equal, which is a measure of the asymmetry associated with the 30◦ tilt.

10.2.2.1 Notch With a 10◦ Tilt

We continue with the problem defined in Fig. 10.13, except that the 5-mm long slot
is tilted 10◦ from the vertical. We use two grids, 16 × 32 × 32 and 32 × 32 × 64, and
extrapolate with a linear error term over these two grids. The formula is
10.2 A Theory of Error Extrapolation 199

Fig. 10.15 Convergence of Convergence of Model Data With Mesh Fineness

the model results for the 30◦ 0.45
slot with mesh fineness at
0.4
7 kHz. The data of Table 10.4 Data
have been smoothed with Extrapo
0.35 64x32x128
cubic splines 64x32x64
0.3 32x32x64
32x32x32
0.25

Z
0.2

0.15

0.1

0.05

0
-30 -20 -10 0 10 20 30
Scan Point (mm)

Fig. 10.16 The measured Measured and model data for 30 degree notch
and extrapolated-model 0.45
responses in the impedance
0.4
plane for the 30◦ notch
0.35

0.3
Measured
Model
0.25
X

0.2

0.15

0.1

0.05

-0.05
-0.02 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
R

4A(h) − A(4h)
f (0) = , (10.18)
3
and the corresponding data are shown in Table 10.5. It is clear, when comparing
the extrapolated data of Table 10.4 with the corresponding data of Table 10.5, that
the asymmetry associated with the 30◦ slot is considerably greater than that with the
10◦ slot, as we would expect intuitively.
200 10 Some Special Topics in Computational Electromagnetics

Table 10.5 Extrapolated Grid Size

results for |Z| at 7 kHz for the
10◦ notch using two grids Scan Pt 16 × 32 × 32 32 × 32 × 64 Extrapolated Value
−28 0.00064 0.00089 0.000973
−21 0.00566 0.008013 0.0088
−14 0.054168 0.078135 0.086124
−7 0.19262 0.28445 0.31506
0 0.26102 0.39007 0.433
7 0.19925 0.29269 0.32384
14 0.058121 0.084437 0.09321
21 0.006087 0.008690 0.00956
28 0.000680 0.000951 0.00104

Table 10.6 Extrapolated Grid Size

results for |Z| at 7 kHz for the
20◦ notch using two grids Scan Pt 64 × 32 × 32 64 × 32 × 128 Extrapolated Value
−28 0.000531 0.000845 0.00095
−21 0.004583 0.007518 0.0085
−14 0.043252 0.073263 0.08242
−7 0.15363 0.27134 0.31058
0 0.20956 0.3763 0.432
7 0.1608 0.28571 0.32735
14 0.049172 0.084589 0.09639
21 0.005265 0.008869 0.01007
28 0.000594 0.000968 0.001093

Table 10.7 Extrapolated results for |Z| at 7 kHz for the 20◦ notch using four grids and algorithm
(7.5)
Grid Size
Scan Pt 8 × 32 × 16 16 × 32 × 32 32 × 32 × 64 64 × 32 × 128 Extrapolated Value
−28 0.000157 0.000263 0.000614 0.000845 0.000938
−21 0.001046 0.002053 0.005365 0.007518 0.008379
−14 0.006663 0.017137 0.051309 0.073263 0.082031
−7 0.013207 0.052769 0.18542 0.27134 0.305687
0 0.013525 0.068195 0.25435 0.3763 0.425
7 0.013263 0.054705 0.19471 0.28571 0.322102
14 0.007017 0.018769 0.058905 0.084589 0.094842
21 0.001132 0.002282 0.006216 0.008869 0.009934
28 0.000167 0.000286 0.000692 0.000968 0.001079

10.2.2.2 Notch With a 20◦ Tilt

We’ll use the same linear extrapolation formula, (10.18), as for the 10◦ notch, but
with grids of 64 × 32 × 32 and 64 × 32 × 128 cells. Table 10.6 gives the results.
Table 10.4 showed the results for a cubic interpolation using four rather fine grids.
If we use the cubic interpolation algorithm of (7.5), we can get by with coarser grids,
as shown in Table 10.7 for the 20◦ notch.
10.2 A Theory of Error Extrapolation 201

Fig. 10.17 Comparing |Z| Response to tilt of notch

for notches of 0◦ , 10◦ , 20◦ , 0.45
and 30◦ tilts
0.4
0
10
0.35 20
30
0.3

0.25

Z
0.2

0.15

0.1

0.05

0
-30 -20 -10 0 10 20 30
Scan Point (mm)

Table 10.8 Asymmetry ratio

for notch tilt Tilt (Asymmetry ratio − 1) (%)
0◦ 0
10◦ 8.2
20◦ 17
30◦ 26.5

Fig. 10.18 Asymmetry ratio Asymmetry Ratio vs Tilt Angle

vs. tilt angle 30

25
(Asymmetry Ratio -1) (%)

20

15

10

0
0 5 10 15 20 25 30
Tilt Angle (degrees)

Figure 10.17 compares the |Z| response to EDM notches at 0◦ , 10◦ , 20◦ , and 30◦
tilts. It is difficult to eyeball these results and draw conclusions as to the slant of each
notch, because each notch has the same volume. If, however, we refer to the raw data
in Tables 10.4–10.6, and consider the “asymmetry ratio” |Z(14)|/|Z(−14)|, then the
results tabulated in Table 10.8 give a clear estimate of the tilt. Indeed, it is a virtually
linear relation, with a slightly increasing slope with angle (see Fig. 10.18).
202 10 Some Special Topics in Computational Electromagnetics

Impedance-plane plot for 0, 10, 20, and 30 degree notches

0.4

0.35

0.3
0
10
0.25 20
30
X

0.2

0.15

0.1

0.05

0
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
R

Fig. 10.19 Illustrating the increase in area with tilt angle enclosed by the impedance-plane plots
of each of the notches

The impedance-plane plots of the responses for the 0◦ , 10◦ , 20◦ , and 30◦ notches
are shown in Fig. 10.19. As we would expect, the area enclosed by each curve
increases with the tilt angle, indicating a greater asymmetry. Note, further, that the
phase angle of the peak response decreases with tilt angle.

10.3 Determining Grid Resolution Requirements

In this section, we will address the question of how to estimate the required grid
resolution to accurately solve a model problem, or, to put it another way, how
do we determine a priori how many cells will be required to solve the problem
accurately. We can always do the recommended thing, and that is to test several grid
resolutions, chosen in a systematic manner, and then check the rate of convergence
of the solution. This process can be followed by an extrapolation, as we have shown
above. Nevertheless, we want to gain insight into the requirements without resorting
at the outset to this empirical approach.

10.3.1 The Volume-Integral Equation

Our theoretical consideration starts with an analysis of the volume-integral equation

that is the basis of VIC-3DR
:
10.3 Determining Grid Resolution Requirements 203


Ja (r)
E0 (r) = − G(r, r ) · Ja (r ), (10.19)
σa (r)

which can be rearranged to read

  
 
Ja (r) = σa (r) E0 (r) + G(r, r ) · Ja(r ) . (10.20)

E0 (r) is the incident electric field, which is known, as is the anomalous conductivity,
σa (r) (for a forward problem). G(r, r ) is the dyadic Green’s function, and Ja (r) is
the unknown anomalous current.
The gridding question can be answered by analyzing the spatial-frequency
content of Ja (r). This requires taking the spatial Fourier transform of (10.20), which
involves the product of two spatially-varying functions, σa (r), and the total electric
field within the brackets. Once we get an estimate of the highest spatial frequencies
that must be present in Ja (r), we can estimate the size of the cells that are required
to produce these frequencies.

10.3.2 Review of Fourier Analysis

We are going to apply Fourier analysis to study the question, using the notion of
spatial frequency in the same manner that temporal frequency is used to answer
resolution questions in system theory. Equation (10.20) shows that we need to
determine the spatial Fourier transform of the product of two functions. This is given
by the convolution in the spatial-frequency domain of the individual transforms, as
we’ll now prove.
    )
1 jk ·r
e− jk·r f (r)g(r)dr = e− jk·r f (r) e G̃(k
)dk
dr
(2π )3

  )
1  − j(k−k )·r
= G̃(k ) e f (r)dr dk
(2π )3

1
= G̃(k )F̃(k − k )dk. (10.21)
(2π )3

Consider, for example, that g(r) is uniform in space, with a value of g. Then G̃(k) =
(2π )3 gδ (k), so that the right-hand side of (10.21) is equal to gF̃(k), as expected.
To simplify our lives a bit, we’ll ignore the integral term in (10.20), and just work
with the approximation, ja (r) ≈ E0 (r)σa (r), in determining the frequency content
of ja (r):

1 ,0 (k )Σ̃a (k − k )dk.
J̃a (k) ≈ E (10.22)
(2π )3
204 10 Some Special Topics in Computational Electromagnetics

Fig. 10.20 Illustrating the Z

bounding box for the 20◦
notch problem of
Sect. 10.2.2.2

σ=1.92E7S/m

0.28mm
4.7463mm

σ=0

X
1.9906mm

We will call Σ̃a (k) the spectrum of the conductivity scene, Ẽ0 (k) the spectrum
of the electromagnetic scene, and the convolution in (10.22) the composite spectral
scene. As a simple example of what to expect, let the conductivity and electromag-
netic spectral scenes in (10.22) each be a “baseband” unit pulse extending from 0
to Kmax , then the composite spectrum will be a unit tent function extending from
0 to 2Kmax . Clearly, the spatial-frequency content has been doubled for J̃a , when
compared to the original spectra.

10.3.3 Application to the 20◦ Notch Problem

The VIC-3D R
-generated bounding box for the 20◦ notch problem of Sect. 10.2.2.2
is shown in Fig. 10.20. We’ll start by computing the spectrum of the conductivity
scene, Σ̃a (kx , kz ). Since our “computational universe” is bounded by the rectangle
X0 = 1.9906, Z0 = 4.7463, we will use a Fourier series representation with discrete
spatial frequencies for all spectra. For the conductivity spectrum we write
 X0  Z0
1
Σ̃a (kx , kz ) = σa (x, z)e− j2π (kx x+kz z) dxdz (10.23)
X0 Z0 0 0
10.3 Determining Grid Resolution Requirements 205

where kx = nx /X0 , kz = nz /Z0 , nx , nz = 0, ±1, ±2, . . .. The inverse relation is

∞ ∞
σa (x, z) = ∑ ∑ Σ̃a (kx , kz )e j2π (kx x+kz z), (10.24)
kx =−∞ kz =−∞

as is well known from the theory of Fourier series.

The equation of the line passing through (1.9906, 0) and (0, 4.7463) is
x/1.9906 + z/4.7463 − 1 = 0, or z = 4.7463 − 2.384x. The unit normal to this line
is an = 0.922ax + 0.387az, so the equation of the line parallel to this line, and
whose distance from the origin is d, is given by the inner-product of an with the
arbitrary vector, xax + zaz : 0.922x + 0.387z = d. Now, the boundaries of the notch
are separated by 0.28 mm, which means that the top boundary of the notch is given
by 0.922x + 0.387z = d + 0.14, and the bottom boundary by 0.922x + 0.387z =
d − 0.14. Thus, the length of the notch in the z direction, at a fixed x coordinate, is
z1 − z2 = 0.28/0.387 = 0.724 mm.
We’ll evaluate (10.23) as an iterated integral, starting with z:
 Z0  z(x)+0.362
1 −1.92 × 107
σa (x, z)e− j2π kz z dz = e− j2π kz z dz
Z0 0 Z0 z(x)−0.362

−1.92 × 107 − j2π kz z(x) sin(2π kz × 0.362)

= e
Z0 π kz
sin(0.153π nz)
= −1.92 × 107e− j2π nz (1−0.502x)
π nz
sin(0.153π nz) j2π nz (0.502x)
= −2.94 × 106e− j2π nz e .
0.153π nz
(10.25)

The number 0.153 is the ratio of the area of the notch in the (X, Z) plane to the total
area of the bounding box.
We complete the calculation of the iterated integral by substituting (10.25) into
(10.23):
 X0
sin(0.153π nz) 1
Σ̃a (kx , kz ) = −2.94 × 106e− j2π nz e j2π nz (0.502x) e− j2π kx x dx
0.153π nz X0 0

sin(0.153π nz) 1 X0 − j2π x(kx −0.502nz )
= −2.94 × 106e− j2π nz e dx
0.153π nz X0 0
 
− j2π (kx −0.502nz )X0 − 1
6 − j2π nz sin(0.153π nz) e
= −2.94 × 10 e
0.153π nz − j2π (nx − nz )
 
sin(0.153 π n ) e − j2π (nx −nz ) − 1
= −2.94 × 106e− j2π nz
z
0.153π nz − j2π (nx − nz )
206 10 Some Special Topics in Computational Electromagnetics

Table 10.9 Comparing nx sin(0.153π nx )/0.153π nx sin(0.612π nx )/0.612π nx

spatial frequency responses
0 1 1
1 0.962 0.488
2 0.853 −0.168
10 −0.207 0.019
100 −0.017 −0.00306
200 0.0096 0.00247
1,000 −0.00069 0

= 0, if nx = nz
sin(0.153π nx)
= −2.94 × 106e− j2π nx , if nx = nz. (10.26)
0.153π nx

The fact that nx = nz means, from (10.24), that the sinusoidal expansion functions
for σa (x, z) have precisely the same spatial frequencies in the x and z directions.
This means that in order to reconstruct the conductivity, we must sample equally
finely in the x and z directions. (This holds only for characterizing the conductivity
scene.) To get an idea of what the sampling requirements are, consider Table 10.9.
At a spatial frequency of 100, the magnitude of the spectrum in (10.26) is reduced
to slightly less than 2 %. Hence, if we are to reconstruct σa (x, z) to this precision we
will need 200 samples in each direction to avoid aliasing. For a flaw with four times
the area, however, we can get by with one-tenth the number of samples for the same
precision.

10.3.4 The Electromagnetic Scene

Now, we’ll turn our attention to the electromagnetic scene for the 20◦ notch problem.
Let the coil lie in region 1, above the workpiece, and let region 2 be the workpiece,
which we assume is a homogeneous half-space. The appropriate axisymmetric
Green’s functions are [27]:
 ∞
  

  1 Re−α0 (z+z ) +e−α0 |z−z |
G11 (r, z; r , z ) = J1 (r l)J1 (rl)ldl, (r , z ) ∈ 1, (r, z) ∈ 1
2π 0 2α0
 ∞
1 T 
G21 (r, z; r , z ) = e(α1 z−α0 z ) J1 (r l)J1 (rl)ldl, (r , z ) ∈ 1, (r, z) ∈ 2, (10.27)
2π 0 2α0
where
α0 − α1 2α0
R= , T= (10.28)
α0 + α1 α0 + α1
10.3 Determining Grid Resolution Requirements 207

are the reflection and transmission coefficients, respectively, α0 = (l 2 − ω 2 μ0 ε0 )1/2 ,

α1 = (l 2 − ω 2 μ0 ε0 + jω μ0 σ )1/2 , l 2 = kx2 + ky2 is the spatial wave-number in Fourier
space, σ is the conductivity of the host medium, r2 = x2 + y2 , and J1 (rl) is the
first-order Bessel function.
The φ -directed electric field within region 2, due to the axisymmetric coil, is
given by
 
E2 (r, z) = − jω 2π μ0 G21 (r, z; r , z )Jc (r , z )r dr dz
coil
 ∞  L+H  R2 )
T 
= − jω μ0 Nc eα1 z Jr (rl) e−α0 z dz J1 (r l)r dr ldl
0 2α0 L R1
 ∞ 
T e−α0 L (1 − e−α0H ) I (R1 l, R2 l) α1 z
= − jω μ0 Nc e J1 (rl)ldl,
0 2α0 2 α0 l
(10.29)
-b
where I (a, b) = a zJ1 (z)dz. We assume that a current of one ampere is passing
through coil, whose turns density is Nc . L is the lift-off parameter of the coil, H the
height, and R1 and R2 the inner- and outer-radii of the coil, respectively.
The expression within the square brackets of (10.29) is the Bessel transform,
Ẽ2 (l, z), of the electric field. It is clearly an explicit function of the depth, z(≤ 0).
It is the relationship between the electric field and z that we need to sample in order
to get an estimate of the number of cells in z that the EM scene will contribute. To do
this, we start by computing the finite Fourier transform in z of Ẽ2 (l, z):
 Z0
1
Ẽ2 (l, kz ) = Ẽ2 (l, z)e− j2π kz z dz
Z0 0
 Z0
− jω μ0 Nc T e−α0 L (1 − e−α0H ) I (R1 l, R2 l)
= e−α1 z e− j2π kz z dz
Z0 2α0 α0 l2 0

− jω μ0 Nc T e−α0 L (1 − e−α0H ) I (R1 l, R2 l) 1 − e−(α1+ j2π kz )Z0

= .
Z0 2α0 α0 l2 α1 + j2π kz
(10.30)

Consider

α1 + j2π kz = (l 2 − k12 )1/2 + j2π kz

2π nz
= (l 2 − ω 2 μ0 ε0 + jω μ0 σ )1/2 + j
Z0
2π nz
≈ (l 2 + jω μ0 σ )1/2 + j , nz = 0, ±1, ±2, . . .. (10.31)
Z0
208 10 Some Special Topics in Computational Electromagnetics

Fig. 10.21 Showing the l

bounding box for Ẽ2 (l, kz ) in
spatial-frequency space.
Ẽ2 (l, kz ) ≈ 0 for spatial lmax
frequencies greater than
(lmax , Nzmax )

Nz
Nz max

Then

− jω μ0 Nc T e−α0 L (1−e−αo H ) I (R1 l, R2 l) 1−e−(l + jω μ0 σ ) Z0 + j2π nz

2 1/2

Ẽ2 (l, kz ) ≈ .
Z0 2α0 α0 l2 2π nz
(l 2 + jω μ0 σ )1/2 + j
Z0
(10.32)

In computing I (R1 l, R2 l) we can make use of

 b
I (a, b) = zJ1 (z)dz
a
 b
dJ0
=− z dz
a dz
 b  b
d
= J0 (z)dz − (zJ0 (z))dz
a a dz
 b
= aJ0 (a) − bJ0(b) + J0 (z)dz, (10.33)
a

where the final integral may be tabulated or easily computed.

The “bounding box” for Ẽ2 (l, kz ) in spatial-frequency, (l, kz ), space is the
smallest rectangle that “supports” Ẽ2 (l, kz ), or, to put it another way, is the smallest
rectangle beyond which Ẽ2 (l, kz ) ≈ 0. Figure 10.21 shows such a bounding box
with boundaries (lmax , Nzmax ). We can get an estimate for Nzmax , by setting l 2 = 0
2π nz
in the denominator of the final term of (10.32), and then arguing that must
Z0
be 50 (say) times as large as ( jω μ0 σ )1/2 to ensure resolution in z. This yields
2π Nzmax
= 50 × (ω μ0 σ )1/2 . Using the values shown in Fig. 10.20, together with
Z0
10.4 Modeling Multiscale Problems 209

2π
a frequency of 7 kHz, we have (ω f μ0 σ )1/2 = 1030, and = 1319, which yields
Z0
Nzmax = 39. At this frequency, the bounding box is about 3.5 skin-depths deep, so
we can argue that we need about 10 cells per skin depth just to account for the
electromagnetic scene. We’ll come back to this.
We can estimate the extent of the bounding box in (kx , ky , kz )-space by simply
2π Nxmax 2π Nymax
postulating that = lmax , = lmax , where, as before, X0 and Y0 are
X0 Y0
the lengths of the physical bounding box in the x and y directions, respectively. This
X0 lmax Y0 lmax
gives the plausible results that Nxmax = , Nymax = , from which we also
2π 2π
Nxmax Nymax
conclude that the cells should satisfy a “similarity relationship”: = .
X0 Y0
Now we can put everything together, and say that, as far as gridding requirements
for Ja (r) are concerned, this can be determined by convolving the spectrum of the
conductivity scene with the spectrum of the EM scene, or J̃a (k) = Σ̃a (k) ⊗ Ẽ2 (l, kz ).
From our previous comments about the extension of the frequency response that
comes from convolutions, we can estimate the composite frequency response to be
(Nxmax
c + Nxmax
e , Nymax
c + Nymax
e , Nzmax
c + Nzmax
e ), where the superscript “c” denotes the
conductivity scene, and “e” denotes the electromagnetic scene. In other words, each
scene contributes to the number of cells required for Ja (r).
The gridding requirement for an accurate impedance calculations requires an
extra convolution, as the following argument shows. We have

δZ = Ja (r) · E0 (r)dr

≈ σa (r)E20 (r)dr, (10.34)

where we are using the same approximation as in (10.22). Thus, in order to get
an accurate estimation of the integrand in (10.34), we should sample at the rate of
(Nxmax
c + 2Nxmax
e , Nymax
c + 2Nymax
e , Nzmax
c + 2Nzmax
e ), which will determine the number
of cells in the grid.

10.4 Modeling Multiscale Problems

10.4.1 The Multiscale Algorithm

The use of integral equations and anomalous currents allows us to efficiently

remove “background effects” in either forward or inverse modeling. Consider the
anomalous region within a background host, as shown in Fig. 10.22. Let σb (r) be the
conductivity when the flaw is absent. Outside the background region, σb (r) is equal
to the host conductivity, σh . Inside the background regions it varies with position, r.
210 10 Some Special Topics in Computational Electromagnetics

Fig. 10.22 Defining an

anomalous region within a
background host. σb (r) − σh Host: σh
is the anomalous conductivity
associated with the “large”
background and σ (r) − σb (r)
is the anomalous conductivity
associated with the “small” Large (Background) Small (Flaw)
flaw Anomaly: σb(r)−σh Anomaly:
σ(r)−σb(r)

Let σ (r) be the conductivity when the flaw is present. Outside the background
region, σ (r) is equal to the host conductivity. Inside the background region, but
outside the flaw, σ (r) = σb (r). Inside the flaw, σ (r) is not equal to σb (r), but varies
with position.
First, consider the unflawed background region. The anomalous current for this
problem satisfies

Jb (r) = (σb (r) − σh )(Ein (r) + E(r)[Jb]) (10.35)

Next, consider the flawed background region and define anomalous currents

J f (r) = (σ (r) − σb (r))(Ein (r) + E(r)[J f ]) (10.36)

Jd (r) = (σb (r) − σh)(Ein (r) + E(r)[J f ] + E(r)[Jd ])
+(σ (r) − σb (r))E(r)[Jd ]. (10.37)

The anomalous current, Ja = Jd + J f , satisfies Ja = (σ (r)− σh )(Ein + E(r)[J f ]+

E(r)[Jd ]).
The change in impedance due to the flaw is

Ein · Ja − Ein · Jb = Ein · (J f + Jd − Jb)

= Ein · J f + Ein · Jint, (10.38)

where we have defined Jint = Jd − Jb.

Now, from (10.35) and (10.37), Jint satisfies the equation

Jint = (σb (r) − σh )(E(r)[J f ] + E(r)[Jint ]) + (σ (r) − σb(r))E(r)[Jd ]

= (σb (r) − σh )(E(r)[J f ] + E(r)[Jint ])
+(σ (r) − σb (r))E(r)[Jint ] + (σ (r) − σb(r))E(r)[Jb ]
= (σb (r) − σh )E(r)[J f ] + (σ (r) − σb(r))E(r)[Jb ] + (σ (r) − σh)E(r)[Jint ].
(10.39)
10.4 Modeling Multiscale Problems 211

From (10.35), (10.36), and (10.39) we obtain the uncoupled integral equations
Jb (r)
Ein (r) = − E(r) [Jb ] (10.40)
σb (r) − σh
J f (r) ! "
Ein (r) = − E(r) J f (10.41)
σ (r) − σb (r)
Jint (r)  
Eef (r) = − E(r) Jint , (10.42)
σ (r) − σh
where the effective incident field, Eef (r), is given by
σb (r) − σh ! " σ (r) − σb (r)
Eef (r) = E Jf + E [Jb ]. (10.43)
σ (r) − σh σ (r) − σh

We construct Jint (r) from two currents, Jintf (r) and Jb (r), defined on the flaw
int

and background grids, respectively. With Jint (r) = Jint

f (r) + Jb (r), we have
int

   
f (r)
Jint b (r)
Jint
Eef (r) = + − E(r) Jint − E(r) Jint
b . (10.44)
σ (r) − σh σ (r) − σh f

b (r) and J f (r) in terms of tent functions that are defined on the
We will expand Jint int

background and flaw grids, respectively.

b (r) gives
Testing (10.44) with the expansion functions for Jint

Eef,b = [Qb − Gb ] · Jint(b) + [Qb f − Gb f ] · Jint( f ) . (10.45)

ef,b
Eef,b is the vector of moments, Eklm , for the effective field, evaluated on the
background grid. Jint(b) and Jint( f ) are the vector expansion coefficients for Jintb (r)
and Jint
f (r), respectively. Testing (10.44) with the expansion functions for J int (r)
f
gives

Eef, f = [Q f − G f ] · Jint( f ) + [Q f b − G f b ] · Jint(b), (10.46)

ef, f
where Eef, f is the vector of moments, Eklm , for the effective field, evaluated on the
flaw grid.
The various matrices in (10.45) and (10.46) are defined here: (Qb , Gb ) and
(Q f , G f ) are the tridiagonal and Töplitz/Hankel matrices with which we are familiar,
defined on the background and flaw grids, respectively. Q f b is a sparse matrix
involving integrals of 1/(σ (r) − σh ) multiplied by the product of an expansion
function for Jint b (r) and an expansion function for J f (r). Q
int b f is the transpose of

Q f b . G f b (Gb f ) is the transfer matrix that multiplies the expansion coefficients for
b (r) (J f (r)) to give the moments of the fields that they produce, evaluated on the
Jint int

flaw (background) grid. Gb f is the transpose of G f b .

212 10 Some Special Topics in Computational Electromagnetics

We will use an approximation to simplify the evaluation of the moments of the

(x)ef, f
effective field. When evaluating Eklm , we replace the values of the ratios, (σb (r)−
σh )/(σ (r)− σh ) and (σ (r)− σb (r))/(σ (r)− σh ), in flaw cells klm and k + 1lm with
x, f f x, f b
the averages, Vklm and Vklm , over those two cells, defined by
 
x, f f 1 σb,klm − σh,klm σb,k+1lm − σh,k+1lm
Vklm = +
2 σklm − σh,klm σk+1lm − σh,k+1lm
 
x, f b 1 σklm − σb,klm σk+1lm − σb,k+1lm
Vklm = + . (10.47)
2 σklm − σh,klm σk+1lm − σh,k+1lm
Treating the y and z directions similarly, we get

Eef, f = V f b · G f b · J(b) + V f f · G f · J( f ), (10.48)

where the diagonal matrices, V f b and V f f , are given by

⎡ x, f b ⎤
V 0 0
V f b = ⎣ 0 V y, f b 0 ⎦
0 0 V z, f b
⎡ ⎤
V x, f f 0 0
Vff = ⎣ 0 V y, f f 0 ⎦. (10.49)
0 0 V z, f f

The contribution of J f (r) to Eef,b is evaluated in a similar way and is given by

Eef,b [J f (r)] = V b f · Gb f · J( f ), (10.50)

where
⎡ ⎤
V x,b f 0 0
Vbf = ⎣ 0 V y,b f 0 ⎦. (10.51)
0 0 V z,b f

V b f is exactly analogous to V f f , except that the ratios are averaged over the
background cells instead of the flaw cells.
In evaluating the contribution of Jb (r) to Eef,b , we approximate the tent functions
that are defined on the background grid in terms of the tent functions that are defined
on the flaw grid. Then we have

Eef,b [Jb (r)] = Ab f · Eef, f [Jb (r)]

= Ab f ·V f b · G f b · J(b). (10.52)

The sparse matrix, Ab f , holds the expansion coefficients and has the structure
10.5 Notch at a Bolt Hole Benchmark Problems 213

⎡ ⎤
Ax,b f 0 0
Ab f = ⎣ 0 Ay,b f 0 ⎦. (10.53)
0 0 Az,b f

Combining the contributions of J f (r) and Jb (r) to Eef,b , we get

Eef,b = Ab f ·V f b · G f b · J(b) + V b f · Gb f · J( f ). (10.54)

After computing Eef,b and Eef, f , we solve the two coupled equations given in
(10.45) and (10.46) using an iterative procedure. First, we rewrite these equations in
the following form:

[Qb − Gb ] · Jint(b) = Eef,b − [Qb f − Gb f ] · Jint( f ) (10.55)

[Q f − G f ] · Jint( f ) = Eef, f − [Q f b − G f b ] · Jint(b). (10.56)

We first solve (10.55), assuming that Jint( f ) = 0. Then we use the solution, Jint(b) , to
compute the right-hand side of (10.56), and solve this equation for Jint( f ) . Then we
use this solution to update the right-hand side of (10.55) and re-solve. We continue
this process until the norm of the residual vector

R = (Eef,b − [Qb f − Gb f ] · Jint( f ) − [Qb − Gb ] · Jint(b) )

+(Eef, f − [Q f b − G f b ] · Jint(b) − [Q f − G f ] · Jint( f ) ) (10.57)

is sufficiently small.
The algorithm for eliminating the background is:
1. Solve (10.40) for Jb using a coarse background grid, Gb
2. Solve (10.41) for J f using a fine grid, G f , covering only the flaw
3. Solve (10.42) for Jint using (10.55) and (10.56)
and then compute the change in the probe impedance, due to the flaw, as

ΔZflaw = Ein · J f + Ein · Jint(b) + Ein · Jint( f ) (10.58)

where the dot products are the usual expressions for impedances. An attractive
feature of this system is that the flaw and background may be gridded separately
in a manner appropriate to their size and characteristic.

10.5 Notch at a Bolt Hole Benchmark Problems

The advantage of using dual integral equations to compute the change in impedance
due to a small flaw in the presence of a larger background region is to avoid the more
straightforward, but less accurate, method of simply computing Z with and without
the flaw and subtracting the two nearly equal values to obtain the small difference.
214 10 Some Special Topics in Computational Electromagnetics

Fig. 10.23 Benchmark 5 mm

25 mm Coil C1
Problem 1: notch in a bolt
hole through an aluminum 5 mm
block. The coil for the test,
C1, is shown roughly to scale

100 mm Aluminum Alloy

Coil H1
25 mm 8 mm

Aluminum Alloy 4 mm 3 mm

Fig. 10.24 Benchmark Problem 2: notch in a bolt hole through a 4 mm thick aluminum plate.
The coil for the test, H1, is shown roughly to scale

Further, when it comes to inversion problems, we can subtract the background

signal and concentrate on solving the much smaller flaw problem. In the following
discussion, we apply a “graduated-grid” extension of the above algorithm to more
accurately model two benchmark problems. The data for these two problems were
supplied by Prof. J.R. Bowler and Yuan Ji of Iowa State University.

10.5.1 Sketch of Benchmark Problems 1 and 2

Benchmark Problem 1 is shown in Fig. 10.23, and Benchmark Problem 2 is shown

in Fig. 10.24. The two coils that are used in the benchmark problems are described
in Table 10.10. C1 will be used to collect data for Benchmark 1 and H1 for
Benchmark 2.
10.5 Notch at a Bolt Hole Benchmark Problems 215

Table 10.10 Data for Coils C1 and H1: dimensions in mm

Coil Outer radius Inner radius Height No. turns C (pF) L (H) R ( Ω)
C1 4 1.58 1.042 305 13.85 463 × 10−6 19.08
H1 7.38 2.51 4.99 4000 72.8 0.10052 688.4

Fig. 10.25 Response of coil Response of C1 in Air

C1 in air 1600

1400

1200

1000
Ohms

800
R
X
600

400

200

0
100 1000 10000 100000 1e+06
Freq

The response of C1 in air is shown in Fig. 10.25 and of H1 in air in Fig. 10.26.
Though it can’t be inferred from Fig. 10.25 because of the scale of the ordinate,
the slope of the reactance curve at low frequencies yields an equivalent inductance
of 0.463 μH for C1, and the same analysis of Fig. 10.26 yields a low-frequency
inductance of 0.10052 H for H1, as shown in Table 10.10. The low-frequency
resistance values for each coil are also given by the data plotted in Figs. 10.25
and 10.26 and are also shown in Table 10.10. Clearly, the small inductance of C1
has pushed its resonant frequency beyond the limits shown in Fig. 10.25, whereas
the large inductance of H1 produces a relatively low-frequency resonance point.

10.5.2 One-Port Circuit Models

Consider the typical equivalent circuit for a coil in freespace shown in Fig. 10.27.
Because there is only a single input port for the model, we refer to this as a
“one-port circuit model.” By applying the usual algorithm, we can determine the
216 10 Some Special Topics in Computational Electromagnetics

Fig. 10.26 Response of coil Response of H1 in Air

H1 in air 250000

200000

150000

100000
R
X
Ohms 50000

-50000

-100000

-150000
100 1000 10000 100000 1e+06
Freq

Fig. 10.27 A typical Air

equivalent circuit for a coil in
freespace
R0
Zair
Yp
L0

Yp = G + j B

shunt admittance, Yp = G p + jB p , and with this we can determine the equivalent

series impedance of the series-parallel circuit:

1
Zair =
1
Yp +
R0 + jω L0

1
=
1
G p + jB p +
R0 + jω L0

R0 + G p (R20 + ω 2 L20 )
=
(1 + R0G p − ω L0 B p )2 + (ω L0 G p + R0B p )2
10.5 Notch at a Bolt Hole Benchmark Problems 217

Fig. 10.28 Depicting the coil a Rair(ω) + jXclutter(ω)

coupled to the workpiece.
(a) Standard representation.
(b) VIC-3D R
representation L0 L1 R1
Zwkpc
in which δ Z is “air balanced”
Workpiece

b Rair(ω) + jXclutter(ω)

Zwkpc δZ

ω L0 (1 − ω L0B p ) − R20 B p
+j
(1 + R0G p − ω L0 B p )2 + (ω L0 G p + R0B p )2
= Rair (ω ) + jXair (ω )
= Rair (ω ) + j[Xair (ω ) − ω L0 ] + jω L0
= Rair (ω ) + jXclutter (ω ) + jω L0, (10.59)

where we have singled out jω L0 for purposes of introducing the coupled circuit
representation in the next step.
When the coil is coupled to the workpiece, we have the situation depicted in
Fig. 10.28. The top part of the figure is the standard coupled-circuit representation,
whereas the bottom part is the VIC-3D R
representation in which δ Z is “air
balanced.” Thus, in order to get the circuit representation to agree with VIC-
3D R
we must subtract Rair (ω ) + jXclutter (ω ) from the measured impedance, Zwkpc .
We could attach another shunt element, Yq , to the driving-point terminals of
Fig. 10.28 to account for loading effects on the probe that are not accounted for
by inductive coupling. For example, it is possible that the presence of the metallic
host may cause a capacitive coupling that is not accounted for by Yp in the freespace
model of the probe. If that is the case, we can determine Yq by repeating the same
process that was used to determine Yp , but this time with a known δ Z in Fig. 10.28.
Figure 10.29 shows Yp for coils C1 and H1. The slope of a line joining the origin
to the extreme point for B( f ) in the top of Fig. 10.29 yields an equivalent capacitance
of 13.85 pF for C1, which is a reasonable approximation for frequencies greater
than 150 kHz, or so. The slope of B( f ) in the bottom of Fig. 10.29 for 0 ≤ f ≤
320 kHz yields an equivalent capacitance of 72.8 pF. These are the values listed in
Table 10.10.
218 10 Some Special Topics in Computational Electromagnetics

Fig. 10.29 Yp for coil C1 Yp for Coil C1

(top) and H1 (bottom) 5e-05

4e-05

3e-05

2e-05
G

Siemens
B
1e-05

-1e-05

-2e-05

-3e-05
0 50 100 150 200 250 300 350 400 450 500
Freq(kHz)

Yp for Coil H1
0.00025

0.0002

0.00015
Siemens

0.0001
G
B

5e-05

-5e-05
0 50 100 150 200 250 300 350 400 450 500
Freq(kHz)
10.5 Notch at a Bolt Hole Benchmark Problems 219

Table 10.11 NLSE inversion

results for Benchmark 1 data Data set Φ σ (S/m)/sensitivity LO(mm)/sensitivity
sets M1 4.9166 1.8156(7)/2.73(−2) 0.1336/1.262(−2)
M2 4.8827 1.8151(7)/2.74(−2) 0.1364/1.255(−2)
M3 4.6704 1.8178(7)/2.63(−2) 0.1420/1.205(−2)
M4 4.8004 1.8175(7)/2.69(−2) 0.1389/1.236(−2)

Table 10.12 NLSE inversion

results for Benchmark 2 data Data set Φ σ (S/m)/sensitivity LO(mm)/sensitivity
sets M1 0.6385 1.8340(7)/4.03(−3) 0.3480/1.542(−2)
M2 0.8585 1.8420(7)/5.46(−3) 0.3560/2.075(−2)
M3 0.7850 1.8390(7)/4.97(−3) 0.3550/1.899(−2)
M4 0.3887 1.8227(7)/2.41(−3) 0.3254/0.933(−2)

10.5.3 Model-Based Inversion with NLSE

After removing the shunt parasitic admittance element, Yp (which includes

the capacitance, C), of each coil in the usual manner, and then removing the
freespace impedance, R + jω L, we are left with the “air-balanced” change in
impedance, δ Z, when the coil is placed over the workpiece. This is the input to
NLSE2 for inverting each of the data sets known as M1, M2, M3, and M4 for each
of the two benchmark tests. (The data sets correspond to four different locations
on the unflawed workpiece.) Even though each data set contained values of δ Z
at 32 frequencies, logarithmically spaced between 500 Hz and 500 kHz, we used
only five frequencies for the inversions. For Benchmark 1, we used the five highest
frequencies, 205.06, 256.24, 320.2, 400.13, and 500.0 kHz, and for Benchmark 2
the five lowest frequencies, 500, 624.8, 780.76, 975.65, and 1219.2 Hz. The reason
for this separation is due to the fact that the coil for Benchmark 1, C1, has a
very small inductance, as is seen from Table 10.10, whereas that for H1 is quite
large. Therefore, we get more reliable data at high frequencies for C1 and at low
frequencies for H1.
Following the customary procedure for model-based inversions with NLSE, we
establish a 4 × 4 interpolation grid for host conductivity and coil lift-off for each of
the benchmark tests. For Benchmark 1, the values of the parameters at the nodes
of the grid are: σ = 1 × 107 , 2 × 107 , 3 × 107 , 4 × 107 S/m; LO = 0.1, 0.2, 0.3,
0.4 mm. The conductivity nodal values are the same for Benchmark 2, whereas the
lift-off values are 0.2, 0.4, 0.6, 0.8 mm.
The inversion results are shown in Table 10.11 for Benchmark 1, and in
Table 10.12 for Benchmark 2.

2 NLSE, a nonlinear least-squares estimator, will be defined and described in Chap. 12. The material

in this section can be deferred until after reading that chapter, or can simply be accepted on faith,
“in anticipation.”
220 10 Some Special Topics in Computational Electromagnetics

Fig. 10.30 Comparing the Notch at a Bolt Hole Benchmark Problem 1

normalized model results and
data sets for M1 of 0.0006
Expt
benchmark 1 in Table 10.11 0.0005 NLSE

dR/f (Ohm/Hz)
over the entire frequency
0.0004
range
0.0003
0.0002
0.0001
0

1000 10000 100000

Frequency (Hz)

Notch at a Bolt Hole Benchmark Problem 1

0
Expt
-0.0002
-0.0004 NLSE
dX/f (Ohm/Hz)

-0.0006
-0.0008
-0.001
-0.0012
-0.0014
-0.0016
-0.0018
1000 10000 100000
Frequency (Hz)

When we use the coil data of Table 10.10, and the results of data set M1 of
Benchmark 1 in Table 10.11 to compute the forward model, and then plot the
results over the entire frequency range of 500 Hz to 500 kHz, we obtain Fig. 10.30.
(The data are normalized with respect to frequency.) That there is a breakdown in
the measured resistance data at low frequencies (dR goes negative!) was anticipated
because of the considerable uncertainty in Yp for coil C1 at low frequencies in
Fig. 10.29. The corresponding results for M1 of Benchmark 2 in Table 10.12
are shown in Fig. 10.31. The results for the other data sets in Tables 10.11 and
10.12 lie on top of the results shown in Figs. 10.30 and 10.31. The high-frequency
breakdown in the resistance and reactance in Fig. 10.31 was anticipated because of
the uncertainty in Yp for H1 at high frequencies in Fig. 10.29. Note, however, that
the normalized model reactance, dX/ f , due to NLSE in Fig. 10.31 “saturates” at
high frequencies, as expected by classical coupled-circuit theory and Förster plots.
The asymptote is a measure of the coupling of the coil to the host. The asymptotic
value of dR/ f = 0 is reached at a higher frequency in Fig. 10.31, as is also predicted
by coupled-circuit theory.
10.5 Notch at a Bolt Hole Benchmark Problems 221

Fig. 10.31 Comparing the Notch at a Bolt Hole Benchmark Prob. 2

normalized model results and 0.1
data sets for M1 of 0.08 Expt
benchmark 2 in Table 10.12
0.06 NLSE

dR/f (Ohm/Hz)
over the entire frequency
0.04
range
0.02
0
-0.02
-0.04
-0.06
1000 10000 100000
Frequency (Hz)

Notch at a Bolt Hole Benchmark Prob. 2

0.1
0 Expt
NLSE
dX/f (Ohm/Hz)

-0.1
-0.2
-0.3
-0.4
-0.5
-0.6
1000 10000 100000
Frequency (Hz)

The final results of our validation studies for the multiscale algorithm are plotted
in Fig. 10.32 for Benchmark Problem 2 and Fig. 10.33 for Benchmark Problem 1.
Numerical experiments show that accurate results for Benchmark Problem 2 require
a background grid that is much finer near the notch. We can obtain results that
effectively use a graduated grid which becomes progressively finer in regions closer
to the notch. It is clear, however, from varying the number of notch cells that the
impedances we are trying to compute have not converged with respect to number of
cells along the x, y, or z directions for a grid of 32 × 2 × 16 cells.
It is also clear from varying the number of background (bolt-hole) cells that the
impedances have not converged with respect to number of cells along the x, y, or z
directions for a grid of 16× 16 × 4 cells. We can, however, improve these impedance
values by combining them with values obtained from an 8 × 8 × 2 cell graduated
grid for the bolt hole and a 16 × 2 × 8 cell notch grid. This allows us to perform a
linear extrapolation to zero cell dimensions. That is the result labeled “Extrapol’d”
in Fig. 10.32.
222 10 Some Special Topics in Computational Electromagnetics

Fig. 10.32 Comparison of Notch at a Bolt Hole Benchmark Problem 2

measured values of the 1
change in impedance to 0
values computed using
-1
different grids for the bolt Expt

dR (Ohm)
hole in Benchmark -2
Extrapol’d
Problem 2. The scan is along -3
x1.21
Y =0 -4
-5
-6
-7
-30 -20 -10 0 10 20 30
Probe x position (mm)

Notch at a Bolt Hole Benchmark Problem 2

10
Expt
8
Extrapol’d
dX (Ohm)

6 x1.12

-30 -20 -10 0 10 20 30

Probe x position (mm)

As with Benchmark Problem 2, accurate results for Benchmark Problem 1

require a background grid that is much finer near the notch, and so we again
apply the graduated-grid scheme that was used for that problem. We also use
the extrapolation procedure that was used in Problem 2, with the result labeled
“Extrapol’d” in Fig. 10.33.
The response of the small notch is clearly evident in both benchmark problems,
and its shape agrees with the measured data, which were the goals of the algorithm,
but the absolute values differ by 12–39 %, with the smaller errors occurring in
Benchmark Problem 2. The reason for the discrepancies in scale is due to the size
of the problems. Benchmark Problem 1 contains a hole that is abnormally large
compared to the notch, and at the excitation frequency of 5 kHz, this hole requires a
huge number of cells. Benchmark Problem 2 is somewhat more realistic in its size,
and the extrapolated results are closer in scale to the measured data. In any case,
we should note that the raw data shown in Fig. 10.34 has a peak value of resistance
∼950 Ω and reactance ∼3, 000 Ω, which means that the values shown in Figs. 10.32
and 10.33 are stable to within a very small fraction of the peak values. This attests
to the accuracy of the algorithm and to the quality of the benchmark data. Had we
attempted to compute the signal due to the small notch by subtracting two much
larger signals, we would expect to get very noisy results.
10.5 Notch at a Bolt Hole Benchmark Problems 223

Fig. 10.33 Comparison of Notch at a Bolt Hole Benchmark Problem 1

measured values of the 0.2
change in impedance to
values computed using 0.1
different grids for the bolt 0

dR (Ohm)
hole in Benchmark -0.1
Problem 1. The scan is along Expt
Y =0 -0.2
Extrapol’d
-0.3
x1.39
-0.4
-0.5
-30 -20 -10 0 10 20 30
Probe x position (mm)

Notch at a Bolt Hole Benchmark Problem 1

1

0.8 Expt
Extrapol’d
dX (Ohm)

0.6
x1.17
0.4

0.2

0
-30 -20 -10 0 10 20 30
Probe x position (mm)
224 10 Some Special Topics in Computational Electromagnetics

Fig. 10.34 Resistance (top) and reactance (bottom) of XY scan at 5 kHZ

 Part II
Inversion Algorithms and
Signal-Processing
Chapter 11
Examples of Basic Inverse Problems

11.1 Overview of the General Inversion Process

for Parameter Estimation

Now that we have the electromagnetics down pat, let’s look ahead to see where that
leads us. In this part of the book, a general inverse method process is introduced that
provides a framework for incorporating the numerical forward models discussed in
Part I in an algorithm that estimates an unknown set of values of the test. A diagram
of a general inverse method process is presented in Fig. 11.1. The heart of the
inversion process is the inverse method step. Typical parameters to be evaluated
include crack dimensions (e.g., surface length or depth), material properties (e.g.,
conductivity) or state(s) of the nondestructive evaluation system (e.g., probe lift-
off or tilt.) Here, parameter values are estimated using an iterative scheme such as
nonlinear least-squares estimation (NLSE) which equates the model results with
experimental data through adjusting the model parameter values. Such general
inverse methods will be presented in detail in Chap. 12. To ensure accurate
estimation results, statistical estimators can be used to evaluate convergence to
the solution and quantify the error bounds on the results. Noise and uncertainty
analysis is first introduced in Sect. 11.2.2, with a more complete study developed in
Sects. 11.6 and 11.7. The Cramer–Rao Lower Bound, another stochastic metric,
is introduced in Sect. A.1 of Chap. 12. The theory and use of robust statistical
estimation metrics will be presented in Chap. 13.
Although inverse methods have been studied for many years, to some degree
there has been limited progress to transition them to practical applications. There
are three primary reasons for the difficulty in transitioning inverse methods to
NDE applications in particular: long solution times for accurate NDE measure-
ment models, the inherent ill-posedness of certain inverse problems in NDE, and
the lack of robustness of the inverse method schemes to the presence of noise
and uncontrolled variation with in-field NDE measurements. To address these
challenges, a comprehensive approach is proposed in Fig. 11.1 for model-based

H.A. Sabbagh et al., Computational Electromagnetics and Model-Based Inversion, 227

Scientific Computation, DOI 10.1007/978-1-4419-8429-6 11,
© Springer Science+Business Media New York 2013
228 11 Examples of Basic Inverse Problems

Fig. 11.1 General inversion process for parameter estimation

inversion design to ensure the reliability of the inverse method step through (a)
key parameter assessment and model parameterization, (b) model reduction through
surrogate models, (c) model calibration, (d) noise removal, (e) data registration, and
(f) feature extraction techniques.
First, the inverse method development process must always begin with a
full assessment of the key factors controlling the experimental data and driving
inversion performance. All critical factors for the NDT technique, part material,
part geometry, and discontinuity characteristics that control signal and noise must
be identified and well understood. Clearly, the flaw characteristics such as the crack
dimensions are the key factors of interest in the forward and inverse models. This
step provides a critical opportunity to reduce the complexity of the inverse problem
by parameterizing the flaw morphology, potentially addressing some levels of ill-
posedness by simplifying the inverse problems. It is important to also consider all
varying conditions in the NDE measurement that affect the measurement response,
for example lift-off or scan resolution. If such key factors cannot be controlled
during measurements, these parameters should also be incorporated in the inverse
problem with the characteristics of the flaw. An excellent case study supporting this
point is the inversion of inner diameter (ID) pits presented in Sect. 18.3.
Although the volume integral method introduced in Part I is quite efficient
numerically for the simulation of eddy current NDE measurements, for inverse
problem applications it can still be time consuming to perform such repeated
numerical calculations within an inversion scheme. Surrogate models or structures
can be created from the results of the numerical model to greatly improve the
solution time and performance of inverse methods. Frequently, data or look-up
tables are constructed and interpolation schemes are applied in order to provide
a fast means of sampling the model over a wide range of conditions. The use of
surrogate structures, first mentioned in Sects. 1.4 and 6.6, is formally introduced
in Chap. 12. Emerging stochastic simulation methods, such as the probabilistic
11.1 Overview of the General Inversion Process for Parameter Estimation 229

collocation method (PCM), are briefly introduced in Chap. 21 on model-assisted

probably of detection (POD) analysis. These methods will be further developed and
applied extensively in the second volume of this series.
Eddy-current simulations are designed to represent the change in impedance
of an eddy-current probe interacting with a test specimen. However, conventional
eddy-current measurement systems typically operate with the probe balanced over
the test specimen and output voltage data. Thus, a model calibration step is a
necessary step in practice to equate the measurement data with simulated results in a
way that mimics the NDE technique procedure. In many cases, gains and thresholds
are set based on the desired response to a known calibration standard. Parameters
of the model can essentially be fit to obtain the best match with a limited set of
empirical data acquired according to a calibration procedure. Chapter 15 introduces
a method for transforming impedance calculations into voltage measurements using
calibration data with known samples. At this stage in the general inversion process,
a fast surrogate model calibrated according to the inspection procedure is available
for inverse method applications.
There are several key steps for data pre-processing that can greatly improve the
reliability of inverse methods with respect to all possible varying test conditions.
Noise and clutter removal is an important first step in preparing data for inversion.
Typical sources for clutter and noise found in eddy-current scans include probe lift-
off variation, local changes in material properties, sensitivity to sub-surface struc-
tures, thermal variation of measurement components, and measurement (electrical)
noise. For noise removal, a variety of conventional signal and image processing
algorithms can be used to clean up the signals, for example: averaging, high-pass
and low-pass filters, and median filters. However, clutter is defined here as coherent
noise associated with departures in the impedance of the workpiece due to real
non-flaw conditions such as surface conditions (probe lift-off), part (sub-surface)
geometry, and material properties. Such variation cannot be removed through basic
averaging or filtering techniques but typically requires a more sophisticated removal
process. Several special clutter removal strategies are proposed in this part of the
book. An iterative algorithm, reminiscent of the projection onto convex sets (POCS),
is described in Chap. 14 to address coherent noise isolated to select data channels. In
Chap. 15, several model-based clutter fitting and removal approaches are introduced
to address more challenging background feature removal problems.
In addition to noise removal, data registration and feature extraction are two
important steps to ensure the measurement data is of high quality necessary for the
inversion process. Data registration is the process of aligning the raw data signals in
time or images in space to facilitate proper data comparison and inversion with
simulated data. Typically, cross-correlation routines are used to align signals to
the necessary high degree for the inverse method to be successful. Lastly, feature
extraction is a useful step in order to minimize the dimension and size of data
being inverted. The effective use of feature extraction can reduce the likelihood an
230 11 Examples of Basic Inverse Problems

a
INPUT SYSTEM OUTPUT
[KNOWN] [KNOWN] [TO BE DETERMINED]

b
SYSTEM OUTPUT
INPUT
[KNOWN] [KNOWN]
[TO BE DETERMINED]

c
INPUT SYSTEM OUTPUT
[KNOWN] [TO BE DETERMINED] [KNOWN]

Fig. 11.2 System representation of direct and inverse problems: (a) the direct problem; (b) the
signal-detection (communication) problem; (c) the inverse problem

inverse problem will be ill-posed, while reducing calculation time and improving the
inverse solution performance. Several examples demonstrating the important of data
registration and feature extraction steps are included in the remainder of the book.
Nondestructive evaluation (NDE) is to materials and structures what computer-
aided tomography (CAT) scanning is to the human body—an attempt to look inside
without opening the body. It is in nature an inverse problem. Figure 11.2 illustrates
a system representation for three important problems: (a) a direct problem, in
which the input and system are known and the output is to be determined; (b) a
signal-detection (communication) problem, in which the system (a communication
channel) and output are known, and the problem is to determine the input signal;
and (c) the inverse problem, in which the input and outputs are known, and we must
determine the system.
The results that we showed in Part I of this text were direct problems; we
assumed knowledge of the probe and flaw and determined the response of the
probe, namely the driving-point or transfer impedances. The second problem is
dealt with in communication and information theory texts, such as [50], and has
a close relation to inverse problems. In the typical inverse problem of eddy-current
NDE, we assume that impedances (or instrument voltages) are measured, given the
probe and measurement conditions, and the intent is to determine the anomaly that
produced the output. By this we mean to quantify the size, shape, and location of a
flaw, or other departure from an assumed-known condition.
The results of Part I have given us an engine, VIC-3D R
, that is designed
to solve the forward problem, and, by doing so, will provide the engine for
11.2 Thickness Measurements with Eddy-Current Probes 231

solving the inverse problem. In this chapter, we will demonstrate the notion of
eddy-current inverse problems by solving three basic problems: (a) a thickness
measurement, (b) determining a conductivity profile, and (c) a combination of the
two in the semiconductor chip industry. These problems are simple examples of the
application of eddy-current NDE to process control and in-service inspection, and
they also give us our first examples of model-based inversion, in which the physical
problem can be modeled with just a few unknown parameters. We will defer to the
next chapter a detailed discussion of the mathematical algorithms that will be used
in this chapter and throughout the remainder of the book.

11.2 Thickness Measurements with Eddy-Current Probes

The problem is illustrated in Fig. 11.3 [52]. The thickness, t, of the plate is to be
determined from measurements with an air-core probe. The probe lift-off, l, is often
poorly known. This can be due to a nonconducting layer of unknown thickness on
the surface, as illustrated in Fig. 11.3b, or simply because the probe geometry is not
well known. Since the probe response can be very sensitive to this parameter, it is
desirable to have an inversion scheme that does not require the lift-off to be known.
We use a scheme that determines this parameter as well as the plate thickness.
The impedance of the probe is a complicated nonlinear function of several
variables:
Z = g(coil, workpiece, t, l, f ) , (11.1)
where “coil” denotes coil parameters and “workpiece” denotes workpiece
parameters. The plate thickness, probe lift-off and frequency are denoted by
t, l, and f , respectively. The problem is to invert (11.1), obtaining t and l when
Z, f , and the coil and plate parameters are known.

Fig. 11.3 Thickness

measurement and lift-off
problem. (a) Unknown
lift-off. (b) Unknown
nonconducting layer
thickness
232 11 Examples of Basic Inverse Problems

In principle t and l can be determined from a single measurement of impedance,

since the resistance, R, and reactance, X, provide two independent equations.
A more accurate determination can be made, however, when more than one
measurement of Z is performed, thereby providing an overdetermined system of
equations for t and l. Z is most often measured as a function of frequency, since this
is usually the most easily varied parameter. In our analysis, we suppress the variation
of Z which is only due to frequency by dividing all impedance values by the
isolated probe reactance. We denote these normalized values simply by Z, R, and X.
The statement of the problem then becomes: Given Z at several frequencies, and the
coil and plate parameters, determine t and l.
Because the problem is nonlinear, we use a Gauss–Newton iteration scheme to
perform the inversion. At a single frequency, f , the linear approximation to R and X
can be written:
     
R( f ,t, l) R̃( f ,t, l) R( f ,t0 , l0 )
≈ =
X( f ,t, l) X̃( f ,t, l) X( f ,t0 , l0 )
⎡ ⎤
∂R ∂R
⎢ ∂t ∂l ⎥  
⎢ ⎥ t − t0
+⎢ ⎥ . (11.2)
⎣ ∂X ∂X ⎦ l − l0
∂ t ∂ l (t0 ,l0 )

The partial derivatives are computed numerically by

∂Z Z(t + Δt) − Z(t)

≈
∂t Δt
∂Z Z(l + Δl) − Z(l)
≈ . (11.3)
∂l Δl

The left-hand side of (11.2) is taken to be the measured values of resistance and
reactance. An initial value, (t0 , l0 ), is chosen, and (11.2) is solved for (t, l). Iteration
then proceeds by replacing (t0 , l0 ) with (t, l), and re-solving (11.2) until t and l
converge, and
   
Rmeas ( f ) R̃( f ,t, l)
= . (11.4)
Xmeas ( f ) X̃( f ,t, l)

When measurements are performed at N frequencies, the overdetermined system

of equations can be written:
11.2 Thickness Measurements with Eddy-Current Probes 233

⎡ ⎤
∂R ∂R
( f1 ) ( f1 )
⎢ ∂t ∂l ⎥
⎢ ⎥
⎡ ⎤ ⎢ ⎥
ΔR( f1 ) ⎢ ∂X ∂ ⎥
⎢ ( f1 )
X
( f1 ) ⎥
⎢ ΔX( f1 ) ⎥ ⎢ ⎥    
⎢ ⎥ ⎢ ∂t ∂l ⎥
⎢ .. ⎥ ⎢ . . ⎥ t − t0 Δt
⎢ . ⎥ =⎢ .. .. ⎥ ≡A
⎢ ⎥ ⎢ ⎥ l − l0 Δl
⎣ ΔR( fN ) ⎦ ⎢ ∂R ∂R ⎥
⎢ ( f ) ( f ) ⎥
ΔX( fN ) ⎢ ∂t N
∂l
N ⎥
⎢ ⎥
⎢ ⎥
⎣ ⎦
∂X ∂X
( fN ) ( fN )
∂t ∂l (t0 ,l0 )

ΔR( f ) = Rmeas ( f ) − R( f ,t0 , l0 )

ΔX( f ) = Xmeas ( f ) − X( f ,t0 , l0 ). (11.5)

Iteration proceeds as before, except that now we solve the overdetermined system by
finding the least-squares solution that minimizes the norm of the residuals (Gauss–
Newton):
Residual2 ≡ ∑ ΔR( fi )2 + ΔX( fi )2 . (11.6)
i

The least-squares solution can be obtained by solving the normal equations:

⎡ ⎤
ΔR( f1 )
  ⎢ ΔX( f1 ) ⎥
⎢ ⎥
Δt T⎢ .. ⎥
T
A A =A ⎢ . ⎥, (11.7)
Δl ⎢ ⎥
⎣ ΔR( fN ) ⎦
ΔX( fN )

but improved numerical stability can be achieved by using the QR-decomposition

[53], in which an orthogonal matrix, Q, is found that satisfies
 
R
QA = , (11.8)
0
and R is nonsingular and upper-triangular. Then the least-squares solution is
⎡ ⎤
ΔR( f1 )
  ⎢ ΔX( f1 ) ⎥
⎢ ⎥
Δt −1 ⎢ .. ⎥
= R Q1 ⎢ . ⎥, (11.9)
Δl ⎢ ⎥
⎣ ΔR( fN ) ⎦
ΔX( fN )

where Q1 consists of the first two rows of Q.

234 11 Examples of Basic Inverse Problems

Fig. 11.4 Test Case 1: a

model problem with an
aluminum plate. The probe is
cylindrical

Fig. 11.5 Test Case 1:

derivative of impedance with
respect to thickness.

Fig. 11.6 Test Case 1:

derivative of impedance with
respect to lift-off

11.2.1 Test Case 1

Frequencies are selected to maximize the sensitivity to the plate thickness. We

useVIC-3D R
to obtain numerical derivatives of the impedance. For the test case
of Fig. 11.4, these are shown in Figs. 11.5 and 11.6. From these derivative values,
11.2 Thickness Measurements with Eddy-Current Probes 235

Table 11.1 Deduced

thickness and lift-off for Test Frequencies Thickness Lift-off No. of iterations
Case 1. The plate is 0.1 mm 20000.0 0.100003 0.499985 3
thick, and the lift-off is 22963.1
0.5 mm. Starting values: 26365.1
0.2 mm thickness, 0.2 mm 30271.2
lift-off 34756.0
39905.2

we choose frequencies in the range 20–40 kHz, where the variation with thickness
is largest. We note that if the primary interest had been in determining the lift-off,
then higher frequencies would be preferable. O. Baltzersen [55] gives an empirical
formula, attributed to Förster and Libby [56], for the optimum frequency:
700
fopt = , (11.10)
σ tAeff
where f is in kHz, σ is the conductivity of the plate in mega/ohm-m, t is the plate
thickness in mm, and Aeff is essentially the coil diameter in mm. In this case the
formula gives 29 kHz, in good agreement with our choice of frequencies.
Six frequencies are chosen to cover the range of 20–40 kHz. Table 11.1 shows
the thickness and lift-off that are obtained from our inversion procedure. The initial
guess for both the thickness and lift-off is 0.2 mm. Convergence is achieved in three
iterations.

11.2.2 Noise and Uncertainties in Δt

The uncertainties in the thickness that is obtained from the inversion procedure
that we have described can be estimated from the uncertainties in the measured
probe impedances. For small changes in resistance and reactance, the change in the
deduced thickness is given by
N
∂t ∂t
Δt = ∑ ΔRi + ΔXi. (11.11)
i=1 ∂ R i ∂ Xi

If ΔRi and ΔXi have Gaussian distributions with half-width σRi and σXi , then Δt is
Gaussian with width
     1/2
∂t 2 2 ∂t 2 2
σt = ∑ σRi + σXi . (11.12)
i ∂ Ri ∂ Xi
236 11 Examples of Basic Inverse Problems

Fig. 11.7 Test Case 1:

distribution of errors for
impedance, thickness and
lift-off

In the simple case that σRi /Ri = σXi /Xi = σZ /Z, the half-width for the thickness is
given by
  2  2 1/2
σt ∂ t Ri ∂ t Xi σZ
t
= ∑ ∂ Ri t
+
∂ Xi t Z
. (11.13)
i

The partial derivatives in (11.13) can be evaluated by performing the inversion

procedure for values of R and X near the measured values and using the deduced
thicknesses to compute numerical derivatives. Using this procedure to evaluate the
partial derivatives for Test Case 1, we obtain a half-width for the deduced thickness
given by σt /t = 0.0542, assuming a half-width for the measured impedance given by
σz /Z = 0.02. This is the half-width that is obtained with the inversion scheme using
six frequencies. If only the single frequency, 30271.2 Hz, is used for the analysis,
then the half-width is given by σt /t = 0.120. The corresponding probability distri-
butions are illustrated in Fig. 11.7. We see that inclusion of multiple frequencies
substantially reduces the spread in the deduced thickness. The uncertainties in the
determination of the probe lift-off can be estimated in the same manner. The half-
width for the deduced lift-off is given by σl /l = 0.0267 when six frequencies are
used.
If ΔRi and ΔXi are due to systematic errors, so that ΔRi /Ri = ΔXi /Xi =
ΔZ/Z, then
 
Δt N
∂ t Ri ∂ t Xi ΔZ
= ∑ + . (11.14)
t i=1 ∂ Ri t ∂ Xi t Z

If the errors in Test Case 1 are systematic, with ΔRi /Ri = ΔXi /Xi = ΔZ/Z < 0.05,
then the errors in thickness and lift-off, using six frequencies, are given by
11.2 Thickness Measurements with Eddy-Current Probes 237

Fig. 11.8 Test Case 2:

experimental arrangement
with brass plate. The probe is
cylindrical

Δt ΔZ
= −7.36
t Z
Δl ΔZ
= −1.85 .
l Z

If the analysis includes only the single frequency, 30271.2 Hz, then Δt/t is about the
same, while Δl/l is twice as large. Thus, if only one frequency is used in the analysis,
then the systematic errors and statistical errors make comparable contributions to the
thickness error, while the inclusion of six frequencies makes the statistical errors
much less significant. In the six-frequency analysis, changes in reactance at high
and low frequencies produce opposite effects on the deduced lift-off. In this case,
systematic errors can be cancelled by an appropriate choice of frequencies.

11.2.3 Test Case 2

We now consider a second test case, for which we have experimental measurements.
The problem is illustrated in Fig. 11.8. The measurements are those of Burke
(1992, private communication) and cover the range 100–10,000 Hz. The computed
derivatives of impedance with respect to thickness and lift-off are shown in
Figs. 11.9 and 11.10. From these plots, the range 1–2 kHz is chosen for determining
the thickness. The thickness and lift-off values obtained from a number of analyses,
each using a different number of frequencies, are shown in Table 11.2.
In Fig. 11.11 the ratio of the deduced thickness and lift-off to the measured
values are plotted versus the number of frequencies used in the analysis. The
improvement observed in thickness determination with the addition of frequencies
to the analysis suggests that there exists a statistical component of the errors in the
measured impedance values. The remaining 1 % difference between the measured
and deduced thickness is likely due to systematic errors in the measurements or
calculations. The probe lift-off is not determined quite as well as the thickness. This
could be anticipated from Figs. 11.9 and 11.10, which show the impedance to be
less sensitive to the lift-off.
238 11 Examples of Basic Inverse Problems

Fig. 11.9 Test Case 2:

derivative of impedance with
respect to thickness

Fig. 11.10 Test Case 2:

derivative of impedance with
respect to lift-off

11.2.4 Conclusions

This first example illustrates several important features of model-based inversion

that will become manifest throughout the remainder of the book. We have shown
how uncertain parameters, such as lift-off, can be accounted for when determining a
“primary parameter,” such as plate thickness. We simply add such parameters to the
bag of unknowns and solve all simultaneously. There is a limit to this, of course, but
in many practical cases that will be illustrated later, this approach succeeds well.
A second point to be made is that with modern computational tools, such as
VIC-3D R
, one can dispense with “rules of thumb” and other empirical criteria that
saturate the current practice of eddy-current NDE. Such tools provide a means of
quickly and efficiently optimizing probe design and measurements to maximize the
accuracy needed in modern engineering practice.
11.3 Conductivity Profile Measurements 239

Table 11.2 Deduced thickness and lift-off for Test Case 2.

The brass plate is 0.89 mm thick and the lift-off is 2.03 mm. Starting
values: 2 mm thickness, 0 mm lift-off
Frequencies Thickness Lift-off No. of iterations
1584.89 0.8779 2.063 3
1412.54 0.8779 2.062 3
1584.89
1412.54 0.8788 2.063 3
1584.89
1778.28
1258.93 0.8793 2.062 3
1412.54
1584.89
1778.28
1258.93 0.8801 2.065 3
1412.54
1584.89
1778.28
1995.26
1122.02 0.8803 2.065 3
1258.93
1412.54
1584.89
1778.28
1995.26
1000.00 0.8801 2.064 3
1122.02
1258.93
1412.54
1584.89
1778.28
1995.26

11.3 Conductivity Profile Measurements

It is often necessary to inspect materials for hardness, residual stress, fiber-volume

content, etc. These properties may be the results of a manufacturing process, or
the results of exposure to a harsh environment. Whatever the cause, when these
physical attributes can be correlated with electrical conductivity or permeability,
then eddy-current inspection can be used to monitor them. We do not address
the question of correlating conductivity to the material properties, leaving that to
materials scientists. What we will do is show how modeling can (1) tell what is
required of the eddy-current instrument in order to achieve satisfactory performance
and (2) help us understand the measurements we obtain from our instruments.
240 11 Examples of Basic Inverse Problems

Fig. 11.11 Test Case 2: thickness and lift-off versus the number of frequencies used

d(Z/Xfree)/dConduct Conductivity Profile

d(Z/Xfree)/dConduct (m/S)

3e-09 8e+07
region 1 7e+07 True
Conductivity (S/m)

2.5e-09
region 2 6e+07 Initial
2e-09
region 3 5e+07 Final
1.5e-09 region 4 4e+07
1e-09 3e+07
2e+07
5e-10
1e+07
0 0
10000 100000 1e+06 0 0.1 0.2 0.3 0.4 0.5
Frequency (Hz) Depth (mm)

Fig. 11.12 A model problem with three layers over a half-space. Derivative of normalized
impedance with respect to conductivity, versus frequency (left), and conductivity profile. Region 1
lies between 0 and 0.1 mm; region 4 lies below 0.3 mm

Suppose a flat, layered structure has layers with the conductivities shown (solid)
in Fig. 11.12 [54]. We can reconstruct this profile using impedance measurements
from a simple air-core probe. To model this problem, we use VIC-3D R
to compute
the impedance seen by an air-core probe resting on our layered structure. Our probe
has 400 turns, is 1.73 mm in height, and has inner and outer radii of 2.54 and
4.27 mm. To determine the best frequencies for our probe, we look at the derivative
of the normalized impedance with respect to the conductivities of the four regions
of our structure. For all regions, this derivative peaks between 3 kHz and 3 MHz, as
shown in Fig. 11.12. We use this frequency range.
We use the same nonlinear least-squares algorithm (Gauss–Newton) that was
used in the thickness measurements to recover the conductivities from the
impedances. The starting point for our iterative algorithm is a uniform conductivity
11.3 Conductivity Profile Measurements 241

Impedance Noise
0.01
Resistance
0.008 Reactance
0.006
Fractional Noise 0.004
0.002
0
- 0.002
- 0.004
- 0.006
- 0.008
- 0.01
1000 10000 100000 1e+06 1e+07
Frequency (Hz)

Fig. 11.13 Perturbations in the computed impedances at 21 frequencies. This is the input data that
represents the effects of “noise”

of 4 × 107 S/m. Three to six frequencies, in a fairly narrow band, are sufficient
to exactly reconstruct the conductivities, if our computed “noiseless” impedances
are used. To investigate the effect of “noise” in the data, we perturb the com-
puted impedances, which are our input data, by the fractional amounts shown in
Fig. 11.13. This is not representative of any particular noise model, but simply
illustrates the effects that perturbations have on the reconstructions. Additional
frequencies are now required to produce a satisfactory inversion. The conductivities
obtained using 21 frequencies are shown (dashed) in Fig. 11.12. The frequencies
chosen are simply those that form a geometric progression between 3 kHz, and
3 MHz. Ten iterations were required.
We can better understand the quality of the reconstruction we can expect from
our measurements by referring to the derivatives shown in Fig. 11.12. Clearly, at
low frequencies, the impedance is much more sensitive to the conductivity of the
half-space, while at high frequencies it is much more sensitive to that of the top
layer. Hence, our recovered conductivity profile is most accurate in these regions.
Our modeling allows us to test our inversion algorithm, determine the proper
frequencies to measure, establish noise requirements for our measurements, and
understand the quality of reconstruction we can expect.
We believe that this is an important point to stress. Eddy-currents have long been
applied to the measurement of thickness and conductivity, but the techniques have
generally been quite empirical, and used discrete standards. The results of blindly
using the nearest available instrument have often been disappointing. Modeling must
be performed before equipment is purchased, or expensive experiments undertaken.
242 11 Examples of Basic Inverse Problems

11.4 Conductivity and Permeability Profile Measurements

In order to learn more about the problem of simultaneously inverting magnetic

permeabilities and electrical conductivities, we resort to the layered system shown
in Fig. 11.14 [57].
We let the (relative) magnetic permeability of the top layer take on the values 1,
2, 4, 8, and 16 and use VIC-3D R
to compute and plot Förster diagrams1 over the
frequency range of 500 kHz–10 MHz. The result is shown in Fig. 11.15. It is clear
that considerable information about the change in impedance with permeability is
revealed in the low-frequency range near the vertical (normalized reactance) axis.
When we slightly modify the conductivity of the top layer of Fig. 11.14, using
the nominal values of σ = 800, 000, μ = 2, we get the Förster plots of Fig. 11.16,
in which the circles correspond to the unperturbed configuration, and the triangles
to the configuration in which the conductivity is slightly perturbed. Clearly, the
conductivity changes are along the Förster plot, in the direction of frequency.
Next, we repeat the experiment, changing the relative permeability of the top
layer slightly from its nominal value of two, and plot the results in Fig. 11.17.
It appears, now, that the change is roughly normal to the original curve.
Indeed, this is borne out in Table 11.3, which shows the cosine of the angle be-
tween the vectors representing conductivity and permeability changes in Figs. 11.16
and 11.17. The frequency range in Table 11.3 extends from 50 kHz to 1 MHz, and
the nominal value of the relative permeability of the top layer is 16.

Coil excited at many frequencies

σ = 800000 μ = 16 8 mils

σ = 730000 μ = 1

Fig. 11.14 A model configuration to study the problem of simultaneously reconstructing magnetic
permeabilities and electrical conductivities. For this study, the relative permeability of the top layer
varies from 1 to 16

1 Recall
that a Förster diagram is a plot of normalized reactance to normalized resistance, with the
normalizing factor being the freespace reactance of the probe.
11.4 Conductivity and Permeability Profile Measurements 243

Fig. 11.15 Förster plots, as computed by VIC-3D

R
for μ = 1, 2, 4, 8, 16, (bottom to top), over
the frequency range of 500 kHz to 10 MHz. The curves evolve in the clockwise direction with
frequency

Fig. 11.16 Förster plots of the unperturbed configuration of Fig. 11.14 (circles) and the config-
uration in which the conductivity of the top layer is slightly perturbed (triangles). The frequency
range is the same as in Fig. 11.15

Clearly, there is an advantage to working in the lower-frequency range when

trying to discriminate between conductivity and permeability changes, because the
correlation between these two parameters is smallest there.
244 11 Examples of Basic Inverse Problems

Fig. 11.17 Förster plots of the unperturbed configuration of Fig. 11.14 (circles) and the configu-
ration in which the permeability of the top layer is slightly perturbed from its nominal value of two
(triangles). The frequency range is the same as in Fig. 11.15

Table 11.3 The cosine of the

angle between the vectors Freq. (Hz) Cosine
representing conductivity and 50,000 −0.130716
permeability changes in 100,000 −0.206825
Figs. 11.16 and 11.17. The 150,000 −0.270171
nominal value of the relative 200,000 −0.326754
permeability of the top layer 250,000 −0.375093
is 16 300,000 −0.420795
350,000 −0.463919
400,000 −0.503752
450,000 −0.542168
500,000 −0.579675
550,000 −0.615551
600,000 −0.650175
650,000 −0.683748
700,000 −0.716262
750,000 −0.747107
800,000 −0.775634
850,000 −0.803411
900,000 −0.828341
950,000 −0.851567
1,000,000 −0.872808
11.5 A Problem in the Semiconductor Chip Industry 245

Coil excited at many frequencies

σ = 800000 μ = 16 8 mils

σ = 730000 μ=1

Fig. 11.18 A two-layer ferromagnetic system, whose conductivities and permeabilities are to be
reconstructed using multifrequencies. The frequency range extends from 50 kHz to 1 MHz in 19
equal steps

Coil excited at many frequencies

σ = 800900 μ = 16.01 8 mils

σ = 744700 μ = 0.9968

Fig. 11.19 Reconstruction of the system shown in Fig. 11.18, when 20 frequencies are used from
50 kHz to 1 MHz. The impedance data are noiseless

11.4.1 Multifrequency Reconstruction of a Two-Layered

Ferromagnetic System

This problem, which is shown in Fig. 11.18, addresses the issue of reconstructing
magnetic permeabilities, as well as electrical conductivities.
The impedance data are taken over the frequency range of 50 kHz to 1 MHz,
in 19 equal intervals. These data, without the addition of noise, are used to
reconstruct both permeability and conductivity, with the results shown in Fig. 11.19.
The starting points for the iterations were σL = σS = 750, 000, μL = μS = 2, where
the subscript “L” denotes the layer and “S” denotes the substrate.

11.5 A Problem in the Semiconductor Chip Industry

The geometry of the problem is illustrated in Fig. 11.20 (TENCOR, Inc., 1996,
private communication):
246 11 Examples of Basic Inverse Problems

Fig. 11.20 Illustrating the

Aluminum t1
geometry of the problem.
Epilayer t2

t0 Silicon t3

11.5.1 Statement of the Problem

Statement No. 1
Given:
(a) t0 , the total thickness of the package
(b) σAl = 3.57 × 107 S/m, the conductivity of aluminum
(c) σEpi = 10 S/m, t2 = 10 × 10−6 m

Determine:
(a) t1 , the thickness of the aluminum layer
(b) σSi , the conductivity of the silicon substrate

Statement No. 2
Given:
(a) t3 = 550 × 10−6 m, the thickness of the silicon substrate
(b) σAl = 3.57 × 107, the conductivity of aluminum
(c) σEpi = 10 S/m, t2 = 10 × 10−6 m

Determine:
(a) t1 , the thickness of the aluminum layer
(b) σSi , the conductivity of the silicon substrate
The only difference between the two statements lies in the information that is
given about the thickness. Mathematically, the two problems are identical, because
one piece of thickness information is given in each of the statements.
11.5 A Problem in the Semiconductor Chip Industry 247

11.5.2 Method of Approach

We will follow the approach described in Sect. 11.2, but now there are two
approaches that we can take to gather the data. As before, we can use multifrequency
excitation of a coil that is placed at a known lift-off above the aluminum layer
and invert the impedance data at these frequencies to determine t1 and σSi , jointly.
The second possibility is to vary the lift-off of the coil that is excited at a fixed
frequency, measure the impedance at each lift-off value, and invert these data
to determine t1 and σSi , jointly. The latter is a realistic possibility, because the
manufacturing and measurement processes for semiconductor chips are performed
in clean rooms, with well-defined conditions and instruments. The lift-off can be
controlled and measured accurately by ultrasonic means.
We’ll briefly examine the multifrequency case and discuss more thoroughly
numerical experiments with the multi lift-off case. It is convenient for analysis
purposes to normalize impedances with respect to the free-space reactance of
the coil, when data are taken over a broad frequency range. Figure 11.21 shows
derivatives of the normalized resistance and reactance with respect to t1 and σSi ,
as a function of frequency over the range 106 –109 Hz. The nominal values of the
unknowns are t1 = 560 Å = 5.6 × 10−5 mm and σSi = 2 × 104 S/m, with a (known)
lift-off of 0.5 mm.
The results indicate that care must be taken in choosing an operating frequency.
The thin aluminum layer shields the probe coil from the silicon substrate at
frequencies greater than 100 MHz, which is why we see a significant decrease in
the conductivity sensitivity coefficients (the derivatives with respect to σSi ) over
this frequency range in the bottom part of Fig. 11.21.
We’ll turn now to the second possibility, in which lift-off is varied over 16
values, starting at 0.0762 mm and ending at 1.01614 mm in equal logarithmic
increments. We use three different frequencies, 10, 25 and 100 MHz. The sensitivity
coefficients (derivatives) for each of these frequencies are shown in Figs. 11.22–
11.24. It is interesting to note that dR/d σSi , as a function of lift-off, is positive for
10 MHz and “pivots” around the zero-value at a lift-off of 1 mm and is negative
at 100 MHz. It appears that at a frequency somewhat higher than 25 MHz, this
derivative is actually zero for all lift-off values, which suggests that at this frequency
the resistance is insensitive to changes in conductivity.
The input data for the multi lift-off inversion problem are obtained by scanning
a coil of inner radius 1 mm, outer radius 1.2 mm, height 0.2 mm, with 18 turns
above the system shown in Fig. 11.20 at 16 lift-off values, and then recording the
change in impedance from freespace conditions. The freespace resistance is zero, of
course, and the freespace reactance is 73.0313 Ω at 10 MHz, 182.578 Ω at 25 MHz,
and 730.314 Ω at 100 MHz, as computed by VIC-3D R
. The model data are shown
in Table 11.4.
248 11 Examples of Basic Inverse Problems

Fig. 11.21 Derivatives of Al(560 Angs.) , Si(20000 S/m) (0.5mm liftoff)

normalized resistance and 150
reactance with respect to t1 dR/d(Al thickness)
and σSi 100
dX/d(Al thickness)

50

d(Z/Xfree)/dt (1/mm)
0

-50

-100

-150

-200

-250
1e+06 1e+07 1e+08 1e+09
Frequency (Hz)

Al(560 Angs.) , Si(20000 S/m) (0.5mm liftoff)

1e-06
dR/d(Si conductivity)
dX/d(Si conductivity)

5e-07
d(Z/Xfree)/ds (m/S)

-5e-07

-1e-06

-1.5e-06
1e+06 1e+07 1e+08 1e+09
Frequency (Hz)

11.6 Noise Analysis: The Covariance Matrix

The matrix, A, of (11.5) is called the Jacobian of the system, and the derivatives
shown in Figs. 11.22–11.24 form the columns of A:

A = [t, σ ], (11.15)
11.6 Noise Analysis: The Covariance Matrix 249

Fig. 11.22 Derivatives of Al(560 Angs.) , Si(20000 S/m) (10MHz)

normalized resistance and 400
reactance with respect to t1 dR/d(Al thickness)
and σSi . Lift-off is varied at dX/d(Al thickness)
f = 10 MHz 300

d(Z/Xfree)/dt (1/mm)
200

100

-100

-200
0.01 0.1 1 10
Lift-Off (mm)

Al(560 Angs.) , Si(20000 S/m) (10MHz)

2.5e-06
dR/d(Si conductivity)
2e-06 dX/d(Si conductivity)

1.5e-06

1e-06
d(Z/Xfree)/ds (m/S)

5e-07

-5e-07

-1e-06

-1.5e-06

-2e-06

-2.5e-06
0.01 0.1 1 10
Lift-Off (mm)

where t denotes the column-vector containing derivatives with respect to t1 and σ

denotes the column-vector containing derivatives with respect to σSi . Each of these
vectors is of length 32.
The 2 × 2 matrix of the dot-products of the columns of A is called the Fisher
Information Matrix:
250 11 Examples of Basic Inverse Problems

Fig. 11.23 Derivatives of Al(560 Angs.) , Si(20000 S/m) (25 MHz)

normalized resistance and 500
reactance with respect to t1 dR/d(Al thickness)
and σSi . Lift-off is varied at 400 dX/d(Al thickness)
f = 25 MHz
300

200

d(Z/Xfree)/dt (1/mm)
100

-100

-200

-300

-400

-500

-600
0.01 0.1 1 10
Lift-Off (mm)

Al(560 Angs.) , Si(20000 S/m) (25 MHz)

2e-06
dR/d(Si conductivity)
dX/d(Si conductivity)
1e-06

0
d(Z/Xfree)/ds (m/S)

-1e-06

-2e-06

-3e-06

-4e-06

-5e-06
0.01 0.1 1 10
Lift-Off (mm)

 
t·t t·σ
F= . (11.16)
σ ·t σ ·σ

If our data have a uniform variance, σmeas

2 , then the covariance matrix, C, for our

computed parameters is given by

11.6 Noise Analysis: The Covariance Matrix 251

Fig. 11.24 Derivatives of Al(560 Angs.) , Si(20000 S/m) (100 MHz)

normalized resistance and 400
reactance with respect to t1 dR/d(Al thickness)
dX/d(Al thickness)
and σSi . Lift-off is varied at
200
f = 100 MHz

d(Z/Xfree)/dt (1/mm)
-200

-400

-600

-800

-1000
0.01 0.1 1 10
Lift-Off (mm)

Al(560 Angs.) , Si(20000 S/m) (100 MHz)

0
dR/d(Si conductivity)
dX/d(Si conductivity)
-5e-07
d(Z/Xfree)/ds (m/S)

-1e-06

-1.5e-06

-2e-06

-2.5e-06

-3e-06
0.01 0.1 1 10
Lift-Off (mm)

C = F −1 σmeas
2
. (11.17)

where σmeas is the variance in the measurement noise. There should be no confusion
between σ implying a variance and σ representing electrical conductivity. From
(11.16) and (11.17) we get
252 11 Examples of Basic Inverse Problems

Table 11.4 Model input data versus lift-off

10 MHz 25 MHz 100 MHz
Lift-off (mm) δ R (Ω) δ X (Ω) δ R (Ω) δ X (Ω) δ R (Ω) δ X (Ω)
0.0762 6.49585 −2.70823 26.0706 −20.723 109.204 −204.725
0.0905639 6.25278 −2.63751 24.9487 −20.0579 103.151 −196.281
0.107635 5.97935 −2.55661 23.6947 −19.3018 96.5301 −186.805
0.127925 5.67444 −2.46468 22.3065 −18.4487 89.3745 −176.274
0.152039 5.33781 −2.36104 20.7867 −17.4951 81.7439 −164.697
0.180699 4.97041 −2.24523 19.1439 −16.4401 73.7279 −152.135
0.214761 4.57465 −2.11715 17.3937 −15.287 65.4471 −138.7
0.255244 4.15469 −1.97718 15.5598 −14.0442 57.0521 −124.571
0.303358 3.71658 −1.82629 13.6743 −12.726 48.7193 −109.995
0.360541 3.2683 −1.66609 11.7774 −11.3531 40.6436 −95.2812
0.428504 2.81956 −1.49894 9.91533 −9.95243 33.0258 −80.7884
0.509278 2.38128 −1.32789 8.13789 −8.5563 26.057 −66.9023
0.605278 1.96477 −1.15656 6.49358 −7.20021 19.8993 −53.9995
0.719374 1.58069 −0.988919 5.02458 −5.91994 14.6663 −42.4069
0.854978 1.23787 −0.828974 3.76144 −4.74799 10.4069 −32.3626
1.01614 0.942303 −0.680426 2.71921 −3.71 7.09821 −23.9863

   
C11 C12 1 σ · σ −t · σ
= × σmeas
2
. (11.18)
C21 C22 (t · t)(σ · σ ) − (t · σ )(σ · t) −σ · t t · t

The variance of the computed parameters is

σ ·σ
σt21 = C11 = × σmeas
2
(t · t)(σ · σ ) − (t · σ )(σ · t)
t·t
σσ2Si = C22 = × σmeas
2
. (11.19)
(t · t)(σ · σ ) − (t · σ )(σ · t)

and the correlation between the two parameters, which lies between 0 and 1, is the
negative cosine of the angle between the two column-vectors of A:
C12 −t · σ
cor(t1 , σSi ) = √ = = − cos(θ ) (11.20)
C11 × C22 (t · t) × (σ · σ )

Notice that the variances are smallest when the correlation is small, in which case
the variance of each parameter is roughly proportional to the inverse of the derivative
of the impedance with respect to that parameter.
Consider, for example, the conditions at the nominal values of the parameters:
t1 = 5.6 × 10−5 mm and σSi = 2 × 104 S/m. From Figs. 11.22–11.24 we compute
the covariance matrix at f = 10, 25, and 100 MHz to be:
 
0.1392 × 10−4(mm)2 −0.1716 × 104(S/m)(mm)
C[10] = × σmeas
2
−0.1716 × 104(mm)(S/m) 0.2280 × 1012(S/m)2
11.6 Noise Analysis: The Covariance Matrix 253

 
0.1629 × 10−5(mm)2 −0.2185 × 103(S/m)(mm)
C[25] = × σmeas
2
−0.2185 × 103(mm)(S/m) 0.3764 × 1011(S/m)2
 
0.3028 × 10−6(mm)2 −0.6073 × 102(S/m)(mm)
C[100] = × σmeas
2
−0.6073 × 102(mm)(S/m) 0.3019 × 1011(S/m)2
(11.21)

If we assume that the measurements have a noise variance of 0.1 % of the

freespace reactance, then we compute the results for 10 MHz:

σt21 = 0.1392 × 10−4 mm2 σmeas

2

= 0.1392 × 10−4 mm2 × (0.1 × 10−2)2

= 0.1392 × 10−10 mm2
σt1 = 0.373 × 10−5 mm
37.3
% Error =
5.6
= 6.66
σσ2Si = 0.2280 × 1012(S/m)2 σmeas
2

= 0.2280 × 1012(S/m)2 × (0.1 × 10−2)2

= 0.2280 × 106(S/m)2
σσSi = 477.49 S/m
47, 749
% Error =
20, 000
= 2.39
−0.1716 × 104
cor(t1 , σSi ) = √
0.1392 × 10−4 × 0.2280 × 1012
= −0.963 , (11.22)

which means that t1 and σSi are highly anti-correlated.

In a similar manner we compute results for 25 MHz

σt1 = 0.1276 × 10−5 mm

% Error = 2.28
σσSi = 194.01 S/m
% Error = 0.97
cor(t1 , σSi ) = −0.882, (11.23)
254 11 Examples of Basic Inverse Problems

Table 11.5 Results of inversions with four different input noise sets at 10 MHz
Noise set t1 (mm) σSi (S/m) C
0.1403×10−4 −0.1727×104
1 5.72505 × 10−5 19794.9
−0.1727×104 0.2292×1012
0.1423×10−4 −0.1751×104
2 5.99395 × 10−5 19543.5
−0.1751×104 0.2320×1012
0.1412×10−4 −0.1739×104
3 5.80033 × 10−5 19671.8
−0.1739×104 0.2305×1012
0.1360×10−4 −0.1678×104
4 5.22920 × 10−5 20441.1
−0.1678×104 0.2236×1012

and

σt1 = 0.055 × 10−5 mm

% Error = 0.982
σσSi = 173.75 S/m
% Error = 0.869
cor(t1 , σSi ) = −0.635, (11.24)

for 100 MHz. It is clear that there is significant improvement in noise reduction at
the higher frequencies. Furthermore, t1 and σSi become less correlated at higher
frequencies, which leads to a more reliable inversion.

11.7 Noise Analysis: Monte Carlo Method

Monte Carlo methods imply performing a number of numerical experiments with

random input data. In the experiments reported here, we add random noise whose
variance is 0.1 % of the freespace reactance to the input data of Table 11.4. Thus,
the input impedance at a lift-off of 0.0762 mm is:

Z = 6.4958 ± 0.073 + j(73.0313 − 2.70823) ± 0.073 Ω at 10 MHz

Z = 26.070 ± 0.183 + j(182.578 − 20.723) ± 0.183 Ω at 25 MHz
Z = 109.204 ± 0.730 + j(730.314 − 204.725) ± 0.73 Ω at 100 MHz
The noise is simulated by using a FORTRAN random number generator.
Our first experiment is conducted at 10 MHz using four different noise sets, all
with the same variance. The results of the inversions with these four input data
vectors are listed in Table 11.5.
Upon computing the mean and standard deviation of this data set (including
the covariance matrix, C), and then performing the usual variance analysis, as in
A.1 A Generalized Error Analysis 255

Table 11.6 Results of Monte Carlo runs at 10 MHz

Variance from
Standard Standard covariance
Variable Average Error deviation deviation matrix
t1 5.68713 × 10−5 mm 1.6 % 3.25623 × 10−6 mm 5.7 % 6.7 %
σSi 19862.8 S/m 0.69 % 398.947 S/m 2.0 % 2.4 %

(11.20)–(11.22), we obtain the results shown in Table 11.6. The closeness of the
values in the last two columns of Table 11.6 indicates that the Monte Carlo process
was reasonably well represented with these four data sets.

Appendix

A.1 A Generalized Error Analysis

Given a set of measurements, bi , we want to estimate the errors in the parameters,

xi , which are the solution to the equation
Ax = b, (11.25)
from the error estimates for the bi . We will evaluate the covariance matrix, C, for
the xi , which will give us estimates of their errors as well as their correlations.
We assume that the matrix A is m × n with m > n. The least squares solution can
be obtained be solving the normal equations:

AT Ax = AT b. (11.26)

Solving for x gives

x = Db, (11.27)

where the matrix D is defined as

D = [AT A]−1 AT . (11.28)

The errors, Δxi , in x are then given by

m
∂ xi
Δxi = ∑ ∂ bk Δbk
k=1
m
= ∑ Dik Δbk . (11.29)
k=1
256 11 Examples of Basic Inverse Problems

Thus
m m
Δxi Δx j = ∑ ∑ Dik D jl Δbk Δbl . (11.30)
k=1 l=1

Assuming that the errors Δbk are distributed so as to give cancellation of the terms
with k = l (e.g., they have a normal distribution), the expected values of Δxi Δx j are
then given by
m
E(Δxi Δx j ) = ∑ Dik D jk E(Δbk Δbk )
k=1
m
= ∑ Dik D jk σb2k
k=1
m
= ∑ [DS]ik [DS]Tkj
k=1

= Ci j , (11.31)

where S is a diagonal matrix with Skk = σbk , the standard deviation for bk , and C is
the covariance matrix

C = [DS][DS]T = [(AT A)−1 AT S][(AT A)−1 AT S]T. (11.32)

When S = σmeas I, i.e., the standard deviation for all measured values is the same,
then

C = [AT A]−1 AT A[AT A]−1 σmeas

2
= [AT A]−1 σmeas
2
. (11.33)

The expected value of the error for xi is now given by


σxi = Cii (11.34)

and the correlation between xi and x j , which lies between 0 and 1, is given by
Ci j
cor(xi , x j ) =  . (11.35)
CiiC j j
Chapter 12
NLSE: Parameter-Based Inversion Algorithm

12.1 Introduction

Chapter 11 introduced us to the notion of an inverse problem and gave us some

examples of the value of this idea to the solution of realistic industrial problems.
The basic inversion algorithm described in Chap. 11 was based upon the Gauss–
Newton theory of nonlinear least-squares estimation and is called NLSE in this
book. In this chapter we will develop the mathematical background of this theory
more fully, because this algorithm will be the foundation of inverse methods and
their applications during the remainder of this book. We hope, thereby, to introduce
the reader to the application of sophisticated mathematical concepts to engineering
practice without introducing excessive mathematical sophistication.
We separate the discussion of inversion algorithms into two categories: (a)
parameter-based, in which a few parameters are used to define the problem or
anomalous region (as in Chap. 11) and (b) voxel-based, in which the anomalous
region is constructed voxel-by-voxel on a grid. NLSE is the prototype of the former
algorithm; the next volume in this series will introduce us to voxel-based algorithms.

12.2 NLSE: Nonlinear Least-Squares Parameter Estimation

12.2.1 Overview of the Algorithm: Nonlinear Least-Squares

Let
Z = g(p1 , . . . , pN , f ) , (12.1)
where p1 , . . . , pN are the N parameters of interest, and f is a control parameter at
which the impedance, Z, is measured. f can be frequency, scan-position, lift-off,
etc. It is, of course, known; it is not one of the parameters to be determined. To be
explicit during our initial discussion of the theory, we will call f “frequency.”

H.A. Sabbagh et al., Computational Electromagnetics and Model-Based Inversion, 257

Scientific Computation, DOI 10.1007/978-1-4419-8429-6 12,
© Springer Science+Business Media New York 2013
258 12 NLSE: Parameter-Based Inversion Algorithm

In order to determine p1 , . . . , pN , we measure Z at M frequencies, f1 , . . . , fM ,

where M > N:
Z1 = g(p1 , . . . , pN , f1 )
..
.
ZM = g(p1 , . . . , pN , fM ). (12.2)
The right-hand side of (12.2) is computed by applying the volume-integral code to
a model of the problem, usually at a discrete number of values of the vector, p,
forming a multidimensional interpolation grid.
Because the problem is nonlinear, we use a Gauss–Newton iteration scheme to
perform the inversion [97]. First, we decompose (12.2) into its real and imaginary
parts, thereby doubling the number of equations (we assume the p1 , . . . , pN are real).
Then we use the linear approximation to the resistance, Ri , and reactance, Xi , at the
ith frequency:
⎡ ⎤
∂ R1 ∂ R1
⎤ ⎢ ∂ p1 · · ·
⎡ (q) (q) ⎢ ∂ pN ⎥⎥
R1 (p1 , . . . , pN ) ⎢ ⎥
⎡ ⎤ ⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ∂ X1 ∂ X1 ⎥
R1 ⎢ (q) (q) ⎥ ⎢ · · · ⎥ ⎡ (q)
⎤
⎢ X1 ⎥ ⎢ X1 (p , . . . , p ) ⎥ ⎢ ∂ p ∂ ⎥ p − p
⎢ ⎥ ⎢ 1 N ⎥ ⎢ 1 p N ⎥ ⎢ 1 1
⎥
⎢ .. ⎥ ⎢ .. ⎥ ⎢ .. ⎥ ⎢ .. ⎥ , (12.3)
⎢ . ⎥≈⎢ . ⎥+⎢ . ⎥ ⎣ . ⎦
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎣ RM ⎦ ⎢ R (p(q) , . . . , p(q) ) ⎥ ⎢ ∂ RM ∂ R ⎥
pN − pN
(q)
⎢ M 1 N ⎥ ⎢
M ⎥
⎢ ⎥ ⎢ ∂ p1 · · · ∂ pN ⎥
XM ⎣ ⎦ ⎢ ⎥
(q) (q) ⎢ ⎥
XM (p1 , . . . , pN ) ⎢ ⎥
⎣ ∂ XM ∂ XM ⎦
···
∂ p1 ∂ pN (p(q) ,...,p(q) )
1 N

where the superscript (q) denotes the qth iteration and the partial derivatives are
computed numerically by the software. The left side of (12.3) is taken to be the
measured values of resistance and reactance. We rewrite (12.3) as

0 ≈ r+Jp , (12.4)

where r is the 2M-vector of residuals, J is the 2M × N Jacobian matrix of derivatives,

and p is the N-dimensional correction vector. Equation (12.4) is solved in a least-
(0) (0)
squares manner starting with an initial value, (x1 , . . . , xN ), for the vector of
unknowns, and then continuing by replacing the initial vector with the updated
(q) (q)
vector (x1 , . . . , xN ) that is obtained from (12.3), until convergence occurs.
We are interested in determining a bound for the sensitivity of the residual norm
to changes in some linear combination of the parameters. Given an ε > 0 and a unit
vector, v, the problem is to determine a sensitivity (upper) bound, σ , such that

r(x∗ + σ v) ≤ (1 + ε )r(x∗ ) . (12.5)

12.2 NLSE: Nonlinear Least-Squares Parameter Estimation 259

We will derive a first-order estimate of σ . Equation (12.5) is equivalent to

r(x∗ + σ v) − r(x∗) ≤ ε r(x∗ ) . (12.6)

The left-hand side of (12.6) can be approximated to the first order in σ by the first-
order Taylor expansion:

r(x∗ + σ v) − r(x∗) ≈ σ v · ∇r(x∗ ) , (12.7)

where ∇ is the gradient operator in N-dimensional space. We compute the gradient
of the norm of the residual vector:
! "1/2
∇r(x) = ∇ f12 (x) + f22 (x) + · · · + f2M
2
(x)

⎡ ⎤T
∂ f1 ∂ f2M
f1+ · · · + f2M
⎢ ∂ x1 ∂ x1 ⎥
1 ⎢ ⎢ .. ⎥
⎥
=
r(x) ⎢
⎣ ∂f
. ⎥
⎦
1 ∂ f2M
f1 + · · · + f2M
∂ xN ∂ xN
⎡
∂ f1 ∂ f1 ⎤
···
⎢ ∂ x1 ∂ xN ⎥
T
r(x) ⎢ ⎢ ⎥
= .. ⎥
r(x) ⎢
⎣∂f
. ⎥
2M ∂f ⎦ 2M
···
∂ x1 ∂ xN

= eT (x) · J , (12.8)
where the superscript T denotes the transpose of a matrix (or vector) and e(x) =
r(x)/r(x) is a unit vector. Thus, (12.7) becomes

r(x∗ + σ v) − r(x∗) ≈ σ eT (x) · J · v . (12.9)

The factor multiplying σ in (12.9) is the dot product of the two vectors, e(x) and
J · v. Hence, its value is less than the product of the magnitude of each vector, which
means that (12.9) becomes
r(x∗ + σ v) − r(x∗) ≤ σ J · v , (12.10)
because e(x) has unit magnitude. Upon equating the right-hand side of (12.10) to
the right-hand side of (12.6), we obtain the first-order estimate of σ :
 
r(x∗ )
σv = ε . (12.11)
J(x∗ ) · v
260 12 NLSE: Parameter-Based Inversion Algorithm

r(xi) − r(x*i )
r(x*i )

xi
−σI −σS x*i σS σI

Fig. 12.1 Showing sensitivity parameters for two system responses to xi . Response S is sensitive
to xi at x∗i , whereas response I is not

Note that if J(x∗ ) · v is small compared to r(x∗ ), then σ is large and the residual
norm is insensitive to changes in the linear combination of the parameters specified
by v. If v = ei , the ith column of the N × N identity matrix, then (12.11) produces
σi , the sensitivity bound for the ith parameter. Since σi will vary in size with the
magnitude of x∗i , it is better to compare the ratios σi /x∗i for i = 1, · · · , N before
drawing conclusions about the fitness of a solution.
The importance of these results is that we now have metrics for the inversion
process: Φ = r(x∗ ), the norm of the residual vector at the solution, tells us how
good the fit is between the model data and measured data. The smaller this number
the better, of course, but the “smallness” depends upon the experimental setup and
the accuracy of the model to fit the experiment. Heuristic judgement based on
experience will help in determining the quality of the solution for a given Φ .
The sensitivity coefficient, σ , is more subtle, but just as important. It, too, should
be small, but, again, the quality of the “smallness” will be determined by heuristics
based upon the problem. If σ is large in some sense, it suggests that the solution
is relatively independent of that parameter, so that we cannot reasonably accept the
value assigned to that parameter as being meaningful, as suggested in Fig. 12.1,
which shows a system, S, for which the system is sensitive to variable, xi , at the
solution point, x∗i , and another system, I, for which the system is insensitive to xi .
An example occurs when one uses a high-frequency excitation, with its attendant
small skin depth, to interrogate a deep-seated flaw. The flaw will be relatively
invisible to the probe at this frequency, and whatever value is given for its parameters
will be highly suspect. When this occurs we will either choose a new parameter to
characterize the flaw, or acquire data at a lower frequency.
12.2 NLSE: Nonlinear Least-Squares Parameter Estimation 261

These metrics are not available to us in the current inspection method, in which
analog instruments acquire data that are then interpreted by humans using hardware
standards. The opportunity to use these metrics is a significant advantage to the
model-based inversion paradigm that we propose in this book.
If the residual norm is relatively insensitive to changes in some linear com-
bination of the parameters, then the Jacobian matrix at the solution is nearly
rank-deficient, and it may be useful to determine a set of linearly-independent
parameters. The covariance matrix (J T J)−1 can be used for this purpose.

12.2.2 Stochastic Methods for Global Optimization

The problem defined above leads to the global optimization problem: finding the
lowest minimizer of a nonlinear function of several variables that has multiple local
minimizers. In the stochastic approach to global optimization, one applies a strictly
descent local search procedure to a subset of a sample of starting points drawn from
a uniform distribution over R, so as to find all the local minima of r that are
potentially global [58–60]. One such stochastic approach, the multilevel, single-
linkage, will, with probability one, find all relevant local minima of the objective
function with the smallest possible number of local searches [59, 61]. We do not
implement the multilevel, single-linkage approach in this book, but use a uniform
distribution of starting points (in each coordinate direction), as in the multilevel
approach. We do not reduce the sample size, however, but use the entire sample,
comprising, perhaps, 500 points in each variable. We have found that, even with four
variables, the procedure is so fast with modern machines, that it is quite efficient.
Hence, our global optimization algorithm starts by generating a uniform dis-
tribution of 500 points in each coordinate, and then immediately applying the
Gauss–Newton iteration to each of these points. The result is 500 local minima,
which are then ordered to give the smallest to largest values of the norm of the
residual. The location of the smallest of these local minima is presumed to be the
global minimum.

12.2.3 Computation of Function Values

NLSE is a post-processing feature of VIC-3D R

. VIC-3D R
is applied a priori to
compute function values at certain equi-spaced values of the parameters, and these
results are then stored in a table for interpolation. This speeds up the application of
the algorithm, though it will require a large database for interpolation, the size, of
course, depending upon the number of unknowns, and the number of precomputed
function values for each unknown. The order of the interpolator is arbitrary, but we
typically use first-order to fourth-order polynomial-splines for each variable, which
means that we need to compute the norm of the residuals at two to five values of
each variable. The derivatives in the Jacobian matrix are computed by interpolating
in this table of precomputed function values.
262 12 NLSE: Parameter-Based Inversion Algorithm

12.2.4 Application of Statistical Communication Theory

Additional insight can be gained by appealing to the statistical theory of communi-

cation [50]. Pretend that the measured impedance vector, Z = [Z1 , . . . , ZM ], of (12.2)
is a random vector, along with the parameter vector, p = [p1 , . . . , pN ]. Think of Z
as a received message over some communication channel, and p as a transmitted
message. These two random vectors have a joint probability density, PZ,p (Z, p),
which is to be maximized. That is to say, for a given Z, we want to determine that p
which maximizes PZ,p .
From probability theory, we have Bayes rule:

PZ,p = PZ (Z)P(p|Z)
= Pp (p)P(Z|p)
= Pp,Z (p, Z) , (12.12)

where PZ (Z) and Pp (p) are called a priori probability density functions, and
P(p|Z), P(Z|p) are called conditional, or a posteriori, probability density func-
tions. The variable to the right of the vertical bar in these latter functions is
called the conditioning variable; for example, P(p|Z) is called the a posteriori
probability density for p, conditioned on the fact that Z was received (or measured).
This a posteriori probability density function is the object of our interest, as we
shall now see.
From (12.12) we want to maximize
PZ (Z)P(p|Z) = Pp (p)P(Z|p) (12.13)

over p. Because PZ (Z) is independent of p, we can ignore it and maximize the a

posteriori probability density
P(p|Z) ∝ Pp (p)P(Z|p) . (12.14)

That is, we want to choose that “message,” p, that is best associated with the
“received signal,” Z.
If the measured Z of (12.2) is corrupted by “noise,” of density Pn , then we
can replace the conditional density P(Z|p) by Pn (Z1 − g1 , Z2 − g2 , . . . , ZM −
gM |p), where g1 = g(p, f1 ),. . . , gM = g(p, fM ) in (12.2). Because the noise in the
measurements is independent of the transmitted message, p, we can ignore the
conditioning variable, and replace (12.14) by
P(p|Z) ∝ Pp (p)Pn (Z − g(p)) . (12.15)
If we have no prior knowledge of p, or if all transmitted messages are a priori
equally likely, then we can ignore Pp (p) in (12.15) and work with the “likelihood
function,” PN (Z − g(p)). Maximizing the likelihood function over p is called
“maximum likelihood estimation.” We usually have some prior knowledge of p,
however, so we incorporate that knowledge in the a priori function, Pp (p).
A.1 Cramer–Rao Lower Bound 263

If p and n are jointly Gaussian processes, then we have the classical problem of
communicating a Gaussian signal in Gaussian noise, which reduces to a classical
least-squares problem. Typically, we work with the negative logarithm of the a
posteriori density in (12.15), so we need to minimize
− ln P(p|Z) = − ln Pp (p) − ln Pn (Z − g(p)) + f (Z) . (12.16)
Now, we let the M components of the noise vector, n, be statistically independent,
zero-mean, Gaussian random variables, each with variance, σn2 , and the N com-
ponents of p be statistically independent Gaussian random variables, with mean
values p, and each with variance, σ p2 . Then, upon taking the negative logarithm of
the appropriate density functions, and discarding unimportant factors, we replace
(12.16) with the objective function
1 M 1 N
Φ (p|Z) = ∑
2σn2 i=1
(Zi − g i (p)) 2
+ ∑ (pi − pi )2
2σ p2 i=1

|Z − g(p)|2 |p − p|2
= + , (12.17)
2σn2 2σ p2
which is to be minimized over p. Multiply (12.17) by 2σn2 to get the final expression
for the objective function
σn2
Φ (p|Z) = |Z − g(p)|2 + |p − p|2 , (12.18)
σ p2
where we use the same notation for the objective function.
The ratio, σ p2 /σn2 , of the variances is called the signal-to-noise ratio and is
usually known. In the context of least-squares problems, this ratio is known as the
Levenberg–Marquardt parameter, and in mathematical inverse theory it is called the
Tichonov–Miller parameter.
Setting the value of the LM-parameter determines the certainty with which we
choose to assert the a priori constraint on p. If σ p is very small, then the LM-
parameter is large, and we are more certain to impose the constraint. Unless we
know the variances, we cannot determine the LM-parameter at the outset. There are
numerical techniques, such as ridge regression [53] and cross-validation [62], that
can be useful in selecting the LM-parameter.

Appendix

A.1 Cramer–Rao Lower Bound

The Cramer–Rao lower bound (CRLB), being the lower bound on the variance of
any unbiased estimator, plays a role in statistical estimation theory [64, 65] that is
similar to the sensitivity (upper) bound, σ , of (12.11). As with σ , the CRLB finds
application in electromagnetic scattering and inverse problems [66].
264 12 NLSE: Parameter-Based Inversion Algorithm

A.1.1 Inverse Method Quality Metrics

Given the potential of inverse methods, it is important to develop a rigorous method

for quantifying the performance and reliability of inversion schemes [105, 107].
Although empirical studies provide the means for evaluating the quality of NDE
techniques incorporating inverse methods, opportunities also exist with inverse
methods to use the model calculations with quantitative measures to evaluate key
estimation performance metrics without considerable experimental burden.
In estimation theory, the Cramer–Rao Lower Bound (CRLB) provides the
minimum variance that can be expected for an unbiased estimator of a set of
unknown parameters. In other words, the CRLB provides a way of quantifying
the inversion algorithm performance. For Gaussian noise, there is a simple inverse
relationship between the CRLB and the Fisher information [65]:
* + ! " ! "
var θ̂i = Cθ̂ ii ≥ I −1 (θ ii , (12.19)
where C is the covariance matrix, the Fisher information is defined as
 2 
∂ ln f (Z; θ )
I (θ )ii = −E , (12.20)
∂ θi ∂ θ j
θ is the parameter being estimated, and Z is the measurement vector. Fisher
information represents the amount of information contained in a measurement and
depends on the derivatives of the likelihood function which is based on the forward
model and the noise parameters. The variance in a measurement is inversely related
to the amount of information contained in the measurement, so it is not a surprise
that (12.19) shows that the variance in the measurement is greater than or equal to
the inverse of the Fisher information matrix. In eddy-current NDE, the measurement
is often the real and imaginary component of the impedance, Z = [R, X], and the
Fisher information becomes a square matrix with dimensions equal to the number
of parameters being estimated.
The covariance matrix can be evaluated as a performance metric for inverse
methods. First, the diagonal terms of the covariance matrix (the CRLB variances)
provide a metric of sensitivity of a parameter estimated using inverse methods to
measurement variation. Second, the off-diagonal terms represent the interdepen-
dence between select parameters being estimated to measurement variation. The
corresponding metric is the correlation coefficient given by
Ci j
ρi, j =  . (12.21)
CiiC j j
These metrics can be used with parametric studies involving frequency or other
probe parameters to optimize the NDE system design. As a general design rule
for inverse methods, it is desirable to minimize the sensitivity to variation (the
CRLB variances) and to have the correlation coefficient between the parameters
being estimated approach zero.
A.1 Cramer–Rao Lower Bound 265

Another tool used in numerical linear algebra for sensitivity analysis is singular
value decomposition (SVD). SVD essentially provides a measure of sensitivity
of measurements to perturbations in the unknown parameters [106]. To evaluate
the sensitivity of an inverse problem for a set of measurements to changes in fit
parameters, SVD can be applied to the Jacobian matrix where
⎡ ⎤
∂ Z1 ∂ Z1
···
⎢ ∂ θ1 ∂ θn ⎥
⎢ . ⎥
J=⎢
⎢ ..
.. .
. .. ⎥ = UΣ V .
⎥ (12.22)
⎣ ∂Z ∂ Zm ⎦
m
···
∂ θ1 ∂ θn

The condition number (CN) of the matrix is defined as the ratio of the largest and
smallest singular values resulting from SVD. For inversion, CN has been used to
quantify the well-posedness of the inverse problem for select parameters. The ability
to estimate parameters independently increases as the condition number approaches
unity. It should be noted that SVD does not incorporate noise; it depends only on the
noiseless relationship between the measurement output and the parameter changes.

A.1.2 Optimizing Layer Estimation Using Metrics

An inversion experiment is revisited [55] for the purpose of demonstrating estima-

tion theory metrics [107]. In this experiment, the thickness of an AISI-304 stainless
steel plate and probe lift-off were estimated. A thickness and lift-off model, similar
to the one shown in Chap. 11, was used to solve the forward problem. The estimation
procedure is represented in (12.23). The left side is the measured impedance, the
Jacobian is simply the derivative information from the forward model, and the
thickness and lift-off parameters are updated until this equation converges,
⎡ ⎤
∂R ∂R
    ⎢ ∂t ∂l ⎥  
R( f ,t, l) R( f ,t0 , l0 ) ⎢ ⎥ t − t0
≈ +⎢ ⎥ . (12.23)
X( f ,t, l) X( f ,t0 , l0 ) ⎣ ∂X ∂X ⎦ l − l0
∂t ∂ l t0 ,l0

Four scenarios in particular are investigated. Impedance values were generated

for combinations of lift-off values of 0.75 and 1.5 mm and a plate thickness values
of 1.0 mm and 2.0 mm with Gaussian noise of 1 % of the impedance value added
as shown in Fig. 12.2a. For each of these measurements, the NLSE algorithm is
applied to estimate the thickness and lift-off simultaneously. Figure 12.2b shows the
inversion results in the parameter space. Note that for high lift-off, visual inspection
indicates the variance in the estimation is much greater for lift-off and likewise for
the thicker plate, the variance of the estimation of thickness is greater.
266 12 NLSE: Parameter-Based Inversion Algorithm

Fig. 12.2 (a) Distribution of source data in impedance plane, and (b) corresponding estimated
values in lift-off-thickness parameter space

The calculations required for the CRLB involve taking numerical derivatives of
the impedance changes with respect to the parameter changes from the forward
model. These calculations thus require far less computational expense with respect
to Monte-Carlo simulation. Following (12.20), the Fisher information for this
particular case is given by:
 2 + J2

J11 J12 J11 + J22J21
I= 21 . (12.24)
J11 J12 + J21J22 2 + J2
J12 22

The covariance matrix is then calculated from the Fisher information (by (12.19)):

C = σ 2 I−1 . (12.25)
The Jacobian is also decomposed into its singular values and singular vectors in the
form of the right-hand side of (12.22). The ratio of the smallest to largest singular
values provides the condition number.
Figure 12.3 shows the CRLB of the estimation of the thickness and lift-off of a
1 mm thick plate and 1 mm lift-off for multiple frequencies. The agreement between
the CRLB and the Monte-Carlo approach is quite good. This analysis demonstrates
that there is an optimal frequency to achieve highest accuracy in the estimation of
thickness. Estimating conductivity and thickness simultaneously is typically more
ill-conditioned than estimating thickness and lift-off simultaneously. The CRLB
for conductivity and thickness estimation along with the condition number and
correlation number as a function of frequency are all displayed in Fig. 12.4. The
behavior of the CRLB as a function of frequency for estimating conductivity and
thickness simultaneously follows a similar trend and this is expected since the
impedance changes due to conductivity and thickness are similar. The condition
number reaches a maximum around 95 kHz which implies that selectivity is good
and the correlation is zero at this frequency which further confirms that point.
A.1 Cramer–Rao Lower Bound 267

Fig. 12.3 Comparison of variance with varying frequency using CRLB and Monte Carlo methods
for estimating (a) lift-off and (b) thickness, respectively

Fig. 12.4 Comparison of inversion metrics with varying frequency: (a) CRLB variance for
thickness and conductivity estimation and (b) correlation and condition number

A.1.3 Two Examples from Chap. 11

Figure 11.5 of Chap. 11 shows the variation with frequency of the derivative of
impedance with respect to thickness for Test Case 1 (shown in Fig. 11.4). There is a
relatively broad peak in this derivative extending from roughly 20 to 50 kHz. Thus,
we would expect that the optimum range of frequencies to be used for determining
the thickness should lie in this range and that was essentially confirmed when
frequencies in the range of 20–40 kHz were found to give good results.
Figure 12.5(top) illustrates the CRLB thickness response for this same test case,
and we see that there is a relatively broad minimum centered near 50 kHz, but
generally including the same frequency range chosen in executing the test case. This
confirms the consistency between the CRLB result and the information contained in
the first derivative.
268 12 NLSE: Parameter-Based Inversion Algorithm

Fig. 12.5 Comparison of variance with varying frequency for Test Case 1 in Chap. 11, using
CRLB and Monte Carlo methods for estimating (top) thickness and (bottom) lift-off, respectively

As for the lift-off problem in Test Case 1, we see from Fig. 11.6 that the derivative
with respect to lift-off continues to increase with frequency even beyond 100 kHz,
which corresponds to what we might expect for the CRLB associated with lift-off,
A.1 Cramer–Rao Lower Bound 269

Fig. 12.6 Comparison of variance with varying frequency for Test Case 2 in Chap. 11, using
CRLB and Monte Carlo methods for estimating (top) thickness and (bottom) lift-off, respectively

as shown in the bottom part of Fig. 12.5. In the latter figure, we see that the CRLB
never does achieve a true minimum, even out to 150 kHz, but the decrease is leveling
off at this frequency.
270 12 NLSE: Parameter-Based Inversion Algorithm

Similar results are obtained for Test Case 2 of Chap. 11, except that the
frequencies are much lower (see Figs. 11.9 and 11.10). The derivative with respect to
thickness for a 1.0 mm-thick brass plate has a maximum in the vicinity of 2 kHz, and
the derivative with respect to lift-off peaks at approximately 10 kHz, for a nominal
lift-off of 2.0 mm. These results are consistent with the CRLB results shown in
Fig. 12.6. In particular, note that the CRLB for lift-off is virtually flat beyond 8 kHz,
which corresponds with the levelling off of the peak derivative shown in Fig. 11.10.

A.2 Selected Bibliography of Inverse Problems in

Eddy-Current NDE

• H. A. Sabbagh and L. D. Sabbagh, Development of a system to invert eddy-

current data and reconstruct flaws, Final Report: Contract No. N60921-81-C-
0302, Naval Surface Weapons Center, White Oak Labs, Silver Springs, MD
(1982).
• H. A. Sabbagh and L. D. Sabbagh, Inversion of eddy-current data and the
reconstruction of flaws using multifrequencies, Final Report: Contract No.
N60921-82-C-0139, Naval Surface Weapons Center, White Oak Labs, Silver
Springs, MD (1983).
• B. A. Auld, G. Mcfetridge, M. Riaziat, and S. Jefferies, Improved probe-
flaw interaction modeling, inversion processing, and surface roughness clutter,
Review of Progress in QNDE, Vol 4A (1985).
• B. A. Auld, S. Jefferies, J. C. Moulder, and J. C. Gerlitz, Semi-elliptical surface
flaw EC interaction and inversion: theory, Review of Progress in QNDE, Vol 5A
(1986).
• H. A. Sabbagh and L. D. Sabbagh, An eddy-current model for three-dimensional
inversion, IEEE Transactions on Magnetics (1986).
• H. A. Sabbagh, L. D. Sabbagh, Verification of an eddy-current flaw inversion
algorithm, IEEE Transactions on Magnetics (1986).
• L. Udpa and S. Udpa, Solution of inverse problems in eddy-current nondestruc-
tive evaluation (NDE), Journal of Nondestructive Evaluation (1988).
• L. D. Sabbagh, and H. A. Sabbagh, Eddy-current modeling and flaw reconstruc-
tion, Journal of Nondestructive Evaluation (1988).
• H. A. Sabbagh, L. D. Sabbagh, and T. M. Roberts, An eddy-current model
and algorithm for three-dimensional nondestructive evaluation of advanced
composites, IEEE Transaction on Magnetics (1988).
• S. M. Nair and J. H. Rose, Reconstruction of three-dimensional conductivity
variations from eddy current (electromagnetic induction) data, Inverse Problems
(1990).
• J. C. Moulder, E. Uzal, and J. H. Rose, Thickness and conductivity of metallic
layers from eddy current measurements, Rev. Sci. Instrum (1992).
A.2 Selected Bibliography of Inverse Problems in Eddy-Current NDE 271

• S. J. Norton and J. R. Bowler, Theory of eddy current inversion, J. Appl. Phys

(1993).
• O. Baltzersen, Model-based inversion of plate thickness and lift-off from eddy
current probe coil measurments, Materials Evaluation (1993).
• N. J. Goldfine, Magnetometers for improved materials characterization in
aerospace applications, Materials evaluation (1993).
• H. A. Sabbagh and R. G. Lautzenheiser, Inverse problems in electromagnetic
nondestructive evaluation, International Journal of Applied Electromagnetics in
Materials (1993).
• J. R. Bowler, S. Norton, D. J. Harrison, Eddy-current interaction with an ideal
crack. II. The inverse problem, J. Appl. Phys. (1994).
• S. K. Burke, Eddy-current inversion in the thin-skin limit: Determination of depth
and opening for a long crack, J. Appl. Phys. (1994).
• C. Tai, J. H. Rose, and J. C. Moulder, Thickness and conductivity of metallic
layers from pulsed eddy-current measurments, Rev. Sci. Instrum. (1996).
• L. Udpa and S. Udpa, Application of signal processing and pattern recognition
techniques to inverse problems in NDE, Int J. of Applied Electromagnetics and
Mechanics (1997).
• Z. Chen and K. Miya, ECT inversion using a knowledge-based forward solver,
Journal of Nondestructive Evaluation (1998).
• H. T. Banks et al, Nondestructive evaluation using a reduced-order computational
methodology, Inverse Problems (2000).
• H. T. Banks et al, Real time computational algorithms for eddy-current-based
damage detection, Inverse Problems (2002).
• W. Yin, S. J. Dickinson, and A. J. Peyton, Imaging the continuous conductivity
profile within layered metal structures using inductance spectroscopy, IEEE
Sensors Journal (2005).
• W. Yin and A. J. Peyton, Thickness measurement of non-magnetic plates using
multi-frequency eddy current sensors, NDT&E International (2007).
Chapter 13
Robust Statistical Estimators

13.1 Introduction

“Robust” estimators are resistant to outliers in data, contrary to the usual classical
least-squares estimator such as NLSE. We will describe two robust estimators in
this chapter and give an example of the application of them to pre-processing some
experimental data. Both of the robust estimators are taken from [63], and we closely
follow those authors in our presentation.

13.2 Robust Estimators

The mathematical regression problem that we are addressing is based on the linear
model given by yi = θ xi + ei for i = 1, . . . , N. yi are the data points corresponding to
the dependent variable, xi the data points corresponding to the independent variable,
ei the error term, and θ the parameter to be determined.
Central to the notion of robust estimation is the concept of the “breakdown point,”
which we define following [63]: Take any sample X of N data points (xi , yi ), and any
estimator T of the parameter θ . Let β (M, T, X) be the supremum of T (X  )− T (X)
for all corrupted samples X  , where any M of the original points of X are replaced
by arbitrary values. Then the breakdown point of T at X is defined as
 
M
εN∗ (T, X) = min ; β (M, T, X) is infinite . (13.1)
N

In words, εN∗ is the smallest fraction of contaminated data that can cause the
estimator to take on values arbitrarily far from T (X). For least-squares, only one
bad observation is needed to cause breakdown, so that εN∗ (T, X) = 1/N, which tends
to 0 % when the sample size N becomes large.

H.A. Sabbagh et al., Computational Electromagnetics and Model-Based Inversion, 273

Scientific Computation, DOI 10.1007/978-1-4419-8429-6 13,
© Springer Science+Business Media New York 2013
274 13 Robust Statistical Estimators

The best breakdown point that we can expect is 50 %; otherwise it becomes

impossible to distinguish the “good” data from the “bad.” During Phase I we studied
two estimators that possess a 50 % breakdown point: the least median of squares
(LMS) and a scale (“S-estimator”) estimator [63].

13.3 Least Median of Squares Estimator

Definition 13.1. Let (x1 , y1 ), . . . , (xn , yn ) be a sample of regression data. For each
slope, θ , we obtain residuals ri (θ ) = yi − θ xi . Then the LMS-estimator θ̂ that is
defined by

min median(ri2 ) , (13.2)

θ

possesses a 50 % breakdown point, is affine equivariant (that is, it is independent of

the choice of the coordinate axes of the xi ) and has an asymptotic convergence rate
of n−1/3 .
We will define the asymptotic convergence rate shortly.
The LMS estimator is easy to code, and runs quickly. Nevertheless, we inves-
tigated the following scale-estimator that has a 50 % breakdown point, is affine
equivariant, and has a better asymptotic convergence rate of n−1/2.

13.4 Scale (S) Estimator

Our one-dimensional estimator of scale (“S-estimator”) is defined by a function ρ

satisfying [63]:
(R1) ρ is symmetric, continuously differentiable and ρ (0) = 0.
(R2) there exists c > 0 such that ρ is strictly increasing on [0, c] and is constant on
[c, ∞).
Definition 13.2. For any sample {r1 , . . . , rn } of real numbers, we define the scale
estimate s(r1 , . . . , rn ) as the solution of

1 n
∑ ρ (ri /s) = K ,
n i=1
(13.3)

where K is taken to be EΦ [ρ ], and Φ is the standard normal distribution. If there is

more than one solution to (13.3), then we put s(r1 , . . . , rn ) equal to the supremum of
the set of solutions; if there is no solution to (13.3), then we put s(r1 , . . . , rn ) = 0.
13.4 Scale (S) Estimator 275

Definition 13.3. Let (x1 , y1 ), . . . , (xn , yn ) be a sample of regression data. For each
slope, θ , we obtain residuals ri (θ ) = yi − θ xi , from which we can calculate the
dispersion s(ri (θ ), . . . , rn (θ )) by (13.3), where ρ satisfies (R1) and (R2). Then the
S-estimator θ̂ is defined by

min s(r1 (θ ), . . . , rn (θ )) , (13.4)

θ

and the final scale parameter is

σ̂ = s(r1 (θ̂ ), . . . , rn (θ̂ )). (13.5)

An example of a ρ -function for (13.3) is [63]

x2 x4 x6
ρ (x) = − 2 + 4 for |x| ≤ c
2 2c 6c
(13.6)
c2
= for |x| ≥ c.
6

In order to achieve a 50 % breakdown point using the ρ -function of (13.6), we

require that
E Φ [ρ ] 1
= , (13.7)
ρ (c) 2
which can be achieved by letting c = 1.547. In this case, K = c2 /12 = 0.1994341,
in (13.3).
The function defined in (13.6) should be compared with Huber’s classic function
x2
ρH (x) = for |x| ≤ k
2 (13.8)
= |x| for |x| ≥ k.
The Huber function produces an L2 norm (least-squares) for x ≤ k, and an L1 norm
otherwise. Though the Huber function gives rise to an estimator that is robust against
outliers in y, it still possesses a breakdown point of 0 % in the presence of outliers
in x, which are called “leverage points,” and have a significant effect on the fit [63].
The following is an example of the resistance of this S-estimator to “bad data”:
suppose we have 100 data points, (xi , yi ), and 51 of them lie exactly on a straight
line. Then the first 51 residuals vanish in (13.3), which means that ρ (ri /s) = 0, i =
1, . . . , 51. Hence, the largest that the left-hand side of (13.3) can be (49/100)(c2/6),
which is less than K. Hence, there is no solution to (13.3), and we set s = 0. This
means that the estimator chooses the slope of the regression line to be equal to
the slope of the “exact data line.” The estimator is impervious to the remaining 49
points.
276 13 Robust Statistical Estimators

Fig. 13.1 The function 0.2

(13.3) produced by a typical
data set from an inversion 0.15
problem
0.1

Estimator function
0.05

-0.05

-0.1

-0.15

-0.2
0 2 4 6 8 10 12
S (scale)

We sketch, in Fig. 13.1, the estimator function, 1n ∑ni=1 ρ (ri /s) − K, for a typical
data set.
Finally, let us say something about the asymptotic behavior of S-estimators at
the central Gaussian model, where (xi , yi ) are independent, identically distributed
random variables satisfying

yi = xi θ0 + ei . (13.9)

xi follows some distribution H, and ei is independent of xi and distributed according

to Φ (e/σ0 ), for some σ0 > 0, where Φ is the standard normal cumulative
distribution function. Then Rousseeuw and Yohai [63] prove the following two
theorems:
Theorem 13.1 (Consistency). Let ρ be a function satisfying (R1) and (R2), with
the derivative ρ  = ψ . Assume that
(i) ψ (u)/u is nonincreasing for u > 0;
(ii) EH [|x|] < ∞, and H has a density;
(iii) Let (xi , yi ) be i.i.d. according to the model (13.9);
(iv) Let θ̂n be a solution of (13.4) for the first n points, and;
(v) Let σ̂n = s(r1 (θ̂n ), . . . , rn (θ̂n )).
Then

θ̂n → θ0 a.s.
σ̂n → σ0 a.s. (13.10)
13.5 An Application of Robust Estimation 277

Theorem 13.2 (Asymptotic Normality). Let θ0 = 0 and σ0 = 1 for simplicity. If

the conditions of Theorem 13.1 hold, and also
(vi) -ψ is differentiable in all but a finite number of points, |ψ  | is bounded, and
ψ  d Φ > 0;
(vii) EH [x2 ] = 0 and EH [|x|3 ] ≤ ∞.
Then
   
2 −1 
n 1/2
(θ̂n − θ0 ) → N 0, EH [x ] ψ d Φ /( ψ d Φ )
2 2

   
n1/2 (σ̂n − σ0 ) → N 0, (ρ (y) − K)2d Φ (y)/( yψ (y)d Φ (y))2 . (13.11)

These theorems guarantee that, while the S-estimators do not break down easily
when the data are contaminated, they continue to behave well when the data are
not contaminated, that is, when they satisfy the classical assumptions [63]. The
Asymptotic Normality Theorem defines the asymptotic convergence rate that was
alluded to earlier.

13.5 An Application of Robust Estimation

Though our main interest in robust estimators is in their application to the voxel-
based inversion algorithm that will be treated in the second volume of this series,
we will here apply the LMS-estimator to the problem of determining a “baseline”
for experimental data that will be the input to the actual inversion algorithm.
The problem is to determine a constant (the “baseline”) from a set of data that
are virtually constant, except for isolated signals due to flaws, tube supports, and
removal of the probe from the end of the tube. Our interest is in the application to
eddy-current NDE of steam generator tubes in nuclear power plants. The isolated
signals constitute “impulsive noise” and may be considered to be the outliers of the
data set. Clearly, if we averaged the data, the resulting baseline would be strongly
affected by these signals, which would hardly result in a baseline. The current
approach in the industry is simply to “eyeball” the data to produce the baseline,
but this is hardly efficient, and is still subject to errors. What is needed is a reliable
mathematical approach that will automatically produce the answer with no human
intervention or preprocessing.
The baseline estimator was exercised on a set of experimental data, supplied
by the Electric Power Research Institute (EPRI). The data consisted of a set of
approximately 12,000 measurements, taken at regular intervals along the axis of
a tube, using eddy-current instrumentation. The measurements were made at four
different frequencies, and at each frequency, the “baseline” had a different value due
to the way the instrumentation behaved. One of the post-processing procedures for
278 13 Robust Statistical Estimators

these data required that the data at the different frequencies have the same baseline
(the data at two or more different frequencies were scaled and mixed to produce a
composite signal). We used the baseline estimator to “normalize” all frequencies to
the same baseline. The figures below show the experimental data, before and after
using the baseline estimator. The results are shown for the full scan, as well as for
smaller portions of the full scan (Figs. 13.2–13.4).

Fig. 13.2 Full scan of experimental data, at four different frequencies (approximately 12,000
points), before (top) and after (bottom) any “balancing” using the baseline estimator
13.5 An Application of Robust Estimation 279

Fig. 13.3 Portion of the full scan, data points 0–1,000, of experimental data, before (top) and after
(bottom) any “balancing” using the baseline estimator
280 13 Robust Statistical Estimators

Fig. 13.4 Portion of the full scan, data points 1,000–2,000, of experimental data, before (top) and
after (bottom) any “balancing” using the baseline estimator
Chapter 14
Some Special Signal-Processing Algorithms

In this chapter we sketch some rather elegant mathematical theorems that have had a
significant impact on computational aspects of electrical engineering. We have used
them over the years in performing eddy-current inversions and believe that they will
have an expanded role to play in the future development of eddy-current NDE.

14.1 Projection Onto Convex Sets

The mathematical foundation of the theorems is the notion of projection onto convex
sets (POCS) in a Hilbert space, which is proving to be quite valuable in engineering
analysis and design [71–78]. We recall that a Hilbert space is simply a vector space
in which an inner-product is defined.1 Our entire development of volume-integral
equations is done in a Hilbert space, though we don’t emphasize that fact. A set, S ,
is called convex if for every pair of vectors, V and U that lie in S , the straight line
joining these two vectors, V + λ (U − V), 0 ≤ λ ≤ 1, also lies in S . Hyperplanes,
spheres, and cubes are simple examples of convex sets. The projection of a vector, V,
onto a convex set, S , is that point, U, on S that is closest to V. We write U = PS V,
where PS is the projection operator.
Projection operators offer a computationally convenient algorithm for deter-
mining the intersections of sets, which can be translated into solving systems of
linear and nonlinear equations. Consider Fig. 14.1, which shows two convex sets
intersecting at the single point, v∗ . If we define the composition operator2
T = P2 P1 , (14.1)
then the alternating projection scheme shown will converge to v∗ ,
for an arbitrary
starting point, v. Because of the tangency of the two boundaries, however, the
convergence will not be geometric in the vicinity of v∗ , so we use over-relaxation

1 The discussion in this section follows [73], which should be consulted for the details.
2 The order of the Pi ’s is not important.

H.A. Sabbagh et al., Computational Electromagnetics and Model-Based Inversion, 281

Scientific Computation, DOI 10.1007/978-1-4419-8429-6 14,
© Springer Science+Business Media New York 2013
282 14 Some Special Signal-Processing Algorithms

v
C1
P1v

TvT2vT3v v*

C2

Fig. 14.1 Illustrating projections onto two convex sets, and the convergence of the iterates to v∗ ,
the intersection of the sets. The starting point, v, is arbitrary

by extending the projections beyond the boundaries of the sets. This is done by
selecting appropriate “relaxation” parameters, r1 , r2 within the interval 0 < r < 2,
replacing P1 and P2 by
T1 = 1 + r1(P1 − 1)
T2 = 1 + r2(P2 − 1) , (14.2)
and then defining
T = T2 T1 , (14.3)
before proceeding with the scheme shown in Fig. 14.1. We will give in the next
section an example of this over-relaxation procedure in the context of Kaczmarz’
algorithm for solving large systems of linear equations. Following that, we will give
examples of the use of POCS for signal processing and inversions with incomplete
or noisy data.

14.2 Kaczmarz’ Algorithm and the Algebraic

Reconstruction Technique

Suppose that we want to find the minimum-norm solution of

Ax − b2 + λ 2 x2 , (14.4)
14.2 Kaczmarz’ Algorithm and the Algebraic Reconstruction Technique 283

H3 H2 H1
(1)
(2) x
x

(0)
x
(3)
x

(5)
x (4)
(6)
x
x

Fig. 14.2 Geometric interpretation of ART for three equations and two cycles of iterations. Each
equation defines a hyperplane. At each stage, the current solution is projected onto the next
hyperplane. Results after two cycles of iterations with unity relaxations are shown

where A is m× n and λ is the Levenberg–Marquardt (LM) parameter. The Kaczmarz

algorithm is an iterative method of computing this solution, in which the new point
in the sequence of iterates depends only on the most recent point and the next
equation to be satisfied. The basic idea is to satisfy each equation, or hyperplane,
by projecting the previous solution onto the next hyperplane. Figure 14.2 is an
illustration of this geometric interpretation where three equations (hyperplanes)
exist and the process has gone through two iterations. By an iteration we mean
a complete cycle through the algorithm. In this example, rk = 1, which is the
case of unity relaxation. If rk > 1, then we have the case of overrelaxation, and
the projections “overshoot” the hyperplanes; if rk < 1, then we have the case of
underrelaxation, and the projections “undershoot” the hyperplanes. We note that
the relaxation parameter can be changed at each iteration. This may help the
convergence of this algorithm when used on certain problems.
The theoretical basis for this method is that if x(k) and z(k) are sequences
generated by the following algorithm, then x(k) converges to the minimum of (14.4).
We now state the algorithm. Initialization: z(0) is arbitrary, and

λ x(0) = AT z(0) ,
284 14 Some Special Signal-Processing Algorithms

step:

z(k+1) = z(k) + λ d (k) ei

x(k+1) = x(k) + d (k) ai ,

where
(k)
bi − λ zi − (ai , x(k) )
d (k) = rk ,
λ 2 + ai2

and ai is the ith row of A, ei is the ith row of the identity matrix I, i = (k mod m) + 1,
and 0 ≤ lim inf rk ≤ lim sup rk ≤ 2. d (k) is the term which represents a generalization
of projections onto hyperplanes, z(k) is a dual variable, and rk is a relaxation
parameter. For each iteration of the algorithm, 1 ≤ i ≤ m. The method converges
for a given range of LM parameters and a given range of relaxation parameters. It is
clear from the algorithm that the role of the LM parameter is to stabilize d (k) when
ai 2 is very small.
In voxel-based inversion algorithms (which will be the subject of volume 2 of
this series) we acquire data via an array of sensors, each of which can be excited to
serve as a transmitter as well as a receiver. This yields the system (14.5)

∑ ELMJ (v) · JLMJ(v)

(1)
Z1 (v) =
LMJ

∑ ELMJ (v) · JLMJ(v)

(2)
Z2 (v) =
LMJ

..
.

∑ ELMJs (v) · JLMJ(v) ,

(N )
ZNs (v) = (14.5)
LMJ
(n)
where ELMJ (v) is the “incident” field produced by the sensor when it is in its nth scan
position, during the vth view, Zn (v) is the measured impedance data, and JLMJ(v) is
the unknown anomalous current within the flaw that is to be determined.
The minimum-norm solution of (14.5) is given by
⎡ ⎤
Z1 (v)
⎢ ⎥
J̃LMJ (v) = M † (v) ⎣ ... ⎦, (14.6)
ZNs (v)
where M † (v) is the pseudoinverse of the matrix in (14.5). One can use the QR-
decomposition of linear algebra to determine M † (v). This works well for fairly
small problems, but for large, indeed huge, problems Kaczmarz’ algorithm would
be the choice [67–69].
14.3 Analytic Continuation with Constraints 285

We have used ART successfully in [70], in which it reconstructed a number of

flaws in a heat-exchanger tube using a linear model consisting of 1,024 equations
in 1,100 unknowns. In those numerical experiments 0.01 ≤ λ ≤ 0.02, and the
relaxation parameter was fixed to be 0.25 throughout the iterations, during which
the equations were accessed in the order in which they appear in the given system.
In [69] it is claimed that by a careful adjustment of the order in which the equations
are accessed during the reconstruction procedure, and by adjusting the relaxation
parameter at each iteration, ART can produce high-quality reconstructions with
excellent computational efficiency.
The algorithm is robust, and the additional features of the Levenberg–Marquardt
parameter and relaxation parameter give it added strength as an inversion algorithm.
Its computational advantages derive from the facts that no changes are made to the
original matrix, no operations are performed on the entire matrix, only one row of
the matrix is required in a single step, and only x(k) is required for the calculation of
x(k+1) during that step.
Kaczmarz proposed his algorithm in 1937 [79], and it was further developed by
Tanabe in 1971 [80]. G.N. Hounsfield, in his 1972 patent disclosing the invention
of the X-ray computed tomographic scanner [81], used this algorithm, which in this
context is called the “algebraic reconstruction technique (ART)” to distinguish it
from Fourier, Hilbert and Radon transform reconstruction algorithms, or from the
convolution-backprojection algorithm [82].

14.3 Analytic Continuation with Constraints

A significant advantage to using POCS for solving inverse problems or performing

signal-processing is that it easily accommodates the imposition of constraints on the
solution, as long as those constraints can be expressed geometrically as convex sets.
This often is the only way to get a unique solution to an inverse problem.
A typical problem is to reconstruct a function, such as the conductivity of
a flawed region, g(x, y) = σ f (x, y)/σh − 1, given its Fourier spectrum, G(kx , ky ),
restricted to low spatial frequencies. The physical process of field-evanescence acts
as a low-pass filter, thereby restricting the sensed magnetic field at high-spatial fre-
quencies to lie below the system noise level. High-frequency components, however,
are necessary for high-resolution reconstructions. Therefore, it is necessary to try
to regain these missing frequencies in order to achieve the desired resolution. That
this attempt at “super-resolution” is feasible and practical rests upon the notion of
“analytic continuation” [83, 84]. We will state two theorems that are relevant (see
[83, p. 133]):
Theorem 14.1. The two-dimensional Fourier transform of a spatially bounded
function is an analytic function in the (kx , ky )-plane.
286 14 Some Special Signal-Processing Algorithms

Theorem 14.2. If any analytic function in the (kx , ky )-plane is known exactly in
an arbitrarily small (nonzero) region of that plane, then the entire function can be
found (uniquely) by means of analytic continuation.
In order to use these theorems we recall that the support of g(x, y) is bounded.
Thus, by Theorem 14.1 G(kx , ky ) is analytic in the (kx , ky )-plane. If we have
only limited information about G, say only its values at low spatial-frequencies,
Theorem 14.2 tells us that we can uniquely extend G to the entire (kx , ky )-plane.
Once we have continued G to the entire (kx , ky )-plane, we can take the inverse
Fourier transform to recover g. The following algorithm is an application of POCS
that performs analytic continuation under the conditions just stated:
ALGORITHM
Let the function G(kx , ky ) be given over a prescribed region L , and let F be the
Fourier transform. Then starting with:
f0 (x, y) = F −1 [G(kx , ky )];
r = 0;
REPEAT
(1)
fr = P1 fr ;
(2) (1)
fr = P3 fr ;
(1) (2)
Fr+1 = F [ fr ];
(1)
Fr+1 = P2 Fr+1 ;
−1
fr+1 = F [Fr+1 ];
r = r + 1;
UNTIL CONVERGENCE OCCURS.
The important operations in the Algorithm, in addition to the Fourier and inverse
Fourier transforms, are the various projection operators, Pi :

f , (x, y) ∈ S , S = support of f ,
P1 f =
0, otherwise.

G(kx , ky ), (kx , ky ) ∈ L ,
P2 F =
F(kx , ky ), (kx , ky ) ∈ L , where F(kx , ky ) = F [ f (x, y)].
⎧
⎨ −1, if f (x, y) < −1
P3 f = f (x, y), if − 1 ≤ f (x, y) ≤ 0 (14.7)
⎩
0, if f (x, y) > 0.
From its definition, −1 ≤ g(x, y) ≤ 0, which is the reason for the definition of
P3 . Hence, the algorithm successively applies the known properties of the function,
g(x, y). The analyses of [72,73,78] prove that the algorithm converges. In Volume 2
we will apply this algorithm to a problem of flaw reconstruction in graphite-epoxy
advanced composites. Other applications of the algorithm can be found in [85–91].
14.4 Reconstructing Network Functions 287

R3
(PH2,PD2)3
R2

(PH2,PD2)2
R1

(PH2,PD2)1

Rmeas

X
Xmeas X3 X2 X1

Fig. 14.3 Illustrating an iterative method, based on the idea of projections onto convex sets
(POCS), for improving resistance data during the inversion process. It is likely that the points
(R1 , X1 ), (R2 , X2 ), (R3 , X3 ) lie on another convex set (or simply a locally-convex set) that
approaches X = Xmeas , but this has not been proved

14.4 Reconstructing Network Functions

In electric circuit theory a network function is the ratio of the Fourier or Laplace
transform of a response to that of an excitation, when the network is initially relaxed.
Typical network functions are driving-point and transfer impedances or admittances.
The inverse problems described in this text use impedances as input data, so it is
important to know that the quality of these impedance data is sufficient to allow
a good reconstruction. It is abundantly clear throughout this text that measured
impedance data usually have a reasonably good reactance, but the resistance is much
noisier, generally because it is much smaller. Even though the resistance may be
small compared to the reactance, it still plays a crucial role in determining the shape
and location of a flaw. Thus, it is important to develop a means of “reconstructing”
the resistance from the known reactance data in order to complete the inversion
process. The POCS concept, together with NLSE, allows us to do that, as illustrated
in Fig. 14.3.
The nomenclature in Fig. 14.3 refers to a corrosion topology problem that will
be described in Chaps. 16 and 17. PH2 and PD2 refer, respectively, to the height
and diameter of a “pillbox” model of a pit in layer 2 of a double-layer system.
The input data for this problem are the impedances obtained from a 245-point
scan of a probe past the pit. Each point on the X-axis, therefore, corresponds to
a 245-vector of reactance data, with Xmeas corresponding to the original measured
288 14 Some Special Signal-Processing Algorithms

data. A similar interpretation holds for the R-axis and Rmeas . The composite (X, R)
space can be called a state-vector space, or simply a function space. The hyperplane
X = Xmeas is a convex set in this space (any two points in the hyperplane can be
joined by a straight line lying within the plane), and the arrowed trajectories intersect
it orthogonally, so that part of the algorithm is a “projection onto a convex set,” but
we have not demonstrated that the other trajectories are projections onto a convex
set. Nevertheless, because of the similarity between this figure and Fig. 14.1, we
will refer to the iterative algorithm of Fig. 14.3 as the “POCS” algorithm later in the
book. If we must distinguish between Figs. 14.1 and 14.3 to avoid confusion, we
will do so at the appropriate time.
The algorithm proceeds iteratively through a series of steps that sequentially
calls NLSE to perform an inversion, VIC-3D R
to perform a direct calculation, and
an orthogonal projection to return the impedance computed by VIC-3D R
back to
the straight line, X = Xmeas , thus assuring that all further input data to NLSE will
always have X = Xmeas as its reactance data. This ensures that the algorithm will be
grounded in the more reliable component of the original measured data.
Starting with the original measured data, (Rmeas , Xmeas ), NLSE0 computes the
first approximation to the pit dimensions, (PH2, PD2)1 , which is then fed to
VIC-3D R
to produce the first update to the impedance data, (R1 , X1 ). X1 , however,
is projected (orthogonally) back onto the hyperplane, X = Xmeas , so that the new
impedance data for NLSE1 are (R1 , Xmeas ). The output of NLSE1 is (PH2, PD2)2 ,
which is then fed into VIC-3D R
to produce (R2 , X2 ), and then (R2 , Xmeas ) after
projection. The process is continued until some stopping point. We have chosen
the stopping point to be when the β -scale factor in β × Zk+1 − Zk = 0 is unity,
or very close. This criterion states that there is not much difference (in the least-
squares sense) between Zk+1 = (Rk+1 , Xk+1 ) and Zk = (Rk , Xk ), so there is no sense
in running another NLSE step. A by-product of this algorithm for reconstructing the
measured resistance is the reconstruction of the pit, which is what we wanted in the
first place, and for that reason the algorithm will play an important role in Part III of
the text.
Chapter 15
Preprocessing Data and Transformation
of Signal Vectors

15.1 Clutter Modeling and Rejection

By “clutter” we mean a large background signal of known origin, such as probe lift-
off, or a large background signal due to systematic errors of unknown origin, such
as a variation in the material properties of the sample, or variations in the subsurface
structure. In any case, clutter is usually more pernicious to the reliable detection and
reconstruction of a flaw than random (electronic) noise that can be often eliminated
by averaging. Hence, the ability to model clutter and reject it is paramount to the
successful application of inverse methods and is the subject of the first part of this
chapter.

15.1.1 A Benchmark Validation Experiment

Data were taken on an Inconel 600 plate, which is nonmagnetic (relative permeabil-
ity, μ = 1), and has a conductivity of σ = 9.86 × 105 S/m. A notch is introduced
into the “outer diameter” of the plate, which means the surface opposite to the
probe. A coil is excited at 100 kHz and is scanned transversely to the slot, such that
impedances are measured at 250 points, each separated by 4 mils (see Fig. 15.1).
The OD of the coil is 0.112 inch, the ID is 0.034 inch, the height is 0.048 inch, and
there are 131 turns in the coil. The experiment is run with a coil lift-off of 0.015
inch above the plate.
The impedance data were measured by Warren Junker of the Westinghouse
Research Labs, using a Hewlett-Packard 4194A impedance analyzer, and recorded
to six significant digits, which implies a dynamic range of 120 dB. The “host-
only” impedance was measured to be Zhost = 4.05115 + j10.8770 Ω. When Zhost
is subtracted from all 250 data points, the result is the “anomaly signal” shown
in Fig. 15.2. Note the significant “clutter” in the resistance data, produced by a
systematic error.

H.A. Sabbagh et al., Computational Electromagnetics and Model-Based Inversion, 289

Scientific Computation, DOI 10.1007/978-1-4419-8429-6 15,
© Springer Science+Business Media New York 2013
290 15 Preprocessing Data and Transformation of Signal Vectors

Coil Scan Direction

Outer Diameter (OD)

Inconel Plate (Host) Notch (Flaw)

Fig. 15.1 Illustrating the setup for the experiment and model calculations

Junker Data Junker Data

0.005 0.007
aod tslot4.100 aod tslot4.100

0.004 0.006

0.005
0.003
Resistance (Ohms)

Reactance (Ohms)
0.004
0.002
0.003
0.001
0.002
0
0.001

-0.001 0

-0.002 −0.001
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5 −0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5
Probe Position (in) Probe Position (in)

Fig. 15.2 Original data supplied by Warren Junker of the Westinghouse Research Labs. Note the
significant “clutter” in the resistance data, produced by a systematic error

In order to determine the “flaw signal,” we subtract the clutter from the anomaly
signal. This is done by first modeling the clutter with a simple mathematical
expression, in this case by the piecewise linear functions:
0.0045 − 0.00225 × (pos+ 0.5) if pos ≤ −0.1
Rclutter =
0.0036 + 0.0066 × (pos+ 0.1) otherwise
Xclutter = −0.0005 + 0.0005 × (pos+ 0.5), (15.1)
where “pos” designates the probe position. When the clutter signal of (15.1) is
subtracted from the data of Fig. 15.2, we get the “processed” data of Fig. 15.3.
It is now clear that the original data resulted from a scan over a notch-type flaw.
We assume that the flaw is a rectangular paralleliped notch that breaks the back
surface of the workpiece, i.e., the surface away from the probe. The problem then
is to determine the size of the notch, namely its length, width, and height. In order
to do that, we apply the processed data of Fig. 15.3 as the input to NLSE (recall
Chap. 12). The three parameters are, of course, the length, width, and height of the
notch. The results are: length = 0.2277 inch, width = 0.0049 inch, and height =
0.025 inch. When we repeat the inversion experiment, this time holding the height
of the flaw to 0.024 inch, we get length = 0.2424, and width = 0.0050. In the first
case (height = 0.025) the final norm of the residuals is 0.8112711 × 10−2, whereas
15.1 Clutter Modeling and Rejection 291

Junker Data Processed Junker Data Processed

0.001 0.007
aod tslot4.100 aod tslot4.100
0.006
0
0.005
Resistance (Ohms)

Reactance (Ohms)
−0.001
0.004

−0.002 0.003

0.002
−0.003
0.001
−0.004
0

−0.005 −0.001
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5 −0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5
Probe Position (in) Probe Position (in)

Fig. 15.3 Showing the original data with the clutter removed

Results of NLSE Inversion

0.001
Processed Data
Results of NLSE Inversion
NLSE:Depth = 0.025 0.007
NLSE:Depth = 0.024 Processed Data
NLSE:Depth = 0.025
0 0.006 NLSE:Depth = 0.024

0.005
Resistance (Ohms)

Reactance (Ohms)

−0.001
0.004

−0.002 0.003

0.002
−0.003
0.001

−0.004 0

−0.001
−0.005 −0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5

Probe Position (in) Probe Position (in)

Fig. 15.4 Showing the effects of inverting the processed original data to reconstruct the flaw

with the height = 0.024 the norm is equal to 0.8522527 × 10−2, which is 5 % larger.
The effect of this is discussed next.
When we use VIC-3D R
in the “forward mode” to compute the scanned
impedances, given notches of the size just determined, and compare these results
with the input data of Fig. 15.3, we get the results shown in Fig. 15.4. Clearly,
the clutter-removal algorithm, together with VIC-3D R
, has been effective. The
resistance curve for the case of the notch with height = 0.025 falls closer to the
negative peak of the input data than does the curve for the other notch. The reactance
curves of the two notches, however, are virtually indistinguishable. This suggests
that it is the better fit of the resistance data that gives the 0.025-inch notch the smaller
residual-norm compared to the 0.024-inch notch.
Table 15.1 compares the nominal, measured, and computed data for these
notches.
292 15 Preprocessing Data and Transformation of Signal Vectors

Table 15.1 Comparison of dimensions of the reconstructed A4 notches with two different depths.
The measured data were determined from salastic molds
Length Length Length Width Width Width Depth Depth Depth
Nominal Measured Computed Nominal Measured Computed Nominal Measured Computed
(inch) (inch) (inch) (inch) (inch) (inch) (inch) (inch) (inch)
0.25 0.252 0.2277 0.005 0.007 0.0049 0.020 0.0238 0.025
0.25 0.252 0.2424 0.005 0.007 0.0050 0.020 0.0238 0.024

Fig. 15.5 Original measured Results for DOD slot1.200

impedance data and 0.014
VIC-3D R
model results for a resistance (data)
ligatured crack 0.012 reactance (data)
resistance (model)
reactance (model)
0.01
Resistance or reactance (Ohms)

0.008

0.006

0.004

0.002

-0.002

-0.004
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
Probe position (inches)

15.1.2 A Ligatured Outer-Diameter Slot at 200 kHz

We continue with another example from the data provided by the Westinghouse
Research Labs. In this example we consider an “outer-diameter” flaw comprising 5
notches and 4 gaps, which could model a crack with periodic contact points between
the surfaces, i.e., a partially closed crack.
The notches are each 0.041 inches long, 0.0065 inches wide, and 0.0138 inches
deep. The plate is 0.048 inches thick, so this would be an outer-diameter flaw that
extends 28.75 % into the host. The gaps (ligatures) are 0.0066 inches long. Hence,
the crack is 0.2314 inches long and is 11.4 % closed.
The original measured impedance data and VIC-3D R
model results at 200 kHz
are shown in Fig. 15.5. We model the clutter signal by the piecewise linear functions:
15.1 Clutter Modeling and Rejection 293

Results of Processing Results of Processing

0.002 0.005
Processed Data Processed Data
VIC Results VIC Results
0.0015 0.004
Resistance (Ohms)

Reactance (Ohms)
0.001 0.003

0.0005 0.002

0 0.001

−0.0005 0

−0.001 −0.001
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5 −0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5
Probe Position (in) Probe Position (in)

Results of Processing
0.002
Processed Data
VIC Results
0.0015
Resistance (Ohms)

0.001

0.0005

−0.0005

−0.001
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5
Probe Position (in)

Fig. 15.6 Measured data with the clutter removed, together with the same VIC-3D R
model
results. Top left: resistance, top right: reactance. Bottom: Measured resistance data with the clutter
removed and then passed through a Bezier filter

Rclutter = 0.0117 − 0.01719 × (pos+ 0.5) if pos ≤ −0.244

= 0.0073 − 0.00986 × (pos+ 0.244) otherwise
(15.2)
Xclutter = −0.003 if pos ≤ 0.1
= −0.003 + 0.0075 × (pos− 0.1) otherwise
When this signal is subtracted from the measured data of Fig. 15.5, we get the results
shown in the top portion of Fig. 15.6. The Bezier-filtered and processed resistance
data are shown in the bottom portion of Fig. 15.6. The ripples that appear in the
VIC-3D R
-generated model results and in the original data over the peak range are
a manifestation of the existence of the periodic contact points or “ligands.”
At 200 kHz, the free-space impedance of the coil that was used to take these
data is 4.02 + j21.91 Ω, and the on-plate, host-only, impedance is 4.78 + j21.29 Ω.
Figure 15.6 shows that we can easily resolve a value of δ R = 0.0001 Ω (using the
Bezier filter) and δ X = 0.0002 Ω. This means that we have an effective dynamic
range of 20 log(4.78/0.0001) = 93.6 dB for resistance, and 20 log(21.29/0.0002) =
294 15 Preprocessing Data and Transformation of Signal Vectors

100.5 dB for reactance. Hence, we have significantly gained dynamic range by the
process of clutter removal, and what was clearly an undetectable flaw-resistance
value in Fig. 15.5 becomes quite detectable in Fig. 15.6. The results of Fig. 15.6 can
then be used successfully in an inversion process.
If we believe in the “principle of conservation of dynamic range,” that dynamic
range in data can be neither created nor destroyed, then we must conclude that the
Hewlett-Packard 4194A impedance analyzer had the requisite dynamic range of
100dB, but that the original data of Fig. 15.5 obscured this fact because of the clutter.
Once the clutter was removed the “true-anomaly data” were then manifest.

15.1.3 An Automated Clutter Removal Algorithm

The preceding clutter removal process was heuristic and required user interaction.
Now, we want to develop an automated, model-based clutter removal algorithm
that uses a more general higher-order polynomial fit. The proposed clutter rejection
algorithm begins with an assessment of the model data. A series of N simulated
impedance data vectors, consisting of M probe positions, are defined as Zk,i (x j ),
where Ri (x j ) is the resistance component, and Xi (x j ) is the reactance component,
representing the ith model data set and jth probe position. To evaluate background
regions of interest in the experimental data, the model data are evaluated for regions
of scans where the response is invariant to changes in flaw parameters. Figure 15.7a
shows a series of simulated data vectors for the reactance component of eddy-current
impedance measurements for varying pit dimension. It is desired to have model data
with pertinent flaw parameters ranging over all expected levels in application.
A model-based variation measure, which is used to estimate regions of invari-
ance, is given by:
σk (x j )
Gk (x j ) = 1 − , (15.3)
max(σk )
where
&
1 N ! "2
σk (x j ) = ∑
N − 1 i=1
Zk,i (x j ) − Z k (x j ) ; (15.4)

k = 1 corresponds to Z1 (x) = R(x), k = 2 corresponds to Z2 (x) = X(x), and Z k (x j )

is the mean of the appropriate impedance component. Thus, for scan regions with
invariance to flaw size, Gk approaches 1, and for locations with greatest variation
to flaw size, Gk approaches 0. Figure 15.7b shows an example of the model-
based variation measure for the model data shown in Fig. 15.7a. For practical
implementation, a threshold function, G (x j ), can be defined using a minimum
variation criterion, γ , resulting in regions associated with the background (equal
to 1) and flaw (equal to 0). Figure 15.7c displays the threshold function, G (x j ),
defining a select scan region for background clutter fit based on a minimum variation
criterion of γ = 0.995.
15.1 Clutter Modeling and Rejection 295

Fig. 15.7 Plots of (a) simulated data for the reactance component of impedance measurements
for varying pit dimensions, (b) a model-based variation measure of the reactance components as a
function of position, and (c) a threshold function defining a select region for background fit

Using the result of the model-based background region assessment, a fit of the
background clutter can now be performed. To properly represent the background
clutter found in experimental data, a 3rd-order polynomial representation,
fcr (x) = a3 x3 + a2 x2 + a1 x + a0 , (15.5)
is used. The coefficients of the polynomial are evaluated in a least-squares sense
with the experimental data fit over the background data point regions. This function
can then be applied to the experimental data to subtract the estimated error due to
background clutter.
Figure 15.8 presents a plot of the resistance component at f = 1.3 kHz for
the transformed experimental data (exp0 ) for the case of a subsurface artificial
pit (0.062 diameter, 0.062 height) on the back side of the first layer of a two-
layer stack-up of aluminum panels of 0.125 in thickness. Figure 15.8 also displays
296 15 Preprocessing Data and Transformation of Signal Vectors

Fig. 15.8 Plots presenting

the experimental data (exp0 )
a curve representing a
polynomial fit to background
clutter, f cr (x), and resulting
experimental data with clutter
removed (expcr ) for
resistance component at
f = 1.3 kHz

Fig. 15.9 Illustrating the results of applying the automated clutter removal algorithm to various
measured data sets

the curve, fcr (x), representing a polynomial fit to background clutter, and the
resulting experimental data with clutter removal (expcr ). Clearly, the application
of the automated clutter removal algorithm is beneficial in eliminating the severe
background variance across both the background and flaw regions. Figure 15.9
shows some results of applying the algorithm to various measured data.
15.2 Transformations of Signal Vectors 297

15.2 Transformations of Signal Vectors

Suppose that we run two N-point scans with a probe, one at frequency f1 and the
other at frequency f2 . We arrange the output of each scan as the two complex
N-vectors, Z1 and Z2 , respectively. That is, these vectors represent the complex
impedance associated with the two scans. We wish to determine a coupling
coefficient, or scale factor, β , such that
β Z1 + Z2 = 0 . (15.6)
If we form the dot-product of (15.6) with Z1 , we easily solve for β :
Z1 · Z2
β =− , (15.7)
|Z1 |2
where
N
Z1 · Z2 = ∑ Z1∗ (n)Z2 (n) , (15.8)
n=1

and the asterisk denotes the complex-conjugate. This transformation can be used
in a number of ways. The use of multifrequencies was introduced as a method
of eliminating background signals due to tube-support plates in the inspection of
heat-exchanger tubes in nuclear power plants. We use it extensively at a single
frequency to transform instrument voltage readings into an equivalent impedance,
thereby effectively transforming the instrument into an impedance analyzer.
If we wish to mix four frequencies, we first collect the impedance vectors, Z1 , Z2 ,
Z3 , Z4 , at the four frequencies, and then determine the three parameters, β1 , β2 , β3 ,
such that

β1 Z1 + β2 Z2 + β3 Z3 + Z4 = 0 . (15.9)

Upon taking dot-products of (15.9) sequentially with respect to Z1 , Z2 , and then Z3 ,

we form the system of three equations in the three unknowns, β1 , β2 , β3 :

c11 β1 + c12β2 + c13 β3 = d1

c∗12 β1 + c22β2 + c23 β3 = d2
c∗13 β1 + c∗23β2 + c33 β3 = d3 , (15.10)

where

c11 = Z1 · Z1 c12 = Z1 · Z2 c13 = Z1 · Z3

c22 = Z2 · Z2 c23 = Z2 · Z3 c33 = Z3 · Z3
d1 = −Z1 · Z4 d2 = −Z2 · Z4 d3 = −Z3 · Z4 . (15.11)
298 15 Preprocessing Data and Transformation of Signal Vectors

The solution of (15.10) is

d1 (c22 c33 − |c23 |2 ) + d2 (c13 c∗23 − c12 c33 ) + d3(c12 c23 − c13c22 )
β1 =
Δ
d1 (c∗13 c23 − c∗12 c33 ) + d2(c11 c33 − |c13|2 ) + d3(c∗12 c13 − c11c23 )
β2 =
Δ
d1 (c∗12 c∗23 −c∗13 c22 )+d2 (c12 c∗13 −c11 c∗23 )+d3 (c11 c22 −|c12 |2 )
β3 = , (15.12)
Δ

where Δ = c11 (c22 c33 − |c23 |2 ) + c∗12(c13 c∗23 − c12 c33 ) + c∗13 (c12 c23 − c13 c22 ).

15.3 Further Developments and Applications

of Scaling and Transformations

15.3.1 Modeling Differential-Bobbin Probes

The differential bobbin probe is often used to inspect heat-exchanger tubes in

nuclear power plants. The probe comprises two identical bobbin coils that are
coaxial and parallel to each other, as shown in Fig. 18.19. Each coil is connected
to one leg of a bridge circuit, which produces the difference signal.
In modeling such a probe, we simply take the response of a single bobbin coil,
and then interpolate within that response to compute the response of the second coil
(assuming that the scan increment is different from the spacing of the two coils).
The interpolated value is then subtracted from the primary response to model the
response of the differential bobbin. In effect, we have transformed a single “true”
response to create a “virtual” differential-bobbin probe response. As we will show
in Chap. 18, this virtual probe yields excellent results when modeling the real probe
response. This suggests, as an aside, that it may not be necessary to use complicated
probes, together with analog circuitry (such as a bridge circuit), to create a desired
response. It may be much easier (and cheaper) to use a simple probe and accomplish
the desired response numerically.

15.3.2 A General Virtual Probe Model

We can generalize the development of a virtual probe beyond the differential-bobbin

example. Suppose we are trying to invert data taken by a “large” probe, but find
that the resolution is insufficient to reconstruct a “small” flaw. What is needed, we
decide, is a “small” probe. It would be nice if we could use the original probe to
take the data, but develop a virtual probe to transform the measured data into data
15.3 Further Developments and Applications of Scaling and Transformations 299

that would allow the reconstruction of the small flaw. This would obviate the need
to use a different piece of hardware, while still retaining the ability to satisfactorily
invert the original measured data. That this concept is mathematically feasible will
be demonstrated in this section.
Assume that we know the “bounding-box” of the flaw, which will allow us to
establish a flaw grid, but we don’t know what is in the box; i.e., we don’t know
what the flaw is, but only its position and maximum extent. With this assumption,
(L)
we can compute the incident-field moments, E0 , due to the large coil. Similarly,
(s)
we can compute E0 , the incident-field moments produced by the small coil, once
we decide what the parameters of the small coil should be. With these results, we
transform the large-coil incident-field moments into the small-coil moments via our
usual least-squares β -relationship:
(L) (s)
β E0 = E0 , (15.13)
which has for its solution
(L)H (s)
E0 · E0
β= (L)
, (15.14)
E0 2
with the symbol, H, denoting the Hermitian (complex-conjugate) transpose of a
vector or matrix.
If the anomalous-current response of the volume-integral equation to a flaw with
(L) (L) (s)
E0 as the driver is Ja , then the response to the same flaw with E0 as the driver
(s) (L)
is Ja = β Ja . This holds without knowing the nature of the flaw and is a property
of the linearity of the volume-integral equation. Thus, the impedance measured by
the virtual (small) probe is
(s) (s) (s)
Za = Ja · E0
(L) (L)
= β 2 Ja · E0
(L)
= β 2 Za ; (15.15)
(L)
Za is the known (measured) impedance using the large probe. This relationship
must be computed at each scan position, so we assume that the small probe follows
the same scan as the large probe.

15.3.3 Another Virtual Probe Model

We’ll apply the ABCD-matrix theory of Chap. 8 to the situation of a single probe,
rather than a T/R system. A single probe is usually represented as a one-port, but
in order to use ABCD-matrix theory, we must transform it into a two-port. If the
driving-point impedance of the probe is Z2 , then an appropriate two-port is shown
300 15 Preprocessing Data and Transformation of Signal Vectors

1 2

1
1 1 Z2
−
Z3 Z2

Y3 − Y2 Y2

Fig. 15.10 Showing a network representation of one-ports. Top: Representation in terms of

impedances. Bottom: Representation in terms of admittances

as system 2 in Fig. 15.10. The open-circuit driving-point/transfer impedance matrix

of such a one-port is
 
Z Z
Z2 = 2 2 . (15.16)
Z2 Z2
We want to transform this probe into one whose driving-point impedance is Z3 ,
and the transformation equations in terms of the ABCD matrices, M1 , M2 , and M3
is M3 = M1 × M2 , where
 
1 0
M2 =
1/Z2 1
 
1 0
M3 = . (15.17)
1/Z3 1
M1 is, therefore, given by

M1 = M−12 × M3
   
1 0 1 0
= ×
−1/Z2 1 1/Z3 1
 
1 0
= , (15.18)
−1/Z2 + 1/Z3 1
from which we conclude that
15.3 Further Developments and Applications of Scaling and Transformations 301

1 2

1
1 1 Z2 + δZ2
−
Z3 Z2

Y2
Y3 − Y2
1 + Y2 δZ2

Fig. 15.11 Showing a network representation of one-ports when the virtual probe is loaded,
producing a change in the driving-point impedance of δ Z2 . Top: representation in terms of
impedances. Bottom: representation in terms of admittances
⎡ 1 1 ⎤
⎢ 1/Z3 − 1/Z2 1/Z3 − 1/Z2 ⎥
Z1 = ⎣ 1 1 ⎦. (15.19)
1/Z3 − 1/Z2 1/Z3 − 1/Z2
This result yields the network representation labeled “1” in the top part of Fig. 15.10.
Under load, the virtual probe has a driving-point impedance of Z2L = Z2 + δ Z2 ,
where δ Z2 is the change in impedance induced by the load. This situation is
shown in Fig. 15.11. Under this condition, the “real” probe has a loaded admittance
given by
(L) Y2
Y3 = Y3 − Y2 +
1 + Y2δ Z2
Y3 + (Y2Y3 − Y22 )δ Z2
= , (15.20)
1 + Y2δ Z2
from which we get for the change in driving-point impedance of the real probe
1 + Y2δ Z2 1
δ Z3 = −
Y3 + (Y2Y3 − Y22 )δ Z2 Y3
Y22 δ Z2
= . (15.21)
Y3 [Y3 + (Y2Y3 − Y22 )δ Z2 ]
We can use this in several ways: given a measured δ Z3 due to a complex probe,
perhaps with a conducting core, and a model result, Y2 δ Z2 , for the same situation
that produced δ Z3 , we have a free parameter, Y3 , that allows us to use (15.21) to
replace the actual probe with the simpler virtual probe. Note that if we ignore the
parenthetical term in the denominator of (15.21), then we have a linear relation
between δ Z3 and δ Z2 , with Y22 /Y32 playing the role of β in (15.6) or (15.13).
302 15 Preprocessing Data and Transformation of Signal Vectors

deltaZ2 and deltaZ3 deltaZ2 and deltaZ3

0.1 1.2
0
1
deltaZ3
−0.1
deltaZ2
0.8
−0.2
Resistance

Reactance
−0.3 0.6
−0.4 deltaZ3 0.4
−0.5 deltaZ2
0.2
−0.6
−0.7 0
−10 −5 0 5 10 −10 −5 0 5 10
Probe Position (mm) Probe Position (mm)

Fig. 15.12 Response of two probes over a flaw. Probe 2 is an air-core coil, and Probe 3 is the same
coil enclosing a ferrite core
Results for DeltaZ3 Results for DeltaZ3
0.1 1.2
0
1
−0.1
0.8
−0.2
Resistance

VIC-3D
Reactance

Transform
−0.3 0.6
−0.4
VIC-3D 0.4
Transform
−0.5
0.2
−0.6
−0.7 0
−10 −5 0 5 10 −10 −5 0 5 10
Probe Position (mm) Probe Position (mm)

Fig. 15.13 Comparing δ Z3 when computed by VIC-3D

R
and when computed by the transfor-
mation equation (15.21)

Remember, we are treating Y2 and Y3 as complex scalars, whereas δ Z2 and δ Z3

are complex vectors, because they are the results of scans over a flaw. If we take Y2
and Y3 as conditions in freespace, and δ Z2 and δ Z3 as changes due to loading by an
unflawed workpiece, then these two deltas are scalars, also.

15.3.4 An Example

Consider a small air-core coil whose impedance over an unflawed workpiece is Z2 =

2.860 + j43.9184 Ω, and the same coil enclosing a nonconducting ferrite core with
relative permeability of 2,000, whose impedance over the same workpiece at the
same frequency is Z3 = 4.4197 + j57.982 Ω. When each of these probes is scanned
past a pit in the workpiece measuring 2 × 2 × 2 mm, the resulting flaw responses
are shown in Fig. 15.12. Figure 15.13 compares the results of the VIC-3D R
ferrite-
core calculation of Fig. 15.12 with the application of the transformation equation,
(15.21), to the air-core calculation shown in Fig. 15.12.
15.4 A General Transformation Matrix 303

Given the values of Y3 and Y2 in this example, together with the fact that |δ Z2 | is
small, we can ignore the parenthetical expression in the denominator of (15.21), so
that the ratio δ Z3 /δ Z2 ≈ [(4.4197 + j57.982)/(2.8603 + j43.9184)]2 ≈ 1.746. For
this same problem, we get β = 1.7053 − j0.0620, which differs from 1.746 by less
than 3 %.

15.4 A General Transformation Matrix

In developing the idea of the virtual probe, we came across a transformation matrix
that is more general than the simple product of a vector with the scalar, β , yet
should be very easy to implement numerically. The mathematical tool that is basic
to our idea is the Householder transformation matrix that finds significant use in
computational linear algebra.1 Our presentation follows the footnoted reference
closely.
The Householder matrix, P, has the form

P = I − 2w · wH , (15.22)

where w is a complex unit vector. As written, the second term in (15.22) is the outer
product of the two vectors.2 The inner or scalar product of two vectors, a and b is
written aH · b. The matrix P is clearly Hermitian and orthogonal, because

P2 = (I − 2w · wH ) · (I − 2w · wH )
= I − 4w · wH + 4w · (wH · w) · wH
=I. (15.23)

Therefore P = P−1 = PH , which establishes orthogonality. Consider the subspace of

vectors that are orthogonal to w; any vector in this subspace remains invariant under
P, but any vector, v = |v|w, that is oriented in the direction of w is inverted, P · v =
−v. As a final geometric interpretation of (15.22), consider P · x = x − 2w · wH · x,
where x is arbitrary. Rearranging, and taking the norms of each side, we have
4(w · wH · x)H · (w · wH · x) = 4|wH · x|2
= (x − P · x)H · (x − P · x)
= |x|2 − 2xH · PH · x + |P · x|2
= 2|x|2 − 2|x|xH · a
= 2|x|2 (1 − cos φ ) , (15.24)

1 See W. H. Press, B. P. Flannery, S. A. Teukolsky, W. T. Vetterling, Numerical Recipes: The Art of

Scientific Computing, Cambridge University Press, 1986.
2 This is also called the direct or tensor or Cartesian product.
304 15 Preprocessing Data and Transformation of Signal Vectors

Fig. 15.14 Illustrating the

use of a “Householder Filter”
to simulate the coupling
R0
between the driving-point
Zair
terminals of a probe and the Householder Filter in Air
VIC-3D R
model. The L0
low-frequency parameters, R0
and L0 , have been
independently measured

R0

L0

Zload Householder Filter in Air

δZ

where a is a unit vector that points in the direction of the image vector, P · x, and we
have used the fact that PH = P = P−1 , which implies that |P · x| = |x|. φ is the angle
between x and its image and is a measure of the amount that x has been rotated by
the orthogonal transformation. Equation (15.24) is the law of cosines in hyperspace.
Now, let u be any vector, and rewrite P as

u · uH
P = I− , (15.25)
N
1
where N = |u|2 . We’ll use this operator to rotate a vector, x, into another vector
2
oriented in the direction of the unit vector, a. To that end, let u = x + |x|a. Then
u
P·x = x− · (x + |x|a)H · x
N
u
= x − (xH · x + |x|2 cos α )
N
u
= x − xH · x(1 + cos α )
N
2uxH · x(1 + cos α )
= x−
(x + (xH · x)1/2a)H · (x + (xH · x)1/2 a)
2uxH · x(1 + cos α )
= x−
(xH · x + 2(xH · x) cos α + xH · x)
= x−u
= −|x|a . (15.26)
15.4 A General Transformation Matrix 305

Thus, if our target vector is A = |A|a, then we have

|A|
A=− P·x , (15.27)
|x|
as the required transformation of x into A that replaces the simpler scaling by β .
The advantage of introducing P, besides the fact that it is orthogonal, is that it is
easily implemented as a stable computational algorithm.
We propose to use this transformation in the same way that we use the β
transformation; we first determine the parameters of the orthogonal transformation
by selecting a standard “source” vector, xs , with its corresponding standard “target”
|As |
vector, As , and establish the parameters, , us = xs + |xs |as and Ns which
|xs |
determine the standard Ps , according to (15.25). We then use Ps , together with
an arbitrary source vector, x, to determine the corresponding target vector, A, by
applying (15.27).

15.4.1 Application to Probe Characterization

We will apply the theory of the Householder transformation matrix to create a

filter to transform measured impedance data into model data for the purpose of
characterizing probes over a broad frequency range. Consider Fig. 15.14, which
shows a situation in which a probe driving-point impedance is measured over a
broad frequency range, and the intention is to isolate, and eventually remove, the
coupling filter between the input terminals and the output circuit that represents the
VIC-3D R
model. It is assumed that the low-frequency parameters, R0 and L0 , have
been independently measured.
In this case, the standard source vector, xs = Zair (ωi ), i = 1, · · · , N, and the
standard target is As = R0 + jωi L0 , i = 1, · · · , N. Once the standard Ps is created, we
use (15.27) with x = Zload (ωi ), i = 1, · · · , N, and compute A = R0 + jωi L0 + δ Z(ωi ),
from which δ Z(ωi ) is determined by subtracting R0 + jωi L0 from A. It is implicitly
assumed that the presence of the workpiece does not significantly change the
coupling network that is represented by the Householder filter in air.
 Part III
Applications
Chapter 16
Modeling Corrosion and Pitting Problems

16.1 Introduction and Overview of Approach

Given the material properties of typical aircraft structures and environmental

conditions experienced over a service life, corrosion in aircraft structures is a major
issue. For the United States Air Force alone, the total annual cost for corrosion
management is on the order of a billion dollars. Avoidance of even a small portion
of this annual expenditure would result in significant savings. To improve the
management of corrosion, limits of existing inspection tools must be addressed.
In particular, the ability to characterize the micro-topography of corrosion that
forms fatigue and stress corrosion cracking is critical [114]. With this capability,
repairs could be better directed to corrosion damage that only exceeds defect
criteria with minimal aircraft disassembly. This chapter explores the problem of
characterizing the surface topology of corrosion at faying surfaces in multilayered
aircraft structures, with the focus on quantifying the size of corrosion pits in both
first and second layers.
Although many corrosion techniques for aircraft structures only use a single NDE
method, such as eddy-current, improvements may be achieved through the use of
measurement data from multiple methods. Research has investigated data fusion of
ultrasonic and eddy-current (EC) methods [115] and conventional and pulsed eddy-
current techniques [116]. In particular, ultrasonic and eddy-current methods are
quite complementary for multilayer corrosion problems, where ultrasonic methods
provide excellent resolution of corrosion in the first layer, while eddy-current
measurements are sensitive to the presence of corrosion in both the first and second
layers. The approach described in this chapter uses ultrasonic data to characterize
the first layer corrosion and thus simplify the second layer characterization problem
using eddy-current.
In particular, a pragmatic approach is presented concerning the fusion of
ultrasonic and eddy-current data, where ultrasonic measurements are used to
accurately characterize first layer corrosion topology and thus minimize the number
of unknowns in the second layer eddy-current inversion problem. This approach

H.A. Sabbagh et al., Computational Electromagnetics and Model-Based Inversion, 309

Scientific Computation, DOI 10.1007/978-1-4419-8429-6 16,
© Springer Science+Business Media New York 2013
310 16 Modeling Corrosion and Pitting Problems

contrasts with traditional data fusion where pixel-by-pixel fusion of measurement

data is used. In reality, subsurface features such as pits are spread across image data
and are dependent upon probe parameters. Thus, models in conjunction with inverse
methods are needed to best extract features associated with the dimensions of the
pit spread across the image data.
For the experiments reported in this chapter, artificial corrosion samples using
drilled holes with conical tips to represent real pits were constructed to investigate
optimizing the model-based inverse method approach of this book. Simultaneous
acquisition of both ultrasonic and eddy-current data was performed using an eddy-
current probe integrated with ultrasonic transducers.

16.2 Modeling the Corrosion Topology Problem

The problem that we are addressing is shown in Fig. 16.1. A probe, comprising a
coil with a rather complex ferrite core, is scanned past a truncated cylindrical pit of
diameter, PD, embedded in a quarter-inch aluminum slab. The pit extends only into
the upper-half of the slab and has a height of PH1. For this analysis we will assume
that the frequency of excitation is 2,200 Hz.

16.2.1 A Simplified Probe Model

It will facilitate running the many forward models that will be required for inverting
data if we can replace the ferrite core in the probe model with a simpler probe
that does not have a ferrite core. We have run a number of NLSE calculations
to determine that coil whose impedance most closely approximates the measured
results of a pit that is 75 % through-wall (PH1 = 2.38125 mm in Fig. 16.1), with a
diameter of 0.125 in (PD = 0.125 in. in Fig. 16.1). The results, which are shown
in Fig. 16.2, were obtained with a coil whose inner radius is 0 mm, outer radius is
2.75 mm, height of 2.54 mm, containing 2,500 turns, and tilted at an angle of 2◦

Z
Core
Scan
Coil
X

PD

Pit PH1
Aluminum : σ = 30% IACS 6.35mm

3.175mm

Fig. 16.1 The basic corrosion topology problem

16.2 Modeling the Corrosion Topology Problem 311

Comparison of Measured and Model Data Comparison of Measured and Model Data
0.02 0.16
measured measured
model 0.14 model
0
0.12
-0.02
Resistance (Ohms)

Reactance (Ohms)
0.1

-0.04 0.08

-0.06 0.06

0.04
-0.08
0.02
-0.1
0

-0.12 -0.02
-15 -10 -5 0 5 10 15 -15 -10 -5 0 5 10 15
Probe Position (mm) Probe Position (mm)

Fig. 16.2 Comparison of measured and model data for a probe that is tilted at an angle of 2◦

about the Y -axis of Fig. 16.1. This probe model captures the peaks and valleys of
the measured data reasonably well and will be used when creating the interpolation
table for inverting the data for the class of problems defined in Fig. 16.1.

16.2.2 Data Transformation

The experimental data for this setup are recorded as A/D counts (or “voltage”) in
two channels that are 90◦ out of phase with each other. We have done some initial
processing on this data by first defining the “baseline value” of each channel and
then subtracting this baseline from the scanned data of each channel. The baseline
value is taken to be the left-hand value of the data, because there are no other pits or
anomalies to interact with this value (we are well away from any edges of the host
plate). We call these data the “host-balanced” data, and they are shown in Fig. 16.3.
The measured data are clearly asymmetrical, even though the pit and probe
are symmetrical. As we discussed above, we have created an accurate, though
simplified, model of the probe, and have shown that if the probe is tilted about
2◦ about the Y -axis of Fig. 16.1, then the model also produces asymmetric lobes
similar to those of Fig. 16.3. It remains, now, to transform data, measured in A/D
counts or “voltage” into impedances, so that they can be further processed using
VIC-3D R
. We use the transformation of signal vectors described in Chap. 15
to accomplish this. The actual data used to determine β were for the same flaw
(PD = 0.125 in, PH1 = 2.38125 mm) as above, and the value obtained was
β = 6.5276 × 10−6 + j8.2605 × 10−6. When the measured data of Z1 (see Fig. 16.3)
are multiplied by β , we get the “β -normalized” measured data shown in Fig. 16.4.
312 16 Modeling Corrosion and Pitting Problems

Fig. 16.3 The measured Measured Input Data (Arbitrary Units)

“host-balanced” data 50
Channel 1
Channel 2
0

-50

Channel 1 or Channel 2
-100

-150

-200

-250

-300

-350

-400

-450
-15 -10 -5 0 5 10 15
Probe Position (mm)

Fig. 16.4 The β -normalized Beta-Normalized Input Data

measured “host-balanced” 0.004
data R
X
0.003
Resistance or Reactance

0.002

0.001

-0.001

-0.002

-0.003
-15 -10 -5 0 5 10 15
Probe Position (mm)

16.3 The Inverse Problem

The inverse problem can be addressed in terms of Fig. 16.1, in which two 1/8-inch
panels are placed back-to-back, with corrosion pits emanating from the bottom of
the first panel upward, and from the top of the second panel downward. The first pit
is called pit 1, and the second, pit 2. The problem is to determine the height of each
pit (PH1 and PH2), as well as the diameter (PD1 and PD2). We confine ourselves
here to pits that emanate upward only; hence, PH2 = PD2 = 0.
16.3 The Inverse Problem 313

Fig. 16.5 A 5 × 5-grid that is Radius (mm)

suitable for interpolation with
polynomials of order 1–4 in 1.75
each variable

1.40

1.05

0.7

0.35 Height (mm)

0.7 1.2 1.7 2.2 2.7

16.3.1 The Interpolator for NLSE

From here on we work with only the radius and height of the pit. The interpolation
table is defined with respect to the grid shown in Fig. 16.5. Because this is a 5 × 5
grid, we can use polynomial splines of order 1–4 for interpolation in each variable.
The interpolating functions (sometimes called “blending” functions) associated
with each of the 25 nodes of Fig. 16.5 are shown in Figs. 16.6–16.10. These
functions were computed using the same probe-flaw model that produced the model
data of Fig. 16.2.
It is clear that the challenge in the inversion process comes not only from the
nonlinear relationship between the impedances and the two parameters, PD and
PH1, but also from the significant change in appearance of the resistance scan-data
with these two parameters. It appears that the inversion might be less confounding
if we restricted ourselves to only reactance data; this will be discussed shortly.

16.3.2 Results at 2,200 Hz

We have run a number of inversions, and we show six that are typical of the results.
In each of the following tabulated results, we show the norm of the residuals, Φ ,
the computed values of pit radius, R, and pit height, H, and the sensitivity of the
solution, ΣR , ΣH , to each of these variables. These data are tabulated as a function
of the order of the interpolating polynomial. Because there are five nodes in each
variable, the interpolating polynomial in each variable can range in order from 1
to 4.
Table 16.1 shows the results of the inversion using the higher-order spline
interpolator when applied to the model input data of Fig. 16.2.
314 16 Modeling Corrosion and Pitting Problems

Nodes With PH1=0.7 Nodes With PH1=0.7

0.005 0.012
0.35 0.35
0.0045 0.70 0.70
1.05 0.01 1.05
0.004 1.4 1.4
1.75 1.75
0.0035 0.008
Resistance (Ohms)

Reactance (Ohms)
0.003
0.006
0.0025

0.002
0.004
0.0015

0.001 0.002

0.0005
0
0

-0.0005 -0.002
-15 -10 -5 0 5 10 15 -15 -10 -5 0 5 10 15
Probe Position (mm) Probe Position (mm)

Fig. 16.6 Interpolating (“blending”) functions for the case PH1 = 0.7. The numbers indicate the
radius of the corresponding node in Fig. 16.5. Left: resistance; Right: reactance

Nodes With PH1=1.2 Nodes With PH1=1.2

0.01 0.035
0.35 0.35
0.70 0.70
0.008 1.05 0.03 1.05
1.4 1.4
1.75 0.025 1.75
Resistance (Ohms)

0.006
Reactance (Ohms)

0.02
0.004
0.015
0.002
0.01
0
0.005

-0.002 0

-0.004 -0.005
-15 -10 -5 0 5 10 15 -15 -10 -5 0 5 10 15
Probe Position (mm) Probe Position (mm)

Fig. 16.7 Interpolating (“blending”) functions for the case PH1 = 1.2. The numbers indicate the
radius of the corresponding node in Fig. 16.5. Left: resistance; Right: reactance

The original data of Fig. 16.11, pd-031 ph1-25 ph2-0 f1, show a clear system-
atic error that we call generically “clutter” (see Chap. 15). The inverted results when
using these cluttered data as input to NLSE are shown in Table 16.2. In order to
determine the “flaw signal,” we must remove the clutter from the original measured
data. This is done, as described in Chap. 15, by modeling the clutter as a simple
mathematical expression:
Rclutter = −0.001 + (0.0085/17.907) × (pos+ 10.033)
Xclutter = 0.00026 − (0.00246/17.907) × (pos+ 10.033) , (16.1)
where “pos” denotes the probe position.
16.3 The Inverse Problem 315

Nodes With PH1=1.7 Nodes With PH1=1.7

0.015 0.07
0.35 0.35
0.70 0.70
0.01 1.05 0.06 1.05
1.4 1.4
1.75 0.05 1.75
Resistance (Ohms)

Reactance (Ohms)
0.005
0.04
0
0.03
-0.005
0.02
-0.01
0.01

-0.015 0

-0.02 -0.01
-15 -10 -5 0 5 10 15 -15 -10 -5 0 5 10 15
Probe Position (mm) Probe Position (mm)

Fig. 16.8 Interpolating (“blending”) functions for the case PH1 = 1.7. The numbers indicate the
radius of the corresponding node in Fig. 16.5. Left: resistance; Right: reactance

Nodes With PH1=2.2 Nodes With PH1=2.2

0.02 0.14
0.35 0.35
0.01 0.70 0.70
1.05 0.12 1.05
0 1.4 1.4
1.75 0.1 1.75
Resistance (Ohms)

Reactance (Ohms)

-0.01
0.08
-0.02

-0.03 0.06

-0.04
0.04
-0.05
0.02
-0.06
0
-0.07

-0.08 -0.02
-15 -10 -5 0 5 10 15 -15 -10 -5 0 5 10 15
Probe Position (mm) Probe Position (mm)

Fig. 16.9 Interpolating (“blending”) functions for the case PH1 = 2.2. The numbers indicate the
radius of the corresponding node in Fig. 16.5. Left: resistance; Right: reactance

After removing the clutter by subtracting (16.1) from the measured data, we get
the results shown in Fig. 16.12, which are a clear improvement. The inverted results
for the new data are shown in Tables 16.3 and 16.4.
In Fig. 16.13, we compare the original pd-031 ph1-50 data with the recon-
structed data computed from the fourth-order result of Table 16.5. Clearly, the
reactance data, which are much cleaner, drive the inversion.
Now, we will test pd-125 ph1-75 ph2-75 f1 as the first part of a two-step
strategy to size the corrosion in layer 2. We will treat this input, since it is at the
high frequency of 2,200 Hz, as containing only a pit in layer 1, so we will use
the machinery already set up for a two-parameter problem, PH1 and PD, to invert
316 16 Modeling Corrosion and Pitting Problems

Nodes With PH1=2.7 Nodes With PH1=2.7

0.05 0.25
0.35 0.35
0.70 0.70
1.05 1.05
0 1.4 1.4
0.2
1.75 1.75
Resistance (Ohms)

Reactance (Ohms)
-0.05
0.15

-0.1

0.1
-0.15

0.05
-0.2

-0.25 0
-15 -10 -5 0 5 10 15 -15 -10 -5 0 5 10 15
Probe Position (mm) Probe Position (mm)

Fig. 16.10 Interpolating (“blending”) functions for the case PH1 = 2.7. The numbers indicate the
radius of the corresponding node in Fig. 16.5. Left: resistance; Right: reactance

Table 16.1 Results of higher-order spline interpolator when applied to the

model input data of Fig. 16.2. ΣV denotes the sensitivity of the solution to the
variable, V . The fourth-order result has an error in radius of about 2% and in
height of about 0.27%
Polynomial order Φ R (mm) ΣR H (mm) ΣH
1 0.0278 1.651 0.02259 2.317 0.01275
2 0.0199 1.6283 0.01530 2.3554 0.00984
3 0.0191 1.6289 0.01469 2.3554 0.009546
4 0.01323 1.620 0.010 2.3747 0.00702

these data (Table 16.6). Then we will use the parameters obtained for PH1 and PD
as constraints for determining PH2. For this purpose we will need pd-125 ph1-
75 ph2-75 f2, which are data at the lower frequency of 1,300 Hz. Further, we will
develop an entirely new interpolation table for operation at this frequency, as will
be described in the next section.
The results for pd-125 ph1-75 ph2-75 f1 are shown in Table 16.7 and confirm
the reasonableness of our strategy of using high frequencies to reconstruct flaws in
layer 1.

16.3.3 Results at 1,300 Hz

Our intention is to use the lower frequency to reconstruct flaws in layer 2. To this
end we introduce a three-parameter interpolation grid and table for PH1, PH2, and
PD. The grid is a three-dimensional version of Fig. 16.5, in which the third variable,
16.3 The Inverse Problem 317

Fig. 16.11 Original data for Data for PD=.031, PH1=25, PH2=0
PD = 0.03125 in., PH1 = 0.008
25% through-wall = resistance
reactance
0.79375 mm
0.006

Resistance/Reactance (Ohms)
0.004

0.002

-0.002

-0.004
-15 -10 -5 0 5 10 15
Probe Position (mm)

Table 16.2 Results of higher-order spline interpolator when applied to

the original input data of Fig. 16.11. ΣV denotes the sensitivity of the
solution to the variable, V . The fourth-order result has an error in radius
of about 75% and in height of about 9.7%
Polynomial order Φ R (mm) ΣR H (mm) ΣH
1 0.0495 0.7005 1.0324 0.7000 0.9457
2 0.0495 0.7067 1.2621 0.7000 1.0554
3 0.0495 0.7087 1.2629 0.7000 1.1122
4 0.0494 0.6945 1.2043 0.7166 1.4225

Table 16.3 Results of higher-order spline interpolator when applied to the

processed input data of Fig. 16.12. ΣV denotes the sensitivity of the solution
to the variable, V . The fourth-order result has an error in radius of about
15.5% and in height of about 0.5%
Polynomial order Φ R (mm) ΣR H (mm) ΣH
1 0.01419 0.4896 0.4502 0.7000 0.4811
2 0.01419 0.5029 0.4874 0.7000 0.5406
3 0.01419 0.5050 0.4718 0.7000 0.5723
4 0.01416 0.4581 0.4030 0.7985 0.7539

PH2, is normal to the page. This variable will have the same five nodes as PH1,
which means that the grid will have five copies of Fig. 16.5, giving a total of 125
nodes.
Our first result is for pd-125 ph1-75 ph2-75 f2 and is shown in Table 16.8. The
sensitivity parameter gives us insight into the quality of the results. The smallest
318 16 Modeling Corrosion and Pitting Problems

Processed Data for PD=.031, PH1=25, PH2=0 Processed Data for PD=.031, PH1=25, PH2=0
0.0025 0.002
R X
0.002
0.0015
0.0015

Reactance (Ohms)
0.001
Resistance (Ohms)

0.001
0.0005

0 0.0005

-0.0005
0
-0.001

-0.0015
-0.0005
-0.002

-0.0025 -0.001
-15 -10 -5 0 5 10 15 -15 -10 -5 0 5 10 15
Probe Position (mm) Probe Position (mm)

Fig. 16.12 Data of Fig. 16.11 after the systematic error “clutter” has been removed. Left:
resistance; Right: reactance

Table 16.4 Results of higher-order spline interpolator when applied to the

input data of pd-125 ph1-75 data2.in. ΣV denotes the sensitivity of the
solution to the variable, V . The fourth-order result has an error in radius of
about 9.2% and in height of about 3.5%
Polynomial order Φ R (mm) ΣR H (mm) ΣH
1 0.2022 1.4539 0.1407 2.4119 0.1175
2 0.1999 1.4552 0.1454 2.4364 0.1090
3 0.1998 1.4508 0.1445 2.4417 0.1141
4 0.1988 1.4422 0.1453 2.4640 0.1159

value of Φ occurs with the first-order polynomial, and the results for PD and PH1
are reasonable. The error in PD is 6.7%, and in PH1 6.9%. Note that the sensitivity
of these two values at the solution is considerably smaller than that for PH2. This
indicates that PD and PH1 are reasonably well determined, but PH2 is not. The
solution is insensitive to PH2 because the frequency is too high for PH2 to be
resolved. This is what we suspected at 2,200 Hz and holds true also for 1,300 Hz.
The result for pd-031 ph1-25 ph2-25 f2, however, is a little bit more encourag-
ing, as shown in Table 16.9. The values given with the fourth-order polynomial for
the three unknowns are in reasonable agreement with their nominal values, but the
sensitivities for these parameters, though a little large due to the quality of the input
data, are all comparable. This suggests that the value of PH2 can be resolved, even
at 1,300 Hz, because PH1 is small.
16.3 The Inverse Problem 319

Comparison of Original and Reconstructed Data Comparison of Original and Reconstructed Data
0.0035 0.007
Original Original
0.003 Reconstructed 0.006 Reconstructed

0.0025 0.005
Resistance (Ohms)

Reactance (Ohms)
0.002
0.004
0.0015
0.003
0.001
0.002
0.0005
0.001
0
0
-0.0005

-0.001 -0.001

-0.0015 -0.002
-15 -10 -5 0 5 10 15 -15 -10 -5 0 5 10 15
Probe Position (mm) Probe Position (mm)

Fig. 16.13 Comparison of original pd-031 ph1-50 data with the reconstructed data computed
from the fourth-order result of Table 16.5. Left: resistance; Right: reactance

Table 16.5 Results of higher-order spline interpolator when applied to the

input data of pd-031 ph1-50 data2.in. ΣV denotes the sensitivity of the
solution to the variable, V . The fourth-order result has an error in radius of
about 12.9% and in height of about 2.1%
Polynomial order Φ R (mm) ΣR H (mm) ΣH
1 0.01404 0.4520 0.1041 1.4504 0.2823
2 0.01394 0.4394 0.1023 1.5452 0.2731
3 0.01392 0.4467 0.1053 1.5307 0.2726
4 0.01383 0.4480 0.1044 1.5534 0.2761

Table 16.6 Results of higher-order spline interpolator when applied to the

input data of pd-031 ph1-75 data2.in. ΣV denotes the sensitivity of the
solution to the variable, V . The first-order result has an error in radius of
about 0.9% and in height of about 13.4%
Polynomial order Φ R (mm) ΣR H (mm) ΣH
1 0.0374 0.3932 0.0457 2.7 0.2441
2 0.0376 0.3968 0.0501 2.7 0.2264
3 0.0377 0.3986 0.0519 2.7 0.2185
4 0.0380 0.4060 0.0575 2.7 0.1896

16.3.4 Frequency Mixing

If we take the (first-order) solution for R in Table 16.8 and average it with the
(fourth-order) solution in Table 16.4, we get 1.5719, which is in error by less than
1% from the nominal value of 1.5875 mm. Similarly, when we average the solutions
for PH1 from the same tables, we get 2.3405, which is in error by 1.7% from the
320 16 Modeling Corrosion and Pitting Problems

Table 16.7 Results of higher-order spline interpolator when applied to the

input data of pd-125 ph1-75 ph2-75 data2.in. ΣV denotes the sensitivity of
the solution to the variable, V . The first-order result has an error in radius of
about 11.8% and in height of about 3.1%
Polynomial order Φ R (mm) ΣR H (mm) ΣH
1 0.24544 1.40 0.1599 2.4559 0.1538
2 0.24553 1.387 0.1728 2.4858 0.1439
3 0.24558 1.3832 0.1720 2.4929 0.1479
4 0.24580 1.370 0.1723 2.5191 0.1474

Table 16.8 Results of higher-order spline interpolator when applied to the input data of pd-
125 ph1-75 ph2-75 f2. ΣV denotes the sensitivity of the solution to the variable, V . The first-order
result has an error in radius of 6.7% and in PH1 of 6.9%. PH2 is essentially undetermined
Polynomial order Φ R (mm) ΣR PH1 (mm) Σ PH1 PH2 (mm) Σ PH2
1 0.12293 1.7015 0.05962 2.2169 0.12605 0.7 3.5745
2 0.12344 1.7091 0.04978 2.2255 0.16515 0.7 3.5464
3 0.12346 1.7136 0.04671 2.2147 0.16735 0.7 3.5473
4 0.12358 1.7232 0.03667 2.2120 0.18108 0.7 3.5622

Table 16.9 Results of higher-order spline interpolator when applied to the input data of pd-
031 ph1-25 ph2-25 f2. ΣV denotes the sensitivity of the solution to the variable, V . The fourth-
order result has an error in radius of 11.8%, in PH1 of 11.6% and 4.8% in PH2
Polynomial order Φ R (mm) ΣR PH1 (mm) Σ PH1 PH2 (mm) Σ PH2
1 0.011785 0.35 0.8655 0.7019 0.8907 0.8817 4.9636
2 0.01180 0.35 0.9287 0.7095 1.0495 0.7434 5.1100
3 0.01182 0.3580 0.8385 0.7 1.1195 0.8822 5.01
4 0.011779 0.35 0.6277 0.7016 1.7667 0.8337 5.3144

nominal value of 2.38125 mm. This indicates that we can often use “frequency
mixing” to get improved results when compared to the results at the individual
frequencies.

16.3.5 Inverting Only Reactance Data

The resistance data for pd-031 ph1-25 ph2-00, shown in Fig. 16.11, are clearly quite
noisy and play a significant role in producing the bad results shown in Table 16.2
for the estimate of the radius. This suggests that we remove the resistance data and
invert only the reactance data. The result of this experiment is shown in Table 16.10.
The estimates of the height and height-sensitivity remain about the same as before,
but the estimates of the radius and radius-sensitivity are improved considerably.
We should not think, however, that the resistance should be discarded under
all conditions. In several of our experiments, we found that discarding “good”
16.4 A New Flaw Model 321

Table 16.10 Results of higher-order spline interpolator when applied to

only the reactive part of the original input data of Fig. 16.11. ΣV denotes
the sensitivity of the solution to the variable, V . The fourth-order result has
an error in radius of about 12% and in height of about 11.8%
Polynomial order Φ R (mm) ΣR H (mm) ΣH
1 0.013566 0.3500 0.4804 0.7016 0.9694
2 0.013560 0.3503 0.5238 0.7000 1.1000
3 0.013561 0.3500 0.5438 0.7007 1.1724
4 0.013558 0.3501 0.6585 0.7000 1.5759

Fig. 16.14 A new flaw Conical ‘Tip’ Flaw

model that accounts for the
conical tip produced by the
drill bit. The usual truncated
Equivalent Pit : Equal Volume
cylindrical pit model is shown Equal Centroid
in the case that the pit is not
deeper than the tip. The pit
model is assumed to have the
same volume and centroid as
the conical tip

resistance data degraded otherwise “good” solutions. It appears that resistance data
stabilize the inversion process, especially in determining the height of a pit. Later,
we will give an example of the use of the projection algorithm, POCS, to clean up
the resistance and yield an improved inversion.

16.4 A New Flaw Model

The pits that have been used in obtaining data that are described in this chapter
are the result of drilling into the aluminum hosts. The drill bits have pointed tips,
so the result of the drilling is to produce a pit that will have a right-conical tip at
its top. The resulting departure from the assumed truncated cylinder geometries is
not significant in most cases, but will be quite significant if the desired “pit” is quite
shallow. In this case we must modify our model of the pit to accommodate this
geometry. We will call this a new flaw model.
Figure 16.14 illustrates the new flaw “tip” model together with an equivalent
truncated cylinder “pit,” in the case that the pit is not deeper than the conical tip.
The truncated cylinder is assumed to have the same volume and centroid as the
conical tip.
322 16 Modeling Corrosion and Pitting Problems

16.5 Low-Frequency Models and Experiments

In order to get good resolution for very deep-lying pits in the 0.125-inch panels, it
is necessary to go to lower frequencies than we have used earlier. In this section we
will describe the results of experiments that use larger-diameter probes of 0.44 and
0.57 inch, excited at 360 and 720 Hz for this purpose. We will carry out a few model
calculations to determine such things as the dynamic range required of instruments
in this frequency range.
Figure 16.15 shows two cases that we will consider. The dimensions of the pits
(mm) are calculated using the “equivalent-pit” model of conical-tip flaw shapes,
as was described in the previous section. The coil diameter, D, is either 0.44 inch
or 0.57 inch. The flaw model on the left is called HAL 5 7 44 57.vic, and the
one on the right is called HAL 5 13 44 57.vic, which follows the notation of the
experiments.
Table 16.11 lists the experimental conditions and the names of the corresponding
VIC-3D R
models. The pit dimensions are nominal lengths; the actual dimensions
of the pits (hole diameter and lengths) are calculated as explained in the preceding
section.
Figure 16.16 illustrates the frequency response of the isolated pit in layer 1 as
a function of the coil diameter. Note that we have extended the scan range from
−19 mm to +19 mm in order to start and finish over unflawed host material.
We use the results of Fig. 16.16 to first determine the dynamic range required to
detect the isolated pit. To do this we compare the peak-to-peak response of the probe
as it scans the pit to the impedance of the probe over the unflawed host material. The
base-10 logarithm of this number is roughly a measure of the number of significant
digits that must be recorded by the test instrument, and is called the dynamic range.

D
0mm

1.458
P1 P1 0.5512
−3.175

P2 2.126

1.5444
−6.35

Fig. 16.15 Showing two low-frequency flaw models. The coil, whose diameter, D, is either
0.44 inch or 0.57 inch, is scanned past two independent flaw systems. The dimensions of the
cylindrical pits (mm) are calculated using the equivalent-pit model. The model on the left is
called HAL 5 7 44 57.vic, and the one on the right is HAL 5 13 44 57.vic, in accordance with
the notation of the experiments
16.5 Low-Frequency Models and Experiments 323

Table 16.11 Definition of experimental conditions for various VIC-3D R

models. The nominal
pit dimensions are lengths; the actual dimensions are calculated as explained in the preceding
section
VIC-3D R
-model Panel thick (in) Hole Diam (in-nom) Pit 1 (in-nom) Pit 2 (in-nom)
5 13 44 57 0.125 0.0625 0.03125 0
5 6 44 57 0.125 0.0625 0.03125 0.03125
5 8 44 57 0.125 0.0625 0.03125 0.0625
5 7 44 57 0.125 0.0625 0.03125 0.09375

Response_5_13_44_57 Response_5_13_44_57
0.001 0.006
5.588/360 5.588/360
7.239/360 7.239/360
5.588/720 0.005 5.588/720
0.0005 7.239/720 7.239/720
Resistance (Ohms)

0.004
Reactance (Ohms)
0
0.003

0.002
-0.0005

0.001
-0.001
0

-0.0015 -0.001
-20 -15 -10 -5 0 5 10 15 20 -20 -15 -10 5 0 5 10 15 20

Probe Position (mm) Probe Position (mm)

Fig. 16.16 Response of HAL 5 13 44 57. The title denotes coil diameter, 0.44 or 0.57 inch,
whereas the legend denotes coil radius in mm, and frequency, 360 or 720 Hz

Table 16.12 Dynamic range requirements in dB for measuring response of the isolated pit in layer
1, as a function of coil diameter and frequency
Radius (mm) Freq (Hz) Zhost Zpeak−peak Dynamic range (dB)
5.588 360 1.9617 + j37.691 0.0007 + j0.0007 69 + j94.7
7.239 360 4.1607 + j51.4588 0.00125 + j0.0013 70.4 + j92.0
5.588 720 5.7508 + j73.3232 0.0015 + j0.0028 71.7 + j88.4
7.239 720 11.445 + j98.2759 0.0021 + j0.005 74.7 + j85.9

This number can also be expressed as A/D bits by dividing by 0.3. The dynamic
range is sometimes expressed in decibels by dB = 20 log(ratio).
Table 16.12 tabulates the results of the calculations. The dynamic range require-
ments are not particularly severe, with the clear advantage going to the 0.57 inch coil
excited at 720 Hz. The dynamic range in reactance is a rather modest 85.9 dB, which
translates into 4.3 significant digits or an A/D converter with 14.3 bits of precision.
Typical instruments start at 16 bits, which give about 4.8 significant digits, or 96 dB
dynamic range.
Figures 16.17–16.20 illustrate the responses of the other model experiments.
Figure 16.20 compares the responses at the single frequency of 720 Hz and coil
324 16 Modeling Corrosion and Pitting Problems

Response_5_6_44_57 Response_5_6_44_57
0.0015 0.01
5.588/360 5.588/360
7.239/360 0.009 7.239/360
0.001 5.588/720 5.588/720
7.239/720 0.008 7.239/720
0.0005 0.007
Resistance (Ohms)

Reactance (Ohms)
0 0.006
0.005
-0.0005
0.004
-0.001 0.003

-0.0015 0.002
0.001
-0.002
0
-0.0025 -0.001
-20 -15 -10 -5 0 5 10 15 20 -20 -15 -10 -5 0 5 10 15 20
Probe Position (mm) Probe Position (mm)

Fig. 16.17 Response of HAL 5 6 44 57. The title denotes coil diameter, 0.44 inch or 0.57 inch,
whereas the legend denotes coil radius in mm, and frequency, 360 Hz or 720 Hz

Response_5_8_44_57 Response_5_8_44_57
0.004 0.016
5.588/360 5.588/360
7.239/360 0.014 7.239/360
0.003 5.588/720 5.588/720
7.239/720 7.239/720
0.012
0.002
Reactance (Ohms)
Resistance (Ohms)

0.01
0.001
0.008
0
0.006
-0.001
0.004
-0.002 0.002

-0.003 0

-0.004 -0.002
-20 -15 -10 -5 0 5 10 15 20 -20 -15 -10 -5 0 5 10 15 20
Probe Position (mm) Probe Position (mm)

Fig. 16.18 Response of HAL 5 8 44 57. The title denotes coil diameter, 0.44 inch or 0.57 inch,
whereas the legend denotes coil radius in mm, and frequency, 360 Hz or 720 Hz

diameter of 0.57 inch. This figure shows that the responses are well distinguished,
which means that at this frequency, and using this coil configuration, we should be
able to invert the measured impedance data and resolve the layer-two pits. For this
reason we recommend the use of the 0.57-inch coil operating at 720 Hz.
Finally, note in all of these figures, especially in the reactive components, the
significant dip that occurs when the coil is centered over the flaw. This is another
manifestation of the response that we have seen earlier, in that a very large coil will
have a very significant valley when the coil is so positioned. This valley will serve
as a “pointer” to the pit and should help resolve nearby pits in the lateral direction.
16.5 Low-Frequency Models and Experiments 325

Response_5_7_44_57 Response_5_7_44_57
0.005 0.02
5.588/360 5.588/360
0.004 7.239/360 7.239/360
5.588/720 5.588/720
0.003 7.239/720 0.015 7.239/720

Reactance (Ohms)
Resistance (Ohms)

0.002

0.001 0.01

-0.001 0.005

-0.002

-0.003 0

-0.004

-0.005 -0.005
-20 -15 -10 -5 0 5 10 15 20 -20 -15 -10 -5 0 5 10 15 20

Probe Position (mm) Probe Position (mm)

Fig. 16.19 Response of HAL 5 7 44 57. The title denotes coil diameter, 0.44 inch or 0.57 inch,
whereas the legend denotes coil radius in mm, and frequency, 360 Hz or 720 Hz

Response of Second-Layer Response of Second-Layer

Pits: D=0.57inch, F=720Hz Pits: D=0.57inch, F=720Hz
0.005 0.02
0% 0%
25% 25%
0.004 50% 50%
75% 0.015 75%
0.003
Resistance (Ohms)

Reactance (Ohms)

0.002
0.01

0.001

0 0.005

-0.001
0
-0.002

-0.003 -0.005
-20 -15 -10 -5 0 5 10 15 20 -20 -15 -10 -5 0 5 10 15 20
Probe Position (mm) Probe Position (mm)

Fig. 16.20 Response of second-layer pits. The first-layer pit is fixed at a nominal value of 25%
through-wall. The legend denotes percent through-wall of the second-layer pits. The coil diameter
is fixed at 0.57 inch, and the frequency is fixed at 720 Hz

16.5.1 Inversion of Measured Data

Now, we return to the situation depicted in Fig. 16.15 and invert measured data taken
on this system. We label the various configurations by their nominal values, in which
the diameter of each pit is “fixed” at 0.0625 inch. In truth, the diameters vary with
depth, as indicated in Sect. 16.4, but since the “β -scaling” is done using the usual
nominal values, that’s what we’ll use. The depth of P1 is fixed at 25% through-
326 16 Modeling Corrosion and Pitting Problems

Table 16.13 Actual ph2(nominal) pd1(mm) ph1(mm) pd2(mm) ph2(mm)

parameters for the various
pit-models 00 1.4588 0.5512 0
25 1.4588 0.5512 1.458 0.5512
50 1.4588 0.5512 1.5265 1.3386
75 1.4588 0.5512 1.5444 2.126

cd-57_f720_ph2-75 cd-57_f720_ph2-75
0.0004 0.002
measured measured
0.0003 model model

0.0002 0.0015
Resistance (Ohms)

Reactance (Ohms)
1e-04

0 0.001

-0.0001

-0.0002 0.0005

-0.0003

-0.0004 0

-0.0005

-0.0006 -0.0005
-20 -15 -10 -5 0 5 10 15 20 -20 -15 -10 -5 0 5 10 15 20
Probe Position (mm) Probe Position (mm)

Fig. 16.21 Comparison of model and measured data for the ph2-75 pit. The data are taken with
the 0.57-inch probe at 720 Hz. Left: resistance; Right: reactance

wall, which is equal to 0.79375 inch, and the depths of P2 are 0% through-wall
(0 mm), 25% through-wall (0.79375 mm), 50% through-wall (1.5875 mm), and 75%
through-wall (2.38125 mm). Data were taken at 360 and 720 Hz using ferrite-core
probes of nominal diameters 0.44 and 0.57 inch. We’ll describe the results for the
0.57-inch diameter probe at 720 Hz, which we label cd-57 f720, here.
In order to generate the most accurate interpolation table, which will be used
with NLSE to invert the data, we tuned the model by using the more accurate data
of Sect. 16.4 for the largest pit, and introducing a simple “fast-probe” model to
facilitate generating the interpolation table. The inner radius of the probe is 0 mm,
the outer radius is 5.588 mm, the height is 2.54 mm, and there are 1,075 turns. These
parameters were chosen to give the best fit to the ph2-75 model and measured data.
The actual parameters for the various pit-models are shown in Table 16.13.
Figure 16.21 shows the result of tuning the probe and ph2-75 model to the
measured data. We emphasize that only the ph2-75 measured data were used in
determining the fast-probe model. Figures 16.22–16.24 demonstrate that the other
measured data were scaled consistently, so that they agree with the fast-probe
model and the actual pit parameters tabulated in Table 16.13. No attempt has been
made to model the asymmetries in the measured impedance data (especially in the
resistance), as was done earlier when we “tilted” the probe model.
We assume that P1 is known (perhaps from UT data), so the problem is to
reconstruct P2. We do this by first generating an interpolation table based on the grid
16.5 Low-Frequency Models and Experiments 327

cd-57_f720_ph2-00 cd-57_f720_ph2-00
0.0002 0.0008
measured measured
model 0.0007 model
1e-04
0.0006
0
Resistance (Ohms)

Reactance (Ohms)
0.0005

-0.0001 0.0004

-0.0002 0.0003

0.0002
-0.0003
0.0001
-0.0004
0

-0.0005 -0.0001
-20 -15 -10 -5 0 5 10 15 20 -20 -15 -10 -5 0 5 10 15 20
Probe Position (mm) Probe Position (mm)

Fig. 16.22 Comparison of model and measured data for the ph2-00 pit. The data are taken with
the 0.57-inch probe at 720 Hz. Left: resistance; Right: reactance

cd-57_f720_ph2-25 cd-57_f720_ph2-25
0.0002 0.0012
measured measured
model model
1e-04 0.001
Resistance (Ohms)

0 0.0008
Reactance (Ohms)

-0.0001 0.0006

-0.0002 0.0004

-0.0003 0.0002

-0.0004 0

-0.0005 -0.0002
-20 -15 -10 -5 0 5 10 15 20 -20 -15 -10 -5 0 5 10 15 20
Probe Position (mm) Probe Position (mm)

Fig. 16.23 Comparison of model and measured data for the ph2-25 pit. The data are taken with
the 0.57-inch probe at 720 Hz. Left: resistance; Right: reactance

of Fig. 16.25. In creating this table, we fixed P1 using the parameters of Table 16.13,
while allowing the P2 parameters to vary according to Fig. 16.25. The results are
shown in Tables 16.14–16.17.
In order to validate our interpolation table, we inverted the model data of
Figs. 16.23 and 16.24, and obtained the results shown in Tables 16.18 and 16.19,
respectively. The results, especially for the model ph2-50 data, indicate that the table
is correct. The discrepancies in the results for the ph2-25 data may be due to the fact
that we did not put enough cells in the grid of the smaller pit while computing the
response with VIC-3D R
.
328 16 Modeling Corrosion and Pitting Problems

cd-57_f720_ph2-50 cd-57_f720_ph2-50
0.0004 0.0016
measured measured
0.0003 model 0.0014 model

0.0002 0.0012
Resistance (Ohms)

Reactance (Ohms)
1e-04 0.001

0 0.0008

-0.0001 0.0006

-0.0002 0.0004

-0.0003 0.0002

-0.0004 0

-0.0005 -0.0002
-20 -15 -10 -5 0 5 10 15 20 -20 -15 -10 -5 0 5 10 15 20
Probe Position (mm) Probe Position (mm)

Fig. 16.24 Comparison of model and measured data for the ph2-50 pit. The data are taken with
the 0.57-inch probe at 720 Hz. Left: resistance; Right: reactance

Fig. 16.25 A 5 × 5 PH2(mm)

interpolation grid for
inverting P2 U V W X Y
2.8

P Q R S T
2.1

K L M N O
1.4

F G H I J
0.7

0.0 A B C D E PD2(mm)
0.4 0.8 1.2 1.6 2.0

Table 16.14 Results for ph2-00 57 720. σ denotes the sensitivity of the
parameter at the solution point
Order Φ PD2/σ PH2/σ
1 0.2183248(−2) 1.200/0.9015 0.1514/0.4280
2 0.2179565(−2) 1.414/1.1797 0.1107/0.3084
3 0.2178009(−2) 1.473/1.2648 0.1024/0.2840
4 0.2172408(−2) 1.974/2.0086 0.0640/0.1757

Tables 16.20–16.23 display the results of inverting the measured data, but with
the resistance data suppressed.
16.5 Low-Frequency Models and Experiments 329

Table 16.15 Results for Order Φ PD2/σ PH2/σ

ph2-25 57 720. σ denotes the
sensitivity of the parameter at 1 0.1884036(−2) 2.0/0.3593 0.3222/0.1631
the solution point 2 0.1877925(−2) 2.0/0.3631 0.3123/0.1574
3 0.1877159(−2) 2.0/0.3638 0.3107/0.1594
4 0.1875607(−2) 2.0/0.3706 0.3039/0.1601

Table 16.16 Results for Order Φ PD2/σ PH2/σ

ph2-50 57 720. σ denotes the
sensitivity of the parameter at 1 0.2083930(−2) 1.201/0.0749 2.2859/1.1356
the solution point 2 0.2082139(−2) 1.198/0.0800 2.2972/1.1161
3 0.2082175(−2) 1.198/0.0793 2.2981/1.1196
4 0.2081406(−2) 1.196/0.0792 2.3090/1.1211

Table 16.17 Results for Order Φ PD2/σ PH2/σ

ph2-75 57 720. σ denotes the
sensitivity of the parameter at 1 0.2875837(−2) 1.492/0.1077 2.1000/0.7593
the solution point 2 0.2878217(−2) 1.502/0.1049 2.0782/0.8968
3 0.2878120(−2) 1.502/0.1051 2.0783/0.9077
4 0.2877737(−2) 1.500/0.1050 2.0872/0.9029

Table 16.18 Results for the Order Φ PD2 PH2

ph2-25 model data of
Fig. 16.23 1 0.9096408(−4) 1.565 0.5172
2 0.2538597(−4) 1.600 0.4860
3 0.2558663(−4) 1.604 0.4840
4 0.1026699(−4) 1.533 0.5176

Table 16.19 Results for the Order Φ PD2 PH2

ph2-50 model data of
Fig. 16.24 1 0.1036414(−4) 1.554 1.3017
2 0.7781143(−6) 1.528 1.3393
3 0.1187602(−5) 1.526 1.3423
4 0.9500942(−6) 1.527 1.3401

Table 16.20 Results for Order Φ PD2/σ PH2/σ

ph2-00 57 720 with
resistance data suppressed. σ 1 0.9401064(−3) 0.8000/0.2364 0.4476/0.4364
denotes the sensitivity of the 2 0.9390695(−3) 0.9116/0.2063 0.3310/0.3223
parameter at the solution 3 0.9387865(−3) 1.0314/0.2415 0.2536/0.2494
point 4 0.9373913(−3) 1.2189/0.3158 0.1756/0.1750

Table 16.21 Results for Order Φ PD2/σ PH2/σ

ph2-25 57 720 with
resistance data suppressed. σ 1 0.1097437(−2) 1.2536/0.1016 0.7000/0.2057
denotes the sensitivity of the 2 0.1089299(−2) 1.8330/0.1851 0.3681/0.1080
parameter at the solution 3 0.1087329(−2) 2.0000/0.2094 0.3241/0.0960
point 4 0.1077701(−2) 2.0000/0.2113 0.3179/0.0957
330 16 Modeling Corrosion and Pitting Problems

Table 16.22 Results for Order Φ PD2/σ PH2/σ

ph2-50 57 720 with
resistance data suppressed. σ 1 0.1307271(−2) 2.000/0.1152 0.7231/0.1479
denotes the sensitivity of the 2 0.1303683(−2) 2.000/0.1145 0.7222/0.1294
parameter at the solution 3 0.1303743(−2) 2.000/0.1144 0.7221/0.1289
point 4 0.1304370(−2) 2.000/0.1139 0.7220/0.1306

Table 16.23 Results for Order Φ PD2/σ PH2/σ

ph2-75 57 720 with
resistance data suppressed. σ 1 0.1761228(−2) 2.000/0.1035 1.1302/0.1993
denotes the sensitivity of the 2 0.1741217(−2) 2.000/0.1027 1.1035/0.2017
parameter at the solution 3 0.1739535(−2) 2.000/0.1023 1.1035/0.2018
point 4 0.1742446(−2) 2.000/0.1011 1.1044/0.2011

16.5.2 Comments and Conclusions

The fourth-order result of Table 16.14 indicates that PH2 is quite small, which
agrees with the actual parameter for ph2-00 in Table 16.13. The values of PD2 in
Table 16.14 are not well determined, of course, because of the small value of PH2.
The results for ph2-75, shown in Table 16.17, also agree quite well with the actual
parameters listed in Table 16.13, for both PD2 and PH2.
The results for PD2 and PH2 for ph2-25 and ph2-50, given in Tables 16.15
and 16.16, respectively, though not accurate, display an interesting feature: the
volume of each of the pits is reproduced accurately. Consider the fourth-order result
for ph2-25 in Table 16.15; we have for the volume, PH2∗ π ∗PD22 /4 = 0.9547 mm3 .
The “true” volume for this pit is computed from Table 16.13 to be 0.9203, which
differs by 3.6% from the “measured” volume. Similarly, the “measured” volume of
ph2-50, as given by the fourth-order result of Table 16.16, is equal to 2.5940 mm3 ,
whereas the true volume is equal to 2.45 mm3 , and differs from the measured volume
by 5.9%.
Tables 16.20–16.23 show that suppressing the resistance data (which are quite
ugly when compared to the reactance data in Figs. 16.21–16.24) has a moderately
stabilizing effect on PD2, in that all of the results cluster closely to each other
(except for ph2-00), as is the case with the actual data of Table 16.13, but, of course,
the measured PD2’s are not close to the actual values. The PH2’s follow an orderly
ascent for ph2-00 to ph2-75, but for the latter four the values are off by a factor of
about 2. So, it is difficult to conclude precisely the role played by the resistance data,
except to say that it is useful, and it needs to be measured more accurately than it
is now.
An alternative inverse problem is to choose a value of PD2 for each flaw, and then
fix that parameter during the inversion and determine the optimum estimate of the
height, PH2, of each flaw. This corresponds to our notion, expressed earlier in this
section, of fixing a priori the lateral resolution of a simple or complex pit, and then
determining the height of the pit. The question is how to determine the resolution.
One way of doing this, if we simply don’t want to postulate a lateral resolution, is to
16.5 Low-Frequency Models and Experiments 331

Table 16.24 Results of applying the POCS algorithm

to ph2-50 57 720. β0 is the scale factor relating the
impedance after the first iteration to the original data, β1
is the scale factor relating the impedance after the second
iteration to that after the first, and β2 is the scale factor
relating the impedance after the third iteration to that after
the second. We do not go any further because β2 is very
close to unity. The “Error” is between the final computed
geometry and the “Nominal” geometry
NLSE STEP (PH2,PD2) β
0 (2.052,1.24) 0.98757 − j0.072423
1 (1.7884,1.3052) 0.98852 − j0.013285
2 (1.5528,1.3660) 0.99960 − j0.014508
3 (1.2625,1.4891)
Nominal (1.3386,1.5265) Error: (6.0%, 2.5%)

take, say, 500 random samples of PD2 and PH2, and determine Φ for each of these
samples. Then we order Φ from smallest to largest in the ensemble, choose PD2 for,
say, the smallest 10 values of Φ , and average the results to give us a prior estimate
of PD2. The inversion is then completed as a one-dimensional problem for PH2. We
will call this the “exhaustive-search” method.
We did this for ph2-25 and ph2-50, and got for ph2-25 values of PD2 = 1.324,
PH2 = 0.5997, and for ph2-50, PD2 = 1.24, PH2 = 2.052. The results for ph2-25
are not in bad agreement with Table 16.13, being off by about 10% in diameter and
about 9% in height. The errors in PD2 and PH2 for ph2-50 are, respectively, 23%
and 53%. Clearly, the measured data for ph2-50 are not in good agreement with the
data that should correspond to the pit whose dimensions are given in Table 16.13.
In general, however, we can conclude that the use of a 0.57-inch probe, operating
at 720 Hz is well suited for characterizing second-layer pits in a system comprising
two 0.125-inch panels.

16.5.3 Applying the POCS Algorithm

As we have stated many times before, the resistance data are the least reliable and
most prone to producing poor reconstructions. This is certainly true of the data set
ph2-50 57 720. The POCS algorithm of Fig. 14.3 was developed to ameliorate this
problem and was applied to this data set.
The results of running the POCS algorithm on ph2-50 57 720 are listed in
Table 16.24.
The input impedance data generated by the POCS algorithm are shown in
Fig. 16.26. Note that the resistance data approach the “model” with each iteration.
The reactances behave in a similar manner, but do not change as strongly from
iteration to iteration. Only the measured reactance data are used in each iteration. It
332 16 Modeling Corrosion and Pitting Problems

Application of POCS to ph2-50 Application of POCS to ph2-50

0.0004 0.0016
measured measured
0.0003 1 0.0014 1
2 2
3 3
0.0002 0.0012 model
model
Resistance (Ohms)

Reactance (Ohms)
1e-04 0.001

0 0.0008

-0.0001 0.0006

-0.0002 0.0004

-0.0003 0.0002

-0.0004 0

-0.0005 -0.0002
-20 -15 -10 -5 0 5 10 15 20 -20 -15 -10 -5 0 5 10 15 20
Probe Position (mm) Probe Position (mm)

Fig. 16.26 Input impedance data generated by the POCS algorithm applied to ph2-50. The
“measured” and “model” data are repeated from Fig. 16.24. The numbered data correspond
to (PH2, PD2)1 , (PH2, PD2)2 and (PH2, PD2)3 given in Table 16.24. Left: resistance; Right:
reactance

is clear that the improvement in the reconstruction is due to the improved resistance
data during the iterations. Finally, we stress that we never use the “model” data
because the model is not known to us a priori. Indeed, determining the model is the
reason for doing the inversion in the first place.

16.6 Results of a 2D Circumferential Sweep Feature

Extraction Algorithm

We have developed a new feature extraction algorithm for axisymmetric pits, in

which averaging in the circumferential direction around the pit center is proposed
to form the “feature vector” as a function of radial position. This method is made
possible because all of the data are taken as 2D raster scans. Figure 16.27 compares
the result of the earlier data for the ph2-25 pit, shown in Fig. 16.23, which we call
“1D” with the new data, called “2D”, together with the model results for the nominal
ph2-25 pit dimensions shown in Table 16.13. Clearly, the resistance data are much
improved.
Curiously, when we apply the 2D data to the usual inversion algorithm with the
grid of Fig. 16.25, we get results that are not significantly different than those shown
in Table 16.15. This may be due to some peculiar mathematical feature of the grid
when applied to the ph2-25 data. We recall in Table 16.18 that even the model
ph2-25 data did not produce as good an inversion as did the model ph2-50 data
in Table 16.19.
If, however, we use the exhaustive-search algorithm, then we get a significant
improvement when using the 2D data over the 1D data. For example, we ran the
16.6 Results of a 2D Circumferential Sweep Feature Extraction Algorithm 333

Response of Feature Vectors for ph2-25 Response of Feature Vectors for ph2-25
0.0002 0.0012
1D 1D
2D 2D
1e-04 model 0.001 model
Resistance (Ohms)

0 0.0008

Reactance (Ohms)
-0.0001 0.0006

-0.0002 0.0004

-0.0003 0.0002

-0.0004 0

-0.0005 -0.0002
-20 -15 -10 -5 0 5 10 15 20 -20 -15 -10 -5 0 5 10 15 20
Probe Position (mm) Probe Position (mm)

Fig. 16.27 Comparison of processed data with model data when 1D and 2D feature vectors are
used. Left: resistance; Right: reactance

Table 16.25 Results for ph2-25 57 720 2DFV. This is the

first iteration in the application of the projection algorithm.
σ denotes the sensitivity of the parameter at the solution point
Order Φ PD2/σ PH2/σ
1 0.1200470(−2) 1.2811/0.1072 0.7000/0.2088
2 0.1169658(−2) 1.8294/0.1840 0.3834/0.1126
3 0.1165389(−2) 2.0000/0.2086 0.3366/0.0995
4 0.1141903(−2) 2.0000/0.2082 0.3297/0.0982

random-PD2 algorithm twice, getting PD2 = 1.4866 the first time, and PD2 =
1.4758 the second. The results for PH2 and Φ in each case were PH2 = 0.5444,
Φ = 0.1200699 × 10−2 for the first, and PH2 = 0.5508, Φ = 0.1200684 × 10−2 for
the second. These results are in excellent agreement with the nominal values of PD2
= 1.458 and PH2 = 0.5512 that are listed in Table 16.13. The corresponding results
for the 1D-data of Sect. 16.5.1 were PD2 = 1.324, PH2 = 0.5997.
Furthermore, when we force PD2 = 1.458 in NLSE, we get a value of PH2 =
0.5193 when using the 1D-data, and PH2 = 0.5618 when using the 2D-data. The
former error is 5.79%, whereas the latter error is 1.92%, so we see that the 2D-data
are much better.

16.6.1 Application of the POCS Algorithm

Now we’ll go back and apply the projection (POCS) algorithm of Fig. 14.3 to the
2DFV data. The result of the first iteration is shown in Table 16.25. We will use
the second-order result for generating the input for the second step, because the
334 16 Modeling Corrosion and Pitting Problems

Table 16.26 Results for PD2 = 1.8294, PH2 = 0.3834, when

projected back to the original reactance data. This is the
second iteration in the application of the projection algorithm.
σ denotes the sensitivity of the parameter at the solution point
Order Φ PD2/σ PH2/σ
1 0.9301147(−3) 1.2916/0.0805 0.7000/0.1594
2 0.9131988(−3) 1.7256/0.1300 0.4241/0.0975
3 0.9121440(−3) 1.8594/0.1451 0.3793/0.0871
4 0.8992282(−3) 2.0000/0.1615 0.3348/0.0775

Table 16.27 Results for PD2 = 1.7256, PH2 = 0.4241, when

projected back to the original reactance data. This is the third
iteration in the application of the projection algorithm. σ
denotes the sensitivity of the parameter at the solution point
Order Φ PD2/σ PH2/σ
1 0.9289556(−3) 1.2921/0.0829 0.7000/0.1591
2 0.9128185(−3) 1.7188/0.1291 0.4270/0.0981
3 0.9118760(−3) 1.8409/0.1428 0.3852/0.0885
4 0.8996030(−3) 2.0000/0.1614 0.3350/0.0775

first-order result is “trapped” at the node PH2 = 0.7 of the interpolation grid of
Fig. 16.25. This often happens with first-order results because the interpolation
table is linear in this order, which means that there will be discontinuities in the
derivative at each node. This discontinuity is suspected to be the culprit in trapping
the solution at a node. The third- and fourth-order results are artificially constrained
at the boundary PD2 = 2.0 of the grid, so we reject those solutions, as well.
When we use PD2 = 1.8294, PH2 = 0.3834 to generate our next input
impedance, which is then projected back to the original reactance data and fed into
NLSE, we get the results shown in Table 16.26.
The third iteration, which uses the results for PD2 = 1.7256, PH2 = 0.4241 as its
input data (after projection), is shown in Table 16.27. Clearly, the 2nd-order results
are converging slowly to the nominal values of PD2 = 1.458, PH2 = 0.5512.

16.6.2 Validation with Real Corrosion Pits

In addition to the examples using artificial pits as described in the earlier sections of
this chapter, validation of this methodology has also been made using a sample with
real corrosion pitting. The pits were generated by an accelerated corrosion process:
ASTM B117 (salt fog test chamber at 95 ◦ F, modified Exco solution, pH 3.2) in
7075-T6 aluminum. Figure 16.28 shows the methods of imaging these pits.
Figure 16.29 shows various images of pits as obtained by the methods shown
in Fig. 16.28. One of the more significant pitting regions was selected from a set
16.6 Results of a 2D Circumferential Sweep Feature Extraction Algorithm 335

Fig. 16.28 Sketch illustrating the methods of imaging real corrosion samples. (a) Laser pro-
filometry, (b) ultrasonic/eddy-current (far 1st layer), (c) eddy-current/ultrasonic (near 2nd layer),
(d) eddy-current/ultrasonic (near 2nd layer with 1st layer simulated pit)

Fig. 16.29 Real corrosion pits. Top left: photograph, Top right: laser profilometry image, Bottom
left: ultrasonic image (pits in far side of single layer), Bottom right: eddy-current image (pits in
near side of second layer, no corrosion in first)

of samples evaluated using laser profilometry (top right of Fig. 16.29). Ultrasonic
time-of-flight data were also acquired to evaluate the pit height and diameter (bottom
left of Fig. 16.29). The pit height was determined to be 0.020 by laser profilometry
and 0.026 using an ultrasonic method. However, there may be some error in the
laser profilometry results due to the presence of corrosion by-product. From the
image data for each method, an estimate of equivalent pit diameter was also made,
varying from 0.040 by laser profilometry and 0.100 using ultrasonic method. It is
336 16 Modeling Corrosion and Pitting Problems

Fig. 16.30 Real corrosion images. Left: eddy-current (far 1st layer at 6.2 kHz), Right: eddy-current
(near 2nd layer at 1.7 kHz)

likely that the pit diameter is closer to the value determined by laser profilometry
given beam spread in the ultrasonic measure. Using the proposed methodology for
second-layer pit characterization with eddy-current data at a single frequency of
6.2 kHz (bottom right of Fig. 16.29), the pit depth was estimated to be 0.029 and
the pit diameter was estimated to be 0.058. These results are in good agreement
with the known values for the pit and provide validation of this general approach. In
particular, the pairing of the iterative inversion (POCS) scheme for noise rejection
with an exhaustive search varying initial conditions was found to be beneficial for
achieving these results (Fig. 16.30).
Chapter 17
Applications to Aerospace Structures

17.1 Inspection of Fastener Sites in Aircraft Structures

A problem of particular interest in the field of nondestructive evaluation is the

inspection of fastener sites in aircraft structures for fatigue cracks. An important
class of structures comprises plane-parallel layered media representing joints in
both fuselage and wing locations. The layer stackup typically consists of two to
four panels and often includes thin layers of a nonconducting sealant or adhesive
between the panels. Additional complexity for computational electromagnetics
concerns the fastener shape (countersunk or buttonhead fastener) and material type
(where ferrous materials are prevalent). Also, fatigue cracks of complex morphology
provide a particular challenge for representation using numerical methods. Prior
work has investigated the problem of modeling an eddy-current inspection of
fastener sites in multilayered structures for fatigue cracks using a volume-integral
equation approach [102]. Although good agreement was achieved with experimental
results, simplications in the model representation were used. In this chapter, the
capability of the volume-integral equation approach with spatial decomposition
algorithms is demonstrated to fully address the problem of ferromagnetic fastener
sites in multilayered structures with gaps between the layers and cracks emanating
from the holes.

17.2 The Cessna Sandwich

To demonstrate the capability of the volume-integral equation method, the “Cessna

Sandwich” example is presented (Amos, J., Pendse, V., Shashikiran, A.: Cessna
Aircraft Co., private communication). Figure 17.1 illustrates the geometry of
the Cessna-Sandwich, which consists of three layers of aluminum comprising
two different aluminum alloys (conductivities), two “air gaps” representing the
nonconducting sealant layers, and a titanium rivet-insert with a countersunk head

H.A. Sabbagh et al., Computational Electromagnetics and Model-Based Inversion, 337

Scientific Computation, DOI 10.1007/978-1-4419-8429-6 17,
© Springer Science+Business Media New York 2013
338 17 Applications to Aerospace Structures

Probe

0mm
6.35mm
1mm Head_Ti: σ = 1.098E6 S/m

7075 :
σ = 2.325 E7 S/m
4.064mm 2.032
Gap:
2.282
Shank_Ti:σ =1.098E6 S/m

Clip_2024:
σ = 1.728 E7 S/m 7.628mm

4.822
Gap:
5.072

Skin_2024:
σ = 1.728 E7 S/m

8.628

Fig. 17.1 The Cessna-Sandwich problem

and a shank connecting all layers [102]. In the model, the layers and gaps make
up a uniform background or host (in x and y directions) with the fastener being an
“anomaly” requiring a local volume element mesh. Because the anomaly extends
through several layers of material with different electromagnetic properties, we must
use VIC-3D R
that has been augmented with the spatial-decomposition algorithm
(SDA) of Chap. 10.
Two model calculations were run based upon Fig. 17.1, the first with both
gaps filled with air, and the second with the top gap filled with 7075-aluminum
(σ = 2.325 × 107), and the other gap filled with 2024-aluminum (σ = 1.728 × 107).
The probe was a simple air-core coil with inner and outer radii of 3.02 and 5.14 mm,
respectively, and a height of 2.48 mm. Both model calculations were run with a
coil lift-off of 1.0 mm, and a frequency of 2,500 Hz. Five spatial-decomposition
grids were used to represent the “anomaly” regions, each with 8 × 8 × 2 cells,
yielding a total of 1,920 unknowns. The results are shown in Figs. 17.2 and 17.3.
The plot labeled “Gap” in Fig. 17.2 corresponds to the air-filled gaps of Fig. 17.1,
and the plot labeled “NoGap” corresponds to the case in which the gaps are filled
17.2 The Cessna Sandwich 339

Fig. 17.2 Model results for Model Results for the Cessna Double-Sandwich
the Cessna Double Sandwich
1
of Fig. 17.1
R:Gap
0.8 R:NoGap
X:Gap

Resistance/Reactance (Ohms)
X:NoGap
0.6

0.4

0.2

-0.2

-0.4

-0.6
-25 -20 -15 -10 -5 0 5
Probe Position (mm)

Fig. 17.3 Convergence of the Convergence With Grid Refinement

Cessna-Sandwich Problem
1
with grid refinement. The
coarsest grid is 4 × 4 × 2 cells R:4x4x2
0.8 R:8x8x2
per SDA grid, and the finest
R:16x16x2
grid is 16 × 16 × 2 cells per
Resistance/Reactance (Ohms)

0.6 X:4x4x2
SDA grid. There are five grids X:8x8x2
in all. A linear extrapolation 0.4 X:16x16x2
of the 4 × 4 × 2 and 8 × 8 × 2
results is within a percent or 0.2
so of the 16 × 16 × 2 results
0

-0.2

-0.4

-0.6

-0.8
-25 -20 -15 -10 -5 0 5
Probe Position (mm)

with aluminum. The solution time for the “Gap” run with 26 probe scan points is
about 6 min on an AMD/Athlon machine, whereas the “NoGap” run took about
2 min. Figure 17.3 shows convergence of the Cessna-Sandwich Problem with grid
refinement.
340 17 Applications to Aerospace Structures

Fig. 17.4 Diagrams of fastener site FEM model with two gap layers between three aluminum
panels

For the VIC-3D R

calculations shown in Fig. 17.2, it is clear that the effect of
the air-filled gaps is to increase somewhat the peak magnitude of the impedance
response when compared to the system with aluminum-filled gaps. The results in
Fig. 17.2 are changes in the driving-point impedance of the coil due to the presence
of the anomaly, in this case the rivet, compared to the “host-only” impedance in
the absence of the anomaly. As such, it is not unusual to have negative values of
resistance. Consider, for example, the results at a probe position of −4.0 mm, for
which δ Zgap = −0.22 + j0.41 Ω and δ Zno gap = −0.195 + j0.35 Ω. In the former, the
presence of the rivet insert reduces the resistance of the host structure with the gap,
compared to the situation with aluminum filling the original gaps. On the other hand,
the presence of the gaps increases the stored magnetic energy, which is manifested
in the slightly larger value of reactance.

17.2.1 Comparison with Finite-Element Method Results

To provide a baseline for this challenge-problem concerning the performance the

volume-integral equation method, an FEM model for the Cessna-Sandwich problem
was implemented in Opera-3D c software package (V9.0) [103]. The numerical
formulation of FEM is well established in the literature [104, 117]. A diagram of
the model is shown in Fig. 17.4. Irregular meshes of tetrahedral elements were used
17.2 The Cessna Sandwich 341

Fig. 17.5 FEM and Results for Cessna Double-Sandwich with Ti Insert
VIC-3D R
model results for 0.6
the Cessna Sandwich with
two airgaps 0.5 VIC3D:R

Resistance/Reactance (Ohm)
X
0.4 Opera3D:R
X
0.3

0.2

0.1

-0.1

-0.2

-0.3
-25 -20 -15 -10 -5 0 5 10 15
Probe Position (mm)

with the finite-element formulation to generally represent the complex geometries.

The mesh for the model required at least two elements per skin depth in the fastener
site region. The lateral dimensions of the plates in the model were extended well
beyond the field generated by the eddy-current coil to minimize the effect of the
part or model domain edges on the measurement response. Continuity conditions
were maintained between all part/air interfaces in the model. Far from the coil and
fastener site, the boundary conditions at the FEM model boundary were defined with
the tangential magnetic field equal to zero. The output of this model is the electric
and magnetic field intensities. Change in impedance can then be calculated using the
change in resistance associated with dissipated energy in the region of the conductor
and the change in inductance related to the stored energy in the whole solution
domain. More information on constructing eddy-current models in Opera3D c can
be found in [103].
Figure 17.5 displays the model results for the Cessna-Sandwich problem with
both airgaps and compares these results with the VIC-3D R
results of Fig. 17.2
labeled “Gap.” Clearly, the FEM response shows the same characteristics as does
the VIC-3D R
-generated solution. Some differences found in the magnitude of the
calculations are likely due to mesh-related error present in the models. Prior work
has shown very good agreement between FEM and VIE for this class of problem
[102]. Computation of the FEM model required 60 h on a 3.02 GHz Pentium 4
with 1 GB of memory. For this complex problem, the advantage of the compact
formulation with volume-integrals (1,920 unknowns) over FEM is obvious.
342 17 Applications to Aerospace Structures

17.3 Simulated Studies on Crack Characteristics:

The Cessna Setups

From the perspective of nondestructive evaluation, the critical requirement of

an NDE model is an accurate representation of the eddy-current measurement
associated with crack detection. Thus, the model must accurately represent the
measurement sensitivity to the crack condition with respect to features such as the
fastener site and gaps between layers. To explore the sensitivity of the numerical
model to varying crack conditions such as crack length and location, simulated
studies were performed using VIC-3D R
. Again, a class of problems based on setup
standards for Cessna (the “Cessna Setups”) was simulated in this study. The Cessna
Setups comprise the complex ring probe shown in Fig. 17.6, the “hi-lock” pin rivet
of Fig. 17.7, and the various test setups shown in Figs. 17.8–17.12. Test setup 0
is the root of the other setups, with each of the others differing by the placement
of the crack (shown in brown). It should be observed that the setups contain
both electrically conducting (lossy) media and ferromagnetic steels. The cracks are
modeled as empty slots.
In modeling the setups, we use the spatial-decomposition algorithm with five
grids, two above the air gap, and two below, with one within the gap. Each grid
has 16 × 16 × 2 cells, yielding a total of 15,360 electric and magnetic currents
to be determined. The ring probe core was modeled with 12,288 electric and
magnetic currents, so that the entire package of unknowns for each setup is 27,648.
The models were run at a frequency of 1.0 kHz.

Balance Coil

Shield
σ = 4.06E6
μ = 40

Cap Cap

Test Coil
Fig. 17.6 The ring probe
17.3 Simulated Studies on Crack Characteristics: The Cessna Setups 343

Fig. 17.7 The hi-lock pin 0.2612"

rivet 0.010"
0.030"

σ = 4.06 E6 S/m
μ = 40 0.112"

0.1635"

0.05" Al

0.00196"

Al: σ=1.728E7 S/m σ =4.06E6 S/m

0.10" μ=40
μ=1

0.00196"
SETUP 0

Fig. 17.8 Test setup 0

0.054"

0.05" Al Crack
Width=0.0107"
0.00196"

Al: σ=1.728E7 S/m σ =4.06E6 S/m

0.10"
μ=1 μ=40

0.00196"
SETUP 1

Fig. 17.9 Test setup 1

17.3.1 Analysis of the Ring Probe Circuit

The test (bottom) and balance (top) coils of the ring probe shown in Fig. 17.6 are
connected in two arms of a bridge within the eddy-current test instrument, as shown
in Fig. 17.13. Zt is the driving-point impedance of the test coil in the presence of the
workpiece (the “loaded impedance”) and Zb is the driving-point impedance of the
balance coil. By the construction of the ring probe, as shown in Fig. 17.6, the cap
344 17 Applications to Aerospace Structures

0.05" Al

0.00196"

Al: σ=1.728E7 S/m σ=4.06E6 S/m

0.10"
μ=1 μ=40 Crack Width=0.0191"

0.00196"
SETUP 2 0.1397"

Fig. 17.10 Test setup 2

0.0620"

0.05" Al Crack
Width=0.0087"
0.00196"

Al: σ=1.728E7 S/m σ=4.06E6 S/m

0.10"
μ=1 μ=40

0.00196"
SETUP 3

Fig. 17.11 Test setup 3

0.05" Al

0.00196"

Al: σ=1.728E7 S/m σ=4.06E6 S/m

0.10"
μ=1 μ=40 Crack Width=0.0108"

0.00196"
SETUP 4 0.1472"

Fig. 17.12 Test setup 4

separating the test and balance coils isolates them electrically, so we assume that the
mutual inductance between the two coils is small enough to ignore. That is, the cap
is a “flux-shunt” that shunts the flux produced by either coil from linking the other
coil. For the same reason, the balance coil is isolated from the workpiece, because
the workpiece is beneath the cap. This means that the driving-point impedance of
the balance coil will always be the freespace impedance, no matter where the probe
is positioned.
17.3 Simulated Studies on Crack Characteristics: The Cessna Setups 345

Fig. 17.13 The eddy-current Rin

test instrument bridge circuit

R0 R0
+
Ein Vout
−
+ −

Zt Zb

The bridge circuit of Fig. 17.13 is designed so that R0 is much larger than either
Zt or Zb . Under this condition, and the condition of zero mutual inductance between
the two coils, it is a simple matter to compute the response of the bridge circuit to be
Ein
Vout ≈ (Zt − Zb ) , (17.1)
R0 + 2Rin
which demonstrates that the bridge circuit acts as an impedance differentiator, and
that the impedance measured is balanced against the freespace impedance of the test
coil. This follows because the freespace impedance of the test and balance coils are
equal, because the coils are identically situated when in freespace. The output of
the test instrument is the “total signal” minus the “freespace signal,” which, in the
terminology of VIC-3D R
, corresponds to the “host + flaw impedance.”
The output of VIC-3D R
is already balanced against both the freespace and host
medium, so it appears that in order to compare the test-output with the VIC-3D R

output, we would have to add the host-only impedance, which will be computed in a
separate model. (All of the VIC-3D R
calculations will be done without the need to
model the balance coil.) Then we will scale the experimental data so that they agree
with the VIC-3D R
-generated model data and produce a scale-factor, β , which will
then be used on all other problems at the same frequency.
The practical upshot of all of this is that because the subtraction (balancing) of
the impedances is done analogically (in a bridge circuit), the probe requires two
identical coils, with identical shielding. If the balancing were done numerically, one
would need only the test coil, which would result in a much less expensive probe.

17.3.2 Model Results

Data for determining the presence of a flaw are obtained by varying the lift-off of the
ring probe over the range 0, 0.25, 0.7493, and 0.9398 mm. By “lift-off” we mean
the position of the bottom of the probe above the workpiece. The solution of the
346 17 Applications to Aerospace Structures

Fig. 17.14 Final scaled Final Scaled Model Curves for SETUPS 0-5
model results for Setups 0–5,
0.35
using Setup 0 as the 0
reference. The horizontal and 1
vertical axes are both 0.3 2
3
measured in instrument volts, 4
V. The parametric points on 0.25 5
the curves are lift-off values

Channel 1 (V)
of 0, 0,25, 0.7493, and
0.9398 mm (right to left) 0.2

0.15

0.1

0.05

0
-3 -2.5 -2 -1.5 -1 -0.5 0 0.5
Channel 2 (V)

problem with 27,648 unknowns takes about 35 min for each lift-off, giving a total
solution time of about 2.3 h for each setup. About 1Gb of storage for auxiliary files
was required.
VIC-3D R
computes impedances, but the industry typically uses analog
instruments to measure data, so the final results of the computation are transformed
into “instrument volts,” as shown in Fig. 17.14. Channel 1 is called the vertical
channel (on the oscilloscope) and indicates the difference signal between flawed
and unflawed rivet. Channel 2 is the horizontal channel and corresponds to lift-
off. The markers correspond to the four lift-off values of 0(mm), 0.25, 0.7493, and
0.9398, from right to left. Both channels (axes) are measured in instrument volts.
We showed in Chap. 15 how impedance calculations can be transformed into
voltage measurements using calibration data with known samples (see also [102]).
Setup 0 is the reference, so it is zero for each lift-off along the negative real axis.
The remaining curves indicate the presence of a flaw and yield information that can
be used in an inversion procedure.
It is clear that Setup 2 produces the largest signal, even though the flaw is in the
second layer. This is due to the fact that the flaw in this setup has a larger volume
than any of the other flaws. Furthermore, it is clear that the flaws in Setups 1 and
3, even though they are in the upper layer, are more difficult to distinguish than the
other two flaws. This is due to the fact that these two flaws have a smaller volume
than those in Setups 2 and 4. Setup 5 is the same as Setup 4, except that the length of
the crack is 0.0986 in., instead of 0.1472 in. The response of Setup 5 is about 0.67
times the response of Setup 4, which is consistent with the fact that the response is
roughly proportional to the volume of the defect.
17.4 Further Tests with the SDA and “Weak-Host” Layers 347

Table 17.1 Convergence of

σ3 Iterations δR δX δZ
Half-Sandwich No. 1 with σ3
105 106 −0.85638 1.8332 2.0234∠115.04
104 186 −0.86202 1.8394 2.0314∠115.11
103 361 −0.86425 1.8425 2.0351∠115.13
100 422 −0.86506 1.8438 2.0366∠115.14
10 428 −0.86513 1.8439 2.0368∠115.14
1 431 −0.86516 1.8439 2.0368∠115.14

These studies indicate that we are capable of modeling and inspecting

fastener-crack problems without removing the fastener, in contrast to the standard
practice which uses a split-D probe, of the type shown in Fig. 6.16 of Chap. 6, that
is inserted within the hole after the fastener is removed. Problems of this type, in
which the fastener is removed before inspection, are called bolt-hole eddy-current
(BHEC) problems, and we have already seen two examples of them in Figs. 10.23
and 10.24 of Chap. 10. In those examples, of course, the probe was a simple coil
that was scanned along the surface, and not inserted within the hole.
Lift-off data are generally too sparse to allow a good inversion. At the time that
the manuscript for this book was submitted to the publisher, the authors began
a program that will lead to the validation of inversions of fastener-flaw systems
without the need to remove the fastener, and using a surface probe of the type
shown in Figs. 10.23 and 10.24. The anticipated success of this project would lead
to a considerable simplification, with significant cost-savings, of the fastener-flaw
inspection problem.

17.4 Further Tests with the SDA and “Weak-Host” Layers

17.4.1 Half-Sandwich No. 1

The setup for this exercise is shown in Fig. 17.15. The insert can be ferromagnetic,
and we have run tests with σ = 1.79 × 106 S/m and μ = 200. In those runs, we
used three SDA grids, each with 32 × 32 × 4 cells, yielding over 62,000 unknowns.
The convergence to 1 % took 251 iterations. With a nonpermeable insert, the number
of unknowns is reduced to 33,024, and convergence to 1 % required 431 iterations.
Apparently the permeability yields a better conditioned system. In those runs, the
middle layer conductivity was σ3 = 0, which made it a “weak-host” layer.
Our interest, now, is in testing the effect of varying σ3 over several decades
in order to determine the effect of a nonzero conductivity in speeding up the
convergence of a problem in which σ = 0. For this exercise we made the insert
nonpermeable, keeping the conductivity equal to 1.79 × 106 S/m, and used a
convergence criterion of 1 %. The convergence results with respect to σ3 are shown
in Table 17.1.
348 17 Applications to Aerospace Structures

8x8

σ=1.876Ε7
μ=1 0.8128

Insert
σ, μ

σ3 0.25
μ=1

σ=1.876Ε7
μ=1
0.8128

Fig. 17.15 Setup for Half-Sandwich No. 1. All dimensions are in mm

The immediate observations are that the model is robust and consistent, in the
sense that there is a monotonic increase in the number of iterations required to
achieve 1 % convergence. This is an indication that the problem becomes slightly
less well conditioned as σ3 decreases, which is to be expected. Further indication of
the consistency in the results is that the magnitude of δ Z increases monotonically
with a decrease in σ3 , as we would expect because the electrical contrast of the
anomaly increases with a decrease in σ3 in the middle slab.
The most important observations, however, are that the phase of δ Z is virtually
independent of the value of σ3 , which means that the magnitude of δ Z, alone, is
sufficient to determine the effects of changing the conductivity of the slab, and
furthermore, the magnitudes change by less than 1 %, which means that we could
safely replace a slab of air with one of σ3 = 105 S/m and get the same answer, while
achieving a significant reduction of computation time. This could be extremely
important when doing a large scan. We shall see how this can be very effective
in the next example.

17.4.2 Half-Sandwich No. 2

This sandwich is shown in Fig. 17.16. Each of the three SDA grids has 16 × 16 × 2
cells, which produces a total of 768 electric unknowns and 498 magnetic unknowns.
17.4 Further Tests with the SDA and “Weak-Host” Layers 349

0.25 x 1.25

σ=1.876E7
μ=1 0.156
σ3 Insert 0.004

σ=1.875E6
σ=1.876E7
μ=200
μ=1 0.58

Fig. 17.16 Setup for Half-Sandwich No. 2. All dimensions are in inches

Table 17.2 Convergence of Half-Sandwich No. 2 with σ3 . The first four

entries are computed, and the fifth is extrapolated using the first three
entries
σ3 Iterations δ R × 10−5 δ X × 10−5 δ Z × 10−5
104 520 −5.012 −3.8032 6.2916∠ − 142.81
103 598 −5.4627 −4.1561 6.864∠ − 142.74
100 1,332 −5.9106 −4.4743 7.4132∠ − 142.87
0 4,268 −5.2723 −4.0658 6.6579∠ − 142.36
1(Extrap.) −6.798 −5.0066 8.4427∠ − 143.63

We do another convergence study with respect to σ3 , with the results shown in

Table 17.2. We note that even with a much smaller number of unknowns than in
Half-Sandwich No. 1, the convergence of the iterative solver to a precision of 1 %
requires a significantly larger number of iterations. We attribute this to the very
small transition layer depth compared to the top and bottom of the sandwich (this
conjecture will be supported in the next example: “Half-Sandwich No. 3”).
The first four entries in Table 17.2 are computed, while the fifth is extrapolated
using a quadratic algorithm that uses the first three entries. We’ll discuss that shortly,
but we want to first comment on the nature of the computed results. The first
three entries show a consistent monotonic increase in magnitude, while their phases
remain virtually equal, as we saw in Half-Sandwich No. 1. The fourth result, in
which σ3 = 0, does not fit the pattern, and that is due to the fact that our convergence
criterion of 1 % is too stringent, given the relatively poor conditioning of the
operator (the large number of iterations required to attain 1 % convergence indicates
this). Tightening the criterion to 0.5 % or smaller to get consistency would defeat the
purpose of extrapolating the better-conditioned models to obtain the desired model.
The extrapolation algorithm used to obtain the final result is quadratic and will
be developed here. Let y = ax2 + 2bx + c be the quadratic relation between the
impedance component and x = log σ3 , and let y2 , y3 , and y4 be values of the
impedance at x2 , x3 , and x4 , respectively (the subscript denotes the logarithm of σ3 ).
We then have
350 17 Applications to Aerospace Structures

y2 = 4a + 4b + c
y3 = 9a + 6b + c
y4 = 16a + 8b + c. (17.2)

These are quickly solved to yield

y4 − 2y3 + y2
a=
2
−5y4 + 12y3 − 7y2
b=
4
c = 6y2 − 8y3 + 3y4, (17.3)

where c is the sought-for solution for σ3 = 1, which is close enough to zero for our
purposes.
Referring to Table 17.2, we take y4 to be the entry for either δ R or δ X for σ3 =
104 , and so on for y3 and y2 . Using these values, we compute the entries for δ R, δ X,
and δ Z in the fifth row “(Extrap.)”. Our strategy, then, is to do the VIC-3DR
model
scans using σ3 = 10 S/m, and then multiplying these results by the ratio
4

8.4427e− j143.63
= 1.3419e− j0.82
6.2916e− j142.81
= 1.3418 − j0.0192 . (17.4)

17.4.3 Half-Sandwich No. 3

This is the same as Half-Sandwich No. 1, except that the gap thickness is 0.025 mm,
and the top and bottom halves of the sandwich are 0.9253 mm thick, yielding the
same overall height of the sandwich of 1.8756 mm. Our interest is in determining
the effect of gap thickness on convergence. For Half-Sandwich No. 3, after 1,001
iterations, the convergence criterion was reduced to only 6.559 %, which clearly
shows the deleterious effect that a very thin gap has on convergence. This supports
our contention that the slow convergence of Half-Sandwich No. 2 is largely due to
the very thin airgap.

17.5 Comments and Conclusions

We have shown that the volume-integral approach has significant advantages over
other numerical methods such as the finite-element method, in that the formulation
of the numerical model is much simpler in the former than in the latter. Furthermore,
the solution time for VIC-3D R
is extremely short for many problems in NDE,
17.5 Comments and Conclusions 351

because the formulation in terms of the Galerkin variant of the method of moments
on a regular grid results in operators that have very special structures; they are
either three-dimensional convolutions, or two-dimensional convolutions and one-
dimensional correlations, which means that we can use three-dimensional FFT’s
to evaluate them in a conjugate-gradient search algorithm. The use of a highly
irregular mesh in the finite-element technique does not allow a similar advantage
in the solution process.
This advantage accrues from the very different nature of the physics that goes
into the formulation of the mathematical models. In volume-integral equations, as
well as boundary-integral equations, the unknowns are anomalous currents that are
supported in a compact domain, namely the domain of the anomaly; in the example
of the Cessna Series, the anomaly is the rivet or rivet and crack. In finite-element or
finite-difference methods the unknowns are the electric and magnetic fields, which
extend to infinity. This has two disadvantages; it increases the number of unknowns
and requires some method of approximating the “boundary-at-infinity” in order to
truncate the problem domain. This increases the complexity in simply defining the
model and presents an extreme challenge to prospective users who are not skilled in
computational electromagnetics. Furthermore, the FEM method is not particularly
well suited to solve typical problems in NDE, because the anomalies, such as rivets
or cracks, require a very complicated mesh, with a large number of very irregular
cells.
Finally, we conclude from the model calculations of the Cessna Series that the
spatial decomposition algorithm is a very efficient method of solving problems
in which an anomaly extends through layers with different electric and magnetic
properties. The spatial decomposition algorithm formulation was also demonstrated
to be valid for cases where a conducting anamoly is present in a nonconducting layer
such as air. This approach greatly expands the capability of volume-integral equation
methods for complex problems in computational electromagnetics, and in particular
nondestructive evaluation. Lastly, rapid solution of the forward problem for com-
putational electromagnetics will also be beneficial for the practical application of
advanced inverse method techniques for quantitative nondestructive evaluation of
material discontinuities.
Chapter 18
Applications to Nuclear Power

The nuclear-power industry faces the serious challenge of convincing a skeptical

public and regulatory agencies that it can operate safely and efficiently. Nondestruc-
tive evaluation (NDE) plays a significant role in this task, and computer modeling is
playing a significant role in NDE. The industry now realizes the value of using such
modeling to replace expensive experimental tests, as well as to design equipment,
and interpret results. Eddy-currents have a traditional place in the inspection of
steam generator tubing, and the industry seeks improved tools for such inspections.
In this chapter we demonstrate the value of model-based inversion in providing
that tool. The problems that we solve are similar to those in the preceding chapters
of this book, with the main difference being the need to model workpieces with
cylindrical geometries rather than planar. The tools for carrying out this model were
developed in Chap. 9.
We start the chapter with a discussion of an extremely important problem in
the NDE of heat-exchanger tubes in nuclear power plants: the ability to model and
reconstruct pitting and corrosion phenomena on the inner diameter of these tubes.
We then extend our study to other classes of anomalies in tubes, while continuing to
emphasize the role of model-based inversion in solving these problems.

18.1 Modeling Pitting and Corrosion Phenomena

in Heat-Exchanger Tubes

18.1.1 An Ellipsoidal Model for Pits

The photomicrographs shown in Fig. 18.1 illustrate the results of long-term,

under deposit pitting corrosion on the inner-diameter of a section of a component

H.A. Sabbagh et al., Computational Electromagnetics and Model-Based Inversion, 353

Scientific Computation, DOI 10.1007/978-1-4419-8429-6 18,
© Springer Science+Business Media New York 2013
354 18 Applications to Nuclear Power

Fig. 18.1 Photomicrographs showing the effects of long-term, under deposit pitting corrosion of
a heat exchanger tube whose nominal wall thickness was 49–51 mils. Upper-left: little corrosion;
Upper-right: remaining wall of 29 mils; Lower-left: remaining wall of 20 mils; Lower-right: actual
pits, with an indication of 88 % through-wall within the middle third of the sample shown, based
on nominal 0.049 inch

cooling heat exchanger tube whose nominal wall thickness is 49–51 mils.1 They
further suggest that we can construct a reasonable model of corrosion pits as
three-dimensional semiellipsoids that originate in the inner-radius of tube walls.
Figure 18.2 illustrates the model, together with its three defining parameters, the
semiaxes, A, B, and C. By modeling the pit in this manner, we cast the problem into
the model-based inversion paradigm. In this case, we hope to determine the three
semiaxes, thereby solving the inverse problem. Clearly, C determines the depth of
the pit.
Furthermore, the photographs give us insight into the sizes of typical pits, which
can then be translated into values of A, B, and C for typical pits. Indeed, from
these photographs we see that the pits appear to be circular, when looking into the
inner surface of the tube, with a nominal radius of 0.0625 inch. Hence, we model
these pits by setting A and B equal to 0.0625. We then choose C to be equal to
the difference between 0.049 inch (the nominal wall thickness) and the minimum
measured wall thickness, as given in the photomicrographs. We assume that the pits
are empty, but, it would be quite easy to model pits filled with corrosion deposits,
whether conducting or magnetic (or both).

1 These photographs and the data concerning them are courtesy of Exelon PowerLabs.
18.1 Modeling Pitting and Corrosion Phenomena in Heat-Exchanger Tubes 355

TUBE AXIS
B

TUBE WALL

B
INNER SURFACE
OF TUBE WALL

Fig. 18.2 Illustrating a model of pits as three-dimensional semiellipsoids that originate in the
inner-radius of a tube wall. The ellipsoid is defined by its three semiaxes, A, B, and C

Table 18.1 Measured pit data for a component cooling heat exchanger tube
Minimum Calculated % thru-
measured wall based on Eddy current
Indication remaining wall (in) Pit depth (in) nominal 0.049 in % thru wall
No. 3 0.023 0.026 53 59
No. 4 0.015 0.034 69.4 88
No. 5 0.020 0.029 59.2 89
No. 7 0.016 0.033 67.3 65
No. 8 0.015 0.034 69.4 97
No. 9 0.017 0.032 65.3 100

18.1.2 Computing Impedance Signatures with VIC-3D

R

The pit data of Table 18.1 are measurements taken on the component cooling heat
exchanger tubes of Fig. 18.1. The first calculations that we make using VIC-3D R

are to determine the impedance signature versus depth for the various pits labeled
“Indication Nos. 3–9,” omitting No. 8, because it has the same depth as No. 4. In all
cases, we fixed A and B to be equal to 0.0625 inch, as explained above. The results
are shown in Fig. 18.3.
356 18 Applications to Nuclear Power

Fig. 18.3 Impedance Impedance Signature vs. Pit Depth

signature versus pit depth for 0.5
the five pits labeled 0.026 in
0.4 0.029 in
“Indication Nos. 3, 4, 5, 7, 9.” 0.032 in
The semiaxes A and B are set 0.3 0.033 in
equal to 0.0625 inch for all 0.034 in

Reactance (Ohms)
five, and C is equal to the pit 0.2
depth 0.1

-0.1

-0.2

-0.3

-0.4

-0.5
-0.05 -0.04 -0.03 -0.02 -0.01 0 0.01 0.02 0.03 0.04 0.05
Resistance (Ohms)

Clearly, the depth, as measured by the semiaxes, C, is quite distinguishable in

these curves because the slope of the corresponding Lissajous figure2 changes with
C. This is the standard eddy-current inspection technique and is based on visual
observation of the curve produced in an oscilloscope of an analog instrument.
It suffers from two disadvantages: (a) noise will usually obliterate the curve, making
it impossible to define a clear slope for the curve and (b) without having prior
knowledge of the lateral dimensions of the pit, the A and B semiaxes, it is impossible
to precisely define C. This is clear from Table 18.1, wherein the eddy-current
technique generally predicted a pitting depth that was larger than true.
In carrying out the model calculations, we assume that the tube is made of 90–10
Copper–Nickel, whose electrical conductivity is σ = 5.277 × 106 S/m (9.1 % IACS),
and is nonmagnetic. The outer diameter of the tube is 0.75 inch, and the nominal
wall thickness is 0.049 inch. The coil parameters are given in the table and the
frequency of operation is assumed to be 70kHz.

Coil spacing 0.060 inch

Coil height 0.055
Coil width 0.060
Coil OD 0.576
Coil ID 0.466
Coil turns 156

2 Lissajousfigures are normally associated with “differential-bobbin” probes, as described in

connection with Fig. 18.19.
18.1 Modeling Pitting and Corrosion Phenomena in Heat-Exchanger Tubes 357

Table 18.2 Estimate of A, B, and C for the semiellipsoidal representation of the pits
labeled Indication Nos. 3–9
Indication Original model data (in) Reconstructed (inverted) data (in)
No. 3 A = 0.0625, B = 0.0625, C = 0.026 A = 0.0643, B = 0.0634, C = 0.0263
No. 4 A = 0.0625, B = 0.0625, C = 0.034 A = 0.0644, B = 0.0639, C = 0.0325
No. 5 A = 0.0625, B = 0.0625, C = 0.029 A = 0.0645, B = 0.0637, C = 0.0285
No. 7 A = 0.0625, B = 0.0625, C = 0.033 A = 0.0645, B = 0.0639, C = 0.0316
No. 8 A = 0.0500, B = 0.0700, C = 0.034 A = 0.0550, B = 0.0686, C = 0.0328
No. 9 A = 0.0625, B = 0.0625, C = 0.032 A = 0.0645, B = 0.0639, C = 0.0309

18.1.3 The Inversion Problem

Now we consider the inversion problem in detail. We take the model data of
Fig. 18.3 and use them to simulate measured impedance data, which are then
submitted to NLSE. The output of the estimator is an estimate of the semiaxes,
A, B, and C of the ellipsoidal pit, as displayed in Table 18.2. It only takes a few
minutes to compute a table for the inversion algorithm, which is then used for all
pits. Subsequently, it takes about 2–4 s to compute the results for each pit.
Indication No. 8 was handled differently from the others. Because the X-Y
cross-section of this pit was not circular, we found it more reliable to acquire data
for it using a surface pancake probe scanned in a two-dimensional raster pattern.
These data were then submitted to the estimator exactly as were the other five.
The results displayed in Table 18.2 indicate that the algorithm for nonlinear
estimation is reliable and robust. The maximum error occurs in A for Indication
No. 8, and is only 10 %, which is quite tolerable for this dimension (the width
of the pit). The important point to be made from this study is the reliability in
reconstructing C, the semiaxes that determines the depth of the pit. Indeed, the
results of the algorithm indicate that the computation is more sensitive to depth
then to the lateral dimensions, which is exactly what we’re looking for!

18.1.4 Inversion of a Complex Pit

Figure 18.4 illustrates a complex pit, comprising three intersecting semiellipsoids.

The A and B semiaxes are 0.0625 inches, as before, and the C semiaxes are 0.029,
0.026, and 0.034 inches. In this exercise, we assume that the A and B semiaxes
are known, and the job is to determine the profile-in-depth of the pit, that is, to
determine the three C-axes. The input data to NLSE at six frequencies, 10, 20,
40, 70, 200, and 700 kHz are shown in Fig. 18.5. The results of applying NLSE
at these frequencies are shown in Table 18.3. The interesting result is the average
over the six frequencies. Clearly, the reconstruction in depth is robust, especially
when frequency scanning is incorporated along with spatial scanning. The middle
lobe, which is effectively lost in the presence of its two larger neighbors, is still quite
accurately reconstructed.
358 18 Applications to Nuclear Power

Tube Axis
Bobbin
Coil
0.0625 in 0.0625 in 0.0625 in 0.0625 in

0.026 in
0.029 in 0.034 in

Complex Semi−Ellipsoidal Pit

Tube Wall

Fig. 18.4 Illustrating a complex pit that comprises three intersecting semiellipsoids, whose
dimensions are shown

Impedance Response of Complex Pit Impedance Response of Complex Pit

0.9 8
10 kHz 10 kHz
0.8 20 kHz 7 20 kHz
40 kHz 40 kHz
0.7 70 kHz 70 kHz
200 kHz 6 200 kHz
0.6
Resistance (Ohms)

700 kHz 700 kHz

Reactance (Ohms)

5
0.5
0.4 4

0.3 3
0.2
2
0.1
1
0
-0.1 0

-0.2 -1
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
Probe Position (in) Probe Position (in)

Fig. 18.5 The model input impedance data to NLSE at 10, 20, 40, 70, 200, and 700 kHz: resistance
(left), reactance (right)

Table 18.3 Estimation of Freq (kHz) Result

three C-axes for the complex
semiellipsoidal pit model. 10 (0.026, 0.022, 0.032)
The row labeled “Freq. Avg.” 20 (0.026, 0.023, 0.032)
is the average of the results 40 (0.026, 0.023, 0.032)
over the six frequencies 70 (0.028, 0.023, 0.033)
200 (0.032, 0.026, 0.036)
700 (0.033, 0.030, 0.037)
Freq. Avg. (0.029, 0.025, 0.034)
True (0.029, 0.026, 0.034)
18.1 Modeling Pitting and Corrosion Phenomena in Heat-Exchanger Tubes 359

Stainless−Steel Block

Semi−Ellipsoidal Pit

Coil

Acrylic Box

Fig. 18.6 A multifrequency validation test. The coil is excited at 10 frequencies between 0.1 and
1.0MHz

18.1.5 A Multifrequency Benchmark Inversion Test

In order to validate our inversion algorithm that utilizes NLSE, we perform a

multifrequency experiment on a block of stainless steel that contains a machined
semiellipsoid, whose semiaxes were nominally 0.0625×0.0625×0.026 in. The data
are obtained by simply laying the stainless-steel block over a coil that is embedded
in an acryllic block, as in Fig. 18.6, and then exciting the coil over the frequency
range of 0.1–1 MHz, in 100 kHz steps. The “pit” is positioned as accurately as
possible to be directly above the coil, whose dimensions are 15 mils inner radius,
48 mils outer radius, 7 mils height, and is wound with 19 turns over two layers.
The measured dc resistance of the coil is 0.741 Ω, and the measured dc inductance
is 0.51092 μH.
The first task is to characterize the host material and to determine the lift-off of
the coil over the host. This is done by taking multifrequency data when the coil is
over the unflawed workpiece, and then introducing these data into NLSE. The results
are that σhost = 1.1763 × 106 S/m, lift-off = 10.201 mils. These, then, are the “host
parameters” for the pit-inversion process.
The multifrequency input (impedance) data obtained when the pit is over the coil
are shown in Fig. 18.7, and these data are presented to NLSE in order to determine
the X,Y coordinates of the center of the coil from the center of the pit, as well as the
values of the three semiaxes. The result of the inversion gives X and Y coordinates of
6.711 mils and 6.695 mils, respectively, for the location of the probe from the center
360 18 Applications to Nuclear Power

Fig. 18.7 Multifrequency Multifrequency Impedance Input Data

input data for the benchmark 0.3
inversion test R
0.25 X

0.2

Resistance or Reactance (Ohms)

0.15

0.1

0.05

-0.05

-0.1

-0.15

-0.2

-0.25
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Frequency (MHz)

of the pit. The dimensions of the pit semiaxes are: A = 0.0737 in., B = 0.0737 in.,
C = 0.0279 in. The solution is least sensitive to the lateral coordinates of the probe
relative to the pit, and most sensitive to the depth of the pit, as we noted before.
This validates the NLSE inversion process, as it is applied to semiellipsoidal pit
models. Further, this process illustrates another method for obtaining data when
a physical scan is difficult or impossible. It would be impossible to infer the
solution for five parameters by using analog instruments and checking traces on
an oscilloscope. This, again, demonstrates the value of model-based inversion in
eddy-current NDE.

18.1.6 Modeling Wall-Thinning Effects

Wall-thinning is often associated with corrosion effects, and it is necessary to

distinguish wall-thinning from such things as pitting. Following the same strategy
used in the complex pit model, we can model wall-thinning in a tube by stacking
axisymmetric rings contiguous to each other, as shown in Fig. 18.8. In this case,
we agree that the axial resolution for wall-thinning will be 0.2 inch, and we take
enough rings to cover the region of interest. The radius of each ring determines
the skyline profile of the inner radius of the tube. Because we assume that bobbin
coils are being used in gathering data, it follows that we can only assume the wall-
thinning to be axisymmetric. In the model shown in Fig. 18.8 the inverse problem
18.1 Modeling Pitting and Corrosion Phenomena in Heat-Exchanger Tubes 361

0.004" 0.008" 0.004"

0.049"

" "
0.2 0.2 0.2"

0.326"

Tube Axis

Fig. 18.8 Modeling localized wall-thinning by using a series of axisymmetric rings

Fig. 18.9 Input data for Model Wall-Thinning Input Data

inverse problem of 4.5
wall-thinning R
4 X

3.5
Resistance/Reactance (Ohms)

2.5

1.5

0.5

-0.5

-1
-0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8
Probe Position (mm)

is to determine the radii of the three rings. The model input data are the scanned
impedances associated with the configuration shown in the figure, and are shown
in Fig. 18.9.
In order to invert these data using up to a fourth-order polynomial, we first
establish a 5 × 5 × 5 interpolation table, in which each wall-loss variable has nodes
at 0.0005, 0.003, 0.0055, 0.008, and 0.0105 inch. The blending functions associated
with the 1st and 125th nodes are shown in Fig. 18.10, along with the input data of
Fig. 18.9.
The results of the NLSE inversion are tabulated in Table 18.4.
362 18 Applications to Nuclear Power

Blending Functions and Input Data Blending Functions and Input Data
0.2 6
Data
BF1
0 5
BF125
Resistance (Ohms)

Reactance (Ohms)
-0.2 Data
BF1 4
BF125
-0.4
3
-0.6

2
-0.8

-1 1

-1.2 0
-0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8
Probe Position (mm) Probe Position (mm)

Fig. 18.10 Illustrating the 1st and 125th blending functions, together with the input data

Table 18.4 Results of the NLSE inversion for the wall-thinning problem
Order Φ Wall-loss1(in)/sensitivity Wall-loss2(in)/sensitivity Wall-loss3(in)/sensitivity
1 0.05056 0.004028/0.6127(−4) 0.008/0.8104(−4) 0.004028/0.6127(−4)
2 0.03609 0.003939/0.3948(−4) 0.00797/0.5103(−4) 0.003939/0.3948(−4)
3 0.04160 0.0039175/0.4593(−4) 0.00795/0.5578(−4) 0.0039175/0.4593(−4)
4 0.04734 0.0039085/0.5220(−4) 0.007945/0.5905(−4) 0.0039085/0.5220(−4)

18.2 Model-Based Inversion of Measured Impedance Data:

The Carderock Problem

Consider a probe that is to be operated as a differential-bobbin within a 70/30 Cu-

Ni tube. The inner radius of the tube is 0.2475 in and the outer radius is 0.3125 in.
The tube’s conductivity is 2.61 × 106 S/m, and its relative magnetic permeability
is unity. As we pointed out in Sect. 15.3.1, in modeling such a probe, we simply
take the response to a single (absolute) bobbin coil, and then interpolate within
that response to compute the response of the second coil, from with the differential
response is computing by subtraction.
When the absolute coil and its connecting cable are scanned over a 0.25-inch,
through-wall round hole in the same tube at 88 kHz, the resulting measured data
are shown in Fig. 18.11, along with the results from a VIC-3D R
-model of the
3
problem. Except for the host-only offset in the two data sets, the results are
in excellent agreement, the l2 -norm error being 1.55 %. The host-only offset is
probably attributable to uncertainty in the original data of the probe in freespace,

3 Themeasured data and parameters of this experiment were provided by J. Liu and N. Trepal of
the Naval Surface Warfare Center, Carderock Division.
18.2 Model-Based Inversion of Measured Impedance Data: The Carderock Problem 363

Model and Experimental Results Model and Experimental Results

for the 0.25-in Hole for the 0.25-in Hole
7.25 22.8

7.2 Model
Exp 22.6 Model
7.15 Exp
22.4
Resistance(Ohms)

7.1

Reactance(Ohms)
7.05
22.2
7
22
6.95

6.9 21.8
6.85
21.6
6.8

6.75 21.4
-0.6 -0.4 -0.2 0 0.2 0.4 0.6 -0.6 -0.4 -0.2 0 0.2 0.4 0.6
Probe Position (in) Probe Position (in)

Fig. 18.11 Model and experimental data for a 0.25 inch, through-wall round hole model of a pit
at 88 kHz

Balanced Results for the 0.25-in Hole Balanced Results for the 0.25-in Hole
6.9 22.8

6.88 Model 22.7

Exp
22.6 Model
6.86 Exp
Resistance(Ohms)

Reactance(Ohms)

22.5
6.84
22.4
6.82
22.3
6.8
22.2
6.78
22.1

6.76 22

6.74 21.9
-0.6 -0.4 -0.2 0 0.2 0.4 0.6 -0.6 -0.4 -0.2 0 0.2 0.4 0.6
Probe Position (in) Probe Position (in)

Fig. 18.12 Model and experimental data for a 0.25 inch, through-wall round hole model of a pit
at 88 kHz after the host-only offset has been removed

without the cable.4 Figure 18.12 compares the responses when the model results are
“balanced” against the measured data by forcing the host-only offsets to be removed.
Now the l2 -norm error is less than 0.1 %.
A comparison of the differential-bobbin responses is shown in Fig. 18.13. It is
clear from this figure that noise in the measured data will produce a Lissajous figure

4 See Appendix A.1 for a discussion of probe + cable effects.

364 18 Applications to Nuclear Power

Fig. 18.13 Comparing the Differential Bobbin Results for the 0.25-in Hole
differential-bobbin 0.8
experimental and VIC-3DR
VIC-3D
responses Experiment
0.6

0.4

0.2

X(Ohms)
0

-0.2

-0.4

-0.6

-0.8
-0.15 -0.1 -0.05 0 0.05 0.1 0.15
R(Ohms)

Fig. 18.14 Interpolation Diameter (inch)

table for the round-hole
problem. The nodes are 0.4 13 14 15 16
numbered within the grid

11 12
0.3 9 10

5 6 7 8
0.2

1 2 3 4
0.1 Wall Loss (%)
25 50 75 100

that will not allow a very accurate assessment of the size and depth of the pit. This
is the reason for doing model-based inversion, as we alluded to in the preceding
section.
The first step in applying model-based inversion to this problem is to develop
an interpolation table for the variables of interest. Figure 18.14 illustrates one such
table for the variables of diameter and wall-loss (depth) of a round-hole. We use
VIC-3D R
to generate the standards (blending-functions) associated with each node
of the interpolation table. These functions are shown in Figs. 18.15–18.18.
18.2 Model-Based Inversion of Measured Impedance Data: The Carderock Problem 365

Blending Functions for Nodes 1-4 Blending Functions for Nodes 1-4
6.83 22.1

6.825
22.08 1
6.82 1 2
2 3
22.06 4
Resistance(Ohms)

6.815 3

Reactance(Ohms)
4
6.81
22.04
6.805
22.02
6.8

6.795 22
6.79
21.98
6.785

6.78 21.96
-0.6 -0.4 -0.2 0 0.2 0.4 0.6 -0.6 -0.4 -0.2 0 0.2 0.4 0.6
Probe Position (in) Probe Position (in)

Fig. 18.15 Blending functions for nodes 1–4

Blending Functions for Nodes 5-8 Blending Functions for Nodes 5-8
6.88 22.5
22.45
6.86 5
6 22.4 5
7 6
6.84 8 22.35 7
Resistance(Ohms)

Reactance(Ohms)

8
22.3
6.82
22.25
22.2
6.8
22.15
6.78 22.1
22.05
6.76
22
6.74 21.95
-0.6 -0.4 -0.2 0 0.2 0.4 0.6 -0.6 -0.4 -0.2 0 0.2 0.4 0.6
Probe Position (in) Probe Position (in)

Fig. 18.16 Blending functions for nodes 5–8

The results when the experimental data of Fig. 18.12 are used as the input to
NLSE are shown in Table 18.5. The heading labeled “Order” denotes the order of
the interpolating polynomial and Φ denotes the norm of the residuals. Recall that
“sensitivity” is the parameter that determines the sensitivity of Φ to that unknown—
the smaller the value of “sensitivity” the more sensitive to that parameter. In addition
to Φ , the sensitivity parameter is important in determining the quality of the
solution, especially in making a judgement as to the importance of an unknown
parameter, or in determining the confidence one has in the value of that parameter
at the solution.
The corresponding results when the model data of Fig. 18.12 are used as the
input to NLSE are given in Table 18.6. The purpose of showing these results is to
366 18 Applications to Nuclear Power

Blending Functions for Nodes 9-12 Blending Functions for Nodes 9-12
6.92 23.2

6.9 9
10 23
6.88 11 9
12 10
22.8
Resistance(Ohms)

Reactance(Ohms)
6.86 11
12
6.84
22.6
6.82
22.4
6.8

6.78 22.2
6.76
22
6.74

6.72 21.8
-0.6 -0.4 -0.2 0 0.2 0.4 0.6 -0.6 -0.4 -0.2 0 0.2 0.4 0.6
Probe Position (in) Probe Position (in)

Fig. 18.17 Blending functions for nodes 9–12

Blending Functions for Nodes 13-16 Blending Functions for Nodes 13-16
7 23.8

23.6
6.95 13 13
14 23.4 14
15 15
16
Resistance(Ohms)

23.2
Reactance(Ohms)

6.9 16
23

6.85 22.8

22.6
6.8
22.4

22.2
6.75
22

6.7 21.8
-0.6 -0.4 -0.2 0 0.2 0.4 0.6 -0.6 -0.4 -0.2 0 0.2 0.4 0.6
Probe Position (in) Probe Position (in)

Fig. 18.18 Blending functions for nodes 13–16

Table 18.5 Response of NLSE when experimental data are used

Order Φ Diameter/sensitivity Depth/sensitivity
1 0.2116 0.2479/0.1466(−1) 0.0601/0.4877(−2)
2 0.2123 0.259/0.1487(−1) 0.0588/0.4355(−2)
3 0.2163 0.2594/0.1523(−1) 0.05895/0.4467(−2)
Nominal 0.25 0.065

establish a standard for comparison. The fact that Φ in Table 18.5 is only about an
order of magnitude larger than that in Table 18.6 indicates that the results shown in
Table 18.5 are acceptable.
18.3 Model-Based Inversion of Measured Instrument Data... 367

Table 18.6 Response of NLSE when model data are used

Order Φ Diameter/sensitivity Depth/sensitivity
1 0.6894(−1) 0.2443/0.4083(−2) 0.065/0.1655(−2)
2 0.1767(−1) 0.2509/0.1018(−2) 0.065/0
3 0.181(−1) 0.2513/0.1054(−2) 0.065/0.3913(−3)
Nominal 0.25 0.065

Table 18.7 Response of NLSE after first iteration of POCS

Order Φ Diameter/sensitivity Depth/sensitivity
1 0.127 0.2536/0.7522(−2) 0.065/0.3393(−2)
2 0.1230 0.2448/0.7617(−2) 0.06386/0.2997(−2)
3 0.1231 0.2477/0.7865(−2) 0.06293/0.2799(−2)
Nominal 0.25 0.065

Even though these results are acceptable, we can improve them by performing
the iterative-refinement algorithm that is based upon projections onto convex sets
(POCS) (recall Chap. 14, Fig. 14.3). We start the next iteration of POCS by
computing the impedance response to a hole that has a diameter of 0.2479 inch and a
depth of 0.0601 inch. These are the estimated parameters shown in Table 18.5. Then
we take the computed resistance as the next estimate of the “correct” resistance,
but replace the computed reactance by the original measured reactance. This keeps
us in touch with the original measurements. Using the “corrected resistance” and
the original reactance, we rerun NLSE and get the new inversion results shown in
Table 18.7. Clearly, this is a significant improvement over the previous results of
Table 18.5, both with respect to a smaller Φ , and for a more accurate estimate of
the diameter and depth of the hole, with smaller sensitivity values. Note that in all
of our inversion experiments, we find that the sensitivity in the depth parameter is
better (smaller) than in the diameter. This is typical for problems, such as corrosion
modeling, in which these are the two parameters of interest.

18.3 Model-Based Inversion of Measured Instrument Data:

The EPRI ID Pits Benchmark Test

Figure 18.19 illustrates the model of the EPRI ID Pits test problem.5 The parameters
of the coil are estimates, only. Each coil is a bobbin coil, and their individual
responses are called “absolute.” When the two coils are connected in a bridge circuit,
they form a differential bobbin, whose response is essentially the derivative of the

5 The test setup and measured data for this section were supplied by K. Krzywosz of the Electric
Power Research Institute (EPRI) as part of its benchmark validation test for sizing inner-diameter
(ID) pit models.
368 18 Applications to Nuclear Power

0.028 ID PIT

0.472 OR
IR

0.030 0.060 0.030

SS: σ = 1.4E6 , μ = 1

OR = 0.445, IR = 0.415, Turns = 55

All dimensions in inches

Fig. 18.19 Schematic of the EPRI ID Pits test problem. The parameters of the coil are
estimates, only

Table 18.8 Labeling the Pit no. Diameter (in) Height (in) Wall loss (%)
999-series of round-bottom
pits 1 0.0625 0.007 25
2 0.0625 0.014 50
3 0.0625 0.021 75
4 0.0625 0.028 100
5 0.09375 0.007 25
6 0.09375 0.014 50
7 0.09375 0.021 75
8 0.09375 0.028 100
9 0.125 0.007 25
10 0.125 0.014 50
11 0.125 0.021 75
12 0.125 0.028 100

absolute response. In VIC-3D R

we simulate the differential-bobbin response in a
post-processing filter, without the need to actually create a physical differential-
bobbin probe with the accompanying bridge circuit, or other circuit that performs
the differencing operation. The results of this section are obtained by means of the
simulated differential-bobbin probe.
Calibration standard data are taken at 1 MHz, 800 kHz, 700 kHz, 600 kHz,
500 kHz, 400 kHz, 300 kHz, and 100 kHz in differential and absolute modes on
a series of flaws with 25 %, 50 %, 75 % and 100 % wall loss. The 999-series
comprises round bottom pits of 1/16 , 3/32’ and 1/8 ’ diameters, the 888-series
comprises elongated pits 1/32 W×3/16L, the 777-series comprises elongated pits
1/16W×3/16 L, and the 666-series comprises elongated pits 1/32 W×1/8 L.
The 999-series is identified in Table 18.8. We will apply VIC-3D R
to develop
model-based “equivalent pillbox” standards for this series at 1 MHz, 800 kHz,
600 kHz, and 400 kHz. By an equivalent pillbox, we mean a truncated right-circular
cylinder that is defined by its diameter and height, and that approximates the true
18.3 Model-Based Inversion of Measured Instrument Data... 369

Fig. 18.20 A 4 × 4 Wall Loss (%)

interpolation table for the
999-series at 1 MHz, 100
800 kHz, 600 kHz, and
400 kHz. This table will allow
us to a third-order
(cubic-spline) interpolation
75

50

25 Diameter
1/32 1/16 3/32 1/8 (inches)

“round-bottom” circular pit in a manner to be described later. We add another

standard for a 1/32”-diameter pit in order to generate at 4 × 4 interpolation table
for inverting data using these model-based standards. Such a table (see Fig. 18.20)
will allow us to do up to a third-order (cubic-spline) interpolation.

18.3.1 Scaling the Measured Instrument Data

The measured data are instrument voltages, which means that they must be
transformed into impedances in order to be used as a source for inversion with VIC-
3D R
. Following the discussion in Chap. 15, we accomplish this by using a general
linear filter,
Z = β1 (C1 + jC2 ) − β2, (18.1)

relating the impedance, Z, to the voltages, C1 and C2 , of instrument channels 1 and 2,

respectively.
The four parameters given by the two complex numbers, β1 and β2 , are frequency
dependent and are determined by fitting the instrument voltages measured for pit 12
at each frequency to the impedances calculated by VIC-3D R
for nominal values
of pit 12 dimensions. A value of β2 = 0 was found to give a good match between
channel data and computed impedances at all frequencies. Thus, we have a two-
parameter transformation that simply scales the complex voltage, C1 + jC2 , and
rotates it in the complex plane. These two operations account for the amplification
and phase shift that the instrument applies to the probe signal.
370 18 Applications to Nuclear Power

18.3.2 Feature Extraction with Clutter Removal and Scan-Step

Correction

The preprocessing and feature extraction of the experimental data are a key part of
the inversion process. The quality of the inversion results from NDE measurement
scans using in-field equipment can be improved by using the appropriate prepro-
cessing steps. Understanding the characteristics of the measurement data and how
they were acquired is important in designing the preprocessing steps. Here, a series
of steps were implemented to automate the extraction and filtering of the pit features
found in the EPRI scan data to locate the exact center of the pit responses. A cross-
correlation of the measured data with a selected calibration pit response curve was
used to locate the exact center of the pit responses. Subsequent steps addressing
background noise filtering and scan-step correction are presented in detail below.
Lastly, symmetry was used to average the experimental response about the pit.
A novel background clutter removal algorithm was developed to compensate for
systematic measurement noise in scan data. Typical sources for clutter and noise
found in eddy-current scans are probe lift-off variation, local changes in material
properties, sensitivity to sub-surface structures, thermal variation of measurement
components, and electrical noise. Figure 18.21a displays an example where the
background noise of the original experimental response is evaluated, using the
procedure described in Chap. 15, providing efficient removal of background clutter
from the experimental data. Compensation for the systematic background noise
will reduce the role that the tails of the pit response play in the least-squares
error minimization and emphasize the center pit region achieving a more accurate
model fit.
A method was also developed to correct for the significant variation in the scan
step size found between different pits in a scan. Since the data were acquired using
a constant rate of acquisition (1,000 Hz), it was observed that the pull velocity of the
probe through the tube actually varied by as much as 30 % with respect to the mean
rate. This resulted in uncertainty in the actual scan step size in the neighborhood
of each pit in the test sample. Thus, it was concluded that the local scan step size
must be estimated for each pit. To determine the local scan step size, an inversion
process was designed that iteratively fit both the equivalent depth and diameter of
the pit and the scan step size. Figure 18.21b displays the reactance response for pit 7
for the measured data with and without scan-step correction and the corresponding
simulated data. Since the scan step size more greatly affects the lateral fit of the data,
while the depth and diameter of the pit more greatly impact the fit in both magnitude
and at the pit center, the inversion of these three parameters simultaneously was
found to be well-posed over the parameter range tested in the EPRI study.
18.3 Model-Based Inversion of Measured Instrument Data... 371

Fig. 18.21 (a) Reactance response for pit 6 demonstrating background clutter removal.
(b) Reactance response for pit 7 demonstrating scan-step correction (for 008–999 at 1.0 MHz)
372 18 Applications to Nuclear Power

Table 18.9 Summary of Pillbox Model Results. “Multi” denotes a multifrequency combination of
all four frequencies, using unnormalized data. The sensitivity to the solution of the final (“Multi”)
estimates of D and H are also given
1 MHz 800 kHz 600 kHz 400 kHz Multi
Pit D H D H D H D H D H
Sensit Sensit
1 0.0433 0.0084 0.0469 0.007 0.0470 0.007 0.0438 0.007 0.0463 0.007
0.0082 0.0028
2 0.0550 0.0108 0.0554 0.0105 0.0558 0.0101 0.0534 0.0103 0.0557 0.0103
0.0044 0.0015
3 0.0578 0.0179 0.0574 0.0175 0.0570 0.0171 0.0538 0.0180 0.0576 0.0173
0.0053 0.0029
4 0.0699 0.028 0.0695 0.028 0.0692 0.028 0.0689 0.028 0.0695 0.028
0.0033 0.0022
5 0.0513 0.007 0.0510 0.007 0.0502 0.007 0.0484 0.007 0.0509 0.007
0.0083 0.0024
6 0.0897 0.0106 0.0904 0.0103 0.0952 0.0092 0.0964 0.0087 0.0911 0.0101
0.0051 0.0010
7 0.0999 0.0173 0.1002 0.0169 0.0985 0.0164 0.0975 0.0162 0.0990 0.0166
0.0033 0.0010
8 0.0992 0.028 0.0990 0.028 0.0987 0.028 0.0984 0.028 0.0990 0.028
0.0044 0.0017
9 0.1021 0.007 0.1015 0.007 0.1004 0.007 0.0973 0.007 0.1012 0.007
0.010 0.0014
10 0.1244 0.0132 0.125 0.0128 0.125 0.0124 0.125 0.0118 0.125 0.0127
0.0036 0.0006
11 0.125 0.0184 0.125 0.0180 0.125 0.0177 0.125 0.0178 0.125 0.0180
0.0046 0.0009
12 0.125 0.0268 0.125 0.0268 0.125 0.0267 0.125 0.0265 0.125 0.0267
0.0071 0.0015

18.3.3 Summary of Multifrequency Results

We consider the inversion of the transformed (scaled) data at all four frequencies,
1MHz, 800kHz, 600kHz, and 400kHz, first individually, and then collectively in
one NLSE run. In the multifrequency run, the data are not normalized to account for
frequency differences. The results are summarized in Table 18.9.
The distribution of pit diameters and depths as listed in Table 18.9 are compared
to the nominal values in Fig. 18.22. It appears that the mechanism that produces
the round-bottom of the shallowest pits, namely 1, 5, and 9, does not allow a fully
developed pit to form, and this may account for the smaller than expected diameters
of these pits. This is reminiscent of the need for a “new flaw model” for pits, as
described in Sect. 16.4.
18.3 Model-Based Inversion of Measured Instrument Data... 373

0.03
0.14
nominal nominal
VIC-3D 0.025 VIC-3D
0.12
Pit Diameter (in)

Pit Depth (in)

0.1 0.02

0.08
0.015
0.06
0.01
0.04
0.005
0.02

0 0
-4 -2 0 2 4 6 8 10 12 14 -4 -2 0 2 4 6 8 10 12 14
Pit No. Pit No.

Fig. 18.22 Distribution of computed and nominal pit diameters and depths

The final results for each pit response at 1 MHz, when the diameter and height
shown in Table 18.9 under “Multi” are used, are plotted in the next figures.

Results for Pit 1 at 1MHz Results for Pit 1 at 1MHz

0.03 0.2

0.15
0.02 Measured
Measured 0.1 Computed
Computed
Resistance(Ohms)

Reactance(Ohms)

0.01
0.05

0 0

-0.05
-0.01
-0.1
-0.02
-0.15

-0.03 -0.2
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
Probe Position (in) Probe Position (in)

Results for Pit 2 at 1MHz Results for Pit 2 at 1MHz

0.04 0.4

0.03 Measured 0.3

Computed Measured
0.02 0.2 Computed
Resistance(Ohms)

Reactance(Ohms)

0.01 0.1

0 0

-0.01 -0.1

-0.02 -0.2

-0.03 -0.3

-0.04 -0.4
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
Probe Position (in) Probe Position (in)
374 18 Applications to Nuclear Power

Results for Pit 3 at 1MHz Results for Pit 3 at 1MHz

0.08 0.5

0.4
0.06
Measured
Measured 0.3 Computed
0.04 Computed
Resistance(Ohms)

Reactance(Ohms)
0.2
0.02
0.1

0 0

-0.1
-0.02
-0.2
-0.04
-0.3
-0.06
-0.4

-0.08 -0.5
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
Probe Position (in) Probe Position (in)
Results for Pit 4 at 1MHz Results for Pit 4 at 1MHz
0.4 1.5

0.3
Measured 1
0.2 Computed Measured
Resistance(Ohms)

Computed
Reactance(Ohms)

0.5
0.1

0 0

-0.1
-0.5
-0.2
-1
-0.3

-0.4 -1.5
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
Probe Position (in) Probe Position (in)

Results for Pit 5 at 1MHz Results for Pit 5 at 1MHz

0.04 0.2

0.03 0.15

Measured Measured
0.02 0.1
Computed Computed
Resistance(Ohms)

Reactance(Ohms)

0.01 0.05

0 0

-0.01 -0.05

-0.02 -0.1

-0.03 -0.15

-0.04 -0.2
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
Probe Position (in) Probe Position (in)
18.3 Model-Based Inversion of Measured Instrument Data... 375

Results for Pit 6 at 1MHz Results for Pit 6 at 1MHz

0.1 0.8

0.08
0.6
Measured
0.06 Computed Measured
0.4 Computed
Resistance(Ohms)

Reactance(Ohms)
0.04
0.2
0.02

0 0

-0.02
-0.2
-0.04
-0.4
-0.06
-0.6
-0.08

-0.1 -0.8
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
Probe Position (in) Probe Position (in)
Results for Pit 7 at 1MHz Results for Pit 7 at 1MHz
0.15 1.5

Measured
0.1 Computed 1
Measured
Computed
Resistance(Ohms)

Reactance(Ohms)

0.05 0.5

0 0

-0.05 -0.5

-0.1 -1

-0.15 -1.5
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
Probe Position (in) Probe Position (in)
Results for Pit 8 at 1MHz Results for Pit 8 at 1MHz
0.8 2.5

2
0.6 Measured
Measured 1.5 Computed
0.4 Computed
Resistance(Ohms)

Reactance(Ohms)

1
0.2
0.5

0 0

-0.5
-0.2
-1
-0.4
-1.5
-0.6
-2

-0.8 -2.5
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
Probe Position (in) Probe Position (in)
376 18 Applications to Nuclear Power

Results for Pit 9 at 1MHz Results for Pit 9 at 1MHz

0.15 0.6

0.1 0.4
Measured Measured
Computed
Resistance(Ohms)

Computed

Reactance(Ohms)
0.05 0.2

0 0

-0.05 -0.2

-0.1 -0.4

-0.15 -0.6
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
Probe Position (in) Probe Position (in)
Results for Pit 10 at 1MHz Results for Pit 10 at 1MHz
0.15 1.5

Measured
0.1 1 Computed
Measured
Computed
Resistance(Ohms)

Reactance(Ohms)

0.05 0.5

0 0

-0.05 -0.5

-0.1 -1

-0.15 -1.5
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
Probe Position (in) Probe Position (in)

Results for Pit 11 at 1MHz Results for Pit 11 at 1MHz

0.25 2

0.2
1.5
0.15 Measured Measured
Computed 1 Computed
Resistance(Ohms)

Reactance(Ohms)

0.1
0.5
0.05

0 0

-0.05
-0.5
-0.1
-1
-0.15
-1.5
-0.2

-0.25 -2
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
Probe Position (in) Probe Position (in)
18.3 Model-Based Inversion of Measured Instrument Data... 377

Results for Pit 12 at 1MHz Results for Pit 12 at 1MHz

1 4

0.8 Measured
3
Computed Measured
0.6 Computed
2
Resistance(Ohms)

Reactance(Ohms)
0.4
1
0.2

0 0

-0.2
-1
-0.4
-2
-0.6
-3
-0.8

-1 -4
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
Probe Position (in) Probe Position (in)

0.05756
0.0

Cavity

Host: σ = 1.4 E6
0.015
#3 0.017
#2 0.019
#1 0.021

0.028

Fig. 18.23 Layered model of pit 3

18.3.4 The Morphology of “Round-Bottom” Pits

We will develop a layered model6 of pit 3 in order to gain some insight into the
morphology of round-bottom pits. Figure 18.23 illustrates a three-layer model of
pit 3, with a resolution of 2 mils per layer. We used the data of Table 18.9 to
establish the size of the cavity and the position of the three layers. The solution
of the multifrequency inverse problem for the conductivities of the layers is shown
in Table 18.10.

6 See Appendix A.2 for a discussion of layered models of pits.

378 18 Applications to Nuclear Power

Table 18.10 Results of inverting the layered model of Fig. 18.23

Φ σ1 /Sensit σ2 /Sensit σ3 /Sensit D(Fixed)
0.4788 1.4E6/0.2192(1) 0.5939E6/0.2099(1) 0.1151E1/0.1781(1) 0.05756

0.05756
0.0

Cavity
0.01731
Host : σ = 1.4E6
0.017
σ = 0.5939E6 0.019

0.028

Fig. 18.24 Results of the inversion of the layered model of pit 3. The “round-bottom” is shown
as the amorphous shaded layer whose effective conductivity is 0.5939 × 106 S/m. The equivalent
pillbox model is shown as the dotted outline

Clearly, the bottom layer is host material, and the top layer is part of the pit cavity.
The middle layer with an effective conductivity of 5.939 × 105 S/m represents the
rounded bottom of the pit. We indicate in Fig. 18.24 this layer as an amorphous
structure, together with the equivalent pillbox model of the pit, using the Table 18.9
results. Upon applying the volume-fraction concept to the effective “round-bottom”
layer whose conductivity is 0.5939 × 106 S/m, we would find a total round-bottom
+ cavity volume that is virtually equal to the cavity volume of the equivalent pillbox
model. The important thing to note here is that the “rounded-bottom” that we have
just computed gives us a net depth of 0.019 inch for pit 3, which is in better
agreement with the nominal value listed in Table 18.8. (See, also, Fig. 18.22). This
shows the power and flexibility of model-based inversion in characterizing pits by
choosing appropriate parameters for modeling the pit.
Using the “Multi” results for D and H of pit 3 from Table 18.9, and the
results for the round-bottom model of pit 3 from Fig. 18.24, we plot in Fig. 18.25
the corresponding impedances at 1MHz. The round-bottom model gives a more
accurate estimation of the peak and slope through the origin of the resistance
scan. This is what we would expect, because the resistance data are important in
determining the depth of the pit. We have not applied this algorithm to the shallow
pits, 1, 5, and 9, because our interest was in determining the origin of the depth-
discrepancies of the deeper pits when the pillbox model was used.
A.1 Modeling Probes + Cables 379

Results for Pit 3 at 1MHz Results for Pit 3 at 1MHz

0.08 0.5

0.4
0.06 Measured
Measured 0.3 Computed
0.04 Computed
Resistance(Ohms)

0.2

Reactance(Ohms)
0.02
0.1

0 0

-0.1
-0.02
-0.2
-0.04
-0.3
-0.06
-0.4

-0.08 -0.5
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
Probe Position (in) Probe Position (in)
Results for Pit 3 Layered Model at 1MHz Results for Pit 3 Layered Model at 1MHz
0.08 0.5
Measured 0.4
0.06 Computed
Measured
0.3 Computed
0.04
Resistance(Ohms)

0.2
Reactance(Ohms)

0.02
0.1

0 0

-0.1
-0.02
-0.2
-0.04
-0.3
-0.06
-0.4

-0.08 -0.5
-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5
Probe Position (in) Probe Position (in)

Fig. 18.25 Final results for pit 3 at 1MHz. Top: pillbox model. Bottom: layered (round-bottom)
model

Appendix

A.1 Modeling Probes + Cables

In order to characterize a probe for the purpose of modeling probe-flaw responses

and validating benchmark tests, it is necessary to consider the effects of real
conditions under which the ideal probe operates. Consider a probe that is to
be operated as a differential-bobbin within a 70/30 Cu-Ni tube with no defects.
The inner radius of the tube is 0.2475 in and the outer radius is 0.3125 in. The
tube’s conductivity is 2.61 × 106 S/m, and its relative magnetic permeability is unity.
Almost nothing is known about the probe except some external dimensions.
380 18 Applications to Nuclear Power

Table 18.11 Frequency response of a probe in freespace and within a tube

Air Measurements Tube Measurements

f(khz) Z (Ω) Θ (◦ ) X (Ω) R (Ω) L(μH) Z (Ω) Θ (◦ ) X (Ω) R (Ω) L(μH)

10 5.6 36.5 3.3 4.5 52.6 5.7 32.4 3.1 4.8 48.6
50 17.0 73.5 16.3 4.8 51.7 14.2 64.8 12.8 6.0 40.6
100 33.1 80.8 32.7 5.3 52.0 25.8 74.0 24.8 7.1 39.5
200 69.1 84.1 68.7 7.0 54.7 50.7 78.7 49.7 9.9 39.6

Data taken for a single probe of the differential-pair are given in Table 18.11.
Referring to the air measurements, we note that the resistance of the probe varies
with frequency, and the reactance is not proportional to frequency. This suggests
that there is a frequency-sensitive two-port network that connects the probe to the
impedance analyzer. It could be something as simple as a shunt self-capacitance of
the coil, but in this case it is a 25-foot (7.62 m), 5/16-inch coaxial cable. Cables
of this length (or more) are used to connect the probe to the measuring instrument
when inspecting heat-exchanger tubes.
The problem, now, is to characterize a probe coil whose parameters are not
known, and to characterize the cable that connects the coil to the impedance
analyzer. This can be done quite simply, though a bit tediously, with VIC-3D R
.
Trial and error on the turns and inner- and outer-radii of the coil suggest that a good
fit to the measured data is given by: Nturns = 55, IR = 0.19 in. and OR = 0.21 in.
The height of the coil is known to be 0.055 in. When we run VIC-3D R
for this
probe in freespace, we get an inductance of 49.55 μH, which is close to, but smaller
than, the inductance measured at 10 kHz. The freespace resistance of the model coil
is, of course, 0 Ω, so we manually enter a resistance of R0 = 4.5 Ω into the VIC-
3D R
file to see what that gives. We choose this value because it agrees with the
measured resistance at the lowest frequency.
Cable effects can be accounted for by defining four parameters: characteristic
impedance, capacitance per unit length, attenuation in dB/m and length. This
follows from transmission-line theory [98]. Let ZL be the load (terminating)
impedance of a coaxial cable, and Zin be the driving-point (input) impedance of
the loaded cable. Then
ZL + Z0 tanh( jβ l + α l)
Zin = Z0
Z0 + ZL tanh( jβ l + α l)

Z0 = (L/C) = characteristic impedance of the cable
ω
β =
vp
1
vp =  ≈ 2 × 108 m/s. (18.2)
(LC)
A.1 Modeling Probes + Cables 381

Table 18.12 Freespace impedance response of the coil + cable

Frequency
α (dB/m) 104 5 × 104 105 2 × 105
0.0 4.5014 + j3.2325 4.5344 + j16.221 4.6399 + j32.809 5.1008 + j68.743
0.005 4.7197 + j3.2299 4.7749 + j16.208 4.9519 + j32.782 5.7317 + j68.677
0.010 4.9379 + j3.2272 5.0151 + j16.194 5.2628 + j32.752 6.3612 + j68.604
0.015 5.1559 + j3.2243 5.2551 + j16.179 5.5735 + j32.721 6.989 + j68.524

Table 18.13 In-tube

impedance response of the f (Hz) Z (Ω)
coil + cable 104 4.8594 + j2.9665
5 × 104 6.1782 + j12.567
105 7.3656 + j24.325
2 × 105 10.117 + j48.48

L and C are, respectively, inductance and capacitance per unit length of the
cable. The value for phase velocity, vP , is a reasonable approximation for typical
transmission lines.
Typical values for coaxial cables that are used in eddy-current NDE are 50 Ω
for the characteristic impedance, and 100 pF/m for the distributed capacitance. Our
problem calls for a cable length of 7.62 m, which leaves only the attenuation to be
determined. Attenuation, α , is frequency dependent, but VIC-3D R
assumes it to
be constant, so it becomes necessary to run several tests to determine the frequency
response of the cable.
The results for tests of the probe+cable in freespace are shown in Table 18.12.
The diagonal entries in this table agree with the corresponding freespace results of
Table 18.11 within 2 %. This indicates that the attenuation varies with frequency as
shown in the left-hand column. This variation agrees with theory, in that attenuation
always increases with frequency.
When modeling the probe within the tube, we use the same cable parameters
as above, except that we use α = 0.0 only for f = 104 Hz, α = 0.005 only for
f = 5 × 104 , etc. The results for the model calculations for the coil within the tube
are given in Table 18.13. The differences between these values and those shown in
Table 18.11 are 3 %, or less, except for X104 and R105 .
It is difficult to determine an outer radius of a many-turn bobbin coil when it is
within a tube, because the windings on the outer layer of the coil nearest the tube
will not lie smoothly in that layer, as shown in Fig. 18.26. This gives the appearance
of an uneven spacing between the outer radius of the coil and the inner tube wall.
This is not as serious a problem when the coil is in the “pancake” aspect.
382 18 Applications to Nuclear Power

Fig. 18.26 The nonuniform

distribution of turns within a
typical real coil
COIL BOBBIN COIL
TURNS ASPECT

PANCAKE COIL ASPECT

0.028 ID PIT (Layered)

0.472 OR
IR

0.030 0.060 0.030

SS: σ = 1.4E6 , μ = 1

OR = 0.445, IR = 0.415, Turns = 55

All dimensions in inches

Fig. 18.27 A layered model of the ID pit shown in Fig. 18.19

A.2 A “Layered” Model of Corrosion Pits

One possible scheme for determining the shape of a corrosion pit that is assumed to
have a fixed morphology would be to use the layered-pit model shown in Fig. 18.27.
The objective is to determine the conductivity of each layer, and from that result
infer the size, and perhaps shape, of the pit. We assume that the pit has a circular
cross-section, as before, so that the only parameter that defines the model is the
radius of the anomalous region. If the layer has a conductivity of 0S/m, then clearly
that layer is entirely filled by the pit, whereas if the layer has a conductivity of
1.4 × 106 S/m, then the layer is entirely free of the pit, being host material, only.
Anything in between requires further consideration, as we show next.
The unknowns in the inversion process are σ1 , . . . , σ4 , the conductivities of layers
1–4, with 1 corresponding to the inner wall and 4 to the outer wall, as well as
the diameter of the anomalous region. Once the conductivity of a layer has been
estimated by NLSE, we apply a simple volume-fraction computation to determine
the relative volume of cavity to host material within that layer. For example, suppose
layer L has a conductivity of 1.135 × 106 S/m. Then to see how much of the Lth
layer is occupied by the cavity, we compute the volume-fraction of the cavity:
1.135
VFL = 1 − = 0.1893. Thus, the cavity occupies less than 19 % of the Lth
1.40
layer.
Chapter 19
Coupled Problems in Heat-Exchanger Tubes

19.1 Introduction

The preceding chapters of this book dealt largely with problems in which the
host material and any anomalies were pure electrical conductors. The presence of
magnetic permeabilities was merely an interesting side effect to the creation of the
VIC-3D R
model and its solution. There are, of course, many problems in which
magnetic permeability is present with electrical conductivity and must be accounted
for, not only in creating the model but also in understanding the physics of the
solution and its effect on the NDE process. In this chapter we consider magnetic
effects in heat-exchanger tubes, focusing later on ferritic tubes, which are becoming
of increasing importance in the nuclear power industry.

19.2 Reconstruction of a Semiellipsoidal Wear Scar

and Permeable Crust in a Heat-Exchanger Tube

In Figs. 19.1 and 19.2, we model “fretting” damage in a heat-exchanger tube [57].
The damage manifests itself as a “wear scar” surrounded by a “crust.” The crusts of
both figures are slightly magnetic, with that of Fig. 19.1 highly conducting, relative
to the host tube, and that of Fig. 19.2 having the same conductivity as the host tube.
The scars and crusts are modeled as semiellipsoids.
We model a seventeen-point scan over the anomalous region shown in Fig. 19.2
at three frequencies, 100 kHz, 400 kHz and 1.6 MHz, to reconstruct the conductivity,
σ , and permeability, μ , of the crust. The results are shown in Table 19.1. In all cases,
the result is more sensitive to μ than to σ . The computed impedance trajectories that
are the input to the inversion algorithm are shown in Fig. 19.3. The rotation of these
trajectories as a function of frequency is clearly shown.

H.A. Sabbagh et al., Computational Electromagnetics and Model-Based Inversion, 383

Scientific Computation, DOI 10.1007/978-1-4419-8429-6 19,
© Springer Science+Business Media New York 2013
384 19 Coupled Problems in Heat-Exchanger Tubes

TUBE AXIS

COIL

TUBE WALL : σ = 1.02Ε6, μ=1

HIGH−COND
CRUST σ=0, μ=1
WEAR SCAR

Fig. 19.1 A semiellipsoidal “wear scar” and highly-conducting permeable “crust” in the outer
surface of a tube

TUBE AXIS

COIL

TUBE WALL : σ =1Ε6, μ=1

PERMEABLE
CRUST σ=0, μ=1
σ = 1Ε6, μ=1.1 WEAR SCAR

Fig. 19.2 A semiellipsoidal “wear scar” and permeable “crust” in the outer surface of a tube
19.3 Modeling Tube-Support Rings (TSR) 385

Table 19.1 Results (μ , σ )

for wear scar and permeable Freq (kHz) Computed Original
crust 100 (1.056, 0.84E6) (1.1, 1E6)
400 (1.105, 0.85E6) (1.1, 1E6)
1,600 (1.106, 0.86E6) (1.1, 1E6)

Fig. 19.3 Impedance-plane Impedance signature vs. frequency

results for “wear scar” and 0.04
permeable “crust” at three 1.6MHz
frequencies 0.03 400kHz
100kHz
0.02
0.01
Reactance (Ohms)

0
-0.01
-0.02
-0.03
-0.04
-0.05
-0.06
-0.07
-0.02 -0.01 0 0.01 0.02 0.03 0.04 0.05 0.06
Resistance (Ohms)

19.3 Modeling Tube-Support Rings (TSR)

19.3.1 Statement of the Problem

In this section and the next we model the tube support ring problem of Fig. 19.4 that
is quite common in the inspection of nuclear power plants. The major difference
between the problem here and in the next section is the nature of the tube, itself.
Here, it is inconel, which is nonmagnetic, while in the next section it is ferritic,
which introduces additional interesting challenges. In both cases, however the
TSP+tube structure is axisymmetric, which allows us to use the axisymmetric model
in cylindrical geometries that was developed in Chap. 9.
386 19 Coupled Problems in Heat-Exchanger Tubes

19.00

Tube Support Ring (TSR)

6.1
μ = 50, σ = 6.7E6 S/m 1.25
Inconel Tube
μ = 1, σ = 1.1E6 S/m

7.725 8.25
Differential Bobbin Coil
6.225

1.5 1.5
1.5

Inconel Tube
TSR

Fig. 19.4 Illustrating a typical tube support ring problem. All dimensions are in millimeters

19.3.2 Measured Data

The measured data for the problem are taken at 100, 200, 300 and 400 kHz, and
are presented here for convenience in Fig. 19.5.1 Note that the data are plotted
in the impedance plane (real vs. reactive), even though they are instrument A/D
counts (or “voltage”) in two channels that are orthogonal in time; i.e., they are
ninety degrees out of phase with each other. The “Lissajous figure” closed-loop
curve is typical of data that are taken with a differential-bobbin probe. The user
of the instrument is usually instructed to “rotate the impedance-plane plot” so that
it satisfies certain criteria, and this makes the interpretation of the measured data
more challenging. We will use a linear filter to transform the measured voltage data
into impedances in a manner similar to that which was done in the EPRI ID pits
problem of Sect. 18.3.1. There are several important features to glean from this
figure: note the clockwise rotation of the loops with frequency and note that the
loops get smaller with frequency. Indeed, the loop for 400 kHz, actually fits within
the loop for 100 kHz.

1 The statement of the problem described in this section, and the original measured data were
supplied by Prof. S-J. Song of Sungkyunkwan University, South Korea.
19.3 Modeling Tube-Support Rings (TSR) 387

Z-Plane Representation of Data

600
100kHz
200kHz
300kHz
400 400kHz

Reactance (Ohms)
200

−200

−400

−600
−600 −400 −200 0 200 400 600
Resistance (Ohms)

Fig. 19.5 The measured data at the four frequencies

Z-Plane Representation of Model Results

0.8
100kHz
200kHz
0.6 300kHz
400kHz
0.4
Reactance (Ohms)

0.2

-0.2

-0.4

-0.6

-0.8
-0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8
Resistance (Ohms)

Fig. 19.6 The differential-bobbin model data at the four frequencies

19.3.3 Model Data Before β -Scaling

The model results (impedances) for Fig. 19.4 are shown in Fig. 19.6. We have not
yet applied the β -scaling that will bring the impedances to voltage or A/D-counts,
but the important features described above are already apparent. The loops decrease
in size with frequency and are rotated clockwise.
388 19 Coupled Problems in Heat-Exchanger Tubes

Table 19.2 Values of β at Frequency (kHz) β

each frequency
100 −384.6 + j464.63
200 −197.7 + j530.7
300 3.4408 + j587.07
400 260.09 + j562.31

Z-Plane Representation of Beta-Scaled Model Results

500
100kHz
400 200kHz
300kHz
300 400kHz
Reactance (Ohms)

200

100

-100

-200

-300

-400

-500
-500 -400 -300 -200 -100 0 100 200 300 400 500
Resistance (Ohms)

Fig. 19.7 The differential-bobbin model data shown in Fig. 19.6 after scaling by the appropriate
β at each of the four frequencies

19.3.4 Model Data After β -Scaling

Following the usual procedure, we scale the model vector by a factor, β , that
will allow us to transform impedance into voltage (or A/D-counts), or vice-versa.
We have determined β for each of the data and model vectors at the four frequencies,
as listed in Table 19.2. We then apply β at each frequency to the corresponding
model data vector of Fig. 19.6 and derive the scaled-model data vectors shown in
the complex plane of Fig. 19.7. We now note the similarities to the measured data
of Fig. 19.5: the loops rotate clockwise and shrink with frequency, and the loop for
400 kHz fits within that for 100 kHz. Hence, we can conclude that there is consistent
agreement between model and measured data.

19.3.5 Scaling in the Opposite Direction

Up to this point, we scaled the model data of Fig. 19.6 to agree as closely as possible
(in the least-squares sense) with the measured data of Fig. 19.5. The result is shown
in Fig. 19.7.
19.3 Modeling Tube-Support Rings (TSR) 389

Table 19.3 Values of β1 to β4 at each frequency

Frequency (kHz) β1 β2 β3 β4
100 −1.5305(−3) 2.7096(−4) −3.1574(−4) −9.0284(−4)
200 −2.1057(−4) 1.5131(−3) −9.0218(−4) −3.3706(−4)
300 3.4058(−4) 4.2174(−4) −3.2197(−4) −1.5568(−3)
400 −1.0023(−4) 5.7554(−4) 1.1822(−3) −9.4413(−4)

Now, we want to do the opposite and scale the measured data to agree as
closely as possible (in the least-squares sense) with the model data. Furthermore,
we are going to use a more general scaling approach in which four parameters,
β1 , β2 , β3 , and β4 , are used instead of only β1 and β2 . The four parameters are
defined by
R = β1C1 + β2C2
X = β3C2 + β4C1, (19.1)
where R and X are the vectors of the model results and C1 and C2 are vectors of
the measured data in each channel of the instrument (which are supposed to be
orthogonal in time).
Upon taking the dot-product of the first equation by, respectively, C1 and C2 , and
similarly the second equation, we get the two linear systems
(V M) (M) (M)
ΓR1 = β1Γ11 + β2Γ21
(V M) (M) (M)
ΓR2 = β1Γ12 + β2Γ22
(V M) (M) (M)
ΓX1 = β4Γ11 + β3Γ21
(V M) (M) (M)
ΓX2 = β4Γ12 + β3Γ22 , (19.2)

where, if these were random processes, the Γ s would be auto- and cross-correlation
functions. For our purposes, they are simply real numbers that are obtained by taking
the dot products of the appropriate vectors.
The solutions of (19.2) are tabulated for each frequency in Table 19.3.
An impedance-plane plot is shown in Fig. 19.8. Once again we see the strong
similarity between the scaled data and the model data of Fig. 19.6; the curves rotate
clockwise with frequency, and they become smaller and narrower.

19.3.6 An Inverse Problem

When we compare the Lissajous figures of Figs. 19.5 and 19.7, we are liable to draw
the conclusion that the model data cannot be used to characterize the system, if the
only data that are known for the system are given by the measuring instrument,
390 19 Coupled Problems in Heat-Exchanger Tubes

Z-Plane Representation of Beta-Scaled Measured Results

0.8
100kHz
200kHz
0.6 300kHz
400kHz
Reactance (Ohms) 0.4

0.2

-0.2

-0.4

-0.6

-0.8
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
Resistance (Ohms)

Fig. 19.8 The differential-bobbin measured data shown in Fig. 19.5 after scaling by the appropri-
ate β1 –β4 at each of the four frequencies

Table 19.4 Results for the

two variables, σ and μ , of the Freq(kHz) σ /sens μ /sens Φ
TSR 100 6.2224E6/4.36 40/20.08 1.67716
200 6.4227E6/3.84 40/28.13 1.28258
300 6.5032E6/3.72 40/43.14 0.93008
400 6.5410E6/3.81 40/30.37 0.68515

because they are not “perfectly” matched. We will show that the model data and
measured data, after either has been scaled to agree with the other, can be used
together to characterize the system. We will do this by using both in an inverse
problem whose goal is to determine the conductivity and magnetic permeability of
the tube-support ring. The inverse problem will be solved using NLSE in impedance
space, which means that we will use the β -scaled measurements of Fig. 19.8.
When we have two unknowns, the conductivity and magnetic permeability, and
apply NLSE to the scaled measured data at each of the four frequencies, we get the
results shown in Table 19.4. The term “sens” denotes the sensitivity of the parameter
at the solution point: the smaller, the better. Φ is the norm of the residual vector at
the solution point: the smaller the better, again.
Note that the conductivity estimate increases somewhat with frequency. Because
conductivity is more accurately estimated at higher frequencies (see [57] as well
as Chap. 11 of this book) we will take σ = 6.55 × 106 S/m as our estimate of the
conductivity.
The estimate of the magnetic permeability, however, is somewhat suspect,
because the solution is rather insensitive to the parameter. This follows from the
large values of sens associated with μ that are shown in the table. Hence, we will
19.4 Modeling Direct and Inverse Problems in Ferritic Tubes 391

Table 19.5 Results for the

Freq (kHz) μ /sens Φ
single variable, μ , of the TSR
when the conductivity is fixed 100 42.517/0.3345 1.5871
at σ = 6.55 × 106 200 41.603/0.2763 1.2401
300 41.206/0.2434 0.9072
400 41.007/0.2239 0.6710

redo the inversion, this time fixing σ = 6.55 × 106, and solving for only the single
variable, μ . When we use NLSE for this one-variable problem, we get the results
shown in Table 19.5.
Now, we see that Φ is reduced, meaning that the solution is in better agreement
with the measured data, and the sensitivity of the solution for μ is much better.
We also see that the estimated value of μ decreases slightly with frequency. Because
μ is more accurately estimated at lower frequencies (see Sect. 11.4), we choose
μ = 42.517 for our best estimate. Thus, both σ and μ are well estimated, given
their nominal values of 6.7 × 106 and 50, respectively.
In performing this inversion, it is important to understand that we did not
rely upon any particular “feature” of the Lissajous figure at any one frequency.
The approach that utilizes features of measured data is likely to be subject to
significant errors, given the subjective nature of “observing” such features. This
is especially true if the Lissajous figure is corrupted by noise or is embedded in
significant background clutter. The mathematical approach that we have indicated
is much more reliable and objective and can be further enhanced by the application
of data-smoothing and clutter-rejection algorithms.

19.4 Modeling Direct and Inverse Problems in Ferritic Tubes

19.4.1 The Model Problem

Ferritic stainless steels, such as Type 439 or SEACURE, are being increasingly
used in heat-exchanger tubes because of the increased resistance to chloride stress
corrosion and intergranular attack when compared to older alloys, such as Type 304.
This presents interesting modeling opportunities for the eddy-current nondestructive
evaluation of these tubes because their magnetic permeability is ∼60–100.2 In this
section,3 we present examples of direct and inverse problems involving such tubes,

2 The magnetic permeability is a function of the stress-state of the tube, as well as the frequency of
operation.
3 Details of the experiments that are described in this section are given in [112]. Related work on

ferritic tubes can be found in [113].

392 19 Coupled Problems in Heat-Exchanger Tubes

0.75"

Tube Support Plate

1"
μ=50, σ=6.76E6

0.030"
0.0625"
,
1 2 3
0.878" Coil

SEACURE
μ=104
σ=1.552E6

0.75"

Tube Support Plate

1"
μ=50, σ=1.0E6

0.030"
0.0625"
,
1 2 3
0.878"
Coil

SEACURE
μ=68.18
σ=1.372E6

Fig. 19.9 Illustrating a ferritic tube, such as SEACURE, with a tube-support plate (TSP), and three
0.0625-inch through-wall holes. Hole 1 is well away from the TSP, hole 2 is centered under the
TSP, and hole 3 is centered at edge of the TSP. Top: High-frequency model. Bottom: Low-frequency
model. See text for details on parameters

then demonstrate conditions that are peculiar to ferritic tubes, and give insight into
the optimum methods for characterizing the tubes and flaws within them.
The problem configuration is shown in Fig. 19.9. It is similar to the model shown
in Fig. 19.4, except that the tube is ferritic, and there are three holes, one located
away from the tube support plate (TSP),4 one located under the center of the TSP,

4 Tube support plate and tube support ring are synonymous terms.
19.4 Modeling Direct and Inverse Problems in Ferritic Tubes 393

R0

L0
M0
Ζin
L1 R1
Lμ Mμ

Fig. 19.10 A coupled-circuit model of the coil in the presence of the ferritic tube. R0 and L0 are,
respectively, the resistance and self-inductance of the coil in freespace, and Lμ is the increased
inductance of the coil due to the permeability of the tube. L1 is the “virtual” secondary inductance
that accounts for induction effects within the tube, M0 is the mutual inductance between L0 and
L1 , and Mμ is the mutual inductance between Lμ and L1 . R1 is the effective “secondary resistance”
that is due to the electrical conductivity of the tube

and one located at the edge of the TSP. We will characterize the tube and then the
composite structure, including the TSP and holes in the following subsections.

19.4.2 A Coupled-Circuit Model

We have seen several times that there is an advantage to working in the

lower-frequency range when trying to discriminate between conductivity and
permeability changes, because the correlation between these two parameters is
smallest there. We can see this again by referring to the coupled-circuit model of
the coil and tube shown in Fig. 19.10. R0 and L0 are, respectively, the resistance and
self-inductance of the coil in freespace, and Lμ is the increased inductance of the
coil due to the permeability of the tube. L1 is the “virtual” secondary inductance that
accounts for induction effects within the tube, M0 is the mutual inductance between
L0 and L1 , and Mμ is the mutual inductance between Lμ and L1 . R1 is the effective
“secondary resistance” that is due to the electrical conductivity of the tube.
Elementary coupled-circuit theory yields an expression for the driving-point
impedance of the loaded coil:

ω 2 Mμ2 ω 2 M02
Zin = R0 + jω L0 + jω Lμ + + . (19.3)
R1 + jω L1 R1 + jω L1
From this we get the change in impedance due to the presence of the tube:

δ Zin = Zin − R0 − jω L0
ω 2 Mμ2 ω 2 M02
= jω Lμ + +
R1 + jω L1 R1 + jω L1
394 19 Coupled Problems in Heat-Exchanger Tubes

Table 19.6 Inversion results using 100 Hz to 1 kHz data

Location Φ σ /sensit μ /sensit No. pts
A 0.188(−2) 1.372 × 106 /2.90(−2) 68.18/3.76(−3) 500
B 0.397(−2) 1.552 × 106 /5.66(−2) 73.43/8.35(−3) 500

ω 2 Mc2 (R1 − jω L1 )
= jω Lμ +
R21 + ω 2L21
 
ω 2 Mc2 R1 ω 2 L1 Mc2
= 2 + jω Lμ − 2 (19.4)
R1 + ω 2 L21 R1 + ω 2 L21

where Mc2 = M02 + Mμ2 . Again, we see that at low frequencies the effects of
conductivity and permeability of the tube are uncoupled. The first term in (19.4)
is the effective coupled resistance as a function of frequency, whereas the second
term displays the coupled inductance as a function of frequency. Note, in particular,
the quadratic dependence on ω of the resistance at low frequencies, ω << R1 /L1 .
Furthermore, when ω is very small the coupled inductance is simply Lμ , the
inductance of the coil due solely to permeability. The coupling to R1 , which
corresponds to conductivity, is virtually nil. Depending upon the values of the
various parameters, the coupled inductance, while starting out as a positive number,
may go negative. If the permeability of the tube were that of freespace, then Lμ = 0,
and the change in coupled inductance would be negative at all frequencies, in
accordance with Lenz’ law.

19.4.3 Characterizing the Tube: Outer-Coil Experiments

19.4.3.1 Inversion Using 100 Hz to 1 kHz Data

We perform a series of experiments using an outer coil and an inner coil to

characterize the tube. We inverted data taken at two locations with the outer coil:
location A, near one end of the tube, and location B, near the center. The results
of the inversions are shown in Table 19.6. These results have a residual norm
that is quite small, with sensitivities that are consistently small for each estimated
variable. We attribute this to “unraveling” the confounding variables, σ and μ ,
at low frequencies, exactly as our analysis above indicated. The fifth column in
Table 19.6 gives the number of points, out of the original distribution of 500 (recall
Sect. 12.2.2), that are mapped into the global minimum.
The results of the inversion are compared with the original data in Fig. 19.11. It is
clear that the inversion not only gives a good fit to the data, but that the quadratic
depence of δ R on frequency, for low frequencies, holds, as derived in (19.4).
Furthermore, the reactance, δ X, varies linearly with frequency, which means that
19.4 Modeling Direct and Inverse Problems in Ferritic Tubes 395

Inversion Results Using 100Hz to 1kHz Data: A Inversion Results Using 100Hz to 1kHz Data: A
0.07 0.8

0.06 0.7

0.6
0.05 Measured
Model
0.5
R(Ohms)

X(Ohms)
0.04
Measured
Model 0.4
0.03
0.3
0.02
0.2
0.01 0.1
0 0
100 200 300 400 500 600 700 800 900 1000 1100 100 200 300 400 500 600 700 800 900 1000 1100
Frequency (Hz) Frequency (Hz)

Inversion Results Using 100Hz to 1kHz Data: B Inversion Results Using 100Hz to 1kHz Data: B
0.09 0.8
0.08 0.7
0.07 0.6
Measured
0.06 Model
X(Ohms)

0.5
R(Ohms)

0.05 Measured
Model
0.4
0.04
0.3
0.03
0.02 0.2

0.01 0.1

0 0
100 200 300 400 500 600 700 800 900 10001100 100 200 300 400 500 600 700 800 900 1000 1100
Frequency (Hz) Frequency (Hz)

Fig. 19.11 Comparing the model results, based on the results of Table 19.6, with the original
measured data over the frequency range of 100 Hz to 1 kHz. Top: location A. Bottom: location B

we are seeing the effects of Lμ , only, with no coupling to the conductivity in the
tube. This further confirms our model results and demonstrates the advantage of
working at low frequencies to distinguish conductivity and permeability effects.

19.4.3.2 Inversion Using 10–100 kHz Data

When we use the inverted results of Table 19.6 for location A, and generate a
model response over the frequency range 10–100 kHz, we get the results shown
in Fig. 19.12. Because the resistance values are in good agreement over this
frequency range, but the reactances differ considerably, it is clear that Lμ in (19.4) is
underestimated, which in turn suggests that μ is too small. We do another inversion
to determine σ and μ , this time using the measured data for location A shown in
Fig. 19.12. The results of this high-frequency inversion are shown in Table 19.7 and
are plotted in Fig. 19.13.
396 19 Coupled Problems in Heat-Exchanger Tubes

Inversion Results Over 10kHz to 100kHz Inversion Results Over 10kHz to 100kHz
16 7.5

14 7

6.5 Measured
12
Measured Model
6
R(Ohms)

X(Ohms)
Model
10
5.5
8
5
6
4.5
4 4

2 3.5
10 20 30 40 50 60 70 80 90 100 10 20 30 40 50 60 70 80 90 100
Frequency (kHz) Frequency (kHz)

Fig. 19.12 Showing the high-frequency response at location A when the low-frequency parame-
ters of Table 19.6 are used

Table 19.7 Inversion results using 10–100 kHz data

Location Φ σ /sensit μ /sensit No. pts
A 1.399 1.319 × 106 /7.803(−2) 78.636/0.232 426

Inversion Results Using 10kHz to 100kHz Data Inversion Results Using 10kHz to 100kHz Data
18 7.5

16 7

14 6.5
Measured
Model
12 6
R(Ohms)

X(Ohms)

Measured
Model
10 5.5

8 5

6 4.5

4 4

2 3.5
10 20 30 40 50 60 70 80 90 100 10 20 30 40 50 60 70 80 90 100

Frequency (kHz) Frequency (kHz)

Fig. 19.13 Comparing the model results, based on the results of Table 19.7, with the original
measured data for location A over the frequency range of 10–100 kHz

The fact that the norm of the residuals, Φ , is much larger here than in Table 19.6
is not surprising, given the much larger data values here, nor is it surprising that
the sensitivity coefficient for μ is much larger, given our observation that the two
variables have the smallest correlation at low frequencies. The important point to
note is that σ is changed by only 4 %, which is probably in reasonable agreement
with the data, but that μ increases by 15 %. It seems likely that the origin of the
ferromagnetic response is dispersive, which yields a permeability that is frequency
dependent. This frequency dependence is reasonable, given that dipolar effects,
whether electrical or magnetic, have their origins in electrons that are bound to the
19.4 Modeling Direct and Inverse Problems in Ferritic Tubes 397

Results Based on Trial and Error Results Based on Trial and Error
18 7.5

16 7

14 6.5

12 6
R(Ohms)

X(Ohms)
Measured
Model
10 5.5
Measured
Model
8 5

6 4.5

4 4

2 3.5
10 20 30 40 50 60 70 80 90 100 10 20 30 40 50 60 70 80 90 100
Frequency (kHz) Frequency (kHz)

Fig. 19.14 Comparing measured and model results over 10–100 kHz for location A, when the
model comprises σ = 1.372 × 106 S/m, as in the low-frequency regime, and μ = 85, chosen by
trial and error

Results Based on Trial and Error for Location B Results Based on Trial and Error for Location B
16 6.5

14 6

12
5.5
R(Ohms)

X(Ohms)

10
Measured
Model 5
8
4.5
6

4 4

2 3.5
10 20 30 40 50 60 70 80 90 100 10 20 30 40 50 60 70 80 90 100
Frequency (kHz) Frequency (kHz)

Fig. 19.15 Comparing measured and model results over 10–100 kHz for location B, when the
model comprises σ = 1.552 × 106 S/m, as in the low-frequency regime, and μ = 86, chosen by
trial and error

host lattice, and are not free to move as are those electrons in the conduction band
[101]. This suggests that dipoles should share the frequency response of the host
lattice, including resonance and similar dispersive phenomena.
An alternative strategy would be to keep the host conductivity constant at
σ = 1.372 × 106 S/m for all frequencies for location A, and then determine a
value of μ that is consistent with the measured data simply by trial and error.
We have done this, and found that μ = 85 gives excellent results, as shown in
Fig. 19.14. Figure 19.15 shows the corresponding results for location B, using
σ = 1.552 × 106 S/m and μ = 86. In either case, it is clear from Figs. 19.14
and 19.15 that the high-frequency reactance demonstrates the nonlinear behavior
shown in (19.4) that is due to Lenz’ law for coupled circuits.
398 19 Coupled Problems in Heat-Exchanger Tubes

Inner-Coil Responses Inner-Coil Responses

0.8 0.35

0.7
0.3
0.6
0.25 A
0.5
R(Ohms)

X(Ohms)
B
BTSP
0.4
A 0.2 BF
B
0.3 BTSP
BF 0.15
0.2
0.1
0.1

0 0.05
10 20 30 40 50 60 70 80 90 100 10 20 30 40 50 60 70 80 90 100
Frequency (kHz) Frequency (kHz)

Fig. 19.16 Inner-coil responses at four probe locations. A: At end of tube. B: Near center of tube.
BF: At center with flaw (drilled hole). BTSP: Near center with tube support plate over it

19.4.4 Characterizing the Tube: Inner-Coil Experiments

19.4.4.1 Results Using 10–100 kHz Data

We have taken multifrequency data using the inner coil at four locations along the
tube. Location A is at an end, B is near the center, in the absence of any anomalies,
BF is at the flaw, which is a drilled hole, and BTSP is near the center under the tube-
support plate. We show only the high-frequency data from 10 to 100 kHz, because
those are the most reliable.5 This is not a problem, because we are not interested in
distinguishing conductivity effects from permeability, as this has already been done
using the outer coil.
Figure 19.16 shows the impedance responses at these four locations. It is clear
that the response at A is distinguished because of its location near the end of the
tube. It is also clear from the resistance response that the tube-support plate is not as
clearly distinguished from its background (B) as is the flaw. This is reasonable, given
that the skin-effect is pronounced at these frequencies, which effectively shields the
TSP from the coil. The flaw, being a surface-breaking inner-diameter hole, is not
similarly shielded.
We have noted that over the frequency range of 10–100 kHz the permeability
does not remain fixed at its low-frequency value. By setting μ = 104 over this
frequency range, and letting the conductivity retain its low-frequency value of

5 Theinner coil is much smaller than the outer, with fewer turns, which means that it has a smaller
inductance, and is, therefore, prone to produce noisier low-frequency data.
19.4 Modeling Direct and Inverse Problems in Ferritic Tubes 399

0.8 0.35

0.7
0.3
0.6
0.25
R(Ohms)

0.5 Model

X(Ohms)
Model B-data
B-data
0.4 0.2

0.3
0.15
0.2
0.1
0.1

0 0.05
10 20 30 40 50 60 70 80 90 100 10 20 30 40 50 60 70 80 90 100
Frequency (kHz) Frequency (kHz)

Fig. 19.17 Showing the model inner-coil response over the frequency range of 10–100 kHz,
together with the measured data for location B. The model parameters used to obtain this result are
σ = 1.552 × 106 S/m, and μ = 104

σ = 1.552 × 106 for location B, we generate the model response for the inner coil
shown in Fig. 19.17, along with the measured data for location B. The responses are
in good agreement within a few percent over this frequency range.6

19.4.5 Characterizing the Composite Structure

19.4.5.1 Model Results

Now that the host ferritic tube has been characterized, we proceed with the main
thrust of this section: characterizing the composite structure consisting of the inner
coil, tube, through-wall hole, and tube support plate. We’ll start with some model
calculations, because these results will allow us to interpret the measured data that
will be described in the next section. The configuration that produces the model
results of this section is the same as that shown in the top of Fig. 19.9. It uses the
same values of σ = 1.552 × 106 S/m and μ = 104 for the host parameters as in
Fig. 19.17.
In Fig. 19.18, we show the results of a scan past the isolated hole at five
frequencies, 1, 10, 50, 75, and 100 kHz. The important feature here is the monotonic
increase (in the negative direction for R) of the responses with frequency. These
results should be contrasted with the corresponding ones for the isolated TSP, shown
in Fig. 19.19. The frequency responses of the TSP are much more convoluted, in

6 One possible explanation for the difference in the values of the permeabilities of Figs. 19.15
and 19.17 is due to material stress inhomogeneities through the wall thickness. See [101] for a
discussion of magnetoelastic effects on permeability.
400 19 Coupled Problems in Heat-Exchanger Tubes

Pit-Only Responses at Five Frequencies Pit-Only Responses at Five Frequencies

0.0005 0.0035

0 0.003

0.0025
−0.0005 1kHz
1kHz 10kHz
10kHz 0.002 50kHz
R(Ohms)

X(Ohms)
−0.001 50kHz 75kHz
75kHz 100kHz
100kHz
0.0015
−0.0015
0.001
−0.002
0.0005
−0.0025 0

−0.003 −0.0005
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
Scan Position (in) Scan Position (in)

Fig. 19.18 Inner-coil responses at five frequencies for the isolated through-wall hole

TSP-Only Responses at Five Frequencies TSP-Only Responses at Five Frequencies

0.0002 0.0012

0.0001 0.001

0.0008 1kHz
0
10kHz
0.0006 50kHz
R(Ohms)

X(Ohms)

−0.0001 75kHz
1kHz 100kHz
10kHz 0.0004
−0.0002 50kHz
75kHz
100kHz 0.0002
−0.0003
0
−0.0004 −0.0002

−0.0005 −0.0004
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
Scan Position (in) Scan Position (in)

Fig. 19.19 Inner-coil responses at five frequencies for the isolated tube-support plate

the sense that they “flip” at low frequencies. This is typical of an outer-diameter
anomaly and is due to the “competition” between μ and σ in the ferritic tube wall, as
we have seen before. Furthermore, the TSP becomes invisible at higher frequencies
(at least as far as the inner coil is concerned), due to the reduced skin depth. This
reduction with frequency is pronounced due to the large magnetic permeability of
the tube.
The important contrast between Figs. 19.18 and 19.19 is in the low- and high-
frequency responses. At low frequencies, 1–10 kHz, the TSP response overwhelms
the hole response, but the situation is completely reversed in the high-frequency
regime at 50 kHz and above, for which the hole dominates the TSP. This suggests
that any protocol for detecting pits in the presence of tube-support plates will
necessarily require choosing the right frequency range in which to operate.
19.4 Modeling Direct and Inverse Problems in Ferritic Tubes 401

TSP-Pit Responses at Five Frequencies TSP-Pit Responses at Five Frequencies

0.0005 0.0035

0 0.003
1kHz
10kHz 0.0025
−0.0005 50kHz
75kHz
0.002
R(Ohms)

X(Ohms)
100kHz
−0.001
0.0015 1kHz
−0.0015 10kHz
0.001 50kHz
75kHz
−0.002 100kHz
0.0005
−0.0025 0

−0.003 −0.0005
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2
Scan Position (in) Scan Position (in)

Fig. 19.20 Inner-coil responses at five frequencies for the tube-support plate with the hole
centered beneath it

Collar-Pit Response Collar-Pit Response

0 0.0015

−0.0005
0.001
−0.001
R(Ohms)

X(Ohms)

0.0005
−0.0015

−0.002
0
−0.0025
−0.0005
−0.003

−0.0035 −0.001
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2

Scan Position (in) Scan Position (in)

Fig. 19.21 Inner-coil response at 100 kHz for the hole of Fig. 19.18 surrounded by a through-wall
collar whose conductivity is 1.552 × 106 S/m and permeability is 100. The diameter of the collar
is one inch

We plot the combined response of a hole centered under the TSP in Fig. 19.20.
It is clear that the TSP completely dominates the hole response at frequencies
of 10 kHz and lower, but that the hole dominates the high-frequency response, in
accordance with the isolated results of Figs. 19.18 and 19.19.
Later, when interpreting measured scan data, we will make use of a model,
whose result is shown in Fig. 19.21, comprising the through-wall hole of Fig. 19.18
surrounded by a through-wall collar whose permeability is 100 and conductivity is
1.522 × 106 S/m. The purpose of this model is to illustrate the significant effect
that even a small material inhomogeneity, such as magnetic permeability, can
have on the impedance response. Contrast, for example, the reactance response
at 100 kHz in Fig. 19.18 with the reactance response in Fig. 19.21. The former
402 19 Coupled Problems in Heat-Exchanger Tubes

Fig. 19.22 Change in resistance (left) and reactance (right) as a function of frequency and scan
position when the probe is scanned past the hole, located at 0.00”, and TSP, located between 1.580”
and 2.30”. The frequencies are 1, 10, 50, 75, and 100 kHz

does not change sign with scan, whereas the latter does. As we will see in the next
section, the background clutter associated with permeability inhomogeneities must
be accounted for in developing an inspection protocol for ferritic tubes.

19.4.5.2 Measured Data

The model results presented in the last section are useful in interpreting measured
data in which the inner-coil is scanned past the hole and TSP. Figure 19.22 shows
the response of the inner coil when it is scanned past the hole, located at 0.00”, and
the TSP, which starts at 1.580” and ends at 2.300”. The frequencies are 1, 10, 50,
75, and 100 kHz.
These results, while exhibiting considerable coherent background clutter, still
show qualitative agreement with our model computations. For example, consider the
reactance function shown in the right-hand side of Fig. 19.22. The hole portion of
this response shows two negative-going “descenders” bracketing the upward-going
response over the hole. This agrees with the model reactance response shown in
Fig. 19.21, which suggests, but certainly does not prove, that the source of the
coherent clutter could be an inhomogeneous distribution of magnetic permeability.
Furthermore, note that the reactance response associated with the hole increases
monotonically with frequency, exactly as we showed in Fig. 19.18.
19.4 Modeling Direct and Inverse Problems in Ferritic Tubes 403

The reactance function response of the TSP in the right-hand side of Fig. 19.22
is also in agreement with the model calculations; the 1 and 10 kHz responses have
opposite signs, which agrees with the results shown in Fig. 19.19. The fact that
each of these two measured responses has the opposite sign of the corresponding
model responses is due to choosing the model TSP parameter, σTSP , arbitrarily to
be 6.76 × 106 S/m; a more reasonable value of 1.0 × 106 S/m will be demonstrated
shortly. Furthermore, note that the two lowest measured TSP signals (1 and 10 kHz)
are much larger than the measured hole signals at 1 and 10 kHz and is the reason
that we must not use such low frequencies to try to locate a hole in the presence
of the TSP. This, too, was predicted in the model calculations, as shown in
Figs. 19.18–19.20. It is clear that the higher frequency signals, namely 50, 75, and
100 kHz, are associated with an inner-diameter anomaly, since they monotonically
increase with frequency. In fact, judging from the shape of these signals—they are
miniversions of the corresponding through-wall hole signals—the anomaly may
well be a small pit. It is likely, however, that these signals are simply part of the
background clutter, whose source needs to be studied further.
Further validation is given in the series of responses shown in Figs. 19.23–19.26,
which display the results for the isolated hole, the hole centered under the TSP, and
the hole placed at an edge of the TSP. These data were taken with an inner coil that
had more turns, and was somewhat larger than that used in the earlier experiments.
Thus, with the increased sensitivity of the coil, the responses are larger than those
shown previously, but retain the same qualitative features. In all of these figures, the
hole is located at 0.00 inch.
The results shown in these figures confirm our previous conclusions. Note that
the high-frequency response of the hole + TSP system approaches the isolated hole
response, no matter where the hole is located relative to the TSP. On the other hand,
the low-frequency response is dominated by the TSP, no matter where the hole is
located, and furthermore, the low-frequency response exhibits the usual “flipping”
phenomenon that we observed earlier. This is made quite clear in Figs. 19.25
and 19.26.

19.4.6 An Improved Low-Frequency Model

In our discussion of the results of Figs. 19.19 and 19.22, we determined that
the permeability of the TSP was in error, and this was due to the fact that we
had not characterized the TSP, but had simply used a previous model, as shown
at the top of Fig. 19.9. In order to rectify this situation, we reran some of the
models with the TSP and hole at low frequencies, namely 1, 10, and 19 kHz.
The new runs were performed with low-frequency permeabilities of 68.18 and
73.43, instead of the high-frequency permeability of 104. Furthermore, we ran
different host conductivities, namely 1.372 × 106 S/m and 1.552 × 106 S/m, which
we computed earlier to be correct for two measurement points with the outer coil.
404 19 Coupled Problems in Heat-Exchanger Tubes

Fig. 19.23 Change in resistance as a function of frequency and scan for (a) no TSP, (b) TSP over
hole center, (c) TSP over hole edge

Because we had not characterized the TSP independently of the tube, we varied
the TSP parameters during these trials, using combinations of μTSP = 50, and 100,
together with σTSP = 6.76 × 106, 1 × 106, 5 × 105 S/m. The combination of all of
these parameters that gave a good qualitative fit to the measured data is shown in
Table 19.8. These parameters are used in the lower configuration shown in Fig. 19.9.
The results of the runs are shown in the top row of Fig. 19.27, and when they
are compared with the corresponding measured data shown in the bottom row, we
see that there is a good qualitative agreement and that the relative magnitudes of the
reactance responses are in reasonable agreement with the measured data.
19.4 Modeling Direct and Inverse Problems in Ferritic Tubes 405

Fig. 19.24 Change in reactance as a function of frequency and scan for (a) no TSP, (b) TSP over
hole center, (c) TSP over hole edge

19.4.7 Comments and Conclusions

The problem of detecting and characterizing anomalies in ferritic tubes can be

solved using conventional eddy-current technology, with the use of multifrequency
scanned impedance data being the key to success. It is necessary to first characterize
the tube in order to determine the optimum operating frequencies to distinguish
classes of anomalies. Typically, one will want to operate at “high” frequencies in
order to reduce the effects of any external anomalies, such as a tube-support plate,
on ID or through-wall anomalies, such as flaws or pits.
406 19 Coupled Problems in Heat-Exchanger Tubes

Fig. 19.25 Change in resistance as a function of frequency and scan for (a) no TSP, (b) TSP over
hole center, (c) TSP over hole edge (low frequencies, only)

The presence of a large, coherent, background clutter, however, remains an

interesting challenge. Based on a comparison of the model results and measure-
ments, the metallurgical origin of this clutter could be inhomogeneities in the
magnetic permeability of the tube. This suggests that a determined effort be made to
study the statistical nature of this clutter, with the intention of developing a protocol
to mitigate its effects. Methods for doing this will be studied in the second volume
of this series.
A.1 Inverse Method Quality Metrics for the Ferritic Tube 407

Fig. 19.26 Change in reactance as a function of frequency and scan for (a) no TSP, (b) TSP over
hole center, (c) TSP over hole edge (low frequencies, only)

Table 19.8 Optimum

parameters for low-frequency σhost μhost σTSP μTSP
TSP-hole models 1.372 × 106 68.18 1 × 106 50

Appendix

A.1 Inverse Method Quality Metrics for the Ferritic Tube

Following the discussion of inverse method quality metrics and the Cramer–Rao
Lower Bound in Chap. 12, we optimize the estimation of conductivity and per-
meability of the ferritic tube. Figure 19.28 presents changes in the normalized
impedance plane for varying permeability (red) and conductivity (black) levels at
408 19 Coupled Problems in Heat-Exchanger Tubes

TSP-Pit Model Responses TSP-Pit Model Responses

at Low Frequencies at Low Frequencies
0.001 0.0025

0.0005 0.002 1kHz

10kHz
0 19kHz
0.0015
−0.0005
R(Ohms)

0.001
−0.001

X(Ohms)
1kHz
10kHz 0.0005
−0.0015 19kHz
0
−0.002

−0.0025 −0.0005

−0.003 −0.001
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2

Scan Position (in) Scan Position (in)

Fig. 19.27 Low-frequency model response of the TSP and hole when the data in Table 19.8 are
used. Top: Model. Bottom: Measured

four frequencies: (a) 100 Hz, (b) 1.0 kHz, (c) 10 kHz, and (d) 100 kHz. Several
observations can be made from these plots. First, it is helpful to increase frequency
in order to get sensitivity to conductivity changes. At very low frequencies, sensitiv-
ity to changes in permeability is dominant in the reactance term. As the frequency
approaches 100 kHz, it is impossible to distinguish changes in permeability and
conductivity in the eddy-current measurements.
Figure 19.29 presents a comparison of four different estimation metrics for this
inversion problem: (a) CRLB for variation in conductivity estimation, (b) CRLB
for variation in permeability estimation, (c) condition number, and (d) correlation.
Based on the CRLB for optimal sensitivity to permeability, lower frequencies are
better. Note, there is really not much benefit from using 100 Hz versus 1.0 kHz.
A.1 Inverse Method Quality Metrics for the Ferritic Tube 409

Fig. 19.28 Changes in normalized impedance plane for varying permeability and conductivity
levels at four frequencies: (a) 100 Hz, (b) 1.0 kHz, (c) 10 kHz, and (d) 100 kHz. Center values
considered in the study were σ = 1.4 × 106 S/m and μ = 60. The plus sign denotes a change in
response with respect to a change in permeability, and the box denotes a change in response with
respect to a change in conductivity

However, there is clearly an optimal region for estimating conductivity. Based on

the CRLB for variation in conductivity estimation, 6.0 kHz was found to be optimal.
This result was found to be in almost near agreement with the condition number
maximum of 5.0 kHz. Lastly, the correlation between the two estimation parameters
was found to be a problem (approaching 1) at 20 kHz and beyond. Thus, it is
recommended that inspectors work in a frequency range of 10 kHz and below for
simultaneously estimating permeability and conductivity.
410 19 Coupled Problems in Heat-Exchanger Tubes

Fig. 19.29 Comparison of estimation metrics: (a) CRLB for variation in conductivity estimation,
(b) CRLB for variation in permeability estimation, (c) condition number, and (d) correlation
Chapter 20
Applications to NDE of Coatings

20.1 Assessing Thermal Barrier Coatings

Characterizing surfaces and coatings has long been a staple of eddy-current

technology, but with the advent of inverse methods that technology has become
even more powerful. In this chapter we describe a case in which model-based
inversion, coupled with computational electromagnetics, can be effectively used to
solve complex coating problems.
The objective is to take measured impedance data and pass them through the
inversion algorithm, NLSE, and thereby reconstruct the layer(s) of a coating, i.e.,
determine the electromagnetic properties of each layer. We have applied VIC-
3D R
to a similar problem of assessing the PWA286 thermal barrier coating by
eddy-current inversion [99, 100]. Such coatings are important in maintaining the
life of new, high-performance gas turbines that are increasingly being used in
small power plants. Because the work reported in [99] and [100] involved actual
measured laboratory data, and because the techniques used to assess these coatings
are identical to what we are proposing in this book, it is useful to include this
material as a benchmark validation.
Advanced turbines are used in applications ranging from aerospace to land-based
power generators. These turbines are fired at higher temperatures (1,850–1,950◦F)
and utilize optimum cooling of hot section components. Because of the higher
operating temperature, the performance and durability of the first-stage blades has
become one of the prime life-limiting factors. Individual blades are nickel-based
GTD 111 alloys that are protected by sacrificial metallic coatings to extend service
life. The first-stage blades are especially important, and it is desirable to develop
an in situ NDE system to monitor, evaluate, and predict remaining coating life.
The coatings used on the turbine blades include CoCrAlY and NiCoCrAlY, with
a top aluminide coating (GT29+, GT33+, respectively), and a NiCoCrAlY coating,
called PWA 286. Figure 20.1 illustrates an as-coated PWA286 coating on a GTD111
substrate and the same coating after aging.

H.A. Sabbagh et al., Computational Electromagnetics and Model-Based Inversion, 411

Scientific Computation, DOI 10.1007/978-1-4419-8429-6 20,
© Springer Science+Business Media New York 2013
412 20 Applications to NDE of Coatings

β−Phase Depleted Zone 2

(Al2O3)
β−Phase Aluminide

β−Phase Aluminide
(MCrAlY)

β−Phase Depleted Zone 1

Interdiffusion Zone Interdiffusion Zone

GTD-111 Substrate GTD-111 Substrate

(Ni-based superalloy)

Fig. 20.1 An as-coated PWA286 coating on a GTD111 substrate (left) and the same coating after
2,400 h (right)

The primary objective is to estimate the equivalent thickness of the aluminum

beta-phase content of the PWA286 coating. This information is essential to main-
taining the integrity of blades, because it allows the timely repair or refurbishment
of coatings to extend the service-life of operating blades. Further, it is desirable to
obtain the interdiffusion layer thickness, since this information indicates the level
of blade exposure to service temperature. The overall remaining coating thickness
indicates the reduction of the coating thickness caused by the oxidation-induced
degradation of the top beta-phase depleted layer.
As the coating evolves with use, aluminum diffuses out to reform the protective
oxide coating and also diffuses into the substrate, causing a transformation of beta
phase NiAl into a gamma matrix of solid nickel solution. This gamma matrix
manifests itself as a beta-phase depleted zone (called “Zone1”), resulting in a
“standard model” shown in Fig. 20.2. Further, while the conductivity of the various
zones remains fixed over time, the layer-thicknesses of these zones changes, and
it is our job to determine the thicknesses of each layer, because this gives us an
indication of the remaining life of the coating.

20.1.1 Inversion of Impedance Data

Using an HP 3577A Network Analyzer, we took laboratory data on a number

of coupon-samples of PWA286 coatings that had undergone various levels of
deterioration. The measured impedances, from 1 to 50 MHz, are shown in Fig. 20.3.
We then applied the eight-layer algorithm, shown in Fig. 20.4, to invert these data.
20.1 Assessing Thermal Barrier Coatings 413

Conductivity (x105)
8.4
7.8

7.0

3.32
Depth
0
Zone2 Beta Zone1 IZ GTD111

Fig. 20.2 Standard model of the PWA286 thermal barrier coating. The conductivities are deter-
mined from measurements on known samples

Inverting the impedance data via the eight-layer algorithm means assigning a
value to L0 , the thickness of the top layer of Fig. 20.4, and assigning conductivities to
the remaining seven layers, each of which has the same thickness, L, that defines the
resolution of the reconstruction. This is done by a process of nonlinear least-squares
(NLSE), in which the eight variables are chosen to give the best fit to the impedance
data. The inversion is done in several steps, starting with the computation of a table
of data, from which the final solution is determined by interpolation. The process is
quite fast and is quite conservative in computational resources.
The results of the inversion process are fitted to the “standard model” of the
PWA286 TBC, that is shown in Fig. 20.2. The values of the conductivity of the
various layers are determined from inversions on known samples. In this manner,
we are able to determine the thickness of each zone. The results are tabulated in
Tables 20.1 and 20.2.
In several cases, the final computed conductivity profiles were obtained by a
process of “focussing” the eight-layer inversion algorithm. This is accomplished
by starting with a rather coarse grid, which allows the Beta Aluminide Zone and
Interdiffusion and Inner Beta Depleted Zone 1 to be generally located relative to the
GTD111 Substrate. Then we refine the grid between the Beta Aluminide Zone and
GTD111 Substrate to get a more precise value for the various zone boundaries, as
well as a more precise value for the conductivity of Zone 1 and the Interdiffusion
Zone. In this manner, we were able to locate the Interdiffusion Zone and determine
its conductivity for the “as-coated” samples, B1B and B1T. This is a fairly difficult
computation to carry out without such a “multigrid rezoning” technique.
The tabulated results confirm that the inversion algorithm produces excellent
reconstructions, with very good resolutions. In particular, we note the good agree-
ment between the computed and measured thickness of the beta-zone of each
sample. This is the critical datum, when it comes to determining the remaining life
of the thermal barrier coating.
414 20 Applications to NDE of Coatings

Real Part of Normalized Imaginary Part of Normalized

Impedances (Top) Impedances (Top)
0.18 −0.05
b1 b1
b2 b2
b3 b3
b4 −0.1 b4
0.16 b5 b5
w3 w3
w5 w5
−0.15
0.14
−0.2
Resistance

0.12

Reactance
−0.25

0.1 −0.3

−0.35
0.08
−0.4

0.06
−0.45

0.04 −0.5
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50
Frequency (MHz) Frequency (MHz)

Real Part of Normalized Imaginary Part of Normalized

Impedances (Bottom) Impedances (Bottom)
0.18 −0.05
b1 b1
b2 b2
b3 b3
b4 −0.1 b4
0.16 b5 b5
w3 w3
w5 w5
−0.15
0.14
−0.2
Resistance

0.12 −0.25
Reactance

0.1 −0.3

−0.35
0.08
−0.4

0.06
−0.45

0.04 −0.5
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50
Frequency (MHz) Frequency (MHz)

Fig. 20.3 Real (left) and imaginary (right) parts of the normalized change in impedance for the
“top” PWA286 samples (upper two figures) and “bottom” PWA286 samples (lower two figures)
20.1 Assessing Thermal Barrier Coatings 415

σ=0 Al2O3 L0
σ1 L

σ2 L

σ3 L

σ4 L
σ5 L

σ6 L

σ7 L

σ=7.8Ε5 GTD 111 Substrate

Fig. 20.4 The eight-layer inversion algorithm, in which L is the given resolution, and the objective
is to determine L0 and σ1 through σ7

20.1.2 Sample Calculation: White-5 Top

We’ll demonstrate the eight-layer algorithm on the White-5 Top (“w5t”) sample.
The numbers that we get may be a little bit different than what are shown in
Table 20.2 because we’ll follow slightly different paths to the final solution.
The data for the problem are: coil IR = 12.5 mils, coil OR = 45 mils, coil HT = 8
mils, and coil turns = 23. The position of the coil within the probe coordinate system
is [0, 0, 5.7874] mils. VIC-3D R
requires a minimum value for the z-coordinate
of 4 mils, and the coil was recessed 1 mil beneath the surface of its container.
The remaining 0.7874 mils is numerical fitting to give the best response to known
lift-off conditions. The probe origin position is [0, 0, Zone2]; i.e., we model the Zone
2 (“Outer Beta Depleted”) height by the z-coordinate in the probe origin position.
We use the 21 even-valued frequencies between 10 and 50 MHz for the model and
measured input data.
Our first calculation uses Fig. 20.4 with L = 20 microns, so that the grid extends
140 microns into the coating. The result of the inversion produced a value of Zone
2=48.8 microns, but the conductivities of the seven layers were not well determined.
The second calculation uses an eight-layer grid, but now with all eight layers
lying within the structure. Since each layer is 20 microns tall, the grid extends
160 microns into the structure. The coil coordinate system is as before, but the
probe coordinates are fixed at [0, 0, 48.8] microns. The results of this inversion are:
σ1 = 897969.603411648, σ2 = 748852.078029096, σ3 = 576648.447023520, σ4 =
821222.023798830, σ5 = 888060.286078721, σ6 = 300016.292476106,
σ7 = 300026.618414368, σ8 = 796133.184675437.
416 20 Applications to NDE of Coatings

Table 20.1 Comparison of measured and computed coating thicknesses for the
PWA286 test
Average Computed
Sample Coating Coating Thickness Thickness
Identification Location Identification (microns) (microns)
1950F-“as coated” Top Side Inner Diffusion 10.6 10.0
(Blue-1) Inner Beta Depleted – –
Beta Phase Zone 132.0 132.0
Outer Beta Depleted – –
Bottom Side Inner Diffusion 10.1 10.0
Inner Beta Depleted – –
Beta Phase Zone 140.9 142.0
Outer Beta Depleted – –
1950F-250 cycles Top Side Inner Diffusion 22.4 20.0
(Blue-2) Inner Beta Depleted 8.9 9.4
Beta Phase Zone 122.4 120.0
Outer Beta Depleted 14.5 17.7
Bottom Side Inner Diffusion 22.2 22.6
Inner Beta Depleted 8.1 9.1
Beta Phase Zone 114.2 112.5
Outer Beta Depleted 11.6 20.0
1950F-500 cycles Top Side Inner Diffusion 24.4 26.3
(Blue-3) Inner Beta Depleted 14.9 14.4
Beta Phase Zone 94.6 93.0
Outer Beta Depleted 17.9 19.2
Bottom Side Inner Diffusion 24.8 28.2
Inner Beta Depleted 16.8 15.3
Beta Phase Zone 98.0 94.5
Outer Beta Depleted 19.7 19.7

Even though there is considerable oscillation in the results (as is typical for these
problems), we use an averaging procedure, together with mixture theory, to make
sense of them. One reason for using a large number of layers is to allow averaging.
Note that the average of σ3 and σ4 is 6.99 × 105 S/m, which means that Zone 1
covers layers 3 and 4 (at least) (see Fig. 20.2). Furthermore, when we get such
agreement, it means that Zone 1 probably ends at a depth of 80 microns.
We can get more information about the start of Zone 1, and all of Beta by using
mixture theory on σ1 and σ2 . Let L be the volume-fraction associated with σ = 8.4×
105 S/m. Then the law for combining Beta and Zone 1 is: 8.9797 × 105 + 7.4885 ×
105 = 16.4682 × 105 = 16.8 × 105L+ 14 × 105(1 − L), or L = 0.8815. Since the total
distance covered is two layers, the length of Beta is 0.8815 × 40 = 35.26 microns.
The remaining 4.74 microns of the second layer belong to Zone 1, which means that
Zone 1 = 4.74 + 40 = 47.4 microns. Because of the oscillations in σ5 through σ7 ,
we must use a third calculation to compute IZ. It is pretty clear that σ8 is associated
with the GTD111 substrate. Another thing that makes us suspicious of the results
20.1 Assessing Thermal Barrier Coatings 417

Table 20.2 Comparison of measured and computed coating thicknesses for the
PWA286 test
Average Computed
Sample Coating Coating Thickness Thickness
Identification Location Identification (microns) (microns)
1950F-1000 cycles Top Side Inner Diffusion 25.8 26.4
(Blue-4) Inner Beta Depleted 34.0 33.0
Beta Phase Zone 62.6 63.8
Outer Beta Depleted 37.5 41.5
Bottom Side Inner Diffusion 23.7 25.0
Inner Beta Depleted 38.1 40.1
Beta Phase Zone 52.6 53.0
Outer Beta Depleted 33.2 34.8
1950F-1938 cycles Top Side Inner Diffusion 23.6 24.0
(Blue-5) Inner Beta Depleted 118.7 119.0
Beta Phase Zone – –
Outer Beta Depleted – 34.3
Bottom Side Inner Diffusion 28.4 25.0
Inner Beta Depleted 61.5 64.0
Beta Phase Zone 15.2 16.0
Outer Beta Depleted 48.5 46.3
1850F-3500 cycles Top Side Inner Diffusion 13.5 13.7
(White-5) Inner Beta Depleted 46.0 45.2
Beta Phase Zone 35.2 34.8
Outer Beta Depleted 50.2 44.6
Bottom Side Inner Diffusion 14.2 15.1
Inner Beta Depleted 45.9 45.5
Beta Phase Zone 27.6 29.6
Outer Beta Depleted 46.9 53.1

for σ5 through σ7 is that σ6 and σ7 are very close to the lower nodal value in the
two-point interpolation table. Thus, the NLSE algorithm may have “hit the stops”
for these two values.

20.1.3 Application of β -Scaling to the Thermal Barrier

Coating Problem

At this point, we have concluded that Zone 2 is 48.8 micrometers thick, the Beta
Zone is 35.26 micrometers thick, and Zone 1 is 44.74 micrometers thick. We will
apply the linear-filter algorithm to determine the thickness of the remaining zone,
IZ, and to do this we start with the four-layer model shown in Fig. 20.5.
We first define the 16 nodes of an interpolation table in Table 20.3, which shows
the conductivities (divided by 105 ) of layers 1–4, respectively.
418 20 Applications to NDE of Coatings

0mm

Beta: σ = 8.4E5
0.03526

Zone 1: σ = 7E5

0.080
Layer 1: σ = 3.32E5, 7.8E5
0.095
Layer 2: σ = 3.32E5, 7.8E5
0.110
Layer 3: σ = 3.32E5, 7.8E5
0.125
Layer 4: σ = 3.32E5, 7.8E5
0.140

GTD111: σ=7.8E5

Fig. 20.5 Illustrating the four-layer algorithm that will be applied to the determination of IZ of
the thermal barrier coating

Table 20.3 Interpolation table for applying the four-layer algorithm of

Fig. 20.5. The conductivities (divided by 105 ) of layers 1–4, respectively, are
shown
Node Conductivity Node Conductivity
1 3.32 3.32 3.32 3.32 9 3.32 3.32 3.32 7.8
2 7.8 3.32 3.32 3.32 10 7.8 3.32 3.32 7.8
3 3.32 7.8 3.32 3.32 11 3.32 7.8 3.32 7.8
4 7.8 7.8 3.32 3.32 12 7.8 7.8 3.32 7.8
5 3.32 3.32 7.8 3.32 13 3.32 3.32 7.8 7.8
6 7.8 3.32 7.8 3.32 14 7.8 3.32 7.8 7.8
7 3.32 7.8 7.8 3.32 15 3.32 7.8 7.8 7.8
8 7.8 7.8 7.8 3.32 16 7.8 7.8 7.8 7.8

In calculating β according to (15.6), we use the “White 5 Top” data of Fig. 20.3
as the measured (input) data for Z1 , and the model data corresponding to each of the
nodes in Table 20.3 for Z2 . The results are shown in Table 20.4.
The data in the third column of Table 20.4 are entered into an interpolation table
for NLSE. The nonlinear least-squares problem now becomes: min pi |β (pi ) − 1|,
where pi , i = 1, .., 4 are the four conductivities of Fig. 20.5. The result of the NLSE
inversion is shown in Table 20.5. The results agree well with our standard model
of Fig. 20.2, except that p4 = 3.754 × 105 is too small. Because of the relatively
large value of σ4 , however, we conclude that the solution is fairly insensitive to this
parameter, so that we can set its value to 7.8 × 105 to agree with the standard model.
20.1 Assessing Thermal Barrier Coatings 419

Table 20.4 Results for β corresponding to the nodes of Table 20.3. The measured data are the
“White 5 Top” data of Fig. 20.3
Node β |β − 1| Node β |β − 1|
1 1.0032 + j0.0071912 0.00787 9 1.0014 + j0.005348 0.005528
2 1.0025 + j0.0009032 0.002658 10 1.0008 − j0.0005944 0.000997
3 1.0018 + j0.00266 0.003212 11 1.0001 + j0.0011611 0.001165
4 1.0009 − j0.0030114 0.00314 12 0.99929 − j0.0042202 0.00428
5 1.0015 + j0.0041487 0.004412 13 0.99975 + j0.0026443 0.002656
6 1.0008 − j0.0016615 0.00184 14 0.99912 − j0.0028707 0.0030
7 1.000 + j0.00012423 0.000124 15 0.99832 − j0.0011249 0.00202
8 0.99912 − j0.0051704 0.005245 16 0.99756 − j0.0061376 0.0066

Table 20.5 Results of min pi |β (pi ) − 1|. σ denotes the sensitivity of the solution
to the conductivity. The conductivities are normalized to 105
Φ p1 /σ1 p2 /σ2 p3 /σ3 p4 /σ4
0.4899173(−3) 3.334/0.4589 7.8/0.5568 7.535/0.8178 3.754/1.3211

Table 20.6 Results for the “White 5 Top” coating

Layer Average of Measurements (μ ) Computed Results (μ )
IZ 13.5 15
Z1 46.0 44.74
Beta 35.2 35.26
Z2 50.2 48.8
Total 144.9 143.8

This implies that IZ exactly occupies Layer 1 of Fig. 20.5, and the remaining
three layers belong to the GTD111 substrate. Hence, IZ is 15 micrometers thick.
We tabulate the computed results for “White 5 Top” along with the average of
the experimentally measured results in Table 20.6. The results computed by the
β -scale method agree well with the measurements and the results computed earlier
by another method, as can be seen in Table 20.2.
Chapter 21
Model-Assisted Probability of Detection

21.1 Introduction

To validate the performance of inspection techniques, probability of detection

(POD) studies are performed. The preparation of POD samples with real fatigue
cracks and the process to acquire a statistically significant number of measurements
is often very time consuming and expensive, providing a significant burden for
the validation of new inspection techniques. A model-assisted strategy for the
design and execution of POD studies has been proposed to help mitigate the
validation costs and to improve POD evaluation quality by addressing a wider
array of inspection variables [118–121]. Using computer and empirical models
to address variables that cannot be easily recreated in experimental samples is a
significant opportunity. Two methodologies have been proposed to perform model-
assisted POD evaluations: the transfer function approach and the full-model-assisted
approach. The transfer function approach transforms an existing POD model for
one inspection to address another similar inspection based on a limited number
of new experimental measurements studying a specific varying parameter. This
approach has been successfully demonstrated to transform POD results for EDM
notch measurements in engine components [122] and aircraft structures [123] to
parts with real cracking conditions. The full-model-assisted approach uses computer
simulations to model the inspection process and determines the POD for the
inspection technique through a combination of experimental and simulated data.
This chapter will introduce the full-model-assisted POD methodology, discuss
the case study evaluation for the inspection of cracks around fastener sites in a
multilayer aircraft structure, and lastly discuss special topics.

H.A. Sabbagh et al., Computational Electromagnetics and Model-Based Inversion, 421

Scientific Computation, DOI 10.1007/978-1-4419-8429-6 21,
© Springer Science+Business Media New York 2013
422 21 Model-Assisted Probability of Detection

21.2 MAPOD Process

An overview of the MAPOD process is presented in this section based on prior work
[121,124–128]. A block diagram of the model building process for a model-assisted
POD evaluation is presented in Fig. 21.1. This process was developed primarily
through the efforts of Bruce Thompson and Chuck Annis and can be found in parts
in the Appendix H of the MIL-HNBK 1823A [121]. For any POD study, the scope
of the POD study must be assessed and all critical factors for the NDT technique,
part material, part geometry, and discontinuity characteristics that control signal
and noise must be identified. As well, the amount of variability (as distribution
functions) and the impact must be assessed through expert knowledge, experimental
results, and/or model-based studies.
Once key factors have been determined, an assessment can be made on whether
to pursue a model-based POD evaluation based on the cost of a fully experimental-
based study and the performance of available NDE simulation tools. Knowledge
of the accuracy and speed of NDT simulations is a critical part of this assessment
process. Thus, an evaluation of the model quality is often necessary at this stage.
For many applications, there will be a mix of simulated and empirical studies that
will provide the greatest coverage of the key factors for an inspection technique.
For example, certain parameters such as the dimensions and location of a crack in
the test part can be more easily controlled and varied through simulated studies.
As well, rare events or factor conditions can be readily assessed and incorporated
in the MAPOD evaluation. Alternatively, noise data from material specimens and

Fig. 21.1 Model-Assisted POD model building process with complete approach to uncertainty
propagation in MAPOD from MIL-HNBK 1823A, Appendix H (2009) [121]
21.2 MAPOD Process 423

measurement system noise can be evaluated quickly through low cost experimental
studies. From this perspective, a model-assisted POD evaluation has the potential
to not only reduce the cost of POD study through reduced test sample requirements
but also to improve the quality of the assessment by including physics-based models
that fully address the factors driving the evaluation process.
A critical component of a MAPOD evaluation is the assessment and propagation
of variability in parameter conditions and the uncertainty in the POD assessment.
Experimental-based POD evaluations must consider the impact of limited samples
on the confidence bounds for a POD model fits. Labels were added to Fig. 21.1 to
present a complete perspective on uncertainty propagation during the POD study.
As part of quantifying the effects of changing the key factors, the amount of
variation of each factor must be well understood [124, 125]. Variability in NDE
measurements is prevalent due to varying part geometries, material properties,
surface conditions, flaw morphology, NDE hardware, and human factors. A typical
representation of variability for an input factor or condition is as a probability
density function (pdf). In addition, confidence bounds for the pdf may be considered
as variation measures of the pdf distribution parameters. Multivariate distributions
would thus include model parameters and a variation (covariance) matrix.
Before models can be directly applied in POD evaluations, two steps are needed
to ensure their performance. First, model calibration involves model adjustments
to variables in a way that mimics the NDE technique procedure. In many cases,
gains and thresholds are set based on the desired response to a known calibration
standard. Parameters of the model are essentially fit to obtain the best match with a
limited set of empirical data acquired according to a calibration procedure. Bayesian
calibration approaches have been implemented to facilitate this analysis with limited
data [137]. Second, model validation ensures that models are in agreement with
well-controlled studies for the appropriate range of conditions expected in practice.
Note, uncertainty bounds on the measurements and numerical model error [129]
must be tracked and extended to any model-assisted evaluation.
Stochastic numerical models provide a means to efficiently represent random
variables in physical systems without excessive computational overhead. New
efficient methods such as the probabilistic collocation methods (PCM) will en-
able the greater use of stochastic studies with existing NDE models [130].
Fundamentally, this POD evaluation including uncertainty bounds becomes a two-
level analysis. Parameter variability is associated with the inherent variability or
randomness of a factor. Alternatively, uncertainty is associated with imperfect
knowledge, often requiring the need for more or better quality data. In practice,
parameter variability and uncertainty can be represented by the random variables
of a statistical distribution and their associated confidence intervals. Methods such
as second order probability analysis can be applied [131]. MAPOD demonstrations
using this approach have recently been performed [134, 136].
424 21 Model-Assisted Probability of Detection

Lastly, at the final stage of the POD evaluation, experimental and theoretically
assessed “â vs. a” models must be resolved. Here, an evaluation of the full noise and
signal distributions as a function of crack length is performed and the call criteria is
applied to evaluation POD and probability of a false call (POFC) or false call rate.
Uncertainty (or credibility) bounds on the POD evaluation must be calculated given
limited empirical samples, factor variability, uncertainty propagation, and model
error. Bayesian methods are ideally suited to incorporate empirical data with NDE
models that include prior information/distributions. A typical example of applying
Bayesian methods is through the addition of new empirical data to evaluate the
posterior distribution, a refinement to the original prior distribution is achieved.
Numerical methods such as Markov Chain Monte Carlo (MCMC) methods and
Bayes Factors can be applied to perform this evaluation [125, 132]. Note, care must
be taken to ensure that all assumptions applied in a Bayesian analysis are valid.
Although this process provides significant complexity to address the most complex
inspection problems, one must keep in mind the goal to simplify.

21.3 Case Study for MAPOD Evaluation of NDI of Fastener

Sites for Sub-surface Fatigue Cracks

The MAPOD case study problem presented here is the detection of cracks under in-
stalled countersunk fasteners in an airframe structures (see Fig. 21.2) [124,133,134].
For this problem, the key factors to include in the evaluation were identified from
experience with the inspection problem and prior experimental studies. The probe
characteristic response, lift-off, and scan resolution are important factors concerning
the NDE measurement. The surface condition of the samples, thickness of the
layers, type of fastener, fastener fit, and distance between adjacent fasteners and
edges are identified as significant factors related to part geometry and material.
The dimensions, aspect ratio, location around fastener site, and morphology are

Fig. 21.2 Diagram of a fastener site with (a) a first layer corner crack, (b) a second layer corner
crack, and (c) a second layer through crack
21.3 Case Study for MAPOD Evaluation of NDI of Fastener Sites... 425

important flaw characteristics to be considered. Also, the design of the algorithm

for extracting the crack measure from the image data will also play a critical role in
the results. Evaluation of the quality of the physics-based models has been the focus
of previous work [135]. Because VIC-3D R
has the capability to handle the generic
two-layer structural problem of interest, it was used for this study.
Details on the sample characteristics and eddy-current measurements for the
model are as follows. The dimensions for the top and bottom layers measured
3.96 mm (0.156) and 2.54 mm (0.100), respectively. Conductivities of 1.87 × 107
S/m for the aluminum layers and 1.79 × 106 S/m for the titanium fasteners were
used in the model. The size of the hole was set to a radius of 4.04 mm (0.159),
respectively. The probe frequency was set to 600 Hz and the permeability of 1,000
was used for the ferrite cup core. The coil dimensions had a height of 6.0 mm
(0.236), an inner radius of 3.0 mm (0.118), and an outer radius of 6.0 mm (0.236).
A corner crack was modeled for the first layer. Two crack conditions in the second
layer were simulated: a through crack and a corner crack with the aspect ratio length
to width to be 1:1. The transition from through to corner crack was assumed to be at
2.5 mm (0.100). For the simulations, all cracks were treated as notches of a finite
width of 0.25 mm (0.010), so it was assumed that there is no electrical contact
between crack faces. Crack lengths in the experimental samples were available
between 0.000 and 0.169.
The simulated and experimental data for a given probe and inspection geometry
were equated using a calibration procedure developed in prior work [135].
Figure 21.3 presents the calibration results demonstrating good agreement between
the model and simulated data for the no-crack fastener site. A Monte Carlo
simulation using 5,000 data points was performed using the noise measurement
distribution for the fastener site response with no cracks and combined with the
corner and/or through crack models to populate the full-model-assisted data sets
for POD evaluation. Figure 21.4a and b present the resulting distribution of data
points generated by the Monte-Carlo simulation. The POD curves were then fit for
the experimental and model-assisted data sets using the conventional maximum
likelihood approach for hit–miss data. The confidence bounds were determined
using the loglikelihood ratio method. Figure 21.5a and b present a comparison of the
POD results for the experimental and full-model-assisted approaches for first and
second layer cracks, respectively. For the first and second layer crack inspection,
both results are near the confidence bounds of the empirical data and thus near
agreement with the experimental solution. Note, these results resolve an error in an
assumption of the crack location (to be in second layer but in reality the first) for
the experimental data presented in prior results [133, 134]. For first layer cracks,
the curve generated with the MAPOD approach is near the upper 95 % confidence
bounds of the empirical data, while for second layer cracks, the curve is near the
lower bound.
These results are encouraging but indicate that the MAPOD evaluation is not
perfect. Potential sources for this remaining difference may include error in the
model/experiment fit or subtle details of the probe that are not being modeled
precisely. In general, there is good agreement between the two data sets with
426 21 Model-Assisted Probability of Detection

some error for select experimental data points. However, the transition region for
the crack response just exceeding the noise threshold is not in perfect agreement.
This difference is the primary source of differences in the two POD calculations.
The presence of coherent noise related to surface features (poor paint quality) was
observed for select experimental data. In particular, three of the four outliers noted
in Fig. 21.5a and b were found to be correlated with the presence of surface related
noise. The other outlier was found to be in close proximity to a steel fastener
and experienced greater noise as well. It is possible to actually shift the crack
measurement either out or in from the optimal location to minimize the effect of
the surface noise.

21.4 Bayesian Methods for Estimating Uncertain Parameters

in POD/MAPOD Evaluation

To fully address the challenge of performing a quality evaluation with limited

experimental samples, a model-based assessment must incorporate the variations
of the most significant input factors and appropriately integrate simulated and
experimental results. As part of this process, there are several critical components
in the evaluation that concern the propagation of varying conditions and uncertainty
through the model, or the updating of parameter estimates in the POD assessment.
Bayesian analysis is a statistical approach for evaluating a posterior distribution
of model parameters through combining information in the prior distribution with
new evidence according to Bayes theorem. Bayesian methods can be leveraged here
for the evaluation of three key components of the methodology: (1) evaluation of
variability for an input factor or condition is as a probability density function (pdf),
(2) model calibration [137], and (3) revised measurement model estimates with
experimental data. In practice, the posterior distribution can be evaluated, providing
a refinement to the original prior distribution through numerical methods such as
Markov Chain Monte Carlo (MCMC) simulation [137–139]. In prior work on NDE
reliability, Bayesian methods have been generally proposed [140, 141] with several
examples for the evaluation of statistical POD models for hit–miss data [142, 143]
and ahat-versus-a reliability studies [144]. In particular, Bayesian methods are
quite useful to address complex estimation problems and when prior information
is available. For example, Li, Meeker and Hovey [144] used Bayesian methods
for simultaneously evaluating the noise interference model for POD and the crack
distribution.
The estimation of both the calibration parameters of a physics-based model and
the distribution of select model factors, including crack aspect ratio and lift-off was
recently studied [145]. Code was developed to call a surrogate measurement model
representing the numerical results from VIC-3D R
. The following model estimation
problem was studied:
21.4 Bayesian Methods for Estimating Uncertain Parameters in POD/MAPOD Evaluation 427

Fig. 21.3 Calibration fit of model with respect to experimental data for titanium fastener site (in
Volts). Comparison plots include (a) the measurement plane, (b) the horizontal component and (c)
the vertical component responses

â = β0 + β1 f (a1 , β2 ) + ε , (21.1)

where ε ∼ N(0, σε2 ), f () is a function call for a physics-based model, β0 and β1 are
model calibration parameters, and β2 is a random variable associated with crack
aspect ratio (b/a). An initial study was performed where the random variables
428 21 Model-Assisted Probability of Detection

Fig. 21.4 Monte Carlo simulation results for varying length of (a) first layer (corner) cracks and
(b) second layer (through and corner) cracks

β0 , β1 , and β2 were estimated using a Bayesian (MCMC) approach. A plot of the

simulated data is given in Fig. 21.6 for corner flaws located at the faying surface
of the first and second layer. Initial results were found to be mixed due to the ill-
posedness of estimating β1 and β2 simultaneously.
To fully evaluate this model, there is also the need to provide a true estimate of
variance for crack aspect ratio random variable. Parametrized eddy-current models
21.4 Bayesian Methods for Estimating Uncertain Parameters in POD/MAPOD Evaluation 429

Fig. 21.5 POD evaluation results for experimental and full-model-assisted POD studies for
(a) first layer corner cracks and (b) second layer cracks

that include aspect ratio as a random variable as shown in Fig. 21.6 can reasonably
address both the mean response and nonconstant variance trends observed in
experimental results. The evaluation of these stochastic model parameters can
be achieved through evaluation of hierarchical Bayesian models. Gelman et al
introduced hierarchical Bayesian models for these classes of problems [146] and
430 21 Model-Assisted Probability of Detection

Fig. 21.6 Simulated results from the surrogate eddy-current model response as a function of crack
length based on VIC-3D R
numerical simulations for first and second layer cracks at a fastener site.
The aspect ratio is considered a Gaussian random variable with the mean and standard deviation
prescribed

presented several examples implemented using WinBUGS. Based on the eddy-

current case study problem, an example is presented here to investigate the ability to
simultaneously evaluate the model fit parameters and the variance terms for both the
noise in the measurement and stochastic variance in the model slope. The physics-
based hierarchical NDE measurement model is given as follows:

â = β0 + β1 f (a1 , β2 ) + εâ, (21.2)

where εâ ∼ N(0, σâ2 ), β2 ∼ N(μβ2 , σβ22 ), f () is a function call for a physics-based
model, β0 , β1 are model calibration parameters, and β2 is a random variable
associated with the crack aspect ratio (b/a) with unknown mean and variance.
A simplified test case hierarchical NDE measurement model is given as follows:

â = β0 + (β1 + η )a1 + εâ , εâ ∼ N(0, σâ2 )

(21.3)
η = εη , εη ∼ N(0, ση2 ),

where β0 and β1 are offset and mean slope terms of the model, respectively, η is a
random variable associated with varying-slope in the model, and ση2 is the variance
in slope parameter. Here, the random variable, η , is used to simulate the increasing
variance with increasing flaw size present in the physics-based model shown in
Fig. 21.6. The goal of this study is to assess how accurately these four parameters,
β0 , β1 , ση2 and, σâ2 , can be simultaneously estimated with respect to known values.
Two example estimation problems are presented in Fig. 21.7. Test case values
were selected to closely represent examples in prior experimental and simulated
21.4 Bayesian Methods for Estimating Uncertain Parameters in POD/MAPOD Evaluation 431

Fig. 21.7 Hierarchical model test cases estimating calibration parameters and variation in mea-
surement noise and stochastic model slope. Test cases shown (a) for strong model slope variation
and (b) for varying both slope and measurement noise significant factors

results in Fig. 21.6. The first case (a) investigated the condition where variance as a
function of flaw size dominates measurement noise (i.e., the variance independent
of flaw size). Using only 100 samples, the estimates for the two variance terms,
and, were found to be 0.2580 and 0.00139, respectively, close to the exact values of
0.300 and 0.0010. However, the 95 % credible bounds for both estimates just missed
containing the true values. This error may have been due to the specific random
sample used with the limited number of points or could be something systematic in
the estimation problem. Note, the estimates for the calibration parameters, β0 and
β1 , were found to be in good agreement with the true values for these case studies.
The variance terms appear to be the more challenging parameters to estimate in the
hierarchical model.
A second case in Fig. 21.7b investigated the condition where variance as a
function of flaw size is on a similar order as the measurement noise (variance
independent of flaw size). Using only 100 samples, the estimates for the two
variance terms, ση2 and σâ2 , were again found to be quite close to the exact
values of 0.100 and 0.0050, respectively, and the credible bounds included the
true values of the parameters. Both variance parameters are slightly overestimated,
while the calibration parameter β1 was slightly underestimated. All in all, these case
studies demonstrate the potential of simultaneously estimating the model calibration
parameters, model random variables, and measurement error.
432 21 Model-Assisted Probability of Detection

21.5 Model-Assisted Approach for Evaluating Localization

and Characterization Capability of NDE Techniques

There is a critical need to certify the capability of nondestructive evaluation

techniques to perform damage characterization. To achieve the grand objectives
of condition-based maintenance plus prognosis (CBM+), the location and size
of damage at any length scale must be determined. Conventional eddy-current
methods have been widely employed for the inspection of fastener sites and engine
components. For some applications, inspectors sometimes apply simple empirically
based formulas to equate the eddy-current response to flaw size. However, real
cracks have a variety of characteristics that impact eddy-current response such as
varying crack location, aspect ratio, contact condition (residual stress), location in
complex parts, and surface morphology that will impact the eddy-current response.
Often, the source for fatigue crack growth in early life is due to unexpected man-
ufacturing anomalies. Such conditions include poorly aligned panels or misdrilled
holes, leading to fastener sites with asymmetric fits and irregular stress conditions.
Thus, more sophisticated inverse methods have the potential to evaluate crack size
in the presence of these varying conditions. However, proper performance validation
of inverse methods must be demonstrated under real test conditions.
Although standards currently exist for evaluation of measurement systems [147],
there are limits to their application for assessing NDE characterization perfor-
mance. For example, conventional ANOVA-based methods do not address complex
underlying relationships well, often wrongly assume parameters are independents
neglecting joint probability, and typically do not naturally address both aleatory and
epistemic uncertainty. Thus, there is a need to develop a comprehensive procedure
for the evaluation of NDE characterization error (CE). This section will briefly
discuss an emerging approach for experimental and model-based procedures.
From the perspective of quantifying the reliability of NDE systems, one must
evaluate the relationship between the accuracy in estimating the damage or material
state estimates (â) with respect to the actual condition (a). An evaluation of the
characterization error (CE), ê = â − a, for all critical location and sizing estimates
is necessary. Characterization error with confidence bounds is the metric that is
proposed for use by condition-based maintenance and prognosis programs to help
evaluate remaining life and determine necessary maintenance actions. This problem
of evaluating characterization error with confidence bounds as a function of a critical
parameter such as flaw size is shown in Fig. 21.8a. This evaluation is generally
similar to the current procedure found in MIL-HDBK-1823A for the evaluation of
the relationship between an NDE measurement (â) and a critical flaw size (a) as
shown in Fig. 21.8b. The foundation for the experimental-based CE procedure will
be MIL-HDBK-1823A.
In practice, NDE and structural health monitoring (SHM) systems may measure
multiple values associated with damage location and size. The challenge here is to
develop a multidimensional characterization error model with confidence bounds,
ê j (ak ), where j = 1 · · · M, for locations and/or sizing parameters being estimated,
21.5 Model-Assisted Approach for Evaluating Localization and Characterization... 433

Fig. 21.8 Evaluation of (a) characterization error (ê j ) with respect to damage conditions (ak ) is
analogous to (b) â-vs-a POD analysis

Fig. 21.9 Model-assisted process for NDE characterization error (CE) evaluation

with k = 1 · · · N key factors, for example the critical flaw size. Gaussian process (GP)
models are ideally suited to evaluate this complex multidimensional relationship and
provide a useful model to support CBM+. A Gaussian Process model is a principled
probabilistic approach to representing data as a collection of random variables with
the property that the joint distribution of any finite subset is a Gaussian function.
This task will be to develop an empirical-based evaluation procedure following
MIL-HDBK-1823A but leveraging Gaussian process models.
Although experimental data are necessary in any measurement system evalua-
tion, due to the large number of factors that often must be addressed and the need
for a statistically significant number of test specimens, a model-assisted approach is
proposed as an alternative to an empirical-based evaluation of characterization error.
This process of model-assisted POD evaluation [121] may be adapted to address
a model-assisted approach to characterization error evaluation. A modified process
diagram for model-assisted NDE characterization error (CE) evaluation is presented
in Fig. 21.9. Throughout this process, proper error evaluation and uncertainty
434 21 Model-Assisted Probability of Detection

propagation must be managed. Three critical components of this evaluation include:

(1) assess key factors (as joint distributions) of the characterization (inversion)
process, (2) perform uncertainty propagation with stochastic numerical models,
and (3) revise model estimates with experimental data using Bayesian methods.
To facilitate the evaluation, the following tools will be leveraged in this process:
(a) Bayesian methods to estimate the joint distribution of unknown/uncertain
factors, (b) Stochastic Numerical Methods to facilitate uncertainty propagation of
input parameter distributions through the model, (c) Gaussian Processes to represent
characterization error (CE) as a range of random variables with a multivariate
normal distribution, and (d) Bayesian Calibration to revise the GP model using
Bayesian method with empirical data. Although the development and integration
of these computational tools are expected to facilitate the model-assisted CE
evaluation, this process is still complicated with respect to conventional reliability
analysis methods. As part of the procedure, a rule-of-thumb will be emphasized
throughout the process that focuses on the key drivers of error and uncertainty in the
evaluation.
References

1. Burrows, M.L.: A theory of eddy-current flaw detection. PhD Thesis, University of Michigan,
University Microfilms, Inc., Ann Arbor, Michigan (1964)
2. Förster, F., Breitfeld, H.: Theoretische und Experimentelle Grundlagen der zerstörungsfreien
Werkstoffprüfung mit Wirbelstromverfahren, (Parts I and II). Z. Metallkd. 43(5), 163–180
(1952)
3. Förster, F., Breitfeld, H., Stambke, K.: Theoretishe und Experimentelle Grundlagen der
zerstörungsfreien Werkstoffprüfung mit Wirbelstromverfahren, (Parts III–VII). Z. Metallkd.
45(5), 166–199 and 221–226 (1954)
4. Dodd, C.V., Deeds, W.E.: Analytical solutions to eddy-current probe coil problems. J. Appl.
Phys. 39(6), 2829–2838 (1968)
5. Luquire, J.W., Deeds, W.E., Dodd, C.V.: Alternating current distribution between planar
conductors. J. Appl. Phys. 41(10), 3983–3991 (1970)
6. Cheng, C.C., Dodd, C.V., Deeds, W.E.: General analysis of probe coils near stratified
conductors. Int. J. Nondestr. Test. 3, 109–130 (1971)
7. Nestor, C.W., Jr., Dodd, C.V., Deeds, W.E.: Analysis and computer programs for eddy current
coils concentric with multiple cylindrical conductors. Report No. ORNL-5220, Oak Ridge
National Laboratory, Oak Ridge, TN 37830, July 1979
8. Deeds, W.E., Dodd, C.V., Scott, G.W.: Computer-aided design of multifrequency eddy-
current tests for layered conductors with multiple property variations. Report No. ORNL/TM-
6858, Oak Ridge National Laboratory, Oak Ridge, TN 37830, October 1979
9. Miller, E.K.: Model-based parameter estimation in electromagnetics: III–applications to EM
integral equations. Appl. Comput. Electrom. 10(3), 9–29 (1995)
10. Murphy, K., Sabbagh, H.A.: A boundary-integral code for electromagnetic nondestructive
evaluation. In: Conference Proceedings: 12th Annual Review of Progress in Applied Compu-
tational Electromagnetics, Applied Computational Electromagnetics Society, 18–22 March
1996, pp. 171–178
11. Xie, H., Song, J., Yang, M., Nakagawa, N.: A novel boundary integral equation for surface
crack model. In: Thompson, D.O., Chimenti, D.E. (eds.) Review of Progress in Quantitative
Nondestructive Evaluation, vol. 29, pp. 329–336. American Institute of Physics, Melville
(2010)
12. Harrington, R.F.: Field Computation by Moment Methods. The Macmillan Company,
New York (1968)
13. Berreman, D.W.: Optics in Stratified and Anisotropic Media: 4 × 4-Matrix Formulation.
J. Opt. Soc. Am. 62(4), 502–510 (April 1972)
14. Altman, C., Schatzberg, A.: Appl. Phys. B 28, 327–333 (1982)
15. Altman, C., Schatzberg, A.: Appl. Phys. B 26, 147–153 (1981)

H.A. Sabbagh et al., Computational Electromagnetics and Model-Based Inversion, 435

Scientific Computation, DOI 10.1007/978-1-4419-8429-6,
© Springer Science+Business Media New York 2013
436 References

16. Altman, C., Schatzberg, A., Suchy, K.: IEEE Trans. Antenn. Propag. AP-32(11) (November
1984)
17. Schatzberg, A., Altman, C.: J. Plasma Phys. 26(Part 2), 333–344 (1981)
18. Suchy, K., Altman, C.: J. Plasma Phys. 13(Part 3), 437–449 (1975)
19. Krowne, C.M.: IEEE Antennas and Propagation Symposium Digest, Boston, MA, 25–29 June
1984, pp. 569–572
20. Krowne, C.M.: IEEE Trans. Microw. Theor. Tech. MTT-32(12), 1617–1625 (December
1984)
21. Krowne, C.M.: IEEE Trans. Antenn. Propag. AP-32(11), 1224–1230 (November 1984).
22. Roberts, T.M., Sabbagh, H.A., Sabbagh, L.D.: Electromagnetic interactions with an
anisotropic slab. IEEE Trans. Magn. 24(6), 3193–3200 (November 1988)
23. Roberts, T.M., Sabbagh, H.A., Sabbagh, L.D.: Electromagnetic scattering for a class of
anisotropic layered media. J. Math. Phys. 29, 2675–2681 (December 1988)
24. Roberts, T.M.: Explicit eigenmodes for anisotropic media. IEEE Trans. Magn. 26(6),
3064–3071 (November 1990)
25. Bowler, J.R., Sabbagh, L.D., Sabbagh, H.A.: A theoretical and computational model of eddy-
current probes incorporating volume integral and conjugate gradient methods. IEEE Trans.
Magn. 25(3), pp. 2650–2664 (May 1989)
26. Sabbagh, H.A., Bowler, J.R., Sabbagh, L.D.: A model of eddy-current probes with ferrite
cores. Nondestr. Test. Eval. 5(1), 67–79 (1989)
27. Sabbagh, H.A.: A model of eddy-current probes with ferrite cores. IEEE Trans. Magn. MAG-
23(3), 1888–1904 (May 1987)
28. Sabbagh, H.A., Sabbagh, L.D., Bowler, J.R.: A volume-integral code for eddy-current
nondestructive evaluation. Int. J. Comput. Math. Elec. Electron. Eng. 9(Suppl. A), 67–70
(1990)
29. Bowler, J.R., Sabbagh, L.D., Sabbagh, H.A.: Eddy-current probe impedance due to a surface
slot in a conductor. IEEE Trans. Magn. 26(2), 889–892 (March 1990)
30. Rao, S.M., Wilton, D.R., Glisson, A.W.: Electromagnetic scattering by surfaces of arbitrary
shape. IEEE Trans. Antenn. Propag. AP-30(3), 409–418 (May 1982)
31. Aubin, J-P.: Approximation of Elliptic Boundary-Value Problems. Wiley-Interscience,
New York (1972)
32. Glisson, A.W., Wilton, D.R.: Simple and efficient numerical methods for problems of
electromagnetic radiation and scattering from surfaces. IEEE Trans. Antenn. Propag. AP-29,
593–603 (1980)
33. Wertgen, W., Jansen, R.H.: Efficient direct and iterative electrodynamic analysis of geomet-
rically complex MIC and MMIC structures. Int. J. Numer. Model. Electron. Network. Dev.
Field. 2(3), 153–186 (September 1989)
34. Yaghjian, A.D.: Electric dyadic Green’s functions in the source region. Proc. IEEE. 68,
248–263 (February 1980)
35. Burke, G.J., Dease, C.G., Didwall, E.M., Lytle, R.J.: Numerical modeling of subsurface
communication. UCID-20439 Rev. 1, Lawrence Livermore National Laboratory, August 1985
36. Catedra, M.F., Gago, E., Nuño, L.: A numerical scheme to obtain the RCS of three-
dimensional bodies of resonant size using the conjugate gradient method and the fast Fourier
transform. IEEE Trans. Antenn. Propag. 37(5), 528–537 (May 1989)
37. Peter, A., Zwamborn, M., van den Berg, P.M., Mooibroek, J., Koenis, F.T.C.: Computation of
three-dimensional electromagnetic-field distributions in a human body using the weak form
of the CGFFT method. Appl. Comput. Electrom. 7(2), 26–42 (Winter 1992)
38. Zwamborn, P., van den Berg, P.M.: The three-dimensional weak form of the conjugate
gradient FFT method for solving scattering problems. IEEE Trans. Microw. Theor. Tech.
40(9), 1757–1766 (September 1992)
39. Sabbagh, H.A.: Splines and their reciprocal-bases in volume-integral equations. IEEE Trans.
Magn. 29(6), 4142–4152 (November 1993)
40. http://www.sabbagh.com
41. Dongarra, J.J., Moler, C.B., Bunch, J.R., Stewart, G.W.: LINPACK Users’ Guide. Society for
Industrial and Applied Mathematics, Philadelphia (1979)
References 437

42. Hestenes, M.: Conjugate Direction Methods in Optimization. Springer, New York (1980)
43. Sarkar, T.P.: Application of the Conjugate Gradient Method in Electromagnetics and Signal
Processing. Elsevier, New York (1991)
44. Catedra, M.F., Torres, R.P., Basterrechea, J., Gago, E.: The CG-FFT Method: Application of
Signal Processing Techniques to Electromagnetics. Artech House, Boston (1995)
45. Peterson, A., Ray, S., Mittra, R.: Computational Methods for Electromagnetics. IEEE,
New York (1998)
46. Chew, W.C., Jin, J.M., Michielsssen, E., Song, J.M.: Fast and Efficient Algorithms in
Computational Electromagnetics. Artech House, Boston (2001)
47. Harrington, R.F.: Time-Harmonic Electromagnetic Fields. McGraw-Hill, New York (1961)
48. Collin, R.E.: Field Theory of Guided Waves. McGraw-Hill, New York (1960)
49. Sabbagh, H.A., Sabbagh, E.H., Murphy, R.K.: Recent advances in modeling eddy-current
probes. In: Thompson, D.O., Chimenti, D.E. (eds.) Review of Progress in Quantitative
Nondestructive Evaluation, vol. 21, pp. 423–429. American Institute of Physics, Melville
(2002)
50. Wozencraft, J.M., Jacobs, I.M.: Principles of Communication Engineering. Wiley, New York
(1965)
51. Umashankar, K.R., Nimmagadda, S., Taflove, A.: Numerical analysis of electromagnetic
scattering by electrically large objects using spatial decomposition technique. IEEE Trans.
Antenn. Propag. 40(8), 867–877 (August 1992)
52. Murphy, K., Sabbagh, H.A., Treece, J.C.: Thickness measurements with eddy-current probes:
a simple inversion problem. In: Thompson, D.O., Chimenti, D.E. (eds.) Review of Progress
in Quantitative Nondestructive Evaluation, vol. 13, pp. 927–934. Plenum Press, New York
(1994)
53. Lawson, C.L., Hanson, R.J.: Solving Least Squares Problems. Prentice-Hall, Inc., Englewood
Cliffs (1974)
54. Murphy, K., Sabbagh, H.A., Treece, J.C.: Some inversion problems in nondestructive
evaluation. In: Thompson, D.O., Chimenti, D.E. (eds.) Review of Progress in Quantitative
Nondestructive Evaluation, vol. 14, pp. 857–861. Plenum Press, New York (1995)
55. Baltzersen, O.: Model-based inversion of plate thickness and liftoff from eddy current probe
coil measurements. Mat. Eval. 51, 72–76 (1993)
56. Förster, F. Libby, H.: Electromagnetic Testing, 2nd edn, pp. 178–179. The American Society
or Nondestructive Testing, Columbus (1986)
57. Sabbagh, H.A., Murphy, R.K. Sabbagh, E.H.: Advances in modeling eddy-current NDE of
ferromagnetic bodies. In: Thompson, D.O., Chimenti, D.E. (eds.) Review of Progress in
Quantitative Nondestructive Evaluation, vol. 22, pp. 383–389. American Institute of Physics,
Melville (2003)
58. Rinnooy Kan, A.H.G., Timmer, G.T.: Stochastic global optimization methods. Part I:
clustering methods. Math. Program. 39, 27–56 (1987)
59. Rinnooy Kan, A.H.G., Timmer, G.T.: Stochastic global optimization methods Part II: multi
level methods. Math. Program. 39, 57–78 (1987)
60. Byrd, R.H., Dert, C.L., Rinnooy Kan, A.H.G., Schnabel, R.B.: Concurrent stochastic methods
for global optimization. Math. Program. 46, 1–29 (1990)
61. Nakhkash, M., Huang, Y., Fang, M.T.C.: Application of the multilevel single-linkage method
to one-dimensional electromagnetic inverse scattering problem. IEEE Trans. Antenn. Propag.
47(11), 1658–1668 (November 1999)
62. Golub, G.H., Van Loan, C.F.: Matrix Computations. The Johns Hopkins University Press,
Baltimore (1983)
63. Rousseeuw, P., Yohai, V.: Robust regression by means of s-estimators. In: Robust and
Nonlinear Time Series Analysis: Proceedings of a Workshop, pp. 256–272, 1984
64. Van Trees, H.L.: Detection, Estimation, and Modulation Theory: Part I. Wiley, New York
(1968)
65. Kay, S.M.: Fundamentals of Statistical Signal Processing: Estimation Theory. Prentice-Hall,
Inc., Upper Saddle River (1993)
438 References

66. Devaney, A.J., Tsihrintzis, G.A.: Maximum likelihood estimation of object location in
diffraction tomography. IEEE Trans. Signal Process. 39(3), 672–682 ( March 1991)
67. Herman, G.T., Lent, A., Hurwitz, H.: A storage-efficient algorithm for finding the regularized
solution of a large, inconsistent system of equations. J. Inst. Math. Appl. 25, 361–366 (1980)
68. Censor, Y.: Row-action methods for huge and sparse systems and their applications. SIAM
Rev. 23, 444–446 (October 1981)
69. Herman, G.T., Meyer, L.B.: Algebraic reconstruction techniques can be made computation-
ally efficient. IEEE Trans. Med. Imag. 12(3), 600–609 (September 1993)
70. Sabbagh, H.A., Sabbagh, L.D., Vernon, S.N.: Verification of an eddy-current flaw inversion
algorithm. IEEE Trans. Magn. 22(6), 1881–1886 (November 1986)
71. Papoulis, A.: A new algorithm in spectral analysis and band-limited extrapolation. IEEE
Trans. Circ. Syst. CAS-22, 735–742 (1975)
72. Youla, D.C.: Generalized image restoration by the method of alternating orthogonal projec-
tions. IEEE Trans. Circ. Syst. CAS-25, 694–702 (1978)
73. Youla, D.C., Webb, H.: Image restoration by the method of convex projections: Part 1-Theory.
IEEE Trans. Med. Imag. MI-1, 81–94 (October 1982)
74. Sezan, M.I., Stark, H.: Image restoration by the method of convex projections:
Part 2-applications and numerical results. IEEE Trans. Med. Imag. MI-1, 95–101 (October
1982)
75. Oskoui-Fard, P., Stark, H.: Tomographic image reconstruction using the theory of convex
projections. IEEE Trans. Med. Imag. 7(1), 45–58 (March 1988)
76. Bucci, O.M., D’Elia, G., Mazzarella, G., Panariello, G.: Antenna pattern synthesis: a new
general approach. Proc. IEEE. 82(3), 358–371 (March 1994)
77. Oh, S., Marks II, R.J., Atlas, L.E.: Kernel synthesis for generalized time-frequency distri-
butions using the method of alternating projections onto convex sets. IEEE Trans. Signal
Process. 42(7), 1653–1661 (July 1994)
78. Lent, A., Tuy, H.: An iterative method for the extrapolation of band limited functions. J. Math.
Anal. Appl. 83, 554–565 (1981)
79. Kaczmarz, S.: Angenaherte auflosung von systemen linearer gleichungen. Bull. Pol. Acad.
Sci. Lett. A, 6–8A, 355–357 (1937)
80. Tanabe, K.: Projection method for solving a singular system. Numer. Math. 17, 203–214
(1971)
81. Hounsfield, G.N.: A method of and apparatus for examination of a body by radiation such as
x-ray or gamma radiation. Patent Specification 1283915, The Patent Office (1972)
82. Kak, A.C., Slaney, M.: Principles of Computerized Tomographic Imaging. IEEE Press,
New York (1988)
83. Goodman, J.W.: Introduction to Fourier Optics. McGraw-Hill, San Francisco (1968)
84. Pratt, W.K.: Digital Image Processing. Wiley, New York (1978).
85. Sabbagh, L.D., Sabbagh, H.A., Klopfenstein, J.S.: Image enhancement via extrapolation
techniques: a two dimensional iterative scheme and a direct matrix inversion scheme. In:
Thompson, D.O., Chimenti, D.E. (eds.) Review of Progress in Quantitative Nondestructive
Evaluation, vol. 5, pp. 473–483. Plenum Press, New York (1986)
86. Sabbagh, L.D., Sabbagh, H.A.: Inversion of eddy current data and the reconstruction of flaws
Part 2: inversion of data. In: Thompson, D.O., Chimenti, D.E. (eds.) Review of Progress in
Quantitative Nondestructive Evaluation, vol. 6, pp. 619–626. Plenum Press, New York (1987)
87. Sabbagh, L.D., Sabbagh, H.A.: Eddy-current modeling and flaw reconstruction. J. Nondestr.
Eval. 7(1/2), 95–110 (1988)
88. Sabbagh, H.A., Sabbagh, L.D.: An eddy-current model for three-dimensional inversion. IEEE
Trans. Magn. MAG-22(4), 282–291 (July 1986)
89. Sabbagh, L.D., Sabbagh, H.A.: Eddy current modeling and signal processing in NDE. In:
Chen, C.H. (ed.) Signal Processing and Pattern Recognition in Nondestructive Evaluation of
Materials: NATO ASI Series, vol. F44, pp. 145–154. Springer, Berlin (1988)
90. Theodoulidis, T.P., Poulakis, N., Bowler, J.R.: Developments in modeling eddy current coil
interactions with a right-angled conductive wedge. In: Takahashi, S., Kikuchi, H. (eds.)
Electromagnetic Nondestructive Evaluation X, pp 41–48. IOS Press, Amsterdam (2007)
References 439

91. Sabbagh, H.A., Sabbagh, L.D., Roberts, T.M.: An eddy-current model and algorithm for
three-dimensional nondestructive evaluation of advanced composites. IEEE Trans. Magn.
24(6), 3201–3212 (November 1988)
92. Murphy, K., Sabbagh, H.A. Treece, J.C.: Verification of a model of eddy-current probes with
ferrite cores. In: Thompson, D.O., Chimenti, D.E. (eds) Review of Progress in Quantitative
Nondestructive Evaluation, vol. 13, pp. 1089–1093. Plenum Press, New York (1994)
93. Mittleman, D.M., Jacobsen, R.H., Buss, M.C.: T-ray imaging. IEEE J. Sel. Top. Quant.
Electron. 2(3), 679–692 (September 1996)
94. Sze, S.M.: Physics of Semiconductor Devices. Wiley, New York (1969)
95. Sabbagh, H.A., Murphy, R.K., Woo, L.W., Sabbagh, E.H., Krzywosz, K.: Recent advances
in modeling eddy-current probe-flaw interactions. In: Thompson, D.O., Chimenti, D.E. (eds.)
Review of Progress in Quantitative Nondestructive Evaluation, vol. 15, pp. 331–338. Plenum
Press, New York (1996)
96. Burke, S.K., Ditchburn, R.J.: Mutual impedance of planar eddy-current driver-pickup spiral
coils. Res. Nondestr. Eval. 19, 1–19 (2008)
97. Moré, J.J., Garbow, B.S., Hillstrom, K.E.: USER GUIDE FOR MINPACK-1. ANL-80-74,
Argonne National Laboratory, August 1980
98. Collin, R.E.: Foundations for Microwave Engineering, Chap. 4. McGraw-Hill Book Com-
pany, New York (1966)
99. Sabbagh, H.A., Sabbagh, E.H., Murphy, R.K., Nyenhuis, J.: Assessing thermal barrier
coatings by eddy current inversion. Mater. Eval. 59(11), 1307–1312 (November 2001)
100. Sabbagh, H.A., Sabbagh, E.H., Murphy, R.K., Nyenhuis, J.: Assessing thermal barrier
coatings by eddy current inversion. In: Thompson, D.O., Chimenti, D.E. (eds.) Review of
Progress in Quantitative Nondestructive Evaluation, vol. 21, pp. 722–727. American Institute
of Physics, Melville (2002)
101. Wang, S. Solid-State Electronics. McGraw-Hill Book Company, New York (1966)
102. Knopp, J.S., Aldrin, J.C., Misra, P.: Considerations in the validation and application of models
for eddy current inspection of cracks around fastener holes. J. Nondestr. Eval. 25(3), 123–138
(2006)
103. Carpenter, D.C.: Use of the finite element method in simulation and visualization of
electromagnetic nondestructive testing applications. Mater. Eval. 58(7), 877–881 (2000)
104. Palanisamy, R., Lord, W.: Prediction of eddy current probe signal trajectories. IEEE Trans.
Magn. 16(5), 1083–1085 (1980)
105. Knopp, J.S., Aldrin, J.C., Sabbagh, H.A., Jata, K.V.: Estimation theory metrics in electromag-
netic NDE, electromagnetic nondestructive evaluation workshop proceedings, Seoul, Korea,
10–12 June 2008
106. Trefethen, L.N. Bau, D.: Numerical Linear Algebra. SIAM, Philadelphia (1997)
107. Sabbagh, H.A., Murphy, R.K., Sabbagh, E.H. Aldrin, J.C., Knopp, J., Blodgett, M.: Com-
putational electromagnetics and model-based inversion: a modern paradigm for eddy-current
nondestructive evaluation. Appl. Comput. Electrom. 24(6), 533–540 (December 2009)
108. Sabbagh, E.M.: Circuit Analysis. Ronald Press Company, New York (1961)
109. Ramo, S., Whinnery, J.R., Van Duzer, T.: Fields and Waves in Communication Electronics.
Wiley, New York (1965)
110. Vernon, S.N.: The universal impedance diagram of the ferrite pot core eddy current transducer.
IEEE Trans. Magn. 25(3), 2639–2645 (May 1989)
111. Bowler, J.R., Sabbagh, L.D., Sabbagh, H.A.: The reduced impedance function for cup-core
eddy-current probes. IEEE Trans. Magn. 25(3), 2646–2649 (May 1989)
112. ’Validation of Direct and Inverse Models of SEACURE Ferritic Tubes With Benchmark Data,’
PID069961, Technical Update, October 2010, Electrical Power Research Institute. Prepared
by Victor Technologies, LLC
113. Todorov, E., Levesque, S., Ames, N., Krzywosz, K.: Measurement of magnetic properites of
ferromagnetic tubes for heat exchangers. Trans. Am. Nucl. Soc. 104, 297–298 (26–30 June
2011, Hollywood, Florida)
440 References

114. Scully, J.R.: Hidden corrosion, what should be measured to improve emerging anticipate
and manage strategies. In: 32nd Annual Review of Progress in Quantitative Non-destructive
Evaluation, QNDE-Brunswick, Maine (Key Note Lecture for Symposium Kick-off General
Session) (2005)
115. Tian, Y., Tamburrino, A., Udpa, S.S., Udpa, L.: Time-of-flight measurements from eddy
current tests. In: Thompson, D.O., Chimenti, D.E. (eds.) Review of Progress in Quantitative
Nondestructive Evaluation, vol. 22, pp. 593–600. American Institute of Physics, Melville
(2003)
116. Liu, Z., Safizadeh, M.-S. Forsyth, D.S., Lepine, B.A.: Data fusion method for the optimal
mixing of multi-frequency eddy current signals. In: Thompson, D.O., Chimenti, D.E.
(eds.) Review of Progress in Quantitative Nondestructive Evaluation, vol. 22, pp. 577–584.
American Institute of Physics, Melville (2003)
117. Liu, X., Deng, Y., Zeng, Z., Udpa, L., Knopp, J.: Model based inversion technique of
GMR signals using element-free Galerkin method. In: Conference Proceedings: 24th Annual
Review of Progress in Applied Computational Electromagnetics, Applied Computational
Electromagnetics Society, pp. 221–226 (March 2008)
118. Gray, J.N., Gray, T.A., Nakagawa, N., Thompson, R.B.: Models for predicting NDE reliabil-
ity, nondestructive evaluation and quality control, Metals Handbook 17, pp 702–715. ASM
International, Ohio, 1989
119. Thompson, R.B.: Using Physical Models of the Testing Process in Determination of
Probability of Detection. Mater. Eval. 69(7), 861–865 (2001)
120. Thompson, R. B.: A unified approach to the model-assisted determination of probability of
detection. Mater. Eval. 66, 667–673 (2008)
121. U.S. Department of Defense, Handbook, Nondestructive Evaluation System Reliability
Assessment, MIL-HDBK-1823A, 7 April 2009
122. Smith, K.D., Thompson, R.B., Brasche, L.: Model-Based POD: Successes and Opportunities,
1st Meeting of the MAPOD Working Group, Albuquerque, New Mexico, 23–24 September
2004. Web site: http://www.cnde.iastate.edu/MAPOD/
123. Harding, C., Hugo, G., Bowles, S.: Model-assisted Probability of Detection Validation of
Automated Ultrasonic Scanning for Crack Detection at Fastener Holes, Proceedings of the
10th Joint FAA/DoD/NASA Conference on Aging Aircraft, Palm Springs, CA, 16–19 April
2007
124. Aldrin, J.C., Knopp, J.S., Lindgren E.A., Jata, K.V.: Model-assisted probability of detection
(MAPOD) evaluation for eddy current inspection of fastener sites. Rev. progr. Quant.
Nondestr. Eval. 28, AIP, 1784–1791 (2009)
125. Aldrin, J.C., Medina, E.A., Lindgren, E.A., Buynak, C., Knopp, J.: Case studies for model-
assisted probabilistic reliability assessment for structural health monitoring systems. Rev.
progr. Quant. Nondestr. Eval. 30, AIP, 1589–1596 (2011)
126. Aldrin, J.C., Medina, E.A., Lindgren, E.A., Buynak, C.F., Knopp, J.S.: Protocol for reliability
assessment of structural health monitoring systems incorporating model-assisted probability
of detection (MAPOD) approach. In: Chang, F.-K. (ed.) Proceedings of the 8th International
Workshop on Structural Health Monitoring, Stanford, 13–15 September 2011
127. Aldrin, J.C., Medina, E.A., Santiago, J., Lindgren, E.A., Buynak, C.F., Knopp, J.S.:
Demonstration study for reliability assessment of SHM systems incorporating model-assisted
probability of detection approach. In: Thompson, D.O., Chimenti, D.E. (eds.) Review of
Progress in Quantitative Nondestructive Evaluation, vol. 31, pp. 1543–1550. American
Institute of Physics, Melville (2012)
128. Aldrin, J.C., Sabbagh, H.A., Murphy, R.K., Sabbagh, E.H., Knopp, J.S., Lindgren, E.A.,
Cherry, M.R.: Demonstration of model-assisted probability of detection evaluation method-
ology for eddy-current nondestructive evaluation. In: Thompson, D.O., Chimenti, D.E. (eds.)
Review of Progress in Quantitative Nondestructive Evaluation, vol. 31, pp. 1733–1740.
American Institute of Physics, Melville (2012)
129. Oberkampf, W.L., Roy, C.J.: Verification and Validation in Scientific Computing. Cambridge
University Press, New York (2010)
References 441

130. Knopp, J.S., Aldrin, J.C., Blodgett, M.P.: Efficient propagation of uncertainty in simula-
tions via the probabilistic collocation method. In: Chady, T., Gratkowski, S., Takagi, T.,
Udpa, S.S. (eds.) Electromagnetic Nondestructive Evaluation (XIV), pp. 141–148. IOS Press,
Amsterdam (2011)
131. Frey, H.C.: Quantitative analysis of uncertainty and variability in environmental policy mak-
ing. Fellowship Program for Environmental Science and Engineering, American Association
for the Advancement of Science, Washington, DC (1992)
132. Mahadevan, S., Rebba, R.: Validation of reliability computational models using Bayes
networks. Reliab. Eng. Syst. Saf. 87, 223–232 (2005)
133. Knopp, J.S., Aldrin, J.C., Lindgren, E., Annis, C.: Investigation of a model-assisted approach
to probability of detection evaluation. Rev. Prog. Quant. Nondestr. Eval. 26, 1775–1782
(2007)
134. Aldrin, J.C., Knopp, J.S.: Modeling and simulation for nondestructive testing with applica-
tions to aerospace structures. Mater. Eval. 66(1), 53–59 (2008)
135. Knopp, J.S., Aldrin, J.C., Misra, P.: Considerations in the validation and application of models
for eddy current inspection of cracks around fastener holes. J. Nondestr. Eval. 25(3), 123–138
(2006)
136. Dominguez, N., Feuillard, V., Jenson, F., Willaume, P.: Simulation assisted POD of a phased
array ultrasonic inspection in manufacturing. Rev. Progr. Quant. Nondestr. Eval. 31, AIP,
1765–1772 (2012)
137. Kennedy, M.C., O’Hagan, A.: Bayesian calibration of computer models. J. R. Stat. Soc. B 63,
425–464 (2001)
138. Gelman, A., Carlin, J.B., Stern, H.S., Rubin, D.B.: Bayesian Data Analysis. CRC, London
(2003)
139. Christensen, R., Johnson, W., Branscum, A.: Bayesian Ideas and Data Analysis: An Introduc-
tion for Scientists and Statisticians. CRC, Boca Raton (2010)
140. Meeker, W.Q., Escobar, L.A.: Introduction to the use of Bayesian methods for reliability data.
In: Statistical Methods for Reliability Data, pp. 343–368. Wiley, New York (1998)
141. Thompson, R.B.: A Bayesian approach to the inversion of NDE and SHM data. Rev. Progr.
Quant. Nondestr. Eval. 29, AIP, 679–686 (2010)
142. Leemans, D.V., Forsyth, D.: Bayesian approaches to using field test data in determining the
probability of detection. Mater. Eval. 62, 855–859 (2004)
143. Knopp, J.S., Zeng, L.: Statistical analysis of hit/miss data. Mater. Eval. 71(3), 323–329 (2013)
144. Li, M., Meeker, W.Q., Hovey, P.: Joint estimation of NDE inspection capability and flaw-size
distribution for in-service aircraft inspections. Res. Nondestr. Eval. 23, 104–123 (2012)
145. Aldrin, J.C., Knopp, J.S., Sabbagh, H.A.: Bayesian methods in probability of detection
estimation and model-assisted probability of detection evaluation. In: Thompson, D.O.,
Chimenti, D.E. (eds.) Review of Progress in Quantitative Nondestructive Evaluation, vol. 32,
pp. 1733–1740. American Institute of Physics, Melville (2013)
146. Gelman, A., Hill, J.: Data Analysis Using Regression and Multilevel/Hierarchical Models.
Cambridge University Press, Cambridge (2007)
147. ASTM Standard E2782, 2011: Standard guide for measurement systems analysis (MSA).
ASTM International, West Conshohocken (2011). doi:10.1520/E2782-11, www.astm.org
Index

Symbols biaxial generalized electrical permittivity

10◦ tilt, 198 matrix, 7
20◦ tilt, 200 bistatic, 58
30 degree EDM notch., 197 blending functions, 313
90-10 Copper-Nickel, 356 bobbin coil, 145, 159
absolute, 145
differential, 145
chloride stress corrosion, 391 incident field and moments, 159
bolt hole benchmark problems, 213
problem 1, 214
A problem 2, 214
adjoint operator, 53 boundary-integral method, 5
algebraic reconstruction technique (ART), 282, broadband, 93
285
Ampere–Maxwell law, 41
amperian current, 41 C
analytic continuation with constraints, 285 capacitance vs terminal separation, 99
anomalies, 8 capacitive coupling, 93
anomalous (scattering) currents, 75 capacitive reactance, 108
anomalous magnetic and electric currents, 8 Carderock problem, 362
ART, see algebraic reconstruction technique Cessna Sandwich, 337
asymmetry ratio, 201 Half-Sandwich No. 1, 347
axisymmetric rings, 360 Half-Sandwich No. 2, 348
axisymmetric systems, 167 Half-Sandwich No. 3, 350
azimuthal wave number, 167 Cessna setups, 342
chain parameters, 138
characterization error (CE), 432
B circulant-matrix, 52
background and flaw grids, 211 circular array
Bayes, 262 ZETEC X-Probe, 129
Bayesian calibration, 434 circumferential rings, 145
Bayesian methods, 426, 434 circumferential sweep, 332
Bessel functions, 36, 156 clutter, 314
first kind, 65, 166 automated clutter removal, 294
scaled, 156, 170 coherent background clutter, 402
Bessel transform, 37 modeling and rejection, 289
Bezier, 293 coaxial filamentary current source, 151

H.A. Sabbagh et al., Computational Electromagnetics and Model-Based Inversion, 443

Scientific Computation, DOI 10.1007/978-1-4419-8429-6,
© Springer Science+Business Media New York 2013
444 Index

composite spectral scene, 204 discretized magnetic equation, 50

composite structure, 399 dispersive, 396
condition number, 265 divergence of the currents, 106
condition-based maintenance plus prognosis divergence-conforming, 46
(CBM+, 432 divergence-free, see solenoidal
conductivity profile measurements, 231, 239 divergenceless, 73
conductivity and permeability, 242 driver pickup, 58
multifrequency reconstruction, 245 dynamic range, 293
conductivity volume-fraction, 47
conforming array, 129
conjugate-gradient algorithm, 51, 53 E
constraint electric field, 96 edge elements, 45, 73
convolution, 46 eigenmodes of anisotropic media, 7
convolution and correlation sums eigenvectors V1m , V2m , V3m , V4m , 151
matrix structure, 52 electric–electric matrices, 95
convolutional, 25 electric–magnetic matrices, 95
2D-convolution, 181 electric-dipole mathematical model, 93
layered-space matrices, 50, 95 encircling-coil, 160
correlational, 25 EPRI ID pits benchmark test, 367
layered-space matrices, 50, 95 equation of discontinuity, 9, 12
corrosion, 309 equivalent pillbox standards, 368
corrosion pits, 145, 334 error analysis, 255
fatigue and stress corrosion cracking, 309 error correlation, 255
micro-topography, 309 error estimation, 190
surface topology, 309 error extrapolation, 190
coupled integral equations estimation theory metrics, 265
axisymmetric, ferromagnetic problems, 148 exhaustive-search method, 331
ferromagnetic, 145 extrapolation formula
volume integrals, see volume-integral four-point, 193
equations, coupled three-point, 192
coupled problems in heat-exchanger tubes, 383 two-point, 191
coupled-circuit model, 66, 393
coupling-coefficient, 68
covariance matrix, 264 F
noise analysis, see noise analysis, Förster plot, 69
covariance matrix facet elements, 45
Cramer–Rao lower bound (CRLB), 263 Faraday–Maxwell law, 41
CRLB, see Cramer-Rao lower bound fast-probe model, 326
cross-validation, 263 fastener sites, 337
curl-conforming, 47, 148 fatigue crack, 75, 82, 337
cylindrical eigenvectors, 162 feature extraction, 370
algorithm, 332
FEM model, 340
D ferrite-core probes, 71
damage characterization, 432 impedance, see impedance, ferrite core
data fusion, 309 probes
differential-receive ferrite-core probe, see split-D, 81–91
ferrite-core probes, split-D ferritic, 385
dipolar effects, 396 stainless steels, 391
direct coupling, 59 SEACURE, 391
direct problems, 230 Type 304, 391
ferritic tubes, 391 Type 439, 391
discretize, 45 ferromagnetic, 145
discretized electric equation, 49 ferromagnetic fastener sites, 337
Index 445

ferromagnetic response, 396 Fourier-domain, 16

ferromagnetic system spectral-domain, 24
multi-frequency reconstruction, see electric–magnetic, 10, 27
conductivity profile measurements, layered-space, 23
conductivity and permeability, magnetic–electric, 28
multi-frequency reconstruction spectral-domain, 16
field grid, 183 magnetic–magnetic, 16
fineness of a mesh, 190 spectral-domain, 26
finite-element method, 340 mixed, 27
Fisher information, 266 grid resolution requirements, 202
Fisher information matrix, 249
Fourier integral, 163
Fourier series, 163 H
Fourier transform, 8 Hall-effect probe, 183
discrete, 51 Hankel, see correlational
FFT, 51 Hankel functions, 156, 166
two-dimensional, 181 scaled, 156, 170
spatial, 203 heat-exchanger tubes, 353
Fourier trasform Helmholz’ theorem, 73
finite, 207 hi-lock pin rivet, 342
Fourier-domain, see spectral-domain higher-order spline interpolator, 313
freespace chain matrix, 139 Householder matrix, 303
frequency mixing, 319
frequency scanning, 357
frequency-response loci, 68 I
fretting damage, 383 impedance
full circuit capacitance values, 103 circular coil, 64
fuselage and wing locations, 337 driving-point, 57, 62
bobbin coil, 161
loaded dipole, 96
G ferrite-core probes, 62
GaAs, 111 impedance-plane plot, 389
Galerkin’s method, 45, 49 normalized, 68–70
Gauss’ divergence theorem, 54 open-circuit transfer, 137
Gauss-Newton, 257 transfer, 57, 59
Gaussian distributions, 235 universal impedance diagram, 70
Gaussian noise, 263 incident term, 153
Gaussian processes, 263, 434 inconel, 385
Gaussian signal, 263 infinite-space matrices, 50, 95
general axisymmetric model, 145 infinite-space term, see infinite term
general transformation matrix, 303 integro-differential equation, 38
generalized permittivities, 7 intergranular attack, 391
giant magnetoresistive (GMR) sensor, 183 intergranular attack (IGA), 145
graduated grid, 214, 221 internal-coil, 160
graduated grid scheme, 222 inverse Fourier transform, 35
graphite-epoxy advanced composites, 286 inverse method quality metrics, 407
Green function inverse problems, 227, 230
electric–electric, 146 eddy-current, 231
infinite-space, 158 conductivity profile measurement, see
matrix element, 158 conductivity profile measurement
layered-space, 158 thickness measurement, see thickness
matrix element, 158 measurement
Green’s dyad, 7–33 ferritic tubes, 391
electric–electric, 10 inversion of a complex pit, 357
446 Index

inversion process transmission coefficients, 20

metrics, 260 metallic traces, 94
isotropic media, 10 method of finite-elements, 5
iterative method, 51 method of moments, 4, 38
discretizatio, 148
minimum-norm solution, 284
J model-assisted probability of detection
Jacobian, 248, 258 (mapod), 421
model-based parameter estimation, 5
modeling probes + cables, 379
K
monte carlo, 266
Kaczmarz’ algorithm, 282
noise analysis, see noise analysis, monte
Kirchoff’s voltage law, 66
carlo
multi lift-off, 247
L multifrequency, 359
layered models of pits, 377, 382 multifrequency data, 359
layered tube multifrequency results, 372
axisymmetric, 151 multilayered aircraft structures, 309
layered-space terms, 153 multilayered structures, 337
least-squares, 233 mutual inductance
Gauss-Newton iteration, see NLSE between the coil and anomaly, 66
normal equations, 233 between the coil and host, 66
Levenberg-Marquardt parameter, 263, 285
ligands, 293
ligatured outer-diameter slot, 292 N
likelihood function, 262 N-port theory of the T/R array, 135
linear extrapolation, 196 naked terminal, 99
Lissajous figure, 356, 386 capacitance, 107
load parameters, 31 NDE of coatings, 411
loaded terminal near-surface residual stress, 93
capacitance, 107 necklace, 129
network functions, 287
NLSE, 219, 232, 240, 257, 287, 290, 357, 359,
M 360, 365, 367, 372, 390, 411
magnetic permeability, 383 interpolator, 313
magnetic polarization vectors, 95 model-based inversion, 219
magnetic sensor, 183 noise analysis
magnetic–electric operator, 148 covariance matrix, 248
magnetite, 145 monte carlo, 254
magnetization vector, 73 noise and uncertainties, 235
magnetoelastic effects, 399 non-magnetic, 145
mapod process, 422 nuclear power, 353
material anisotropy, 8
maximum likelihood estimation, 262
Maxwell’s equations, 7 O
matrix formulation, 8 Opera-3D c , 340
eigenvalues, 9 out-going cylindrical wave, 168
eigenvectors, 9 outer-coil experiments, 394
mode coefficients, 20
reflection coefficients, 20, 30
scattered fields, 21 P
spectral-domain parasitic chain matrix, 139
four-vector matrix differential equation, parasitic effects, 139
8 periodic azimuthal array, 129
Index 447

permeability volume-fraction, 48 ring probe, 342

permeable crust, 383 robust statistical estimators, 273
piezoresistive effect, 93 roll-expanded transition zones, 145
pitting, 309 round-bottom pits, 369
pitting and corrosion phenomena, 353 morphology, 377
plane-parallel layered media, 337
POCS, see projection onto convex sets
point-source of electric or magnetic current, 10 S
printed-circuit receive-sensors, 135 scan-step correction, 370
probability density functions scattered fields in the vicinity of a probe, 183
a posteriori density, 262 self-impedance, see impedance, driving-point
a priori density, 262 semiaxes, 354
probability distributions, 236 semiconductor chip industry, 231, 245
probe + cable effects, 363 semiellipsoids, 354
probe characterization, 305 sensitivity bound, 258, 260
probes sensitivity coefficient, 260
differential probe, 60 signal-detection, 230
differential-bobbin, 298, 356, 367 singular value decomposition (SVD), 265
responses, 363 skin effect, 69
eddy-current probes slide functions, 98
conductivity profile measurement, see sludge, 145
conductivity profile measurement solenoidal, 71
ferrite-core probes, see ferrite core source grid, 183
probes spatial decomposition, 177
planar spiral-coil, 93 algorithms, 177, 337
thickness measurement, see thickness technique, 177
measurement spatial-frequency content, 203
printed-circuit probes, 129 spectral-domain, 35
projection onto convex sets (POCS), 281, 282, spectrum of the conductivity scene, 204
288, 321, 331, 333, 336, 367 spectrum of the electromagnetic scene, 204
proximity sensor problem, 191 spherically symmetric, 36
pulse function, 45, 46 spline-generated basis-functions, 45
pulsed eddy-current techniques, 309 split-core differential, see ferrite-core probes,
split-D
statistical errors, 237
Q steam generator tubing, 145
QR-decomposition, 233 stochastic methods for global optimization,
261
stochastic numerical methods, 434
R stochastic simulation methods, 228
racetrack coils, 191 stray capacitive effects, 93
radiation condition at infinity, 168 surface admittance, 31
ramp functions, 98 surface impedance, 31
reaction concept, 57 surrogate models, 228
reaction-formula, 161 SVD, see singular value decomposition
receiver, 58 systematic errors, 236
reciprocal network, 137
reciprocal two-ports, 139
reciprocity theorem, 57 T
reflection coefficients, 154 Töplitz, see convolutional
double reflection coefficient, 154 tangent coil, 87
relaxation parameter, 284 tent function, 45, 46
remote-field, 58 terahertz frequencies, 111
ridge regression, 263 testing, 49
448 Index

testing the integral equations, 47, 49 two-dimensional convolutions and one-

tetrahedral elements, 340 dimensional correlations, 51,
thermal barrier coatings, 411 54
thickness measurement, 231 two-layered spiral coil, 110
test case 1, 234 two-port parameter relations, 138
test case 2, 237
three-dimensional convolutions, 51, 54
three-turn spiral coil, 109 U
through-wall collar, 401 uncertain parameters, 426
THz transmitter, 111 unflawed host, 67
Tichonov-Miller parameter, 263
tilted notch benchmark studies, 197 V
titanium rivet-insert, 337 variance analysis, 254
tomographic scanner, 285 vector potential, 37
transfer marices, 181 virtual probe model, 298, 299
transform to cylindrical coordinates, 36 volume-integral equations, 7
transformation of signal vectors, 289, coupled, 41
297 electric differential, 36
transmitter, 58 vector form, 39
transverse divergence, 40, 49 kernel, 10
transverse electric (TE) modes, 10, 167 volumetric rooftop, 45
wave-admittance, 30 voxel-based, 257
transverse gradient, 40
transverse magnetic (TM) modes, 10, 167
wave-impedance, 30 W
transversely isotropic, 9 wall-thinning effects, 360
tri-diagonal matrices, 50, 95, 150 wave number space, 163
TSP+tube structure, 385 weak-host, 347
tube supports, 145 wear scar, 383