PARSIMONIOUS MODELING OF INELASTIC SYSTEMS

Thesis by

Dar-Yun Chiang

In Partial Fulfillment of the Requirements

for the Degree of

Doctor of Philosophy

California Institute of Technology

Pasadena, California

1993

(Submitted September 30, 1992)

11
 iii

ACKNOWLED GEMENTS

I am heartily grateful to my advisor, Professor James L. Beck, for his continued

guidance and encouragement throughout the course of my study and research at
Caltech. His infectious enthusiasm and continuous availability to advise me with
my work are greatly appreciated.

I would like to thank the California Institute of Technology for the generous
financial support offered to me, which made my study here possible. I also wish
to express my gratitude to the people in Thomas Lab, who created a friendly
environment making my stay here a very enjoyable experience. My special thanks
are extended to Dr. Jay C. Chen, who was at JPL and is now in the Hong-Kong
University of Science and Technology, for his constant help that made my overseas
study here so pleasant and memorable.

My deepest thanks are due to my wife, Tsai-Hsiu, not only for her assistance
in typing this thesis, but also for her patience, for her love and for her sharing
wonderful moments here with me. Finally, I dedicate this thesis to my parents,
Kuan and Hsiu-Lien, who always support me in the pursuit of my career goals.
 lV

ABSTRACT

Analytical modeling of one-dimensional hysteresis and general multi-axial cyclic

plasticity is studied, with particular emphasis on the parsimony of model parameters
and the physical consistency of model behavior. General criteria for good models are
proposed to provide guidelines to the modeling studies conducted in this research.
Various one-dimensional hysteretic models are examined in detail, including both
deteriorating and non-deteriorating models. A general formulation for modeling of
degrading systems is presented based on the formulation of the Distributed-Element
Model (DEM) and the introduction of a damage index function. A new class of
deteriorating Masing models, whose behavior can be completely described by a few
simple mathematical rules and the extended Masing rules, is also developed to sub-
stitute for a special class of deteriorating DEMs, so that their applicabilty to system
identification studies is improved.
The one-dimensional DEMs are extended to the multi-dimensional case for con-
stitutive modeling of cyclic plasticity, while preserving the concept of modeling plas-
ticity by an assemblage of simple ideal elasto-plastic elements. In the generalization,
a new invariant-yield-surface theory is proposed, in which no kinematic hardening
rule is needed to account for the subsequent yielding and strain hardening behavior.
A general theory is also developed to elucidate some important properties of material
behavior based on the proposed multi-dimensional DEMs. The establishment of the
theory provides instructive insight into the elastic-plastic response mechanisms of
real materials under complicated loading conditions. Based on the insight, the Mas-
ing rules for one-dimensional hysteresis are extended to the multi-dimensional case
by introducing a composition of plane-geometry transformations to a response for-
mula developed for initial loading. This transformation method serves as an efficient
way of implementing the classical multi-yield-surface theory with the Mroz kinematic
hardening rule. Validity of the new formulations are confirmed by comparison with
experimental results from the literature.
 v

TABLE OF CONTENTS

Acknowledgements ..... . .... . . ............. .... . . ..... ... .. ........ ....... iii
Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1v
Table of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v
List of Tables and Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii

Chapter 1 Introduction ... ...... .. .. . . .. ..... ...... .. . ..... .... .. . ... ... . 1

Chapter 2 System Identification and Modeling . . . . . . . . . . . . . . . . . . . . . . . . 6

2.1 Introduction ...... .... .. ..... . ........ .. ....... ... .... . .. . . . . ........ 6
2.2 Practical Considerations of System Identification . . . . . . . . . . . . . . . . . . . . . 7
2.3 Criteria for Good Modeling .... ...... . .. .. ............... .. .......... 9
2.3.1 An Illustrative Example of Proper Modeling ............. ..... 11

Chapter 3 Modeling of One-dimensional Hysteretic Systems . . . . . . . . 15

3.1 Introduction ........... .. . ..... .. ......... ... . ........ ....... ....... 15
3.2 Modeling of Hysteretic Systems without Deterioration ... ........... . 16
3.2.1 Hysteretic Models Described Solely by Differential Equations . . 16
3.2.2 The Distributed-Element Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.2.3 The General Class of Masing Models ... . . .. ..... . . .... . ... ... 20
3.2.3.1 Masing's Hypothesis and Extended Rules . . . ........... 20
3.2.3.2 A Special Class of Masing Models .. .... ................ 22
3.3 Modeling of Degrading Hysteretic Systems ....... .............. . ... . 23
3.3.1 The Degrading Bouc-Wen Model .... . ....... .... .... . ... . .... 24
3.3.2 A Class of Deteriorating Distributed-Element Models .. ....... 25
3.3.3 A Class of Deteriorating Masing Models .. .. ..... ... ........ .. 27
3.3.3.1 General Formulation .... . .. ... .. .............. .... . .... 27
3.3.3.2 A Special Class of Deteriorating Masing Models ........ 29
3.3.3.3 Yield-strength Distribution Function ...... . ..... ... .... 32
3.3.4 Other Models for Degrading Systems . . . . . . . . . . . . . . . . . . . . . . . . . 36

Chapter 4 Modeling Based on Endochronic Theory . . . . . . . . . . . . . . . . . . 53

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.2 Endochronic Theory and Its Implementation . . . . . . . . . . . . . . . . . . . . . . . . 54
 vi

4.3 A Modeling Technique for the Endochronic Models .................. 61

4.4 Investigation of Cyclic Hardening Behavior . . . . . . . . . . . . . . . . . . . . . . . . . . 66

Chapter 5 Generalization of Distributed-Element Model to Multiple

Dimensions ... . . ... . . . ....... .. ....... ..... ... . . . . . . . .. . ... .. 72
5.1 Introduction .... .. . . ........ .. . .. . .. . .. . ... .. ..... . . ........ . . . . .. .. 72
5.2 A New Class of Distributed-Element Models for Plasticity . . . ... . . . . . 73
5.2.1 Concept and Theoretical Background ............. .. ... . . . . ... 73
5 .2.2 Mathematical Formulation . . . .................. .. . . ..... .. . . . 76
5 .2.3 Numerical Implementation of t he New DEMs . . . ......... .. . .. 80
5.2.4 An Application to Biaxial Loading ......... . . ... .. ......... ... 82
5.3 Important Properties of the New Multi-dimensional DEMs ...... . ... 84
5.3.1 General Behavior of Ideal Plasticity of A Single Element . . .... 85
5 .3 .2 General Treatment of the Theory of P lasticity . . . . . . . . . ....... 89
5.3.3 Geometrical Considerations of Yield Surfaces for the New DEMs 99

Chapte r 6 Generalized Masing Rules for Cyclic Plasticity . .... . . . .. 123

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
6.2 Extension of 1-D Response Formulas to Higher Dimensions . . . . . ... . 124
6.2.1 T wo-dimensional Bouc-Wen Model ......... ............ . .. . . 124
6.2.2 A Recent P rocedure for Generalizing 1-D Hysteretic Models . 126
6 .3 A Class of Generalized Masing Models for Multi-axial Plasticity Model 130
6 .3.1 A Response Formula for Initial Loading ............. ...... .. 130
6 .3 .2 Response Formulas for Unloading and Reloading Branches . .. 133
6.3.3 R ules for Transient Response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
6 .3.4 Simulation Studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
6.4 Comparison of the Generalized Masing models with the multi-axial DEMs
.............. ... .......... .... ...... . ... . ...... ..... .... . ....... . .. 140

Chapter 7 Summary and Conclusions .. . ....... .... .. . . .. .... .... .. .. 155

References 159
Appendix A: Operator Theory on Convex Sets . . . . . . . . . . . . . . . . . . . . . . . 164
Appendix B: Derivation of Transformation Formulas for Generalized
Masing Rules for Multi-Axial Cyclic R esponse B eh avior
. . .. . .... . . .. ... ... .. . . ........... . .. .. . .... .... ... .... . . ..... 166
 Vll

LIST OF TABLES AND FIGURES

Figure 2.1: Comparison of response histories obtained using the fast algorithm and the
4-th order Runge-Kutta method . ..... .. . . .. . . .. ..... . ... .. .. .. .. . .. 14

Figure 3.1: Hysteretic restoring force behavior of (a) the elasto-perfectly plastic model
(b) the bilinear model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

Figure 3.2: Response behavior of the Bouc-Wen model (a) restoring force diagram
(b) prescribed displacement history . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

Figure 3.3: The unstable drift exhibited by the Bouc-Wen model . . . . . . . . . . . . . . . 41

Figure 3.4: The Distributed-Element Model for one-dimensional hysteresis .. .... 41

Figure 3.5: Masing's hypothesis for cyclic hysteretic loops . . . . . . . . . . . . . . . . . . . . . . 42

Figure 3.6: Hysteretic loops for transient loading ............................ .. . 42

Figure 3.7: Effect of the parameter n of a special class of Masing models [24] on the
smoothness of yielding curves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

Figure 3.8: Effects of the degrading parameters on the behavior of the deteriorating
Bouc-Wen model (from [52]) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

Figure 3.9: Restoring force diagram of a typical breaking element in the maximum-
displacement-controlled deteriorating DEM ... . ... . ...... . .. ........ 45

Figure 3.10: Restoring force behavior of the maximum-displacement-controlled

deteriorating DEM subject to an earthquake excitation (from [22]) 45

Figure 3.11: Simulated restoring force behavior of the proposed deteriorating Masing
model . . .. .. ... ... . ....... . ............. . . . . . .. ........ ........... 46

Figure 3.12: Comparison of the behavior of two matching Masing models . ...... 47

Figure 3.13: Distribution curves described by the generalized Rayleigh distribution

function with different values of n ... ... ... .... .. . . . .. . ... .. ... . ... 47

Figure 3.14: Plastic deformation of the Masing model based on a statistical interpreta-
tion of the Rayleigh yield-strength distribution function . .. . . . .... . 48

Figure 3.15: Drift response of a Masing model based on the Rayleigh distribution
function as compared with an endochronic model . . . . . . . . . . . . . . . . . . 48
 viii

Figure 3.16: Different types of elements used in a model for hysteretic systems with
stiffness degradation (from [13]) . . ............. . . . . ... . . . . ......... 49

Figure 3.17: Configuration and behavior of Gates' degrading model (from [13]) . 50

Figure 3.18: Clough's hysteretic model for stiffness-degrading behavior ...... . .. 51

Figure 3.19: Takeda's hysteretic model for reinforced concrete structural systems 51

Figure 3.20: Saiidi & Sozen's hysteretic model for stiffness-degrading behavior . . 52

Figure 4.1: Behavior of the endochronic model using the kernel function defined by
Eqn. (4.18) . . . . .. ........ . .. . .... ..... .. . .... . ... .. .. .. ............ 69

Figure 4.2: A typical yielding curve for illustration of the proposed modeling technique
based on the endochronic theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69

Figure 4.3: Effect of the parameter p of the proposed class of endochronic models on
the yielding behavior . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

Figure 4.4: Response behavior of a one-dimensional endochronic model and a match-

ing Masing model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

Figure 4.5: Uniaxial cyclic hardening behavior of (a) a Masing model (b) a real
material . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

Figure 5.1: Two different one-dimensional Distributed-Element Models (a) parallel-

series model (b) series-parallel model . ... .. .. .. . ........ .. .... . .. .. 109

Figure 5.2: Illustration of the space-dependent yielding behavior of ideal plasticity 110

Figure 5.3: Invariant yield surfaces nested in the element stress space . . ...... . 111

Figure 5.4: Selection of yield constants for a finite number of elements according to
a specified yield-strength distribution function . . . . . . ....... .. . . ... . 111

Figure 5.5: A flow diagram showing numerical procedure for obtaining stress response
of anN-element DEM . ... . .. .. ..... .. ... . .. .. ..... .. .. . .... . ... .. . 112

Figure 5.6: Prescribed strain loading paths for response studies of the proposed multi-
dimensional DEMs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113

Figure 5.7: Experimentally-observed stress response of copper to the prescribed strain

path given in Figure 5.6(a) ... . . .. .. ..... ... ..... . .. .. . ..... . . . . . . . 114
 lX

Figure 5.8: Experimentally-observed stress response of copper to the prescribed strain

path given in Figure 5.6(b) ....... .. ...................... .. ....... 115

Figure 5.9: Stress response predicted by a new DEM subject to the prescribed strain
path given in Figure 5.6(a) ........... . .............. ... ..... . ..... 116

Figure 5.10: Stress response predicted by a new DEM subject to the prescribed strain
path given in Figure 5.6(b) ..... .... .. ......... ..... .............. 117

Figure 5.11: Response predicted by different plasticity models to the strain path given
in Figure 5.6(b): (a) von Mises' yield surface with Prager's hardening rule
(b) Tresca's yield surface with Ziegler's hardening rule (c) Tresca's yield
surface and limit surface with Mroz' hardening rule .... . ...... ... 118

Figure 5.12: Response behavior of ideal plasticity under a proportional strain loading
path (a) proportional strain path (b) stress response behavior .... 119

Figure 5.13: Definition of the plastic-relaxation stress increment in the uniaxial case
.... .. ............................................................ 120

Figure 5.14: Illustration of the existence of equilibrium points associated with a big
strain cycle (a) a strain cycle (b) the corresponding stress response 121

Figure 5.15: A diagram showing the rotation of coordinate axes which makes the x 1
axis perpendicular to the tangent plane to the yield surface 80. 0 at Qo
.............. . ..... . .. ... ...... . ..... ..... . ...................... 122

Figure 5.16: An illustrative diagram showing non-strict convexity of yield surfaces 122

Figure 6.1: Hysteretic behavior of the two-dimensional Bouc-Wen model (a) propor-
tional displacement path (b) hysteretic restoring force behavior . . . . 142

Figure 6.2: Comparison of the initial response predicted by Eqn. (6.16) and by aDEM
........................................ . .. .... . ...... ............. 143

Figure 6.3: The Mroz kinematic hardening rule for multiple yield surfaces .. . .. 144

Figure 6.4: A biaxial strain path for the study of the proposed response formulas for
initial loading ............. . .... ................ ...... . ...... .... .. 144

Figure 6.5: Experimentally-observed stress response of copper to the strain path given
in Figure 6.4 (from [17]) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145

Figure 6.6: Stress response predicted by Eqn.(6.19) with the prescribed strain path
 X

given in Figure 6.4 •• • 0 0 • • ••• 0 0 0 0 0 •••• 0 • • 0 •• 0 0 0 •• 0 0 • • 0 0 0 0 • •• • • • 0 • •• 146

Figure 6.7: Stress response predicted by Eqn. (6.13) with the prescribed strain path
given in Figure 6.4 ............... .. .. .. ... .. . . . ............ . ... .. . 147

Figure 6.8: Movement of yield surfaces with current stress state moving from A to B
(a) initial configuration (b) current stress state A (c) current stress state
B ...... . .. .. ... . ... ........... . .. . . . .. .. . ..... .. . ........... . .. . . 148
Figure 6.9: Different yield surfaces and shifted 1r planes in the principal stress space
(a) von Mises yield surface (b) Drucker-Prager yield surface .... . .. 149

Figure 6.10: A schematic diagram showing transformation on the 1r plane . ... . 150

Figure 6.11: Illustration of completed loops and numerical difficulty associated with
the transformation approach . .. .... . . .. . .. ...... . . ............ . .. 151

Figure 6.12: Geometrical considerations of transient response (a) unloading from

point A (b) unloading from point B with new center C . . . . . . . . . . . 152

Figure 6.13: Stress response predicted by a generalized Masing model subject to the
strain path given in Figure 5.6(a) . . .... .. . .. ..... . ......... .. .... 153

Figure 6.14: Stress response predicted by a generalized Masing model subject to the
strain path given in Figure 5.6(b) ..... . ............ . ............. 154

Figure B.1: Geometrical configurations before and after transformation .. . .... . 169

Figure B .2: Configurations at different transformation stages . . ..... . . . . . . ... .. 170

Figure B.3: Conditions of the principle of normality on the proposed transformation

..... . . .. . . . . . . .. . .. ... .. ... .. ... . . .. ... ... .. .... .. .. ........ . ... .. 171
 1

CHAPTER 1
INTRODUCTION

Most structures exhibit nearly linearly elastic restoring force behavior under
moderately small loading conditions. However, when subjected to severe excitations
such as strong earthquake ground motions, structures may respond inelastically and
exhibit hysteretic behavior so that the restoring force at a time instant depends not
only on the instantaneous state, but also on the past response history. The study
of nonlinear, hysteretic behavior of mechanical systems has been of great interest to
researchers in many engineering fields, and particularly in earthquake engineering
[4, 9, 10, 13, 18-24].

Structures of simple configurations and homogeneous materials may usually be

approximated by simplifed analytical models so that their response to complicated
external loading can be analyzed more efficiently. For example, normal building
structures under seismic excitations can often be modeled as shear buildings (i.e.,
chain models) so that their response characteristics, such as natural frequencies,
maximum displacement response, etc. , can be estimated efficiently and with rea-
sonable accuracy. When such simplified models are used, the hysteretic response
of structural systems is often described by its overall interstory force-deflection re-
lationship so as to avoid complex stress-strain calculations for which constitutive
equations governing material behavior at a point are needed [23].

Although overall planar force-deflection representations in nonlinear structural

analysis can reflect behavior of structural members or substructures as a whole,
including both material and geometrical effects, they are not suitable for describing
local response behavior in the case of complex mechanical systems or complicated
loading conditions in which responses in different directions may interact signifi-
cantly with one another. For that purpose, one needs to introduce appropriate
constitutive laws depicting stress-strain relations at different material points, from
which local response behavior can then be derived.
 2

As mentioned earlier, linear models for mechanical systems are, in general,

sufficient to represent system response resulting from small excitations. The mathe-
matical representations of such linear models are usually simple and of clear physical
significance. Thus, modeling of linear systems may be thought of as an easy and
straightforward task, if considered solely from the viewpoint of forward analysis. If
system identification is under consideration, however, the choice of a suitable class
of linear models may become crucial to the success of the model identification from
response data. For example, if linear models are to be used in the identification
of structures using earthquake data, one should consider identifiable modal models
with parameters of modal frequencies, dampings, and mode shape components, not
the models in the physical coordinates with parameters of stiffness and damping
matrices, as recommended by Beck (3], so that more reliable results can be obtained
in the case of a limited number of measurement channels. Once the modal parame-
ters are estimated from the earthquake data, they can be used in a subsequent stage
to investigate the generally nonunique inverse problem of going from the incomplete
set of modal parameters to structural stiffness parameters.

In contrast to linear models, nonlinear models are usually more mathemati-

cally involved, especially when hysteretic behavior is taken into account. For the
simplified force-deflection relationship which is a one-dimensional formulation of sys-
tem behavior, numerous models have been proposed ranging from simple ones such
as elasto-perfectly-plastic and bilinear hysteretic models to sophisticated ones like
Takeda's and the Bouc-Wen's models [44, 51] . Among these models, the Distributed-
Element Model (DEM), developed by lwan [19], has been successfully applied to
structural dynamic analysis because ofits physically consistent behavior. The DEM
consists of an assemblage of simple ideal elasto-plastic elements that have different
yield strengths governed by some distribution function. As shown by Jayakumar
[23], the DEM formulation is mathematically equivalent to a general class of Masing
models in which Masing's hypothesis is extended for transient behavior of general
hysteretic response.

An interesting problem that has not been resolved so far is how the Masing
rules can in some way be extended to two or higher dimensions, and how would
the general rules compare with the behavior of a general multi-dimensional DEM,
 3

if available. Furthermore, how would the behavior of such general models compare
with those based on the classical theory of plasticity? If these questions can be
answered clearly, then modeling of general plastic behavior of mechanical systems
can be improved and analysis of complex structures can be performed with more
success than before.

We remark that not all modeling of structural systems can be done based exclu-
sively on theoretical considerations. In most cases, the mathematical models are so
complicated that an empirical approach is needed to identify an appropriate model
from within a prescribed class of models using structural response data. Therefore,
it is of practical importance to build new models also from the system identification
point of view. In general, a good model should be not only mathematically simple,
physically consistent, and computationally efficient, but also parsimonious in the
number of parameters.

This thesis consists of five independent, yet interrelated chapters in addition to

Chapter 1, the introduction, and Chapter 7 in which summary and conclusions of
this research are given. In Chapter 2, important issues and practical considerations
of system identification are discussed with particular emphasis on the process of
model building. Criteria of good models for mechanical systems are also proposed
based on considerations of system identification, which provide useful guidelines for
the modeling studies conducted in this research.

Chapter 3 gives an extensive review and discussion of various models for mod-
eling of one-dimensional hysteretic behavior, including both the models defined by
empirical rules and those by differential equations. In particular, the DEM and
Masing models are described in detail, since they form the fundamental starting
points of the present research. Hysteretic models including strength and/ or stiff-
ness deterioration are also discussed in this chapter. An extension of the hysteretic
response rules based on Masing's hypothesis to the case where degradation effects
are included is proposed using the distributed-element formulation and the intro-
duction of a "damage" function. Explicit mathematical rules are then derived for
a particular class of maximum-displacement-controlled deteriorating DEMs [22] .
With these rules, the numerical implementation of this special class of deteriorating
 4

DEMs becomes simpler as compared with the direct computation of model response
by keeping track of response behavior of all the elements constituting the model.
As a result, identification studies based on such models can then be performed with
more efficiency and higher accuracy. Response behavior of this new class of degrad-
ing Masing models is compared with those of other well-behaved degrading models.
Of particular interest is the response behavior of a class of endochronic models that
are described theoretically by integra-differential equations [46, 47]. The flexible,
physically consistent behavior of the modified one-dimensional endochronic model
motivated the study of the general endochronic models for cyclic plasticity in which
complicated multi-axial loading conditions are considered. This is presented in
Chapter 4.

Although the behavior of the modified endochronic models is govened by some

mathematically involved integra-differential equations, effective procedures for con-
stitutive modeling based on the theory have been established in the past so that
numerical implemention of the models is practically feasible. A new way of modeling
based on the endochronic theory is proposed in Chapter 4, which was inspired by
the modeling technique used in the Masing models. With the introduction of this
new modeling technique, the building process of endochronic models is much simpler
and identification studies of systems modeled by endochronic theory then become
easier in practice. Furthermore, inspired by the study of the endochronic theory
for cyclic hardening behavior, the one-dimensional Masing models (or D EMs) are
extended, in a very effective way, to account for cyclic hardening behavior.

In Chapter 5, the one-dimensional DEMs are generalized to three dimensions

(multi-axialloading case), so that they can be used for constitutive modeling of com-
plex structural systems. Although this work has been pursued by some researchers
in the past [20, 56], limited success was achieved. In the present study, a concept of
nested yield surfaces that are "invariant" (fixed from moving) in the stress space is
proposed for a new class of general multi-dimensional Distributed-Element Models.
This concept is different from that of classical plasticity theory, as will be explained
later in detail in Chapter 5. With the new formulation of the DEM, constitutive
modeling of general structural behavior for cyclic plasticity becomes very simple,
 5

and the associated numerical scheme for implementing the solution algorithm is ef-
ficient as well. The new DEM is shown to not only provide more accurate response
predictions for experimental results compared with models based on the classical
theory of plasticity, but it also serves as a good physical model through which
response mechanisms of complicated plastic behavior can be clearly pictured and
elucidated. A rather complete mathematical work regarding the properties of the
new DEM, such as the existence of equilibrium points and that of a limit surface,
will also be covered in Chapter 5.

Though the general multi-dimensional DEM provides a useful and efficient way
of constitutive modeling for cyclic plasticity, to implement the theory only a limited
number of elements can be introduced due to practical concerns. An interesting
question then remaining is whether some mathematical rules can be found which
are similar to those used in the one-dimensional Masing models so that even in the
general multi-axial loading case, model response can be found without the need of
keeping track of each element's behavior. This problem is solved with success in
Chapter 6. By introducing a formula good for initial response under multi-axial
loading, further unloading and reloading response can then be found by applying a
composition of proper transformations to the state variables involved in the formula.
This method is theoretically equivalent to that utilizing the classical multi-yield-
surface theory with the Mroz kinematic hardening rule [35]. However, it's only with
this new approach proposed here that the response behavior of a model with an
infinite collection of yield surfaces can be analyzed. Computational efficiency is
also preserved in the algorithmic implementation of this new theory, whose validity
is confirmed by modeling some biaxial tension-torsion tests [17, 30) under non-
proportional cyclic loading conditions. A comparison between the models based
on the new approach and the multi-dimensional DEMs is also made at the end of
Chapter 6.

A summary of this research and some general conclusions are presented in

Chapter 7, in which suggestions for further exploration in related subjects are also
given.
 6

CHAPTER 2
SYSTEM IDENTIFICATION AND MODELING

2.1 Introduction

Engineering problems may usually be classified as direct (forward) or inverse

according to the nature and purpose of the analysis. Direct problems are those
of finding the response of systems to specified input excitations; whereas in inverse
problems, the output response to some input is known but either the physical process
(the system) or the input excitation is unknown. System identification may be de-
fined as the process of systematically determining a model of a physical system from
its observed input and output data , and so it falls obviously into the category of
inverse problems. In general, system identification problems can be further divided
into two categories: nonparametric identification and parametric identification. If
the detailed mathematical description of a system is totally unknown or of little inter-
est , then we have a nonparametric identification problem or a so-called "black box"
identification problem, in which a functional relationship between input and output
is to be determined. On the other hand, if some knowledge of the mathematical
structure of the system is available and the problem is that of determining unknown
parameters within the structure, then it is a parametric identification problem. In
engineering applications, the main interest is usually in parametric identification
problems in which an optimal model out of a certain class of models is to be found
for the system under consideration so that prediction of future performance of the
system can be improved accordingly. In a parametric identification problem, the in-
troduction of model structure usually reduces statistical variability of the estimated
model. Also, the identification problem actually becomes a parameter estimation
problem since the mathematical structure of the model is already specified.
In the following sections, practical considerations about system identification
will be discussed with particular emphasis on the process of model building to which
the chapters that follow are closely related. Criteria of good analytical models for
mechanical systems in engineering applications will also be proposed and discussed
in detail from the system identification point of view.
 7

2.2 P ractical Consid erations of System Identification

Identification of structural systems through the use of experimental data is of

considerable importance in many areas of engineering studies, particularly in the
fields of structural vibration and system control. Identification problems usually are
considerably more difficult than forward (response analysis) problems due to the
following reasons:
1) The requirement of well-posed analytical models is more critical to identification
problems than to forward problems.
2) A characteristic feature of identification problems is that the accuracy of iden-
tification results is degraded by a combination of measurement, modeling, and
numerical errors.
3) The problem of existence and uniqueness of solution (i.e., identifiability prob-
lem) is usually very difficult to be resolved due to practical limitations.

The main purpose of a system identification study is to appropriately represent

the physical structure of a system for response prediction, not just to accurately
reproduce the observed data. To this end, the parametric identification approach
is usually adopted in engineering applications due to its capability of making fur-
ther predictions. Three important stages are included in a parametric identification
problem. The first one is model selection, i.e., choosing a mathematical formula-
tion to represent the physical structure of a system. The second stage is parameter
estimation, which is the determination of the "best" parameters for the specified
mathematical structure for the system. The final stage is model validation, in which
some tests on the identified model are conducted to see if the model adequately
represents the system with respect to the desired objectives. Though the nature of
the three stages are quite different, they are all important to the success of identi-
fication studies. Generally speaking, model selection deals with the application of
appropriate physical laws to the systems under consideration. From these laws, some
relations involving parameter variables follow, such as equations for constitutive laws
in modeling of stress-strain relationships of materials. In a parametric identification
problem, the variable parameters of the model are to be estimated based on some
systematic approach so that a "best" model whose parameters are optimal in some
sense, given the input/output (I/0) data, will be identified from the specified class
 8

of models. Thus far, many systematic approaches have been developed for finding
optimal parameters within a class of models. These identification methods almost
always involve minimization of some error criterion functions. Three widely-used
approaches based on different error criteria are as follows:
1) Equation-error method : The discrepancy between the model equation and the
measured I/ 0 data is minimized through a regression analysis technique. This
method is simple and computationally effective. However, complete measure-
ments of all the state variables involved in the model equation are required for
the method to be effective, which is usually a severe restriction in practical
problems.
2) Output-error method: The difference between the output of a system and that
of the model in response to some input is minimized by some functional min-
imization technique. This is probably the most widely-used approach in prac-
tical system identification problems due to its great flexibility and moderate
mathematical tractability.
3) Combined method: One can take a combined equation-error/output-error ap-
proach to perform the identification analysis, such as the Kalman filter method
[25, 28] developed for optimal sequential estimates of parameters and states of
a system. These kinds of sequential estimation methods are good for modern
control problems which require real-time (on-line) analysis capability.

An important aspect of the different identification approaches is that they can

be formulated within a unifying statistical framework [6]. For example, the estimate
obtained from the classical output-error least-square method used extensively in
earthquake engineering can be shown to be equivalent to the classical maximum-
likelihood estimate [33], and to that based on Bayesian statistical inference [7],
under the assumption that the output error can be modeled as a Gaussian white
process.
Probably, the most difficult and important issue of parametric identification is
the problem of identifiability of parameters, which refers to the capability of uniquely
determining the parameters of a model from the available I/ 0 data. The problem of
identifiability can exist even in some of the simplest cases of identification due to the
limitation of available measurement data. For example, when linear models are to be
 9

estimated from seismic excitation and response histories, the stiffness and damping
matrices of the system may not be determined uniquely in typical situations, as
pointed out by Beck [3]. Instead, modal models consisting of dominant modes in
the response should be used for the identification purpose, and if estimation of
system parameters in the physical coordinates is of interest, a further analysis can
always be done in a separate stage utilizing the original data as well as the modal
data already obtained.
Recently, Beck and Katafygiotis [7] addressed the issues of model identifiability
versus system identifiability of optimal parameters from a class of models. For a set of
parameters to be "model identifiable" (globally or locally), there must exist at most
a finite number of sets of parameter values which give "output-equivalent" models
under a specified input. In contrast, the parameters that are "system identifiable"
determine a finite set of optimal models in which the parameters take the most
probable values out of a class of models given a set of I/ 0 data. The problem
associated with system identifiability is obviously much more difficult than that of
model identifiability due to the additional considerations of modeling error involved
in the approximate analytical model and measurement noise contaminated in the
response data.
In this thesis work, we are mainly concerned with modeling of general non-
linear mechanical systems. Although there will be no original theory regarding
system identification techniques proposed herein, some new modeling theories to be
presented later are based on practical considerations of system identification. More-
over, the system identification approach is usually the best way to perform model
validation. For this purpose, one could use a set of test data to identify an optimal
model out of the class of proposed models, and then make predictions of response
behavior observed in another experiment using the identified optimal model. In the
next section, some criteria for good analytical models are proposed and discussed
mainly from the identification point of view.

2.3 Criteria for Good Modeling

In engineering problems, direct or inverse, a good model is always of crucial

importance to the success of investigations. Although models for different problems
 10

depend significantly on the nature of the systems under consideration and may vary
considerably from case to case, there are some general rules that may be consid-
ered as modeling criteria for most practical mechanical problems. The criteria are
proposed as follows:

1) Mathematical Simplicity
A model is a mathematical realization of a physical system which may, in gen-
eral, have rather complicated behavior. However, considering practical issues such
as mathematical tractability, a good model should always be as simple as possible
so that it can have greater applicability to widely-spread engineering problems. For
example, in system identification problems, parsimony in parameters of a model not
only makes numerical computation more easy, but also reduces possible identifiabil-
ity problems.

2) Physical Reality:
A good model should be able to capture most of the important features observed
in the physical system to be modeled. Exhibition of abnormal, nonphysical charac-
teristics can sometimes lead to numerical instability even when the model is used
for numerical solution of well-posed physical problems [40]. Following this criterion,
parameters selected in a good model should all have clear physical significances.

3) Modeling Versatility:
A class of models should be able, or can be easily extended, to account for
various effects exhibited by the physical systems of interest. A versatile mathe-
matical model may also possibly serve different purposes in different engineering
applications.

4) Computational Efficiency :
Practical implementation of a good model should be reasonably simple and
computationally efficient so that applicability of such a model will not be limited
by practical concerns. Computational efficiency for numerical implementation of a
model is of particular importance from the identification points of view, since identi-
fication processes usually require a large amount of iterations of response calculation.

5) Robustness:
 11

The sensitivity coefficients* of the parameters involved in a model should be

neither too high nor too low under practical considerations. This property is referred
to as robustness of a model. If the sensitivity of some parameters is too high, the
result of a forward response analysis may be significantly degraded by model error;
on the other hand, if it is too low, nonunique result of identification may be obtained.

It should be mentioned that in practice it may be difficult to find a model that

meets all the criteria listed above. For example, the widely-used Bouc-Wen model
[51, 52], which is described solely by a differential equation, can be viewed as both
efficient and versatile since it can model various characteristics of hysteresis, and can
be easily extended to include the effects of strength and/or stiffness deterioration.
Furthermore, the model is applicable to response analysis, including random vibra-
tion problems, as well as system identification studies [16, 42). However, this model
may not be simple and realistic enough from some practical points of view. For
example, the model may be, in some case, overparameterized such that appropriate
choice of the parameter values and identifiability problems can be very tough, as
will be discussed later in the chapters that follow. Nevertheless, if a model is always
built with the aformentioned criteria in mind, then reliable behavior and practical
applicability of the model can always be expected.

2.3.1 An Illustrative Example of Proper Modeling

To make the above ideas clearer, in the following we pose an example to show
that a simple, efficient model for forward analysis may lack robustness for identifi-
cation.
Accurate and efficient algorithms have been developed [5] for computing the re-
sponse of a single-degree-of-freedom linear oscillator subjected to an arbitrary forcing
function, in which any desired response quantity can be computed through the use
of a discrete recursive formula based on Duhamel's integral and linear interpolation

* The sensitivity coefficient of a parameter is defined as the change of some re-

sponse quantity relative to the change of the parameter value.
 12

for the excitation between discrete points sampled uniformly in time. For example,
if acceleration response i(t) of a linear oscillator described by

i + 2~w0 x + w3x = r(t) (2.1)

is of interest, where ~ and w 0 are damping ratio and natural frequency of the system,
one can use the following recursive formula:

(2.2)

(2.3)

to compute the acceleration time history, where subscript i denotes the timet= ti,
and the expressions for the coefficients can be found in [5) . In particular, the two
coefficients b1 and b2 are selected such that the transient response due to initial
conditions is determined exactly. This yields

(2.4)

where no= woflt and Dd = )1- eno. Another interpretation of this choice of b1
and b2 is that it ensures that the poles of the transfer function of the oscillator are
also poles of the transfer function of the algorithm [5). It may be noted that this
model based on the algorithm given by Eqns. (2.2), (2.3) is highly computationally
efficient and, in addition, it gives very accurate results. A comparison of a response
history obtained using the fast algorithm and that using the 4-th order Runge-Kutta
method is shown in Fig. 2.1 for a system with w 0 =2Hz and~= 5% subjected to a
scaled El Centro ground motion. The superiority of this algorithm over the Runge-
Kutta method is also prominent in other case studies. Suppose now that this model
is used for an identification study in which the input/ output data, i(t) and r(t) ,
are given, and the system characteristics ~ and w0 are to be identified. According
to Eqn. (2.2), the coefficients b1, b2 , and c1 can be found easily using regression
analysis technique such as the least-square method. Thus, based on Eqn. (2.4) we
can solve for~ and w 0 from b1 and b2 as follows :

1 a1
~= (2.5)
1 + (~) 2 ) wo = ~ flt )
 13

where
(2.6)

This method of identification is also computationally efficient, since there is no

nonlinear optimization process involved in the parameter identification procedure.
However, the identification results using the model based on the foregoing discrete
recursive algorithm will be significantly degraded when there is some noise present.
This is due to a numerical ill-conditioning inherent with the model. It follows from
Eqn. (2.4) that for lightly damped systems, the two coefficients satisfies

(2.7)

1.e., b1 is close to but less than 2, and b2 is close to but greater than -1. If the
identified b1 and b2 satisfy the relations given in (2.7), then the system parameters
~ and wo can have reasonable values. However, due to the presence of measurement
noise or numerical error, it is very possible that the values of b1 and b2 identified
using the least-square method do not meet (2. 7). As a result of this, the parameters
~ and wo computed using (2.5), (2.6) may become totally unrealistic due to the
logrithmic function involved in Eqn. (2.6). For example, if we have b1 = 2.000,
b2 = -1.001 and 6.t = 0.01 sec, then by Eqns. (2.5), (2.6) we find~ ~ 1% and
wo ~ -0.42 Hz, where the negative frequency does not make any sense in reality.
Besides, since ~ and 6.t are usually small in practical dynamic analyses, the error
of wo computed using Eqn. (2.5) may become quite large even if the error of the
identified a 1 is small.

Through the above example we realize that a model which is simple and compu-
tationally efficient may lack robustness for identification due to the high sensitivity
of its parameter values to "noise." However, we remark that if the output-error
method, instead of the equation-error method, is used with the discrete recursive
model, i.e., the optimal parameters ~ and w0 are obtained directly by minimiz-
ing some error function associated with output variables without first finding the
coefficients b1 and b2 , then the lack of robustness of the model vanishes and as
demonstrated by Beck [3], the model actually works well in the modal identification
of multi-degree-of-freedom linear structural systems.
 0.4

0 . 35

0 .3 1
,.
,J

0.2 5 -
'·
0 .2 L l\ I\
0.15 ~ j.h\
/\
I\ I\ i
0 .1
A
0 .05
58
0
0
~ ........
p. ~
til
.... -0 .03
Q
-0. 1

-O. U

·0 .2
\7
-0.25 fast recursiv<;, dt= O .OJ
"'''
i.i
-0 .3
IJ Rungc - Kutta, dt= O .OJ

-0.33 Rungc - Kutta, dt= 0 .002

-0.4 •

Time

Figure 2.1 Comparison of response histories obtained using the fast algorithm
and the 4-th order Runge-Kutta metha
 15

CHAPTER 3
M ODE LING OF ONE-DIMENSIONAL HYSTERETIC SYSTEMS

3.1 Introdu ction

Linear models, though mathematically simple, are only good for representing
structure response resulting from small loadings. When subjected to severe excita-
tions such as strong seismic ground motions, structures usually respond inelastically
so that nonlinear analytical models are required to adequately represent the struc-
tural behavior. In most cases, the response of a system that is stressed beyond
its "yield point)) into the nonlinear regime depends not only on the instantaneous
state, but also on its past history. The history-dependence phenomenon is generally
referred to as hysteresis. The study of analytical modeling of nonlinear, hysteretic
behavior of mechanical systems has thus been a research area of great interest.
In this chapter, we are concerned with one-dimensional hysteretic models that
can be used to describe nonlinear restoring force-deflection behavior or uniaxial
stress-strain relations of structural systems. The more general modeling of con-
stitutive laws of materials will be discussed later on in the following chapters.
The simplest hysteretic models are probably the elasto-perfectly-plastic model and
the bilinear model, as sketched in Fig. 3.l(a) and (b), respectively. These models
have been used extensively in many engineering applications due to their analytical
tractability. However, they are often too simple to yield good approximation to
real systems. Previous studies [23, 51] indicated that the deviation from linearity
around the yield point of a structural system should have a smooth transition in
general as it reflects the effect of the yielding in an assemblage of many structural
members. Furthermore, these simple models do not account for the hysteretic en-
ergy dissipation at small-amplitude cyclic response after the occurrence of yielding,
which can lead to a higher predicted response level than the actual response [39].
Various mathematical models have been proposed for modeling of hysteretic
behavior of structural systems. A thorough understanding of these models helps fur-
ther modeling processes involved in building up new more general models. However,
a comprehensive review of existing nonlinear models will not be presented herein,
 16

since many good references on this topic are readily available, for example, [23,
45] . Instead, we will concentrate on those models that can give smooth transition
from linear into the nonlinear regime. These models include those described solely
by differential equations involving "hidden variables", physically-based Distributed-
Element Models [19, 20], and the Masing models which are formulated based on
Masing's hypothesis and some extended rules [23] .
Since the inelastic response of a structural system is usually accompanied by
stiffness and/ or strength deterioration, it is important to extend the hysteretic mod-
els to account for these effects so that they can be used for modeling of degrading
systems. This will also be investigated in this chapter. A new class of deteri-
orating Masing models will be proposed for identification purposes to substitute
for a class of maximum-displacement-controlled deteriorating Distributed-Element
Models proposed by Iwan and Cifuentes [22].

3 .2 M o d e ling of H yst e r etic System s without D e t eriora tion

3.2.1 H y ster etic M ode ls D escribed Sole ly by Differential Equations

Analytical hysteretic models that are described solely by differential equations

in general have the advantage of good mathematical tractability. A well-known
model in this category is the Bouc-Wen model which was originally proposed by
Bouc and later generalized by Wen [51], who also applied this model to random
response analysis of structural systems [52]. The model is completely described by
the following first-order, nonlinear ordinary differential equation:

(3. 1)

where r is the hysteretic restoring force and x is the displacement of a system.

The parameters A, a, {3, and n control the slope, amplitude, and shape of the
hysteretic loops and the smoothness of yielding. This model can be shown to exhibit
an exponential type of curvilinear behavior. A simulated response of the Bouc-
Wen model is shown in Fig. 3.2(a) , where the model is subjected to a prescribed
displacement history of growing sine waves as shown in Fig. 3.2(b). Note that the
 17

Bouc-Wen model is actually a rate-independent model as Eqn. (3.1) can be put into
the form
~: = A ± (a ± ,B) Tn , (3.2)

depending on the signs of x and T.

Some special cases of the Bouc-Wen model are noteworthy, as they form the
basis of later investigations on constitutive modeling for the general multi-axial
loading cases. One is the Ozdemir's model without "back-stress" [36) which can be
described by the differential equation:

r
To =
x
xo -
Ixox I(To )n '
T (3.3)

where To is the yielding force and x 0 the yielding displacement of a system. Another
way to write Eqn. (3.3) is

(3.4)

Note that To/xo represents the linear (small-amplitude) stiffness of the system.
Another special case of Eqn. (3.1) can be found by putting A= E , a= 1/Z, ,B = 0
and n = 1 to yield
1
r =Ex- zlxiT, (3.5)

or
1
dT = Edx- z ldx iT. (3.6)

Equation (3.6) describes a simple endochronic model* [46, 47) as shown by Bazant
and Bhat [2). The merit of these models described by differential equations lies in
that they are completely defined by a single differential equation so that their ap-
plications to various engineering problems can be made more easily. For instance, it
is possible to find the closed-form solution of the stochastic equivalent linearization
coefficients for the Bouc-Wen model given by Eqn. (3.1) under some mild assump-
tions of a joint Gaussian distribution on the variables involved [52). This is the
main reason why this model has been widely used in random vibration analysis of
hysteretic systems.

* A detailed description of the endochronic model will be given in Chapter 4.

 18

Although the foregoing hysteretic models described solely by differential equa-

tions have already had many engineering applications, they commonly exhibit some
unrealistic characteristics that are inconsistent with the physical behavior of many
materials or structural systems. The main problem from a practical point of view
is the unstable drift exhibited by these models when subjected to small cyclic exci-
tations. This kind of unrealistic behavior may be attributed to the lack of physical
motivation in formulating the models. To examine in detail this unrealistic model
behavior, we note that Eqn. (3.6) can be rewritten as

(3.7)

where the minus sign corresponds to the loading case, while the plus sign corre-
sponds to the unloading case. Also, in Eqn. (3.7), we put EZ = r 0 , the ultimate
strength of the model, since we have dr / dx --+ 0 as r --+ r 0 . Thus, we may observe
from Eqn. (3.7) that the unloading stiffness of the system, given by E(1 + r/r 0 ),
can be much larger than the tangent stiffness of loading, given by E(1- r jr0 ) . This
property yields artificial drift, and also leads to the violation of Drucker's stabil-
ity postulates, since under small cyclic loading, the force-deflection loops are wide
open, as shown in Fig. 3.3, which means that energy is generated instead of being
dissipated through hysteresis. As shown by Sandler [40), such nonphysical behavior
will also lead to numerical instability when the models are used for the numerical
solution of well-posed mechanical problems.

3.2.2 The Distributed-Element Model

In 1926, Masing [31) proposed the so-called Masing's hypothesis for one-dimen-
sional hysteretic behavior of materials by thinking of a hysteretic system as con-
sisting of a collection of many ideal elasto-plastic elements, all of which have the
same elastic stiffness but different yield levels. Later in 1959, Whiteman [54), based
on the same idea, derived the uniaxial stress-strain relation of such a model by
introducing the concept of the "frequency" distribution function of the yield levels
of elements. By postulating a suitable distribution function for the yield levels of
elements, he found that the changes in the hysteresis loops are similar to those that
occur in reality, and the Bauschinger effect can be appropriately accounted for. It
 19

was Iwan [19] who first referred to such models as the Distributed-Element Models
(DEMs) and applied them to structural dynamic analysis. He constructed a model
composed of a set of N elements connected in parallel, each of which consists of
a linear spring with stiffness k/N in series with a slip element (Coulomb damper)
of strength ri / N , as shown in Fig. 3.4. Each element in the assemblage is thus
an ideal elasto-plastic element that has a force-deflection behavior as described in
Fig. 3.1(a). The DEM has been considered as physically motivated, as many real
materials or mechanical systems can be thought of as having a similar structure.
For example, real materials may have a crystalline structure that is made up of a
distribution of slip-planes or dislocations of different slip strengths. Therefore, the
behavior of such a class of models can be expected to be physically consistent, with-
out exhibiting unrealistic characteristics. The restoring force of a DEM consisting
of N elements can be found to be given by

(3.8)

for initial loading, where n is the number of elements in the yielding state. When
the total number of elements N becomes very large, the summation in Eqn. (3.8)
may be replaced by an integral so that Eqn. (3.8) becomes

1oo
r =
i
0
kx
r* </>(r*) dr* + kx
kx
</>(r*) dr*, (3.9)

where ¢(r*) dr* denotes the fraction of the total number of elements with strengths
in the ranger* ~ ri ~ r* + dr* , and the yield-strength distribution function ¢(r*)
satisfies
1 00

</>(r*) dr* = 1. (3.10)

When the loading is reversed after initial loading (i.e., unloading occurs), the force-
deflection relation can be found similarly by keeping track of response behavior of
elements in different states (yielded or elastic) . This procedure can be carried on to
trace out the entire history of hysteresis without the need to introduce additional
rules for different loading conditions. The adaptability of the DEM to transient
loading problems was considered as one of the important advantages of the model.
However, evaluation of the integrals involved in the procedure, such as those shown
 20

in Eqn. (3.9), may not be efficient for numerical solutions. Thus, in practical ap-
plications using the DEM, such as the system identification study performed by
Cifuentes and Iwan [9], one has to introduce a finite number of elements so that
their response can be traced with reasonable computation effort. This would, some-
how, degrade the results of analysis (e.g., the hysteretic yielding response curves
become non-smooth), and, moreover, make the parameter identification more dif-
ficult, unless some additional assumptions are made regarding elements' behavior
so that the number of parameters involved in the model can be appropriately re-
duced. It is important to note that the DEMs actually fall within the general class
of Masing models whose behavior can be described by the Masing's hypothesis and
some extended r ules [23]. This will be elucidated in the next section.

3.2.3 The G e n e ral Class of M asing Models

3.2.3.1 M asing' s H y pothsis and E xte nded Rules

In his original paper titled "Self Stretching and Hardening for Brass" [31],
Masing asserted that if the force-displacement curve for a system at the initial
loading is described by
f(r,x)= O, (3.11)

were r is the restoring force and x the displacement of the system, then the unload-
ing and reloading branches of the steady-state hysteretic response of the system are
geometrically similar to the initial loading curve except for a two-fold magnification,
and are described by
r- ro x- xo) = (3.12)
f( 0
2 ' 2 '
where (x 0 , r0 ) is the load reversal point for that particular loading branch. Note
that the function f should satisfy

f ( -r, -x) = f(r, x) (3.13)

so that the initial force-deflection curve is symmetric about the origin. The above
assertion is usually referred to as Masing's hypothesis for steady-state cyclic hys-
teretic response. A schematic diagram illustrating Masing's hypothesis is shown
in Fig. 3.5. The model behavior obtained using Masing's hypothesis is consistent
 21

with some experimental observations of the Bauschinger effect occurring in some

metals, which indicates that an initial plastic deformation in one direction reduces
the resistance of the material with respect to a subsequent plastic deformation in
the opposite direction. This effect is usually attributed to the residual stresses left
in the material due to initial loading or to the anisotropy of the dislocation field
generated by loading processes [26]. Some properties of response behavior using
Masing's hypothesis were summarized in [23] and are not reiterated here. One
major concern associated with the original Masing's hypothesis is that it is useful
only for steady-state cyclic response or loading between fixed limits. In the case of
transient response, or loading between variable limits, the hypothesis was consid-
ered to be of no help. However, Jayakumar [23] proposed an extension of Masing's
hypothesis by stipulating two general hysteresis rules so that simple and physical
behavior for transient hysteretic response can be obtained accordingly. The two
rules are as follows:

Rule 1: Incomplete Loops

The equation of any hysteretic response curve, irrespective of steady-state or
transient response, can be obtained simply by applying the original Masing rule to
the virgin loading curve using the latest point of loading reversal.

Consider, for example, the hysteretic loops shown in Fig. 3.6. If the virgin loading
curve OA is characterized by Eqn. (3.11), then applying Rule 1, the equation for
the branch curve CD will be

r - rc X - Xc) =
f( 2 ' 2 0. (3.14)

Based on Eqn. (3.14), it is easy to show that if the reloading curve CD in Fig. 3.6 had
been continued, it would have formed a closed hysteresis loop given by ABCDA.

Rule 2 : Completed Loops

The ultimate fate of an interior curve under continued loading or unloading is
such that once the interior curve crosses a curve described in a previous load cycle,
the force-deformation curve follows that of the previous cycle.

Based on Rules 1 and 2, if the transient unloading curve DE in Fig. 3.6 is continued,
it will reach point C and then follow the curve ABC.
 22

An effective algorithm for numerical implementation of these extended Masing

rules was proposed by Thyagarajan [45], and this algorithm has been adopted in
the present study for response simulation of Masing models. In the algorithm,
two load reversal points are removed from the memory list each time an interior
response curve crosses a curve described in a previous load cycle. One important
result regarding the extended Masing rules is that the hysteretic behavior of a
DEM can be completely described by these rules without the need of tracing each
element's behavior. This has been proven by Jayakumar [23], who then proposed a
general class of Masing models based on the extended Masing rules. To specify any
particular model in this class, only its initial loading curve need be prescribed. The
complete hysteretic behavior of this general class of Masing models is then governed
by the extended rules 1 and 2 stated above.
Thus, aDEM can be equivalent to a general Masing model if the yield-strength
distribution function ¢(r*), along with the stiffness constant k, of aDEM is chosen
to match identically the initial loading curve for a general Masing model. Con-
sidering the equivalence of the two classes of models, one may prefer to use the
Masing models in practical applications, since the implementation of the Masing
models is much simpler and more efficient than that of the DEMs. In particular, for
applications, such as system identification studies, that involve a large number of
iterations of model response calculations, numerical efficiency is of vital importance
in the choice of models, provided that the models do not exhibit any unrealistic or
unstable behavior.

3.2.3.2 A Special Class of Masing Models

Following the general class of Masing models based on the extended Masing
rules for transient response, Jayakumar and Beck [24] proposed a special class of
Masing models by defining the restoring force-deformation relation for the virgin
loading by the differential equation:

(3.15)

where k, ru and n are three model parameters which provide sufficient flexibility
to capture the essential features of hysteretic behavior of most structural systems.
 23

The parameters k and ru, respectively, represent the small-amplitude stiffness and
the ultimate strength of the system to be modeled. The additional parameter n is
introduced to model different degrees of smoothness around the yielding point, as
shown in Fig. 3.7, where the case n--+ oo corresponds to the elasto-perfectly plastic
model.
Based on the extended Masing rules, the force-deformation relation for any
loading branch other than the virgin curve is thus defined by the following equation:

dr = k [1
dx
-I r- ro
2ru
In]. (3.16)

It should be noted that the structure of Eqn. (3. 15) is similar to those used in
many other models, including the Bouc-Wen, Ozdemir, and the simple endochronic
models as introduced in Section 3.2. The major difference, though, is that for
the special class of Masing models, Eqn. (3.15) is used only for virgin loading,
not for the complete response history. The major advantage of introducing the
supplementary hysteresis rules is that unrealistic cyclic behavior, such as unstable
drift and nonclosure of hysteresis loops, can be eliminated.

This special class of models has been applied to system identification studies
using inelastic pseudo-dynamic test data from a full-scale, six-story steel structure
[24]. Even though a shear-building approximation was used in the modeling of the
pseudo-dynamic test structure, good results obtained confirm the applicability of
this class of hysteretic models to real structures. Some other special classes of Mas-
ing models can also be proposed by choosing particular yield-strength distribution
functions that satisfy Eqn. (3.10) and are characterized by suitable parameters.
More will be said about this in the next section when the Masing models are ex-
tended to account for the effects of strength and stiffness degradation.

3.3 Modeling of Degrading Hysteretic Systems

The response of mechanical systems to strong excitations can usually be char-

acterized by inelastic behavior through which energy is dissipated due to hysteresis.
Frequently, however, the inelastic response of a structural system is accompanied
by stiffness and/ or strength deterioration due to damage accumulation under cyclic
loads. The effects of deterioration of a system usually include:
 24

1) a loss of stiffness, which often results in an increase in the period of vibration,

2) a decrease in energy dissipation capacity,
3) a redistribution of internal forces, and
4) a reduction of ultimate strength.
Iemura and Jennings [18] have observed a stiffness degradation of more than
50%, based on the change of natural frequencies, by analyzing measurements from
Millikan Library Building, located on the campus of California Institute of Tech-
nology, Pasadena, during the 1971 San Fernando earthquake. Therefore, structures
under strong environmental loads are expected to undergo nonlinear and time-
dependent degrading behavior. The analysis of such a problem is also important
in many other engineering areas, such as damage detection of space structures and
adaptive control of mechanical systems. The difficulty of modeling such hysteretic
degrading systems often lies in the fact that the exact nature of system degradation
depends not only on the structural materials but also on the detailed configurations,
and may vary considerably from system to system.
In this section, we will investigate the extension of the three types of one-
dimensional hysteretic models described in the previous section to account for the
effects of deterioration observed in actual structural systems.

3.3.1 The Degrading Bouc-We n Mode l

The versatile nature and simple analytic form of the Bouc-Wen model as de-
scribed by Eqn. (3.1) has attracted considerable attention from researchers in many
related engineering fields. To make it even more general, Wen [52] extended the
original model to include the effects of strength and stiffness deterioration. T he
modeling technique for incorporating system degradation consists of the introduc-
tion of more control parameters and the selection of a response index on which the
rate of degradation is based. Wen extended Eqn. (3.1) to

(3.17)

where the two additional parameters rJ and v are introduced to control the stiffness
and strength degradation, respectively, by assuming that they are functions of a
 25

properly-chosen response index. In his original work, Wen also chose the parame-
ter A that controls the response amplitude to be a function of the response index
so that the model thus defined can achieve the maximum flexibility in modeling
general hysteretic behavior, including strain hardening or softening effects. The
response index on which the degrading parameters depend should be able to reflect
the severity and duration of the system response and is usually selected as the max-
imun displacement experienced by the model or the total energy dissipated through
hysteresis, depending on the specific structural system being modeled. Fig. 3.8 il-
lustrates the effects of the degrading parameters on the model behavior, in which
the parameters A, 'fJ, v are defined as

A(e) = Ao- DA e,

'fJ(e) = 1.0 + 871 e, v(e) = 1.0 + Ov e,

where e denotes the accumulation of the dissipated hysteretic energy, and the 8's
are constants specified for the desired rate of degradation. Although the degrading
hysteretic model given by Eqn. (3.17) is general and flexible, and it has closed-
form stochastic equivalent linearization coefficients as well, it is in general over-
parameterized, which causes difficulties in choosing appropriate parameter values
because of lack of identifiability when the model is applied to system identification
studies. For example, as reported in the paper by Sues et al. [42], the parameters
identified from simulated response can have values very different from those origi-
nally used in the simulation of response histories, though the response generated by
the identified parameters was found to be almost identical with that generated with
the original set of parameters. Another problem with the model is, as stated earlier
in this chapter, that it exhibits unrealistic drift behavior when subjected to small
cyclic excitations, which leads to a violation of Drucker's postulates of stability.

3.3.2 A Class of Deteriorating Distributed-Element Models

In order to apply the physically-motivated Distributed-Element Model (DEM)

to practical structural problems encountered in earthquake engineering, Iwan and
Cifuentes [22] presented a class of deteriorating DEMs for the overall restoring force
behavior of reinforced concrete structures. This deteriorating model consists of the
 26

same ensemble of linear springs and Coulomb slip elements as shown in Fig. 3.4;
however, some of the elements are allowed to "break" if the absolute value of the
displacements of an element exceeds some value, say J.LYi , J.L ~ 1, where Yi is the
slip (yield) displacement of the element. Once the element "fails", it no longer
contributes to the overall restoring force. Fig. 3.9 illustrates the restoring force
diagram of a typical "breaking" element, and Fig. 3.10 shows an example of the
overall response behavior of such a model subjected to a real earthquake excitation.
Note that the parameter J.L acts as the maximum ductility ratio of an element defined
as the ratio of maximum possible displacement response to the yield displacement of
an element. It was also assumed for simplicity that all the elements of a model have
the same value of J.L. This deteriorating model was shown to be capable of describing
the major features of the restoring force behavior of concrete structures [22] while
still maintaining the inherent simplicity of the original DEM. This model has also
been applied to system identification studies of real structures using earthquake
data [9] to assess the damage suffered by a structure and to predict its future
performance.

Although satisfactory identification results were obtained using this special

class of deteriorating DEMs, some additional assumptions were made in [9] pertain-
ing to the relations among parameters and to the number of elements so that the
actual number of parameters of the model was low enough for reliable identification.
The introduction of these additional assumptions on the class of deteriorating DEMs
reduces the generality and physical reality of the otherwise physically-consistent
models. This point motivated the extension of the Masing rules described in the
previous section to the case where degradation effects are of interest. If we can de-
rive a general class of deteriorating Masing models within which the aformentioned
deteriorating DEMs fall, then not only the numerical implementation of the models
can be greatly simplified, but also the problem of parameter identifiabilty can be
resolved without sacrificing the flexibility and accuracy of the model behavior.
 27

3.3.3 A Class of D eteriorating Masing Models

3.3.3.1 General Formulation

We have mentioned in Section 3.2 the equivalence between the classes of DEMs
and Masing models such that a class of DEMs may be replaced by a class of Masing
models in system identification applications in which parameter identifiability is of
major concern. The extended Masing rules on which the behavior of the general
class of nondegrading Masing models are based provide an effective way of imple-
menting numerically the model behavior. In the case where degradation effects
are to be taken into account, can we still find some appropriate rules so that the
behavior of the DEMs can be found without the need of keeping track of elements'
behavior? This question is answered affirmatively in this section for the class of
maximum-displacement-controlled DEMs mentioned in the previous section.
To begin with, we propose a general formulation for modeling of degrading
systems. Following the integral formulation of the DEM, such as that given by
Eqn. (3.9), a damage index function a = a(r*,x) t can be introduced so that
at displacement x, the fraction of the total number of elements which are "un-
damaged" and with yield strengths in the range [r*, r* + dr*] can be denoted as
a(r*, x)</>(r*) dr* :j:. Thus, for initial loading, the restoring force can be represented
as
oo </>(r*)dr* + 1kx r*a(r* ,x)</>(r*)dr*,
r(x) = kx
1
kx 0
(3. 18)

where we assume that a(r*, x) = 1 for r* 2 kx, which means that elements that
are unyielded must be undamaged. The second term on the right-hand side of
Eqn. (3. 18) is the contribution from elements that are yielded at deformation x,
and the first term denotes the contribution from elements that are still in the
elastic state for which a(·, ·) = 1.

t In general, a can be a function of the history of x, not just x(t).

:j: (1-a) could be viewed as the fraction of elements in the yield range [r*, r*+dr*]
which have failed and completely lost their strength, or that (1- a) is the "average"
partial loss of strength of each element in the yield range [r*, r* + dr*].
 28

Similarly, for unloading from x 0 , we have for -x0 ::; x::; x 0 ,

oo ~kxo
r(x)=kx
1k Xp
¢>(r*)dr*+
k ( xp-x)
2
[r*-k(xo-x)]a(r*,x)¢>(r*)dr*

k(xp - x)
2
-r* a(r* , x) ¢>(r*) dr*, (3.19)
+1
and for x < -xo,

oo
+ 1-kx -r*a(r*,x)¢>(r*)dr*.
r(x) = kx
l-kx
¢>(r*)dr*
0
(3.20)

Very similar equations can also be derived for reloading. Equations (3.18), (3.19)
and (3.20) can be differentiated with respect to x to get equations for the model
"stiffness" as follows:

dr = k r oo ¢>(r*)dr* +1kx r* aa(r*,x) ¢>(r*)dr*, (3.21)

dx l kx 0 ax

for initial loading,

dr 1 oo jkxo
-d =k ¢>(r*)dr*+k a(r* ,x)¢>(r*)dr*
X k k (x -xp )
Xp 2

k k(xp-x)

+ r Xp [r*- k(Xo- x)] aa(r* l X) ¢>(r*) dr* - 1

2
r* aa(T*, X) ¢>(r*) dr* l
J~(xp-x) 2
ax 0 ax
(3.22)
for unloading from xo with -xo ::; x ::; x 0 , and

dr = kloo ¢>(r*)dr*- r -kx r* aa(r*,x) ¢>(r*)dr*, (3.23)

dx -kx lo ax

for unloading from xo with x < -xo .

We may note that the equations for r and dr / dx in the case of unloading from
x 0 with x < -x0 (Eqns. (3.20) and (3.23)) are the same as those for initial loading
in the negative direction (cf. Eqns. (3.18) and (3.21)). This is consistent with the
extended Masing rule 2 regarding completed loops for transient response, as stated
in the previous section. Based on the above general formulation for modeling of
degrading systems, we can propose different classes of degrading models by suitably
 29

choosing the damage index functions a. A specific example of choosing a as a

Heaviside step function will be given in the next section to illustrate the above
formulation.

3.3.3.2 A Special Class of Deteriorating Masing Models

To gain more insight into the above general formulation, we consider the special
case where a specific damage index function is chosen for modeling deteriorating be-
havior of hysteretic systems. If, for example, the maximum-displacement-controlled
deteriorating DEM described in the previous section is to be derived from this gen-
eral formulation , we can choose the following damage index function:

a (r * ,x) = 1- H(kxm(t) - r *) , (3.24)

J-i

where H(-) is the Heaviside step function, xm(t) - ~~: lx(T)I, which is the maxi-
mum displacement magnitude experienced by the model, and J-i is the parameter of
maximum ductility ratio of the model. Note that the maximum possible displace-
ment of an element with yield strength r* is given by J-Lr* jk. Using Eqns. (3.21),
(3.22), and (3.23) with a given by Eqn. (3.24), and defining the "stiffness" function

f(x)- k r=¢(r*)dr*,
Jk x

which is assumed to be differentiable, we can find

-dr = k
dx
1=(
kx
¢r*) d*r -
k-x ¢
2
(kx)
2
- = f(x)+ 2
x f'(x)
J-i
-,
J-i J-i J-i
(3.25)

for initial loading,

dr = k rXJ if>(r*) dr* = f(XQ ), (3.26)
dx } !::..rL J-i
1-'

for unloading from x 0 with lf~ x 0 2

::; x < x 0 , where we note that f(xo/J-L) is a
constant so that the restoring force r is linear over this portion,

dr
dx
=k r=
} ~<:z:o-:z:>
if>(r*) dr* = f(XQ- X),
2
(3.27)
2
 30

for unloading from xo with -x0 ~ x < ~-'~ 2 xo, and

-dr =
dx
k ;=(
-h ~
2
¢> r *) dr * + -k x ¢> (-kx)
- = f (-x ) - -x f '(-x
~
-)
~ ~
(3.28)

for continued loading where X < -xo. Note that in the derivation oafox is zero
unless Xm is increasing during the loading branch under consideration. The above
results lead to the following remarks pertaining to the behavior of the special class
of degrading Masing models:
1) By symmetry of the force-deflection response about the origin, we require that
the stiffness function f be even, i.e., f( -x) = f(x ), and as a result, f'(x) must
be an odd function of x.
2) By comparing with the formulation of the nondegrading case, we can find
that the term -;r f' (~) signifies the effect of stiffness deterioration due to the
breaking behavior of elements. One can also find the corresponding term for
strength deterioration from equations for the restoring-force function r(x).
3) Eqn. (3.28) is equivalent to Eqn. (3.25) with x replaced by -x, which is consis-
tent with the extended Masing rule 2 on completed loops as mentioned earlier.
4) It can be shown that for the case of reloading from -x0 , the result will be
identical to those derived for unloading from x 0 , except for that the term xo;x
is replaced by x-.to .
5) The steady-state response behavior of the maximum-displacement-controlled
deteriorating DEM can be summarized as follows for the case of cycling between
displacement [-xo, xo] with previous maximum displacement amplitude Xm:

dr __ f(x) + _x f'(-x) (for initial loading ) ,

dx ~2 ~

{ -xo < x
x >x>x-~ (for unloading),
= J(Xm), if 0 - 0 f'2x
~ ~ -xo + ~ (for reloading),
(3.29)
__ f(xo- x), 1.f Xo -
2xm
----,;:- > x 2: -Xm
(unloading) ,
2
__ f(x- xo), I.f
2
- Xo +----,;:-
2xm < x ~ Xm
(reloading) ,

=same as initial loading, if lxl = Xm with lxl increasing.

 31

6) The results given in Eqns. (3.29) can also be obtained by directly keeping track
of elements' behavior at different response stages.
7) The behavior for the transient response of the model can be similarly derived
and can be shown to be completely consistent with the extended Masing rules
1 and 2 proposed by Jayakumar (23] for the case of no degradation.
8) Typical behavior of the model response for different loading branches is shown
in Fig. 3.11, where the effects of strength and stiffness deterioration can be
clearly observed. The model is based on a Rayleigh distribution for the yield-
strength distribution function ¢(r) (cf. next section) with k = 20,ru = 1.2,
and J-L = 5.

In summary, we have derived a special class of degrading Masing models, which

is equivalent to the maximum-displacement-controlled deteriorating DEMs pro-
posed by Iwan and Cifuentes (22]. This class of degrading Masing models, however,
can be completely defined by specifying the stiffness function f (or the initial load-
ing curve) and J-L, and the response behavior for other loading branches will follow
the rules given in (3.29) and the two extended Masing rules for transient response
given in Section 3.2. In this way, we not only simplify the numerical implementation
of the special class of DEMs, but also solve the problem of parameter identifiability
without introducing additional assumptions regarding model behavior. In other
words, the applicability of such models to practical engineering problems is greatly
increased.
As a final remark, we note that by choosing other damage index functions
(which is equivalent to choosing some particular element behavior of the DEMs)
based on theoretical results or physical observations, one can come up with different
classes of degrading models that may be suitable for particular structural systems.
An important feature of this modeling approach is that the models thus derived are
all based on the behavior of distributed elements, which is supposed to be physically
consistent, i.e. , will not exhibit any unrealistic characteristics or introduce any
physical or numerical instabilities.
 32

3.3.3.3 Yield- Strength D istribution Function

Similar to the general class of nondegrading Masing models, a class of degrading

Masing models can be completely defined by specifying the initial loading curves
with some appropriate functions. This can, in general, be accomplished in two
ways. One way is to specify the "stiffness" function f(x) as shown in (3.29) directly
in terms of x , or in the form of a differential equation, such as the special class of
Masing models proposed by Jayakumar [24]. This special class of Masing models is
described by the following differential equation for initial loading:

(3.30)

where only three parameters k, Tv. and n are needed for modeling general yielding
behavior. However, the function f(x) cannot, in general, be found explicitly from
Eqn. (3.30) except for the cases where n = 1 and n 2, which results in the
following relationships:
(3.31a)

for n = 1, and
(3.31b)

for n = 2.

Another way of specifying the initial loading curve can be done by choosing
appropriately the yield-strength distribution function </>(T*) . One may have noted
that the distribution function </>(T*) behaves the same as a probability density func-
tion does, as suggested by Eqn. (3.10) . In this study, we propose the use of Rayleigh
distribution described by

2
1r T -7r T )
</>(T) = - -2
exp ( - - 0::; T < oo. (3.32)
2 T 1.1. 4 T2
1.1.
'

And by definition we can find the corresponding "stiffness" function given by

00
[-1r (-)
kx 2] ,
f(x) = k
1
kx
</> (T * )dT* = kexp -
4 Tv.
(3.33)
 33

which is expressed explicitly in terms of x . Also, we can find that the corresponding
restoring force is given by

T(x) = Jo
rxf(~) d1, =Tv. erf
(T.fii kx)
Tv. , (3.34)

where the error function erf(·) is defined as

We remark that the parameter Tv. represents the ultimate strength of the model,
since by Eqn. (3.34), T(x) ---+ Tv. as x ---+ oo (since erf(oo) = 1). The choice of the
Rayleigh distribution is thought to be a natural one, since it distributes within [0,
oo) and has only one single parameter, as Tv. in Eqn. (3.32). From Eqn. (3.34), we
note that the Masing model derived using the Rayleigh distribution involves only
two parameters k and Tv., which is similar to Eqns. (3.30) and (3.31) corresponding
to fixed values of n. It can be shown, by numerical calculation, that the model
using the Rayleigh distribution is very close to Jayakumar's special class of Masing
models with n = 2.5, as illustrated in Fig. 3.12, where initial loading curves of the
two models with the same values of k and Tv. are compared.
A more general distribution function can be proposed, based on the Rayleigh
distribution, as follows:

¢(T) =
2n T2n-l exp [- (7r4TT;2
(~T;r
)n J' 0 :::; T < oo, n > 0, (3.35)

which may be referred to as the generalized Rayleigh distribution. The additional

parameter n is introduced to control the smoothness of transition from elastic to
plastic state. When n = 1, Eqn. (3.35) reduces to the Rayleigh distribution given by
Eqn. (3.32). A plot of the distribution curves described by the generalized Rayleigh
distribution function for different values of n is shown in Fig. 3.13. We find from
Eqn. (3.35) that
(~ k ~ )n],
2 2

f(x) = kexp [- (3.36)

4 Tv.

and the restoring force can also be found in closed form as

T(X) =
2 Tv. r ( VX 2n, 2n n+ 1) + k X
.fii e-1/X
2n
, (3.37)
2
 34

where we defined v = (~k;

ru
)n and

(3.38)

which is the Incomplete Gamma function with parameter a.

Thus, we have defined a general class of degrading Masing models based on the
generalized Rayleigh distribution function. Some major advantages of this class of
models are as follows:
1) There are only 3 parameters k, ru. and n needed for modeling general one-
dimensional non-degrading, hysteretic behavior. For modeling of degrading
systems governed by maximum displacement response, only one additional pa-
rameter J.l is needed. The parsimony and clear physical significance of param-
eters make this class of models excellent for identification purposes.
2) Explicit closed-form representations in terms of x can be obtained for the stiff-
ness function f and the restoring force r. This feature makes the numerical im-
plementation of this class of models computationally efficient, especially when
the displacement history x(t) is prescribed.
3) The mathematically tractable form of the stiffness function f(x), given by
Eqn. (3.36) in terms of the exponential function, allows this formulation to
be handled more easily in the case of system deterioration, since the extended
hysteretic rules for degrading behavior, described by (3.29), involve f'(x) in
addition to f(x).
4) The specification of the yield-strength distribution function ¢(r*) facilitates
the computation of some response quantities of the model based on statistical
results. For example, if the drift (plastic deformation) response d(t) of a non-
degrading DEM based on the Rayleigh distribution is of interest, then we
can simply find the mean (expected) value of the drifts of all the distributed
elements constituting the model. Thus, for initial loading with x > 0,

d(t) = rk x (x- ~) ¢(r) dr

lo k
7rT2) dr
=
1o
kx
(x
r 1fT
- -) -
k
exp
2r~
(
- -
4r~
 35

= x- ru erf ( .Jiikx)
k 2ru

-
r(x)
- x - -k- . (3.39)

This "statistical" interpretation of plastic deformation is thus equivalent to

the conventional formulation of the total-deformation theory, as illustrated in
Fig. 3.14. Also, for unloading or reloading from x 0 ,
kJo:-o:p l

d(t) =do± fo 2

( lx- xo l - ~) </>(r) dr
-_ do ± [I x - xo I - -2ru erf (.Jii klx_- _xo_.:.i)J
- ___:_
k 4 ru
= do ± 2 [x - kru erf .fiikx)]
( 2;:::- (x _ lx ~ xol)
-
=do±2 [x- r(x)],
k (3.40)

where "+" and "-" correspond to the cases where i; > 0 and i; < 0, respec-
tively, and d0 denotes the drift response corresponding to xo. From Eqn. (3.40),
we can derive
(3.41)

Comparing (3.41) with (3.39), one can realize that with the statistical formulation
of the Masing models (or the DEMs), the Masing rules for restoring force response
also apply to the drift response or possibly other response quantities.

Thus, with the formulas given in Eqns. (3.39) to (3.41), we can compute ef-
fectively the drift response history, in addition to other response quantities, of a
hysteretic system modeled by the special class of Masing models. A numerical sim-
ulation using these formulas* is performed for a DEM and the result is shown in

* The numerical implementation of the error function is done by using the Hast-
ing's formula [15]:

where
1
t = , a 1 = 0.254829592, a2 = -0.284496736,
1 + 0.3275911x
 36

Fig. 3.15, where the comparison has been made to a one-dimensional endochronic
model which will be investigated in detail later in Chapter 4. The good agreement
of the drift response histories between the models again indicates the validity of the
"statistical" formulation based on the "probabilistic" distribution function ¢(r) of
the yield strengths of distributed elements.
Based on the previous experience with identification of structural systems us-
ing the general class of Masing models [23], it should be noted that, in practice,
more reliable results of identification can be obtained by fixing the value of the pa-
rameter n in the general model so that the interactions among parameters can be
greatly reduced, which implies that model identifiability is much improved. This is
particularly important for the identification studies in which the system response is
not driven far into the strongly inelastic regime. Therefore, although there is some
loss in the flexibility of the model, it is proposed in later identification studies to
fix the value of n based on appropriate engineering judgement.

3.3.4 Other Models for D egrading Systems

There have been many other models than those described above for modeling
of degrading systems. For earthquake motions, building structures made of rein-
forced concrete often exihibit stiffness deteriorating behavior. Iwan [21] presented
a hysteretic model for stiffness degrading systems which may be thought of as a
subclass of the distributed-element model. This model consists of three types of
basic elements, including theE-type (elastic elements), theY-type (elasto-perfectly
plastic elements), and the C-type (elements exhibiting cracking and crushing like
behavior), as shown in Fig. 3.16. Gates [13] applied this model to earthquake re-
sponse analysis of deteriorating systems by using only one element from each of the
three basic types. The model configuration and its response behavior are shown
in Fig. 3.17, where the contributions from each type of element are also included.
Although this model is capable of modeling a wide range of deteriorating structures,
the model characteristics and its response behavior are considered to be too com-
plicated as far as system identification is concerned. An attempt at deriving the

a3 = 1.421413741 ,a4 = -1.453152027,a5 = 1.061405429.

This formula is accurate over [0, oo) to within ±1.5 x 10- 7 .
 37

corresponding response rules for the model as was done earlier for the maximum-
displacement-controlled DEM was made. However, the result was too complicated
for practical applications, as a complete description of the model behavior required
too many mathematical rules for different response branches.
Clough [10) proposed a stiffness-degrading hysteretic model based on the bilin-
ear hysteretic model. The model behavior is shown schematically in Fig. 3.18. In
this model, stiffness degradation is introduced only as "load reversal" occurs (i.e.,
when the restoring force r changes its sign). This is not completely consistent with
experimental observations which show that stiffness degradation occurs also during
unloading behavior. Takeda et al. [44) presented a rather complicated degrading
model based on their experimental results regarding reinforced concrete behav-
ior. The model behavior is based on a trilinear primary curve which represents the
three stages of uncracked, cracked, and post-yielding response of concrete structural
members. The general behavior of the model is sketched in Fig. 3.19. However, a
complete description of the model behavior requires more than a dozen rules gov-
erning different response branches. In contrast to Clough's model, Takeda's model
takes account of stiffness degradation at both unloading and load reversals. The
slope of an unloading curve after yielding occurs is given by the empirical equation

kunloading = k' ( x: )0.4,

X
(3.42)

where k' is the slope of a line joining the yield point in one direction to the cracking
point in the opposite direction (cf. Fig. 3.19), and Xy and Xm denote, respectively,
the yield deformation and maximum deformation experienced by the system in the
direction of current loading. Although the model was built based on observations
made in many experimental studies on reinforced concrete members, it is, in general,
too complicated for practical applications, especially for identification studies.
A simplified version of Takeda's model was developed by Saiidi and Sozen [38)
in which the trilinear primary curve is replaced by a bilinear curve as shown in
Fig. 3.20. To simplify the model behavior, the largest excursion point in both
directions is viewed as the largest excursion point in either direction. This model
takes into account hysteretic energy dissipation during low-amplitude deformation
if the model has yielded in at least one direction . This is true also for Takeda's and
 38

Clough's model, but not for the elasto-perfectly plastic or bilinear models. This
characteristic is important as reported in [39] for accurate prediction of response
peaks and frequency content of hysteretic systems subject to earthquake excitations.
The unloading slope in the inelastic region of the Saiidi and Sozen's model
is similar to that given by Eqn. (3.42) except that k' is replaced by the initial
elastic slope of the response. It was shown in [39] that Saiidi and Sozen's hysteretic
model incorporates the principal features of hysteresis presented in Takeda's model,
including:
1) dependence of unloading stiffness on the maximum deformation experienced
by the system,
2) stiffness degradation during load reversals, and
3) hysteretic energy dissipation for small-amplitude deformation after yielding.
But the model is considerably simpler than Takeda's, which makes this model more
suitable in practice for determining hysteretic response, or for identifying system
characteristics of reinforced concrete structures.
The investigation of the preceding degrading models is to gain better insight
into the behavior of degrading hysteretic systems and to provide some justification
for the aforementioned degrading Masing models. The maximum-displacement-
controlled degrading Masing model possesses all the three response features stated
above without the need to introduce any additional empirical approximations re-
garding its hysteretic behavior. This indicates that the proposed Masing model
has a physically consistent behavior and is good for modeling of reinforced concrete
structural systems. Another interesting hysteretic model based on endochronic
theory will be presented in the next chapter. This generally-formulated model for
multi-axial cyclic plasticity also serves as a comparison basis for the Masing models
or the DEMs, as we did in Fig. 3.15 for the drift response of a Masing model. Be-
sides, the general, consistent behavior of the endochronic model also motivates the
generalization of the DEMs into multi-dimensional representations for constitutive
modeling of stress-strain relations of materials. This will be covered in the next two
chapters.
 39

I X
I
I
I

(a)

(b)

Figure 3.1 Hysteretic restoring force behavior of (a) the elasto-perfectly

plastic model (b) the bilinear model
 40

2 .5

1.5
Q
:;
1-

u."" 0 .5
!:!)
c:
·c
....0 .0.5
"'0
~

Displacement

(a) restoring force diagram

0.8
0.7

0.6

0.5

0 .4
....
t: 0 .3
0
E
()
0 .2

u 0 .1
(';

a.
tl)

.0.1
0
-0.2

.0.3

.0.4

-0.5

-0.6
.0.7

Time

(b) prescribed displacement history

Figure 3.2 Response behavior of the Bouc-Wen model

 41

Figure 3.3 The unstable drift exhibited by the Bouc-Wen model

n* /N

•
•
•
•
•
•

kN/N

Figure 3.4 The Distributed-Element Model for one-dimensional hysteresis

 42

Figure 3.5 Masing's hypothesis for cyclic hysteretic loops

Figure 3.6 Hysteretic loops for transient response

 12r---------r-------~--------~---------r---------r--------,---------~--------r--------.r--------,

Q)
()

~ 7
~
eo
6
·20
.... ~
Cll 51 II/ I u.J
0
~
4

0
! I
6.3 l I I3 ~ 2~ ! BI ! • I3 ~
Displacement

Figure 3. 7 Effect of the parameter n of a specia l class of Masing models [241

on the smoothness of yielding curves
 44
z
o.., =0.05

-5 5 u

o.,-o.2o z

-5 s u

6.•0.20 z

-5 5 u

Figure 3.8 Effects of the degrading parameters on the behavior of the deteriorating
Bouc-Wen model (from (52])
 45

rt

Figure 3.9 Restoring force diagram of a typical breaking element in the maximum-
displacement-controlled deteriorating OEM

u
~
....Ill
E
::.
(/)
(/)
:II

-,
E
c:
....q) 0

u
...
.2
0
c: ·•OO
-=2
(/)
q)
a:
·-
Figure 3.10 Restoring force behavior of the maximum-displacement-controlled
deteriorating OEM subject to an earthquake excitation (from [22])
 1.2 .------,------,----.-----.-----y----.-----y-----,
k = 20.
0.8 ru = 1.2
fL = 5.0
0.6

Q)
u 0 .4

8 0 .2
~
co
c:
·c .+;:>..
0\
9
Ill · 0 .2
Q)
~ · 0.4

· 0.6

· 0 .8

·I

· 1.2 I 1 I h b I 1 1
-o.>s X •< _Kt Jl 0.15 6. 1 0.15 0.2
k<
Displacement

Figure 3.11 Simulated restoring force behavior of the proposed deteriorating

Masing model based on a Rayleigh distribution function
 47

4 .5

0 3.S
0
...
0
~ 3
bl)
.51-< 2..5

~
- 0
{ /)
0
1.5
2
Rayleigh distribution

Jayalcumar. n=2.5

0 .5

0
r------c~-----.&------.~------~----~
Displacement

Figure 3.12 Comparison of the behavior of two matching Masing models

0 .7

0 .65

0 .6

c o.ss
........0 0..5
0
0 .4S

& 0 .4
c
...0
-
.D
j

' I"'
!l
0 .35

0 .3

0 .25

....
{/)
0.2
a O. IS

0 .1

o.os
0
12
Yield Strength

Figure 3.13 Distribution curves described by the generalized Rayleigh distribution

function with different values of n
 48

Xo X

Figure 3.14 Plastic deformation of the Masing model based on a statistical

interpretation of the Rayleigh yield-strength distribution function

0.9

0.1

0.7

0.6

..,
...
O.l

0.1

~
0.1

a ....
.0.2

-<>.>

....
..,.,
.0.6
- - Masing model
-<>.7
- - endochronic model
..... ········· prescribed displacement
-<>.9

·•
Time

Figure 3.15 Drift response of a Masing model based on the Rayleigh distribution
function as compared with an endochronic model
 X X X

kc
ke ks

f f f

.p..
CD

X X X

E- TYPE ELEMENT Y- TYPE ELEMENT C- TYPE ELEMENT

Figure 3.16 Different types of elements used in a model for hysteretic systems
with stiffness degradation (from [ 131)
 50

SYSTEM

-=:=:::::::::::===--f-==:::::::::== E-TYPE ELEMENT

C~NTRIBUTI~N

z .<
z z 2
/ z :;:bsrz z z z z 7
z Y-TYPE ELEMENT
C~NTRIBUTI~N
I

Figure 3.17 Configuration and behavior of Gates' degrading model (from [13])
 51

Figure 3.18 Clough's hysteretic model for stiffness-degrading behavior

a=(Xy/Xm) 0.4

Figure 3.19 Takeda's hysteretic model for reinforced concrete structural systems
 52

-Xm
Xm X

a=(Xy/Xm) 0.4

Figure 3.20 Saiidi and Sozen's hysteretic model for stiffness-degrading behavior
 53

CHAPTER 4
MODELING BASED ON ENDOCHRONIC THEORY

4.1 Introduction

A simple endochronic model described by a differential equation has been intro-

duced in Chapter 3 (cf. Eqn. (3.5)) as a special case of the Bouc-Wen model. The
original endochronic theory was proposed and developed by Valanis in 1971 [46], and
modified later in 1980 [47] for constitutive modeling in cyclic plasticity. The en-
dochronic theory can be viewed as a generalization of the theory of viscoelasticity in
which the real time variable is replaced by an auxiliary variable, called the intrinsic
time, which is a monotonically increasing measure of deformation history of a mate-
rial. The constitutive law developed based on the endochronic theory can adequately
characterize hysteresis, and strain hardening behavior of some metals, without re-
sorting to the definition of yield conditions, flow rules, or any hardening rules. The
original endochronic theory, of which the model described by Eqn. (3.5) serves as a
special case of a one-dimensional formulation, was shown to violate Drucker's pos-
tulates of stability, as discovered by Sandler [40]. This discovery led to a major
modification of t he theory in which the intrinsic time was defined in the plastic-
strain space, instead of the total-strain space [47]. This makes the stiffness at the
onset of unloading identical to the "small-amplitude" stiffness of the initial loading,
and the resulting endochronic formulation shows proper hysteresis-loop closure and
hence more realistic response behavior. Although the endochronic theory was orig-
inally developed for modeling material behavior of some metals, extensions of the
theory have been made for modeling of some other materials, such as concrete [2]
and soils (48].
The major contribution of the endochronic theory is t hat the theory provides a
unifying approach of describing the elasto-plastic behavior of materials without the
requirement of introducing a yield surface and a loading function which distinguishes
between loading and unloading. In the next section, the basic formulation of the
modified endochronic theory and an effective algorithmic implementation of the the-
ory will be described. Moreover, some inherent properties and difficulties associated
with the endochronic model will also be mentioned from some practical points of
 54

view. In Section 4.3, a very effective modeling technique for models based on the
endochronic theory will be proposed to simplify the otherwise complicated modeling
process. Comparison of simulated responses between the endochronic models and
the Masing models will also be made for the uniaxial loading case to examine further
the model behavior. Finally, cyclic hardening behavior exhibited by real materials
and issues pertaining to the modeling of such behavior will be discussed.

4.2 Endochronic Theory And Its Implementation

The endochronic theory was originally derived based on the internal variable
theory of irreversible thermodynamics and the concept of intrinsic time which acts as
a proper measure of material memory of its past deformation history. AB mentioned
earlier, the endochronic theory can be viewed as a generalization of the theory of
viscoelasticity. To demonstrate this, let us consider the one-dimensional Maxwell
model in viscoelasticity [11], given by

1 (}'
dE= E dCY + EZ dt, (4.1)

or equivalently,
8E
1
t - ( t-T)
(}' = Ee z -a dT, (4.2)
0 T

where E is Young's modulus and Z is the relaxation time of the material being
modeled. If we replace the time differential dt by a differential of the intrinsic time
d(, which is defined by
(4.3)

then we get the simple endochronic model given by Eqn. (3.5), or equivalently by

{( ( ') 8E 1
(4.4)
CY = Jo P ( - ( 8(' d( ,

where
p(() = Ee=i. (4.5)

The integro-differential form of Eqn. (4.4) , in which p(() represents a material func-
tion, is typical for general endochronic models.
 55

A complete set of constitutive equ ations for plastically incompressible, rate inde-
pendent materials based on the modified endochronic theory [47) can be summarized
as follows:
(4.6)

(4.7)

and
z d P
s'I.J· · =
10
I fij I
p(z - z ) -d I dz l
z
(4.8)

where tii and Efj are, respectively, the elastic and plastic components of the total
strain tensor Eij, and Sij is the deviatoric stress tensor defined as

1
s'I.J-
·· - CJ'I.J
·· - -Cikk8·
3 ·
'I.J l
(4.9)

and
d7]
dz = f(7J), !(7J) > 0, (4. 10)

d'Tl., = (d€1?'I.J. d€1?.) 112

'I.J ' (4. 11)

where p(z) and f(7J) are material functions called t he kernel (or memory) function and
the (cyclic) hardening function, respectively. The differential quantity d7J represents
the distance between two consecutive plastic-strain states, so that 'T} defines a memory
path in the plast ic-strain space through which history-dependent effects of a material
are introduced into the endochronic model. Note the resemblance of Eqn. (4.8) to
Eqn. (4.4), where the total-strain increment is replaced by the plastic-strain increment
so as to make the model behavior more physically consistent [47). In the case of
isotropic materials, Eqn. (4.6) becomes

Ev e ~ 2G e (4.12)
Ciij = (1 + v)( 1 - 2v) fkk Uij + Eij '

where E is Young's modulus , v P oisson's ratio, G = E / 2(1 + v), the shear modulus,
and 8ij denotes t he Kronecker delta function, i.e., 8ij = 1, if i = j, and 8ij = 0,
otherwise.
By introducing different forms for the two material functions p(z) and f(TJ),
various types of elasto-plastic behavior of materials can be adequately modeled. An
 56

important feature of the kernel function p(z) in Eqn.(4.8) is that the function must
be singular at the origin, that is
p(O) = oo, (4.13)

so that the model can account appropriately for the elastic behavior at the onset
of initia l loading and unloading response. There are basically two major types of
endochronic models based on the assumed form of the kernel function p(z), which
are now described.
The first one utilizing the Dirac delta function 6(z) is given by

p(z) = s~ 6(z) + P1 (z), (4.14)

where p 1 (z) is a regular function and s~ is a material constant that has the physical
significance of t he initial yield stress in simple tension. This formulation leads from
Eqn. (4.8) to
(4.15)

where
r ij ( z) = }
rz Pl ( z
I dEfj I
- z ) dz 1 dz . (4. 16)
0

Note that Eqns. (4. 10), (4.11) and (4.15) imply

ll s (z)- r (z)ll = s~ f('f]), (4. 17)

1 2
where s and r denote respectively the two tensors Sij and rij, and lls ll - (sij Sij) 1 .

Note, however, that when z = 0, i.e., in the process of initial loading and lls ll < s~,
the model response is governed by purely elastic behavior, such as that given by
Eqn. (4.12) in the case of isotropic materials. From Eqn. (4.17), it can be deduced that
this formulation results in a generalization of the classical theory of plasticity in such
a way that the hardening function f ("7) signifies isotropic hardening behavior (yield-
surface expansion), while the tensor r ij ( z) denotes kinematic hardening behavior
(yield-surface translation). Furthermore, it can be shown (50] that by a suitable
choice of p 1 (z) as a sum of exponential functions, the theory becomes similar to the
classical multiple-yield-surface theory in which nested yield surfaces translate in the
stress space according to some kinematic hardening rule.
 57

Although the above formulation that uses the Dirac delta function in p(z) led
to a generalization of the classical plasticity theory, the original idea of avoiding the
concept of yield surfaces and hardening rules was not completely preserved. Thus,
another type of formulation of the endochronic theory has been developed [49] by
assuming the kernel function as
00

p(z) = L cke-crkz, (4.18)

k=l

where the non-negative constants Ck and ak satisfy

00

L Ck =oo, (4.19)
k=l

so that Eqn. (4.13) is satisfied and the integrability of p(z) can also be guaranteed.
An effective numerical algorithm for implementing the endochronic t heory can be
derived based on the assumption of Eqn. (4.18) for the kernel function [17]. Suppose
that the loading process is divided into many small steps and in each step no load
reversal occurs, t hen Eqn. (4.8) can be written as

Zl p(z- z') df..'f!. 1 Z2 p(z- z') df..'f!.

Sij(z) =
1 0
d t~ dz' +
Z
d t~ dz' + ...
Zl Z

~
df..'f!.
dtJ
Z
Iz=O 1Z1
0
dif_ .
p(z- z') dz' + dtJ I
1 Z2 p(z- z') dz' + .. . ,
Z z=z1 z
(4.20)
1

where an approximation has been made by assuming that dcfj / dz is const ant within
each loading step. With the kernel function defined by Eqn. (4. 18), (4.20) can be
manipulated further to obtain

(4.21)

where tl zi _ Zi - Zi - l and the subscript m denotes the m-th loading step. Note
that in Eqn. (4.21), the infinite sum over k has been approximated by a finite sum
of n terms as a practical consideration. In order to avoid t he numerical difficulty of
 58

small numbers involved in the term e-ak(zm-z,_l), one can, based on mathemat ical
induction, convert (4. 21) into a recursive formula:
n
Sij(Zm) = L st(zm), (4.22)
k=l

and

(4.23)

It should be noted that the aforementioned approximate numerical scheme will result
in an exact solution to the constitutive equations based on the endochronic theory if
the material being modeled does not exhibit cyclic hardening behavior (i.e., f (TJ) =1),
and the deformation history follows a piecewise linear path in the plastic-strain space,
such as the uni-axial loading case. Equations (4.6) to (4.11), with (4.8) replaced by
(4.22) and (4.23), provide a set of recursive formulas for computation of the response
of models based on endochronic theory. The advantage of using this numerical scheme
is that only the values of the response states at the end of the previous loading step
need be stored. Once the loading increment (b.Efj)m (or (b.sij)m) is given, (st)m can
then be determined by referring to (sfj)m-l· Hs u, et al. [17] proposed two efficient
schemes following the above algorithm for either stress-controlled or strain-controlled
response simulations, by making a few more algebraic manipulations on the foregoing
formulas.
In the following, we address some important points regarding the general behav-
ior and properties of the endochronic model. With the kernel function defined by
Eqn. (4.14), Watanbe and Atluri [50] introduced the concept of "limit surfaces" as-
sociated with the endochronic models. When the hardening function f (TJ) saturates
to a limit value, a limit surface exists which can move around in the stress space, as
can be deduced from Eqn. (4.17). As indicated by experimental observations [30],
however, materials such as metals, which have been cyclically stabilized, in general
exhibit response behavior with "fixed" limit surfaces, beyond which the stress state
never goes. Thus, we reject Eqn. (4. 14) and it is interesting to see whether the en-
dochronic model based on the alternative formulation of the kernel function given by
Eqn. (4.18) can exhibit such a physical behavior.
 59

Note that Eqn. (4.23) may be rewritten as

- Im
Ck .6-t:fj
(.6-Stk··] )m= [ -
ak -.6-z - S k·· (Zm
tJ -
I) J(1-e -a:- k l:!.z" ' ) . (4.24)

If the stress state reaches an "equilibrium state", at which the stress increments
approach zero for appreciable unidirectional strain increments (30], i.e.,

(.6-s:j)eq = 0 Vk = 1, 2, ... , n, (4.25)

then by Eqn. (4.24) we have

(4.26)

or
(4.27)

From Eqn. (4.27) , we see that the stress state will remain at the equilibrium
state until the plastic-strain path changes its direction. Noting that in the case of no
cyclic hardening, (.6.z) 2 = .6.t:fj .6.t:fj, we get, from Eqn. (4.27),

(4.28)

where ku is a finite model constant (see Eqn. (4.19)) that can be related to some ma-
terial constant, as will be done later. Equation (4.28) signifies that all the equilibrium
states (Sij) eq form a hypersurface in the six-dimensional stress space (considering the
symmetry of a stress tensor). This hypersurface is actually a limit surface associ-
ated with an endochronic model based on the kernel function given by Eqn. (4.18).
This can be shown as follows: Considering the cyclically stabilized behavior (i.e.,
f(7J) - 1), from Eqns. (4.10) and (4.11) we can put

(4.29)

where nij(z) ~ 1 and satisfies

(4.30)
 60

Thus, it follows from Eqn. ( 4.8) that

(4.31)

Using the tensorial product, we get

Sij(z) Sij(z) = foz p(z- z') nij(z') dz' foz p(z- z") nij(z") dz"

= foz foz p(z-z')p(z-z")nij(z')nij(z")dz'dz"

:::; fo z foz p(z- z')p(z- z") llnij(z')llllnij(z")ll dz'dz"

z 2
=[
1 0
p(z - z') dz'] . (4.32)

Now if the kernel function p(z) is given by Eqn. (4.18), then

[
1z p(z- z') dz']
2
=[ 1zL 00

Ck e-ak(z-z')dz']
2

0 0 k=l

(4.33)

Thus, from (4.32), (4.33) we obtain

(4.34)

This proves that Eqn. (4.28) represents a limit surface associated with an endochronic
model so that no stress state of the model can lie outside the limit surface. In other
words, the set of the stress points associated with different equilibrium states of an
endochronic model represents the limit surface of that model. The issues of existence
and uniqueness of equilibrium states and the associated limit surface will be addressed
in detail in the next chapter where the Distributed-Element Model is generalized to
a multi-dimensional representation.
 61

Another interesting point about the endochronic model is the resemblance of

Eqn. (4.28) to the von Mises yield criterion in the classical theory of plasticity, given
by
2 2
Sij Sij = 3 a0 , (4.35)

where ao is the yield stress of a material in the uniaxial tension test. Actually,
considering the uniaxial tension case where s 11 = 2a11 /3 = -2s22 = -2s 33 , and
Sij = 0, i =I j , Eqn. (4.28) implies

since we have (a11 )eq = a0 . Thus, we obtain the relation:

(4.36)

which relates the model constants to the material constant a 0 . Equation (4.36)
provides a guideline for choosing t he model constants ck and ak, in addition to
the constraints given in (4. 19). This relation motivates a new effective modeling
technique for endochronic models based on the formulation of the kernel function
given by Eqn. (4.18), as will be presented in the next section.

4. 3 A M ode ling Technique for the Endochro nic M ode ls

So far we have formulated a complete set of constitutive equations based on the

endochronic theory which is ready for numerical calculation of model response under
general cyclic loading conditions. To apply an endochronic model to simulation stud-
ies, however, one needs to choose proper values for those model constants involved in
the definition of the kernel function in the model, as given by Eqn. (4.18). Accord-
ing to the previous experience with modeling of the kernel function [17, 49], at least
three exponential terms in the series representation of p( z) are required to adequately
represent the smooth yielding behavior of hysteretic systems. T h e response behavior
of a one-dimensional endochronic model based on Eqn. (4. 18) is illustrated in Figure
4.1, where hysteresis loops of model response using one, two, or three exponential
terms are plotted together for comparison. It can be observed that the one-term
 62

formulation of the kernel function leads to elasto-perfectly-plastic behavior, and t he

three-term model gives a smooth yielding behavior as desired. In previous studies,
for example [17, 49], the model constants involved in the three-term kernel function
were chosen exclusively based on a trial-and-error procedure, which may be very
difficult and inefficient in practical situations. Furthermore, from a system identifi-
cation point of view, such a model involving at least eight model parameters (ck and
ak, k = 1, 2, 3, in addition to at least two other parameters of elasticity, E and v
say) would definitely violate the criteria of simplicity, physicality and robustness for
a good model, as stated in Chapter 2. Thus, it is very desirable that the modeling
procedure for the kernel function of an endochronic model be simplified. In the fol-
lowing, an efficient modeling technique is proposed to define the kernel function of an
endochronic model based on Eqn. (4.18) so that the number of parameters involved
can be reduced.
As mentioned above, in practice, a three-term representation for the kernel func-
tion p( z) is sufficient to yield a smooth yielding curve for general hysteretic response.
Thus, consistent with Eqn. (4.36) , we may introduce the following additional condi-
tions regarding the model constants:

vk = 1,2,3. (4.37)

The generality of the model behavior is not lost by introducing these conditions
because it is the ak that primarily control the shape of the hysteresis loops. In order
to satisfy approximately the first condition in (4.19), we can always choose a 1 to be
a very large number, say
(4.38)

Moreover, as suggested by results of numerical simulations, we may fix the ratio of

a 2 to a 3 so that a smooth yielding curve can always be attained, regardless of the
shape of the hysteresis loops. For this, we can set

(4.39)

as suggested by numerical simulations. Equations (4.37) to (4.39) provide five equa-

tions for the six constants Ck and ak , k = 1, 2, 3, in which only one model parameter
ku is required.
 63

To completely define the six constants, we need one more equation to determine
the absolute magnitude of the constant a3 (or a2). We note that since we defined
a 1 > a2 > a3 in our modeling based on Eqn. (4. 18), a 3 controls the final portion
of the yielding curve near the ultimate stress, as illustrated by the portion ABC in
Fig. 4.2. Thus, we can choose the value of a 3 to match appropriately that portion
in a uniaxial initial loading curve so that it is related to other model parameters. In
the one-dimensional case, using the three-term representation for the kernel function
based on Eqn. (4.18), we can derive the following equation for the n-th branch of the
response curve:

(4.40)

for z > Zn, where Zn corresponds to the n-th load reversal point. Thus, for initial
loading curve (n = 0) , we have

(4.41)

where we made use of Eqn. (4.36). For the final portion of the initial loading curve,
i.e. , a ---7 a 0 , t he variable z, which is a measure of accumulated plastic deformation,
will not be small and hence, by Eqn. (4.37),

(4.42)

where we assumed that e-o 1 z, e-o 2 z « e-o3 z, since a 1 and a 2 are considerably
larger than a 3 . Thus, by (4.41) and (4.42), we can derive

(4.43)

where we use the fact that dz = V3fi dt:P for the initial loading in the positive
direction of the uniaxial case. Thus, for a given yielding curve, we can find the value
of a 3 by matching some point, say B in Figure 4.2, in the final portion of t he curve.
 64

For example, if at point B , as= 0.95a0 and E~ = 1ao/E where 1 > 0, as shown in
Fig. 4.2, then by (4.43) we have

which yields
1.549 E
1 ao
For yielding curves of different shapes (but with the same E and a-0 ) , we may thus
define
1.549 E E
a3 = - - - = (2.Q)P-, (4.45)
1 ao ao
where p is an alternative parameter to 1 for controlling the degree of smoothness
of yielding. This is demonstrated in Figure 4.3, where we note that as p - oo, a
yielding curve corresponding to the elasto-perfectly-plastic behavior is obtained.

To summarize, we have the following result for modeling of the kernel function
using three exponential terms:

3
p(z) = L Ck e-akz, (4.46)
k=l

(4.47)

(4.48)

Thus, only three parameters, E, a-0 , and p, are needed in the modeling process
for general uniaxial hysteretic behavior, as before when the generalized Rayleigh
distribution function was used for the one-dimensional Masing models. In the case
where multi-axial response behavior of isotropic materials is of interest, only one
additional parameter, the Poisson's ratio v, is needed for t he endochronic models
based on the preceding formulation. This makes the modeling of endochronic models
much easier so that they become widely applicable to general multi-axial response
problems of cyclic plasticity, especially in the case where system identification is of
interest.
 65

Note that the parameter a 0 should have been defined as the ultimate stress (or
force) in the uniaxial (or one-dimensional) case where only cyclically stablized be-
havior is to be accounted for. This definition of a 0 may be viewed as an extension
of defining ao as the yield stress of simple tension, since for engineering applications,
structural systems do not, in general, exhibit hysteretic behavior with prominent
yielding point [23]. Another point to remark regarding the identifiability of param-
eters of this class of endochronic models is that under the circumstance where the
model response is not driven into a strong nonlinear regime, the two parameters a 0
and p may not be identified accurately due to their interactive effects on the sys-
tem response. As a consequence, in practical identification studies, we tend to fix
the value of the parameter p (or a 0 ) so that more reliable identification result can
be achieved. For most structural systems, the value of p can be set to be around
1.0 (which yields an endochronic model having uniaxial response behavior close to
that of a distributed-element model based on the Rayleigh distribution for the yield-
strength distribution function, as will be shown later in Fig. 4.4), so t hat only two
parameters E, ao are left in the one-dimensional models based on the endochronic
theory. These two parameters, E and a 0 , have clear physical significance since they
represent the initial stiffness and ultimate strength of a system in the context of the
generalized one-dimensional force-deflection behavior, or the Young's modulus and
the simple-tension yield stress in the context of general plasticity. The two-parameter
endochronic model (excluding Poisson's ratio) based on the three-term kernel func-
t ion formulation can thus be summarized as follows:

E
a2 = 6.0 - ,
ao
c2 = ~aoa2, (4.49)

a3
E
= 2.0-,
ao
C3 = ~aoa3.
It should be noted, however, that the numerical values defined in (4.49) for a2 and
a3 may vary slightly for particular applications in practice so as to reach the best
results.
 66

The main idea presented here is that Eqns. (4.46) to (4.48) set up a simple class of
models based on the endochronic theory that is generally applicable to plasticity prob-
lems. Even in the general multi-axial loading case, modeling of isotropic materials
based on such a class of endochronic models can be based only on the uniaxial initial
loading curve of the material being modeled, as long as the Poisson's ratio is given.
As a special case of the above general formulation, the one-dimensional hysteretic
behavior can be modeled with the tensorial quantities replaced by scalar quantites.
Thus, the uniaxial plastic strain response (or, equivalently, the drift response) using
the endochronic model is readily obtained through the recursive solution procedure
introduced above. A comparison of the drift response of an endochronic model to
that of a Masing model has already been shown in Figure 3.15. A simulated restoring-
force response of a one-dimensional endochronic model based on (4.49) (i.e. , p = 1)
and that of a matching Masing model subject to a prescribed cyclic displacement his-
tory are also compared, as shown in Fig. 4.4, where excellent agreement of response
behavior between the two models is observed. In the example, the two models were
chosen to have the same E and u 0 , and the Masing model was based on a Rayleigh
yield-strength distribution function. It can be noted, however, that the endochronic
model exhibits slightly different response characteristics from those of the Masing
model. The main differences are that for the endochronic model, the small cyclic
loops of transient response may not be "strictly" closed (i. e., the loops may not go
through the associated unloading points even though they are closed), and the geo-
metrical shape of the unloading or reloading branches is not the same as that of the
virgin loading curve in contrast to the Masing model. Nevertheless, the behavior of
the two differently-formulated models is essentially consistent as far as the overall
response is concerned.

4.4 Investigation of Cyclic Hardening Behavior

In the previous section, modeling based on the endochronic theory was mainly
conducted for cyclically stablized behavior of materials, for which the hardening
function in the formulation was taken as unity, i.e., f("l) ::::: 1. To model material
behavior including cyclic hardening (or softening) based on the endochronic theory,
one must appropriately define the hardening function f('fJ) , which should be positive,
 67

and a monotonically increasing function in the case of hardening, or a monotonically

decreasing function in the case of softening. The effect of the hardening function
f (TJ) is to stretch or compress the memory path defined through TJ in the plastic-
strain space, as can be deduced from Eqn. (4.10). Also, from Eqn. (4.17), we see that
if f(TJ) is a monotonically increasing function of ry, the corresponding yield surface
will expand accordingly, which signifies the hardening behavior of materials.
Some particular forms for the hardening function of an endochronic model have
been previously proposed. Valanis and Fan [49] proposed a form described by :

(4.50)

But the harding behavior of such a model in relation to the two parameters {31 and
fJ2 is not very clear. Another form of the hardening function has been proposed [43],
which is described by :

(4.51)

This function form involves only two parameters and exhibits appropriate and gen-
eral hardening behavior with f(O) = 1 and f(oo) = {31 . Physically, the parameter
{31 denotes the ratio of the two ultimate strengths of the model after and before
hardening occurs, and {32 accounts for the rate of hardening.
An important property of the cyclic hardening behavior, as observed from the
experimental result shown in Fig. 4.5(b), is that only the ultimate stress au is changed
during the cyclic process, while the Young's modulus E which governs the initial un-
loading or reloading slope of t he response curves almost stays invariant. Based on
this idea, we can extend the Masing models (or D EMs) to account for the cyclic
hardening behavior in a very effective way by utilizing the hardening function given
in (4.51). For example, if the special class of Masing models based on the Rayleigh
distribution function (cf. Sec. 3.3.3.3), which involves only two parameters E and
au, is to be extended to model cyclic hardening behavior, we need only make the
parameter au an appropriate function of some response quantity, such as the accu-
mulated plastic deformation, so that the ultimate strength au of the model changes
with cyclic response. Fig. 4.5(a) shows a simulated cyclic hardening behavior of a
 68

Masing model based on a Rayleigh yield-strength distribution function , in which the

parameter CJu is modeled as
Clu(17) = CJo J(17), (4.52)
where CJo is the initial strength of the model and !(77) is given by Eqn. (4.51). Note
that in Eqn. (4.52), 17 is defined as the accumulation of plastic deformation, i.e. ,

77 j ldcPI. (4.53)
Recall that the plastic deformation of a Masing model based on a Rayleigh yield-
strength distribution function can be found through Eqns. (3.39) and (3.40). In the
example, the material constants used for the model are E = 16, 700 ksi, CJo = 5 ksi,
and {31 = 4.4, !32 = 14.0. It is clearly demonstrated in Fig. 4.5 that the behavior of
the Masing model based on Eqn. (4.51) for cyclic hardening effect is almost identical
to the experimental result.
Another issue in modeling for cyclic hardening behavior is the effect of non-
proportional (or "out-of-phase") hardening exhibited by real materials. According
to experimental observations, the peak normal stress resulting from nonproportional
hardening is about 40 percent higher than that after uniaxial cycling [30, 43). This
phenomenon is physically complicated and, as a consequence, modeling of this behav-
ior is much more difficult considering the very limited experimental results currently
available. Sugiura, et al. [43) proposed a modified model based on the endochronic
theory that can adequately predict the nonproportional hardening behavior of mate-
rials by introduing a nonproportionality function, which depends in some empirical
way on the nonproportional plastic-strain response path. However, further explo-
ration on this topic is beyond the scope of the current study.

The purpose of this work is not to investigate extensive material behavior in

depth, but to get some physical insight regarding mathematical modeling of general
plasticity behavior of materials. Based on these insights, one can possibly extend the
Distributed-Element Models or the Masing models to account for various hardening
effects, as we did above for cyclic hardening behavior. Also motivated by the close
similarity between the behavior of the one-dimensional endochronic models and that
of the Masing models, generalization of the Distributed-Element Models and Masing
models to a multi-dimensional representation becomes of great interest. These are
to be investigated in the next two chapters.
 69

1.1

16

...
ll

0.1 :,. '

ll 0.6 ...: ,·
,.
~
0
[1.
~
OA

0.1
,.
j
·2 i
...
0
Ill
0.1

ll Q.4
~
Q.6

0.1

········· o ne - term model

1.1
- ·-·-·- two-term model
lA

- 1.6 - - three - te rm m odel

1.1

·l
. Q.

Displacement

Figure 4.1 Behavior of the endochronic model using the kernel function defined by
Eqn. (4. 18)

ao
0.95ao

f.p
B

Figure 4.2 A typical yielding curve for illustration of the proposed modeling
technique based on the endochronic theory
 70

J
:)

...:.> 2
0
u.
.:0
::
·c 0

.0
rJl
-I

-
,y
()
-2

.J
p=0.5
.... p=LO
p=2.0
-S

-6

Displacement:

Figure 4.3 Effect of the parameter p of the proposed endochronic model on

yielding behavior

12~------------~--------------~--------------r-------------~

()
u
...
0
u.
~
c
·c
...0
rJl
-2
()
~ A

-6

. - - endochronic
---····· Masing
- 10

- 12
-I
Displacement:

Figure 4.4 Response behavior of a one-dimensional endochronic model and a

matching Masing model
 71

30

2S

20

15

10
..IJJ.
"'"' 5
.!4
'-'
rtJ 0
rtJ

...g
Cl)
-5

-10

- 15

-20

-2S

-30
-1
Strain(% )

(a)

,. ~~~:L 0.25 "·

fig. 2 Recording of axial cyclic hardening (previously undeformed)

(b)

Figure 4.5 Uniaxial cyclic hardening behavior of (a) a Masing model

(b) a real material (from [30])
 72

CHAPTER 5
GENERALIZATION OF DISTRIBUTED-ELEMENT MODEL
TO MULTIPLE DIMENSIONS

5.1 Introduction

The one-dimensional Distributed-Element Models (DEMs), introduced by Iwan

[19] for structural dynamic analysis, have been investigated in detail in Chapter 3
for both deteriorating and non-deteriorating hysteretic behavior. In order to ex-
tend the one-dimensional DEMs to multiple dimensions for constitutive modeling of
general plasticity, Iwan [20] introduced the concept of a collection of nested yield
surfaces associated with a DEM, which move around in the stress space according
to some kinematic rules so that the Bauschinger effect could be accounted for in a
more realistic way. This class of multi-dimensional models for plasticity based on
the distributed-element formulation provided a conceptual generalization of the cus-
tomary formulation of the incremental theory of classical plasticity. However, the
numerical implementation of such a class of multi-dimensional models involves trac-
ing subsequent yield surfaces and hence is quite difficult and computationally ineffi-
cient. Yoder [56] proposed an alternative version of plasticity theory formulated in
the strain space. That t heory, based on a different class of DEMs from that used by
lwan, is closely parallel to the traditional theory of plasticity, but interchanges the
roles of stress and strain. In contrast to lwan's multi-dimensional model, the model
proposed by Yoder consists of a collection of yield surfaces formulated in the strain
space. However, the same problem pertaining to the numerical implementation of the
model behavior arises.
In the following sections, a new class of plasticity models, also based on the
distributed-element behavior, will be proposed, in which yield surfaces of different
yield levels are introduced for the elements involved in a model. The main idea be-
hind this new class of DEMs is t hat the yield surfaces are defined in the "element"
stress space and are "invariant" , i.e., fixed from moving in t hat space, no matter how
the model response varies. Due to the invariant characteristics of the yield surfaces
thus defined, the theoretical formulation of such models is so simple that there is
 73

no need for any kinematic hardening rules for subsequent yielding behavior. Fur-
thermore, the numerical implementation of the new model is straightforward and
highly efficient, even though quite a few elements are needed for the model to yield
good results in applications. What might be more interesting is that t he behav-
ior of this new class of DEMs provides us with a physical model for understanding
complicated response mechanisms in cyclic plasticity. Some experimentally-observed
material behavior can be adequately elucidated by t he model through the establish-
ment of some relevant properties of t he model behavior. The validity of t his new class
of Distributed-Element Models is confirmed by comparison with experimental results
from the literature. Excellent response predictions using the new models have been
obtained under complicated multi-axial loading conditions.

5.2 A New Class of Distributed- Element Models for Plasticity

5.2.1 Concept and Theoretical Background

Before looking into the generalization of the one-dimensional (1-D) DEMs to

higher dimensions, let us examine two different classes of 1-D DEMs that have been
used before. The two classes of models are composed of collections of elasto-perfectly-
plastic elements in either a parallel-series (P-S) or a series-parallel (S-P) combination,
as shown in Fig. 5.1(a) and (b), respectively. It can be shown [20] that when the
number of elements become very large so that the element strengths a; are described
in terms of some distributed function </>(a*), where </>(a*)da* denotes the fraction of
the total number of elements that have a slip stress between a* and a* + da*, then
the model behavior can be described by

(5.1)

for t he P-S model, or

E= 100

E(a*) ¢(a*) da* (5.2)

for the S-P model. It should be noted, however, that for the S-P model to be physically
consistent, the distribution function has to be singular at the origin , i.e. , </>(0) ~ oo,
so that dEj dait=O =/= 0; otherwise we will have da/dEit=O = oo, which means that the
initial slope of the stress-strain curve is infinitely large. Thus, from both physical and
 74

mathematical points of view, the S-P model is considered to be not so good as the
P-S model for which the distribution function ¢>(CJ*) can be any function that satisfies

100 ¢>(CJ*) dCJ* = 1, (5.3)

even though both models can be shown to exhibit Masing type of behavior [45].
In the following, we will extend the 1-D P-S model to t he general multi-dimen-
sional case so that they can be used for constitutive modeling in cyclic plasticity
problems. The generalization of the S-P model will not be done here, though it can
be treated in a similar way. More will be said about this later.
To account for the general multi-axial response behavior, we need to first define
the basic kinematic behavior of the distributed elements constituting the model. In
this study, we postulate the following rules for the new multi-dimensional DEM :
1) Each element in the model is subject to the same total-strain response as expe-
rienced by the model itself.
2) Each element has the response behavior of ideal plasticity so that its associated
yield surface remains "invariant" in the stress space. In other words, the yield
surface associated with an element is described by a function that depends only
on the element stress.
3) All the elements have the same elastic properties and the associated yield func-
tions have the same mathematical form , but they have different yield constants
which are governed by some distribution function.
4) The stress state of the model is defined as the average of the stress states of all
the elements.
Following these rules, the overall stress of the model can be expressed in terms of the
element stress states as follows :

a-(t) _ 1oo a(k, t) ¢>(k) dk, (5.4)

where a denotes the tensor CJij, and a(k, t) is the corresponding stress state of the
elements having yielding constant k governed by a distribution function ¢>(k). Note
the resemblance of Eqn. (5.4) to Eqn. (5.1). The constant k is related to the yield
function associated with ea.ch element in the model so that the equation

F(a(k), k) = o (5.5)
 75

represents a yield surface associated with an element of yield constant k in the element
stress space. Note that without loss of generality, we can choose

k = CJo(k), (5.6)

where CJo(k) is the uniaxial yield stress of the associated element. The definition of
the yield function defined in Eqn. (5.5) is conceptually the same as that used in the
classical theory of plasticity so as to characterize the general behavior of materials
under multi-axial loading conditions. However, in this new formulation, the yield
surfaces are defined in the element-stress space, not in the model-stress space as in
the classical theory of plasticity. Moreover, since each element in the model has the
behavior of ideal plasticity, the yield surfaces associated with the elements will remain
"invariant" in their space of definition , no matter how t he model behaves. Also, the
stress response of each element remains linearly elastic until yielding occurs, after
which the element stress state will move on the associated yield surface during plastic
flow and will never go beyond it. An important remark regarding the overall model
behavior is that the stress state of the model may possibly lie outside some of the
yield surfaces associated with the elements, which makes the new model distinctive
from those based on classical multi-yield-surface theory. It is also this formulation in
the "invariant-yield-surface" space that makes this new model mathematically simple,
physically realistic, and computationally effective, in contrast to the aforementioned
multi-dimensional DEMs proposed by lwan and Yoder (20, 56].
The theoretical background of this new formulation lies in the deduction that
corresponding to a yield surface in the stress space, there should be a yield surface in
the strain space regardless of what model is being considered. Consider, for example,
the case of ideal plasticity where a yield surfa~e formulated in the stress space always
stays invariant throughout t he deformation history. However, the corresponding yield
surface formulated in the strain space has to move around, along with the current
strain state of the model, so as to account adequately for the Bauschinger effect
exhibited by real materials under cyclic plastic deformations. This space-dependent
yielding behavior is illustrated in Fig. 5.2 for the one-dimensional (uniaxial) case.
With this concept in mind , the formulation of plasticity in either stress or strain
spa~e can be made equivalent to that in the other space, as long as a ppropriate
 76

kinematic behavior of the yield surfaces is taken into consideration. However, it is

obvious that the formulation without the need of kinematic hardening rules (such as
the stress-space formulation in the foregoing example) would be much easier than the
other, especially in the case where multiple yield surfaces are needed in obtaining
model response, as for the models based on distributed elements.

5.2.2 Mathematical Formulation

As mentioned above, the behavior of the new class of multi-dimensional DEMs

for plasticity is based on the element behavior formulated in stress space, instead
of in strain space, so that complicated kinematic hardening rules can be avoided for
response after initial yielding. Thus, the yield surfaces associated with the elements
stay invariant in the stress space under the assumption of ideal plasticity, and they
are nested within one another in the element stress space due to the different yield
strengths of the elements. This is illustrated in Fig. 5.3, where concentric circles of dif-
ferent radii represent yield surfaces of different yield strengths in the two-dimensional
(biaxial) case.
If the yield surface associated with an element with yield constant k is described
by Eqn. (5.5), then under the assumption of ideal plasticity, we have that when
F(a-(k), k) = 0, the plastic flow takes place without limit, and therefore,

(5.7)

for plastic flow. Then, from t he normality rule of plastic flow given by the classical
theory of plasticity, which specifies that the direction of a plastic strain increment is
normal to the yield surface at the current stress point, we have the flow rule for an
element of yield strength k:

(5.8)

where d>.k is a coefficient of proportionality, whose value can be determined as follows.

Firstly, we introduce the general stress-strain relation in incremental form as

(5.9)
 77

where Cijmn denotes the tensor of elasticity constants and it has been assumed that
all elements in the model have the same elasticity constants and identical total-strain
response so that the dependence of Cijmn and dtmn on k can be dropped. Based on
Eqns. (5.7), (5.8) and (5.9), we can find the expression for the coefficient of propor-
tionality as:
(5.10)

Summarizing from the above, we arrive at the following set of constitutive equa-
tions for the new Distributed-Element Model for general plasticity formulated in the
"invariant-yield-surface" space:

a= fooo a(k) <P(k) dk, (5.11a)

and F(a(k) , k) :::; o always. (5.11b)

If F(a(k), k) = o, (5.11c)
fJF
and dF = fJaij(k) daij(k) = 0, (never> 0) (5.11d)

fJF
then daij(k) = Cijmn(dtmn - OO"mn(k) d>.k), (5.11e)

where d>.k is given by Equation (5.10).

If Equation (5.11c) or (5.11d) is violated, then

(5.11!)

Throughout the above, all the derivatives involving F are to be evaluated at the
current value of a(k). Eqn. (5.11!) signifies that the instantaneous element response
will be linearly elastic if the element is not yielded (F(a(k), k) < 0) , or it is subject
to a condition of unloading (dF < 0).
Through the equations in (5.11), the model behavior is completely defined as
long as the mathematical forms of the two material functions , the yield-strength
distribution function <P(k) and the element yield function F(a(k), k), are specified.
The way to define the distribution function ¢(k) is similar to that used for the one-
dimensional DEMs, or Masing models, since the general multi-dimensional model
 78

should reduce to the one-dimensional case as the loading is restricted to be uniaxial.

Thus, similar to Eqn. (5.3), the distribution function satisfies

1= </>( k) dk = 1. (5.12)

Also, by Eqn. (5.11a), using V k a11 (k) = k and aij (k) = 0 if i # 1 or j # 1 (which
signifies that every element is in yielding state under the uniaxial loading condition),
we have
(5.13)

where au denotes the ult imate uniaxial stress of the model. Eqns. (5.12) and (5.13)
provide two conditions for the yield-strength distribution function </>(k) to satisfy. As
a consequence, </>(k) can be chosen as any probability density function t hat has the
mean value a u as a parameter. To this end, the Rayleigh distribution, defined as
2
-1f k )
k exp ( -
</>(k ) = 1f
-- 2
- 2
, (5.14)
2 au 4 au

serves as a good candidate for ¢( k), as already demonstrated in Chapter 3. Addi-

tional parameters may be incorporat ed in the definition of </>(k) so that more general
response behavior can be modeled. For example, if the degree of smoothness of yield-
ing behavior is to be modeled accurately, then the generalized Rayleigh distribution,
as given by Eqn. (3.35), can be used with the introduction of one additional parameter
n.
As mentioned earlier, the definition of the yield function associated with each
element is the same as that used in classical theory of plasticity. There have been
many yield criteria proposed in plasticity theory for various materials. Among them,
t he von Mises yield criterion described by

(5. 15)

is probably the most widely recognized criterion for modeling yielding behavior of ma-
terials due to its physical consistency and mathematical tractability. In Eqn. (5.15),
Sij denotes the deviatoric stress tensor defined as
 79

where bij is the Kronecker delta function. In the present study of constitutive mod-
eling, however, the yield function can be chosen appropriately for the material under
consideration based on any criterion used in plasticity theory or any empirical condi-
tion observed experimentally.
In addition to the two material functions discussed above, one needs to specify
the elastic constants Cijmn involved in the constitutive equations of the model. It is
noted that these constants, assumed to be identical for all elements, are essentially
the same as those of the model itself. This can be proven as follows. When the system
response is very small, the model and all its elements can be assumed to be in purely
elastic states, then, from (5.11a), we have

C7ij = 1
00

C7ij(k) </>(k) dk

= 1
00

Cijmn Emn(k) </>(k) dk

= Cijmn Emn 1 00

</>(k) dk

where we used Eqn. (5.11) and Emn(k) = Emn V k. This property makes the modeling
of this class of multi-dimensional DEMs very straightforward, since the elastic con-
stants associated with various materials have been well documented, or can be found
through simple experiments.
It should be pointed out that although we assume that all the elements in the
DEM are subject to the same total strain increment as experienced by the model
itself, i.e. , dEij(k) = dEij V k, the plastic strain response of the model is given by

as can be derived using Eqns. (5.4) and (5.9) , where dEfj(k) is to be found from
Eqns. (5.8) and (5.10).
In summary, this class of DEMs formulated in the "invariant-yield-surface" space
for cyclic plasticity involves only very few parameters that have clear physical signif-
icance. In the case where isotropic materials are of interest, only four parameters
are sufficient to represent general multi-axial elastic-plastic response behavior: two
 80

parameters (cru and n) are used for describing various shapes of hysteresis loops;
and another two (E and v) are for elastic behavior. As discussed in Chapter 3, the
Rayleigh distribution serves as a good model for the yield-strength distribution func-
tion ¢(k) in many engineering applications. In this case, only three parameters (n is
fixed to be 1) need be specified (or identified, in identification studies), which makes
the modeling process or t he identification procedure even simpler and more efficient.

5 .2.3 Numerical Implementation of the New Distributed-Element Model

The theoretical background and mathematical formulation of the new class of

multi-dimensional DEMs formulated in the invariant-yield-surface space have been
presented in the previous sections. In theory, the model may consist of an infinite
number of elements whose yield strengths distribute according to the specified dis-
tribution function ¢(k), and the model response is found by keeping track of all the
element behavior (cf. (5. 11)). However, to numerically implement the formulation,
one has to introduce a finite number of elements so that the solution algorithm is prac-
tically feasible. In order to preserve the advantages of this simple, physical model, it is
proposed that the introduction of the finite number of elements be made according to
the specified yield-strength distribution function ¢(k), so that the number of parame-
ters involved in t he model does not increase with the number of elements introduced.
In the case where the Distributed-Element Model consists of a finite number of, say
N, elements, the integral operation in Eqn. (5.11a) is replaced by the summation
operation as follows
N
o- = L a(ki) 'lj;(ki), (5. 16)
i=l

where the "weighting function" '1/;(ki) satisfies

N
I: 'lj;(ki) = 1, (5.17)
i=l

in place of Eqn. (5.11). Also, Eqn. (5.12) becomes

L ki 'lj;(ki) = CTu· (5.18)

i=l
 81

In order to obtain smooth response curves, one can choose, without loss of generality,

1
'1/;(ki) = N 'V i = 1, ... , N, (5.19)

and the yield constants ki, i = 1, ... , N are selected based on the specified distribution
function ¢(k), k E [0, oo), so that each time a new element yields, the model loses
1/ N of its initial stiffness. This can be done by dividing the region below the curve
described by the distribution function into N equal-area portions, and selecting ki as
a representative value for the i-th portion, so that Eqns. (5.18) and (5.19) are satisfied,
that is
N
L ki = N O"u· (5.20)
i= l

The aforementioned modeling procedure is illustrated schematically in Fig. 5.4. For

most applications, it suffices to use ten elements or so in representing the new model
in order to get a reasonably smooth hysteresis curve. For example, the yield constants
for the ten elements corresponding to a Rayleigh distribution can be defined as:

k1 = 0.2638, k2 = 0.4601, k3 = 0.6097, k4 = 0.7448, k5 = 0.8767,

(5.21)
k6 = 1.0128, k1 = 1.1612, k8 = 1.3347, k9 = 1.5630, k10 = 1.9732,

where we define ki- ki / D"u·

The numerical procedure for obtaining the stress response of anN-element model
based on the invariant-yield-surface formulation, subject to some prescribed strain
path, can be best described by a flow diagram as shown in Fig. 5.5, where we assume
that the strain increment in each loading step is small; otherwise, some subdivision
of 1::1€ is needed to assure that Eqn. (5.11b) is satisfied appropriately. As can be
seen from the flow diagram in Fig. 5.5, the numerical implementation of this new
multi-dimensional class of DEMs is surprisingly simple and computationally efficient,
due to the formulation in the invariant-yield-surface space, which avoids the usually
complicated kinematic hardening rule required for accounting for the Bauschinger
effect in cyclic plasticity.
 82

5.2.4 An Application to Biaxial Loading

A series of simulation studies on the response of the new model to some prescribed
strain paths are conducted to examine the model behavior in the biaxial tension-
torsion case, for which published work is readily available for comparison. Lamba and
Sidebottom [30] conducted a series of biaxial tension-torsion tests on copper in which
cyclic, nonproportional axial-torsional strain paths were applied to examine material
response behavior. The test samples used were thin-walled hollow cylindrical shafts
and were loaded with combined axial force and torsion. In the experimental studies,
the state of axial stress and shear stress resulting from the applied axial force and
torsion respectively was considered to be uniform in the test region at every t ime
instant. We have the following tensors of stress and strain in the biaxial loading case:

(5.22)

where the coefficient 1 represents the Poisson effect which is a variable when inelastic
deformations are involved in the response. It can be shown [14] that if we assume
incompressibility of plastic deformation, then we have the following expression for 1:
1 1 1 duu
I = -- -(-- v)-, (5.23)
2 E 2 dcu
where v is the Poisson's ratio for linear elasticity.
In the simulation studies, the model used consists of ten distributed elements,
and the Rayleigh distribution is used for describing the yield-strength distribution
in the formulation , so that Eqn. (5. 21 ) defines the yield constants of the elements.
The model parameters used are E = 16,700 ksi, v = 0.33, and u 0 = 30 ksi. The
prescribed strain loading paths are shown in Fig. 5.6, for which the corresponding
experimentally-observed stress responses are available [30], as shown in Figures 5. 7
and 5.8. Note that the loading path sequence in Fig. 5.6(a) is 0-1-0-2-0-3-0-... , so as
to study the property of erasure-of-memory. Also, the stress path resulting from the
repetition of path 0 each t ime is not plotted in Fig. 5.7(a) for clarity.
The stress responses predicted using the DEMs are shown respectively in Figures
5.9 and 5.10, where both von Mises' and Tresca's yield criteria were used in the simu-
lations for comparison purposes. In general, the results obtained in all cases are both
 83

qualitatively and quantitatively consistent with those observed experimentally, and

Tresca's yield condition gives slightly better results t han von Mises' does considering
the value of the ultimate shear stress predicted. Note that Fig. 5.9(a) contains t he
full stress path whereas Fig. 5.7(a) does not. It can be clearly seen in Figures 5.7 and
5.8, that there exists equilibrium points corresponding to those uni-directional strain
paths, at which the stress increments approach zero for appreciable strain increments
that remain uni-directional. In addition, it can be deduced that there exists a limit
surface in the stress space in each of the two loading cases so that stress states can
never go beyond it. Moreover, the property of erasure-of-memory is clearly demon-
strated by the DEM, as one can see that the model is always brought back to the
same state every time the path 0 in Fig. 5.6(a) is t raced.
Other important response features in cyclic plast icity, such as smooth yielding,
nonlinear strain hardening and multi-axial Bauschinger effect are also well demon-
strated by the new DEMs. The computational effort involved in ob taining t he re-
sponse based on the new model is low, since no kinematic hardening rule is required
to account for the subsequent yielding behavior of materials. We remark that models
based on the classical theory of plasticity in general do not predict response behavior
so well as the DEMs do, as we can see in Fig. 5.11, where different yield conditions
together with different kinematic hardening rules were employed to predict the re-
sponse to t he strain loading path given in Fig. 5.6(b) (30]. The dashed curve in each
plot of Fig. 5.11 represents the locus of the center of the yield surface, which is irrel-
evant to the discussion made here. A clear deficiency of the first two models, which
use respectively the von Mises yield condition with the Prager hardening rule and
the Tresca yield condition with the Ziegler hardening rule, t is that the axial stress
predicted does not return to zero. In more specific terms, these two models fail to
demonstrate the behavior of equilibrium points and a limit surface exhibited by real
materials. In Fig. 5.1l(c), a much more elaborated model is used , which employs a
Tresca yield surface, a Tresca limit surface, the Mroz kinematic hardening rule, and

t The Prager hardening rule specifies that the yield surface translates in the direc-
tion of the outward normal at the current stress point, while the Ziegler hardening
rule specifies that t he yield surface moves in t he same direction as the line joining the
center of the yield surface to the current stress point [30].
 84

an empirical nonlinear strain hardening assumption [30), so as to give the plasticity

model a maximum chance of success. Although the result shown in Fig. 5.11(c) is
much better than those shown in Fig. 5.ll(a) and (b), it is still not so good in accu-
racy as that predicted by the DEM, as justified by the exp erimental result shown in
Fig. 5.8(a) .
It should be noted that in the preceding examples, we did not use any system
identification technique to choose optimally the model parameters. Instead, the values
of the three parameters E , v , cr0 were specified directly from the corresponding exper-
imental results [30). This advantage is obviously due to the physical consistency and
the parsimony in parameters of the proposed DEM for cyclic plasticity. In the case
where complex structural systems are of interest, the parameters may b e optimally
identified using structural response data. Furthermore, we can treat some constants
in the yield condition required for the model as parameters to be identified, so that
the "best" result may be achieved in practice.
In the next section, we will address some important properties associated with the
new class of DEMs for general plasticity. Thorough understanding of these properties
helps to explain some material properties and complicated material behavior under
cyclic loading conditions. As a final remark, we note that the new Distributed-Element
Model for multi-dimensional plasticity can be viewed as a statistical mechanical model
which is a generalization of the classical formulation of plasticity theory. In the new
theory, the yield condition for elasto-plastic response characterization and the flow
rule for prescribing plastic strain increment are treated in a statistical sense, so that
the model response is the statistical average of the element response, each of which
follows from the classical theory of plasticity.

5 .3 Important Properties of the New Multi-dimensional DEMs

In the foregoing simulation studies, we have seen the existence of equilibrium

points and a limit surface associated with a new multi-dimensional DEM. Further-
more, the property of erasure-of-memory exhibited by real materials is also demon-
strated by the new model so that excellent results of response prediction have been
obtained when compared to experimental observations. It is of great interest to fur-
ther investigate these general properties of material behavior from a theoretical point
 85

of view, which will surely help the study of complicated response behavior of cyclic
plasticity. A thorough understanding of these properties may also provide useful
insight and guidelines for validating analytical models and for performing analyti-
cal/experimental studies in the related areas of plasticity.

5.3.1 General B ehavior of Ideal Plasticity of a Single Element

In view of the physically consistent behavior of the multi-dimensional DEM,

we would like to further study some relevant properties of the new model, which
consists of a collection of distributed elements of different yield strengths. Before
doing this, let us examine in detail the response characteristics of a single element.
Each element in the model exhibits the behavior of ideal plasticity, and so is governed
by the constitutive equations given in (5.1lb) to (5.11!). If we rewrite the incremental
stress-strain relation (5.9) as

(5.24)

then based upon the normality principle of plastic strain increment, we can derive for
a yielding state from (5.8) and (5.10):

(5.25)

where d>.. 2:: 0 is a constant of proport ionality and F is the yield function.
In order to characterize the complicated elasto-plastic behavior more easily, we
consider the biaxial tension-torsion loading case and employ the well-recognized von
Mises yield criterion. If we denote the axial stress and strain components as cr and
€, and the shear stress and strain components as T and 1, respectively, then we can
derive the following set of equations from the general formulation:

(5.26)

8F 8F
8cr = 2cr, 8T = 6T, (5.27)

dn = E(d€- d€P), dT = G(dl- diP), (5.28)

(5.29)
 86

where d>.. = 0 in the elastic state (F < 0), while in the yielding state (F = 0):
d>.. = (2a)Edc. + (6T)Gd--y
(5.30)
E(2a)2 + G(6T)2

An important remark regarding the biaxial formulation is that in the above we have
defined c. = en, a = au, 1' - 2c. 12 = 2c. 211 whereas T - a 12 - a2 1. It is more
convenient and consistent to adopt these definitions due to the fact that in the general
flow rule given in Eqn. (5.25), the yield function F is considered as a function of nine
stress components, counting a 12 and a 21 as separate variables. Note that the above
definitions imply that

(5.31)

SO that dc.P = (8Fj8a) d).. and d--yP = (8Fj8T) d>...

In order to understand the detailed behavior of an element after yielding occurs,
we first consider a simple case where a proportional strain loading path is prescribed,
as the path 0 - 1 shown in F ig. 5.12(a). The response of an element having perfect
plastic behavior to such a loading path can be depicted in a correponding stress space
as shown in Fig. 5.12(b). If the element yields at point P on path 0- 1, then the
corresponding stress state will just reach the yield surface at, say, P' in the stress
space. As the loading is continued, the stress state will move around on the yield
surface without going beyond it. But then an interesting question arises concerning
which direction the point P' will move under further loading. To answer this, we
have to look back at the constitutive equations which govern the response behavior.
Combining (5.28), (5.29) and (5.30) we get

12EGT
da = E(2a)2 + G(6T)2 (3Tdc. - ad--y),
(5.32)
-4EGa
dT = ( ) G( )2 (3Tdc.- ad--y).
E 2a 2 + 6T

Thus, given strain increments de. and d--y , we are able to determine the direction of
stress increments at a yield state by using (5.32), which leads to the following rules
for response behavior:
 87

(I) If 3Tde > crd"(:

dcr "' T, dT"' -CT, (5.33)

dcr = dT = 0, (5.34)

(III) If 3Tde < CTd"(:

dcr "' - T, dT "' CT, (5.35)

where ""'" means "proportional to". In Rule (II) we have the sit uation th at if the
ratio of de to d1 is kept fixed (e.g., in the case of proportional loading ) and the
stress state satisfies cr /3T = de/ d1, then the stress state will remain invariant unless
t he strain p ath changes its direction. In t his case, we say t hat the response state
reaches an "equilibrium state"* associated with that particular uni-directional strain
path. Mathematically, it can be shown that if the current state is governed by (I)
or (III), then it will approach a state described by (II). Thus , physically, the above
rules signify that a response state of a yielded element moves on the yield surface in
a direction toward an equilibuium state corresponding to the prescribed strain path.
Following the rules we can find that in Fig. 5.12(b), the response state at P' will
move in t he direction of P' Q' and will finally stop at the equilibrium state at Q' ,
as shown schematically in the figure. Similarly, we can determine the directions of
stress increments corresponding to the strain paths in different quadrants as shown
in Fig. 5.12(b), provided that the initial strain loading path remains uni-directional
in each of t he cases. It should be noted, however, that the direction of a stress
increment corresponding to a stress state on the yield surface may b e different from
those given in Fig. 5.12(b) in case of a strain loading p ath that is not virgin loading
or uni-directional. Nevertheless, the direction of a stress increment at a given yield
state can always be determined based on t he foregoing rules.
An important property regarding the element behavior is the existence of equi-
librium states associated with different uni-directional strain paths. The equilibrium
states can be defined as the st ates at which the stress increments a pproach zero
for a ppreciable strain increments that remain uni-directional. This proper ty of re-
sponse in cyclic plasticity has actually been observed in exp erimental studies of some

* A more rigorous definition for equilibrium states will be given later in this chapter.
 88

materials [30]. Each equilibrium state thus defined is associated with a particular
uni-directional strain path. The mathematical aspects of the existence and unique-
ness of an equilibrium state will be treated later on within a general formulation of
plasticity. However, we remark here that the existence and uniqueness of equilibrium
states leads to another important material property called the property of erasure-
of-memory. The property of erasure-of-memory is also deduced from experimental
observations [30] and can be stated as follows: If a material has been stabilized by
"out-of-phase" cycling (i.e., loading with non-proportional strain cycles) and if the
subsequent strain paths remain in the region enclosed by the out-of-phase cycling,
then one "big" strain cycle, which is sufficiently smooth and long so that t here exists
at least one equilibrium state associated with it, will always bring the material back to
the particular equilibrium state that is unique to that big strain cycle. This property
is very useful in conducting experiments on cyclic plasticity, since a single specimen
can always be brought back to the same reference state, and so can be used repeatedly
in charactizing material response to various loading paths. This ensures that more
reliable results can be obtained with considerably less labor and cost.
We remark that for the existence of erasure-of-memory, the out-of-phase stabi-
lization is a prerequisite conditiont since experimental results have shown that the
peak stress (or yield stress) resulting from out-of-phase hardening is about 40% higher
than that from uniaxial cycling, as already discussed in Chapter 4. Hence, if a mate-
rial has not yet been out-of-phase stabilized, its yield condition becomes variant and
depends on the non-proportionality of the loading path. This phenomenon cannot
be characterized by conventional plasticity models+ unless special treatment is made
[43].

t As shown later, t he property of erasure-of-memory is closely related to the exis-

tence and uniqueness of equilibrium states, which in turn depends upon the associated
yield condition.
t Most constitutive theories of cyclic plasticity are concerned with cyclically sta-
bilized behavior. One reason for this is that crack initiation predictions are generally
based on cyclically stabilized states [8].
 89

5.3.2 General Treatment of the Theory of Plasticity

In the following, a general treatment of the classical incremental theory of plas-

ticity is presented and some properties of the formulation are summarized. The
general formulation will then be used to derive important properties associated with
the new class of multi-dimensional DEMs
Let a - (O"ij) be the total stress tensor and f = (Eij) the total strain tensor 0

Define the elastic component and the plastic-relaxation component of the stress
increment tensor, dO-e and d(iP, respectively by:

(5.36)

(5.37)

where C[jkl is an elastic modulus tensor which is independent of response states. The
above definitions can be better understood through the schematic diagram shown in
Figure 5.13, where t he uniaxial stress-strain relation is considered.
Introduce a plastic modulus reduction tensor A ijkl so that

(5.38)

then following from the incremental stress-strain relation:

(5.39)

we can derive that

and
(5.40)

Note t hat the plastic modulus reduction tensor is, in general, a function of not only
the current state, but also the load increment.
The above equations can be put in a vector form as follows. Since O"ij and Eij

are symmetric second-order tensors, they can be written as vectors ~, £ E ~ defined 6

by
(5.41)
 90

(5.42)

so that the values of the inner products between tensors and between vectors are
preserved, and where the superscript T denotes matrix transpose, i.e.,

Thus, Eqn. (5.40) becomes

(5.43)

and Eqn. (5.39) can be rewritten as

(5.44)

where A and c e are the matrices corresponding to the fourth-order tensors A i jkl

and Cfikt so that the equations defined accordingly are consistent, and I is t he
6 X 6 identity matrix. The elastic modulus matrix c e is symmetric because of
the symmetries associated with the tensor Cfikt for elastic behavior. Also, it is a
constant matrix under the assumption t hat the elastic behavior of the material is
linear . It is positive definite if the material is stable to small strain perturbations
(or if a Drucker's postulate holds), which we assume is the case. Equation (5.44)
can be reformulated as

(5.45)

where CP c eA can be referred to as t he plastic modulus matrix.

Equation (5.44) (or (5.45)) gives the general formulation of the basic constitutive
law that we will use in the following, where we establish some theorems pertaining to
the properties of the formulation. With these theorems, some important properties
associated with t he new class of multi-dimensional DEMs can then be derived and
presented in a more efficient way.
 91

[Theorem 1] : Within the classical formulation of plasticity*, the plastic modulus

reduction matrix A and the plastic modulus matrix C P are both of rank one or zero,
corresponding to yielding and elastic behavior respectively.

[Proof]: If the classical theory of plasticity is considered, the yield function can be
assumed to be of the form :

(5.46)

where both isotropic and kinematic hardening are taken into account. In the case
of ideal plasticity, a = Q and k = constant throughout the response history. Based
on the "associated flow rule," which states that the yield function is the same as the
plastic potential function which defines the directions of the plastic strain increments,
we have
oF
dcip1. =d).~, d'"'~ 0. (5.47)
UC7ij

During plastic deformation we require F = 0, and so dF = 0, i.e.,

oF oF p _
~ duij + £l P de - 0. (5.48)
UC7ij uf.ij

Define
(5.49)

where the vector gradient of a scalar function is defined by

oF oF oF T
V'~F(1f,~, ... ) = ( ~'~· ... '!:I") (5.50)
UU! UU2 UUn

if 1f =< u 1 , u 2 , .. . , Un >T . Ftom Eqns. (5.47), (5.48) and (5.49), and the incremental
stress-strain relation
(5.51)

* By classical formulation of plasticity we mean that the elastic-plastic response

behavior is characterized by a yield condition, a flow rule, and a strain hardening
rule. The flow rule relates the increment of plastic strain to the current state and
the stress increment. The strain hardening rule specifies how the yield surface is
changed during plastic flow.
 92

it follows that
d).. = { 0: QT ds_ if QT ds_ > 0;
(5.52)
0 if QT ds_ ~ 0,

where Q- CeQ and 0: = 1/(fl?Q + QTg), which is just a scalar, and hence

(5.53)

Therefore, by comparing (5.53) with (5.44), and using the fact that c e is invertible,
we find
A= { O:QQT if Q~d£ > 0; (5.54)
0 if Q d£ ~ 0.
Thus, we may conclude that A is a 6 x 6 matrix of rank one (only one independent
row or column ) or zero. Also since

(5.55)

CP is of rank one or zero, too. Since yielding is equivalent to a non-zero plastic

strain increment ds_P, from Eqn.(5.43) we see that A and CP are both of rank one
during yielding and are both zero during elastic behavior. We remark that even if a
non-associated flow rule is used, the same conclusion can still be made.

The practical significance of Theorem 1 in the theory of plasticity may be stated

as another theorem as follows .

[Theorem 2] : Within the context of classical plasticity, the incremental plastic

deformation, if it exists, only occurs in a one-dimensional subspace (which changes
with the current yielding stress state) of the six-dimensional space of the total strain
increment.

[Proof]: \Ve know from Theorem 1 that the plastic modulus reduction matrix A is
a 6 x 6 matrix of rank 1 during plastic flow. Thus, there exists {~i : i = 1, 2, ... , 5}
forming a basis for the 5-dimensional null space of A , i.e.,

A -x.l =0 , "i/ i = 1, 2, ... , 5. (5.56)

 93

Equations (5.43), (5.44) and (5.56) imply that there always exists five linearly-
independent strain increment vectors ds_i = dJ.Li~i' dJ.Li > 0, i = 1, . .. , 5, such that

da
-l
· = c ede~l. and dc1?
.!:.l
= A de.!:.l. = 0 ' '-y-' ,;~ -- 1 ' 2 ' • •• '
5' (5.57)

and the corresponding plastic-relaxation stress increment satisfies

(5.58)

Thus, purely elastic behavior always occurs in, at least , a 5-dimensional subspace of
the six-dimensional space of the total strain increment, and then the conclusion of
Theorem 2 follows.

We remark that in the case of ideal plasticity, the plastic modulus reduction
matrix takes the form (cf. Eqn. (5.54) with d = Q. in a)

a bT
A (a, E, da , dE )= - T- (5.59)
-- - - Q Q

during plastic deformation, where Q, fl. are defined as before. Thus, as a result
of Theorem 1, the eigenvalues Ai, i = 1, . .. , 5, of A are zero, and the remaining
eigenvalue ,\6 must be 1, since

6 ~6
'
"'6 = "'""' '
L......t "'i = T r (A) = w i=l
T aibi = 1, (5.60)
i =l QQ

where Tr(·) denotes the trace of a square matrix. Let the corresponding eigenvectors
be ds_i, i = 1, ... , 6, where ds_i, i = 1, ... , 5, give purely elastic behavior, as in the
proof of Theorem 2, then,

which implies
dE.B = d~ , or ds.6 = Q.. (5.61)

Thus, we get purely plastic strain increment in a one-dimensional subspace. In the

case where strain hardening is taken into account, however, dE.B will not, in general,
be fully plastic, since ..\6 may be less than one, as can be deduced from Eqn. (5.54)
and Theorem 1.
 94

[Theorem 3]: If an associated flow rule is used in the formulation based on the
classical theory of plasticity, the plastic modulus matrix CP is symmetric. Also, in
general, Drucker's postulates of stability imply that C P is positive semi-definite.

[Proof): From Eqns. (5.54) and (5. 55) we get

since Q - ce~. Thus, we have

i.e., CP is a symmetric matrix. Drucker's postulate requires that

(5.62)

so by Eqns. (5.43) and (5.44), we have

===} d.£T(l - A )C eA d.£~ 0

===} d.£r(CP- A T c eA ) d£ ~ 0

===} d.£T C Pd.£ - d.£T A Tee A d.£~ 0.

Since c e is posit ive definite, we may conclude that

(5.63)

i.e., CP is posit ive semi-definite. Note that actually (5.63) does not hold for all d.£,
only Q.T d.£ > 0, but C P is itself zero otherwise.

[Theorem 4): The plastic modulus reduction matrix A , formulated based on an

associated flow rule, has, at most , one nonzero positive eigenvalue whose value is
never greater than one.

[Proof]: Define S =(c e) - l/2, which is a symmetric, positive-definite matrix, since

c e is symmetric and positive definite. Now consider the eigenvalue problem:
 95

Following from Theorem 3 which claims that C P is symmetric, positive semi-definite

if an associated flow rule is used, we have .A ~ 0. Let 11.. = S .;£ then

{::=::} scp'J!.. = .AS- 1 JL

{::=::} (Ce)- 1 CPJ! = AJ! (since 8 2 = (Ce)- 1 ) ,

I.e. ,

(since CP = c eA ).
Hence, the eigenvalues of A are non-negative. Furthermore, by Drucker's postulate:

we can derive
ds_T[Ce(l - A) ds_J ~ 0,

==} ds_T (c e - C P)ds. ~ 0,

which implies that the symmetric matrix c e- C P is positive semi-definite. Thus,

det(I- A) ~ 0,

where det( ·) designates the determinant of a matrix. Since the determinant is a

product of all the eigenvalues of the matrix and Theorem 1 implies that A must
have 5 zero eigenvalues, it follows from the above results that there is at most one
nonzero eigenvalue A6 such that

We remark that as in the proof leading to Eqn. (5.61), the case A6 = 1 corresponds
to the behavior of ideal plasticity.
With the general formulation introduced above, we are now in a position to
derive some important properties of the multi-dimensional DEMs. To begin with,
we introduce the following theorem regarding the plastic behavior of the DEMs.
 96

[Theorem 5]: For aDEM, the plastic modulus matrix C P is symmetric and positive
semi-definite, and the eigenvalues of the plastic modulus reduction matrix A are all
non-negative. Besides,
0 ~ Tr(A) ~ 1. (5.64)

[Proof]: In the formulation of the DEM with a finite number of elements, the plastic
strain increment of the model is given by

N
dcP
- = "
0 ' n/o. dc'f!
'f/1. - f., (5.65)
i=l

which can be easily derived from the basic assumptions of the kinematic behavior
of the DEM and the incremental stress-strain relation. By Eqn. (5.43) and the
definition of the plastic modulus matrix, we have

N
A (Q., d~) = L 1/Ji A i (Q.i' d~) (5 .66a)
i=l

and
N

CP(Q., d~) =c eA(Q., d~) L 1/Ji Cf(Q.i, d~) .

=
i =l
(5 .66b)

where A i and Cf are matrices associated with each of the elements in t he DEM.
From the result of Theorem 3, it follows that Cf, \;/ i = 1, ... , N , are symmetric
and positive semi-definite, and so therefore is C P. Also, as in the proof of Theorem
4, we can show tha t all the eigenvalues of A are non-negative. In addition, since
each element in the DEM has the behavior of ideal plasticity, we have Tr (A i) = 0
or 1, depending on whether the element state is elastic or yielded, and then by
Eqn. (5.66a)
N N
0 < Tr (A ) L 1/Ji Tr (A i) < I: 1/Ji 1.
i=l i=l

In order to develop a mathematically rigorous theory on the properties of t he

DEM, we make the following definition regarding the "state of equilibrium":
 97

[Definition 1]: An "equilibrium point (state)" is a response state associated with

a uni-directional strain path ds_ = ~ dt, with ~ =1- 0 and dt > 0, at which

{5.67)

1.e.,
[I - A (Q.,f,Q,~)]~ = 0, (5.68)

where the dependence of A on dt is dropped since we are mainly concerned with

rate-independent plasticity in this research.

The uni-directional strain path that defines an equilibrium point is referred

to as the reference path associated with that equilibrium point. From (5.68), the
reference path is an eigenvector with eigenvalue unity corresponding to A evaluated
at the equilibrium state. The term "equilibrium point (state)" was used because if
we consider the following system

(5.69)

(5.70)

is an equilibrium point of the system described by the ordinary differential equation

(5.69).
One important property associated with equilibrium states follows from the
definition and can be stated as follows.

(Theorem 6]: At an equilibrium state for aDEM (or a classical plasticity model),
the plastic modulus reduction matrix A is of rank one, and the only nonzero eigen-
value of A has a value of one.

[Proof] : By Eqn. (5.68), we know that the matrix I - A (Q., f, Q, ~) must be singular at
an equilibrium state associated with the reference path~. so that A has an eigenvalue
of 1. As a consequence, we may conclude that the matrix A(Q.iq,f:.iq,Q,~) must have
5 zero eigenvalues and an eigenvalue of 1 (Theorem 5), regardless of what ~ might
be. This concludes the proof.
 98

The physical significance of this property is that strain hardening effect must vanish
as an equilibrium state is approached. This can be deduced from the result of
Theorem 4. Furthermore, Theorem 6 implies the following corollary.

[Corollary): Purely plastic deformation at a state of aDEM occurs only in a one-

dimensional subspace of the 6-dirnensional strain-increment space if and only if the
state is an equilibrium point.

This property of the DEM is different from those of the models based on the classical
theory of plasticity, by which purely plastic deformation always occurs in a one-
dimensional subspace due to the use of the principle of normality.
We remark that at an equilibrium state (for any plasticity model) corresponding
to some reference path f, the work done over any strain loading increment d£ = f dt
must be zero, since dq_ is identically zero. This remark, which seems trivial, turns
out to be useful later in deriving the properties associated with the DEMs.
It should be noted that although an equilibrium point is defined to be associated
with a uni-directional strain path, a strain cycle which is sufficient smooth and long
may also have particular equilibrium points associated with itself, as mentioned
earlier in the description of the property of erasure-of-memory. This situation is
illustrated in Figures 5.14(a) and (b) , where the ellipse in the E- 'Y strain space
denotes the prescribed strain cycle with discrete increments, and the corresponding
stress response calculated for a DEM shows two equilibrium points on the ellipse in
the CJ- T stress space with the densest stress increments around them. A big smooth
strain cycle is needed to get the stress response on the limit surface, then whenever
a strain increment matches the direction of the normal to the limit surface at the
current stress point, an equilibrium state is reached.
Two important issues pertaining to the equilibrium points are the existence and
uniqueness of an equilibrium point given a specified reference path. Mathematically,
it is difficult to show directly the existence of an equilibrium state considering the
rather complicated formulation of plasticity involved. Instead, we use some simple
energy arguments to solve the problem, as presented in the following theorem.
 99

[Theorem 7] : For stable materials which have bounded elastic strain energy, given
a specified uni-directional strain path, a corresponding equilibrium point always
exists.

[Proof]: Define the elastic strain energy of a system as

1 e ce
W e =_ 2€ij e 1 ( e)TC e e
ijkl Ekt = 2 £ £ ·

By assumption, we is b ounded so that the elastic strain response Eij is also bounded.
Along a uni-directional strain loading path £ = _g t f Q, where, by definition t is
monotonically increasing (dt > 0) , the elastic strain energy would never decrease
after a certain state at, say, t = k Thus, it requires for all t > ti

(5.71)

For bounded elastic strain energy we must have dWe(t) - 0, as t - t 0 , where t 0

corresponds to some st ate at which dWe(t) = 0, possibly to = oo. By (5.71), we
must have*

i. e.,
daij(to) = Cfjkt dEkt(to) = 0, V i,j.

Thus, following from the definition of an equilibrium point, as in Eqn. (5.67), exis-
tence of the equilibrium state associated with a reference path £ = ~ dt =/= Qis always
guaranteed .

The issue of uniqueness of an equilibrium point associated with a reference path

will be discussed later after we h ave introduced the concept of the limit surface and
its associated properties.

5.3.3 Geometrical Considerations of Yield Surfaces for the New DEM

In defining the kinematics of an element in the DEM, we introduced a yield

condition for characterizing t he general multi-axial elasto-plastic behavior of the

* We can rule out the possibilty that d£e becomes orthogonal to c e£e, since in
(5. 71 ), we have neglected the high-order terms, which never vanish unless d£e = 0.
 100

element. The yield condition has been defined in the same sense as done in the
classical theory of plasticity. In other words, the yield condition for a given material
is essentially the extension of a single yield point in the uniaxial (or one-dimensional)
case to a hypersurface in the six-dimensional stress space (considering the symmetry
of stress tensors). Since the DEM consists only of ideal plasticity elements, we may
concentrate on the corresponding formulation, so that a yield condition is simply
described by
F(aij(k), k) = 0, (5.72)
where k represents a yield constant corresponding to some particular element. For
isotropic materials, since rotating the axes does not affect the yielding behavior, we
can choose the principal stress axes for defining t he coordinate system so that Eqn.
(5.72) may be rewritten as

F(at (k), a2(k), a3(k), k) = 0. (5. 73)

In the (ai. a2, a3) coordinate system, which represents a stress space sometimes
referred to as the Haigh-Westergaard stress space, Eqn. ( 5. 73) specifies a normal
three-dimensional surface that one can easily picture. As discussed in many text-
books of plasticity, e.g., [26, 34], the yield locus that a yield surface intersects with
the 1r plane, a plane passing through the origin and perpendicular to the hydrostatic
axis for which a 1 = a2 = a 3, must be symmetric in the principal stresses. If one
further assumes equal yielding in tension and compression, then the yield locus can
be divided into twelve symmetric sectors, each of 30 degrees. In the following, we
will formulate some important properties related to the yield surfaces of a DEM
based on some basic principles in operator theory.
Recall that we defined a DEM as consisting of a collection of elasto-perfectly-
plastic elements whose yield surfaces are nested within one another and are governed
by yield functions of the same mathematical form so that the yield surfaces may have
similar shapes. To make it clearer, we introduce the following definition.

(Definition 2]: Two hypersurfaces S 1 : F(q_, k 1 ) = 0 and S2 : F (q_, k2) = 0 are said
to be similar (in shape) with dimension ratio kifk 2 , if any ray from the origin that
passes through S 1 at q_ 1 intersects S2 at q_2 such that
 101

Mathematically, if the dimension ratio of two similar surfaces S 2 and S 1 is c > 0,

then by definition we have

F(Qo, ko) = 0 <==> F(CQo, cko) = 0. (5.74)

Thus, we have the following theorem regarding the condition for similar surfaces.

[Theorem 8]: A set of yield surfaces S defined by S = {o- : F(Q., ck0 ) = 0, c > 0}
are all similar with dimensions proportional to c, if the yield function F( ·, ·) is
homogeneous (of any order).

[Proof]: IfF(-,·) is homogeneous of some order, say m, and

F(Qo , ko) = 0, (i)

then
F(CQo, cko) = c= F(Qo, ko) = 0 Vc > 0. (ii)

By (5.74) we may conclude that all surfaces are similar with dimensions proportional
to c.
Based on the above result we now assume that the yield function used to define
the yield surfaces of a DEM is homogeneous so that the nested yield surfaces are
all similar in shape with dimensions proportional to the yield constants k' s. Thus,
the domain of elasticity, ni, in the element stress space for an element with yield
constant ki, defined by
(5.75)

can be expressed as
k0 = 1, (5.76)

or
(5.77)

where flo is the domain of elasticity of some element with yield constant ko = 1, and
it is assumed to be a bounded, convex set. The boundedness follows from the fact
that any real material has finite ultimate strength (peak stress) , and the convexity
 102

follows from the well-known result that a yield surface is convex if Drucker's postu-
lates hold [12, 34]. Since the model response of aDEM can be written as (using the
formulation of a finite number of elements):

N
Q. = ~ 1/JiQ.i, (5.78)
i=l

where N is the total number of elements and

N

I: '1/Ji = 1, '1/Ji ;::: o, (5.79)

i= l

by operator theory on convex sets [27], the set of all model stress points, denoted as
the domain n, is given by

N N N
n= L:('I/Ji n i) = ~('1/Ji kino)= ( ~ '1/Ji ki)no (since no is convex and ki '1/Ji > o)
i=l i=l i=l
N
= kuno (ku =L
i=l
'I/Jiki)- (5.80)

In the derivation we used some fundamental theorems in operator theory on convex

sets. The relevant theory is summarized in Appendix A. By (5.80), the existence
of n is guaranteed as long as ku < oo in the case where N ---+ oo, which may
again be thought of as a condition of finite ultimate strength for any real materials.
FUrthermore, we may conclude that n is similar in shape to no. Thus, the boundary
of n, an, defines a limit surface of aDEM, which can be described by

F(Q., ku) = 0, (5.81)

such that a model stress state can never go beyond the limit surface associated with
the model. This proves the following theorem specifying an important property of
the DEM:

[Theorem 9] : There exists a limit surface associated with a DEM, described by

F(Q., k,J = 0,
 103

where ku = E~ 1 '1/Ji ki, and ki, i = 1, ... , N, are the yield constants of theN elements
constituting the model. The limit surface is similar in shape to the yield surfaces
associated with each of the distributed elements.

In the following , we will derive some important properties related to the equi-
librium points and the limit surface associated with a DEM. First of all, we note
that, from Definition 1 and Theorem 6, at an equilibrium state corresponding to a
reference path d£. = f dt, the plastic-strain response increment will be the same as
the prescribed strain increment, i.e.,

Thus, we have the following theorem pertaining to the equilibrium states of aDEM.

[Theorem 10] : At an equilibrium state of aDEM, all elements in the model are
in corresponding equilibrium states, which lie on the associated yield surfaces at
points having the same outward normal direction as the reference strain path, and
conversely (all elements in equilibrium states implies DEM in equilibrium state).

[Proof): Converse is trivial since if each element is in an equilibrium state, we have

Vi dg:_i = 0 for d§. = fdt , then dg:_ = E~ 1 '1/Ji dg:_i = 0.
Now, if a DEM is in an equilibrium state corresponding to a reference path
f, then the work done over any strain loading increment df = f dt must be zero,
as we remarked earlier. Since the DEM actually consists of an assemblage of ideal
plasticity elements that are subject to the same total strain increment, the sum of
the work done by all the elements must also vanish. Thus, since the incremental
work done by each individual element is non-negative (Drucker's postulate) , we may
conclude that
dg:_f df = 0 Vi= 1, ... , N. (5.82)

Also, by the assumption of ideal plasticity for each element, we have [34)

oo'!'
-t
del!=
-t 0 Vi= 1, .. . , N. (5.83)

It follows from (5.82) and (5.83) that

 104

Since c e is positive definite, we must have d£f = Q Vi, and hence dq_i = Q Vi, which
shows that each element is in an equilibrium state corresponding to reference path
_g. Furthermore, each d£f = _g dt, so by the principle of normality for each element,
the outward normal at each element's equilibrium point is in the direction of _g.

[Theorem 11]: If all the element stress states of aDEM lie on the associated yield
surfaces and line up in the stress space on a ray from the origin, then the stress state
of the DEM is on the associated limit surface.

[Proof]: Firstly, we note that each yield surface associated with an element is t he
boundary, ani, of the domain of elasticity ni of that element, i.e. ,

(5.84)

From Eqn. (5.76), we have

(5.85)

that is, each yield surface is described by

(5.86)

Also, from Theorem 9, the limit surface is the set given by

80. = {kuQ{}: Q.o E 80.o}. (5.87)

Thus, if all the element stress states of a DEM lie on the associated yield surfaces
and line up in the stress space on a ray from the origin, then the element stress
states must be proportional to one another with proportionality constants of yield
strengths, i.e.,
CI ·
- l
= k•Cio
·-'
(for some Q{) E 80.o)

and hence we have

N N
Q_ - "L VJi Q_i = "L VJi (ki Q{))
i=l i=l
N
= (L ki VJi)Q{) = ku Q{)·
i=l
 105

From Eqn. (5.87), the conclusion of the theorem follows.

It is of great importance to note that the limit surface to aDEM is like the yield
surface to a model of ideal plasticity as far as the plastic behavior is concerned. This
can be deduced from the following important theorem which relates the equilibrium
points to the limit surface of aDEM.

[Theorem 12): If the admissible stress region bounded by the limit surface is
convext, then the limit surface associated with aDEM is the set of all the equilibrium
points corresponding to all possible reference paths.

(Proof): It is equivalent to showing that a stress state of a DEM is an equilibrium

state if and only if it lies on the limit surface, which is convex, of the model.
Sufficiency: If Q. is a stress state of aDEM on the limit surface, then from Eqn. (5.87)

Q. = ku Q.o (for some Q.o E ono). (i)

Also, by definition, we have

N N
Q. = L 'I/Ji!Z.i = L 'I/Jiki~i) (~i) E no Vi= l , .. . ,N). (ii)
i=I i=I

If we rotate the coordinate axes so that the XI axis in the stress space is perpendicular
to the tangent plane to the yield surface on0 at the point Q.o, as shown in Fig. 5.15,
and define the XI coordinate of Q.o, (Q.o)I, to be a, a> 0, then from (i)

N
(Q.)I = ku (Q.o h = Q ku = Q L ki '1/Ji · (iii)
i=I

Since each yield surface is convex and so the region no lies completely on one side
of any tangent plane of ano, we can deduce

(iv)

t This is equivalent to the earlier assumption that no is convex, which is actually

a consequence of Drucker's postulates.
 106

It follows from (ii), (iii), (iv) that

i.e. , Qj,i), Vi = 1, ... , N, lie on the tangent plane x 1 = a. Thus, it follows from the
shape similarity of the yield surfaces that all the element stress states are on the
associated yield surfaces at the points having the same outward normal direction
(perpendicular to the tangent plane). By the principle of normality for each ele-
ment, the corresponding plastic strain increments of elements are all in the same
direction, say f , and so is the plastic strain increment of the DEM at Q. (following
from Eqn. (5.65)). This shows that the principle of normality holds for a state of
the DEM on the limit surface. Now if the total strain increment prescribed is in
the direction of f, then the elastic strain increment at Q. must be zero under loading
condition (otherwise, the stress increment will point outward so that the stress state
goes beyond the limit surface), and so therefore is the stress increment. Thus, by
definition, the state Q. must be an equilibrium point associated with the reference
path f.
Necessity: If Q. is an equilibrium point corresponding to a reference path f, then
according to Theorem 10, every element state must lie on its associated yield surface
at the point corresponding to the outward normal direction f . Note that, however,
without the assumption of strict convexity+ of t he yield surfaces, we cannot conclude
that all element stress states are on a line from the origin (so that by Theorem 11 ,
the model stress state is on the limit surface). Nevertheless, we can still argue as
follows. Let~ denote the subset on a yield surface of contant ki, in which all points
correspond to the same outward normal direction, i.e. , ~ lies on a hyperplane in
the stress space, which may be described by a linear function in Q.i, so

6
~ = {fli: F(Q.i) - l:=aj(Q.i)j = ki and F(Q.i,ki) = 0}, (5.88)
j=l

t A region D and its boundary 8!1 are said to be strictly convex, if 8!1 is convex
and there are no two points on 8!1 that have the same outward normal direction.
 107

where j denotes the j-th component of a vector, so that the vector gradient '\1!!...F is
a constant vector throughout t he region ~. T hus, it follows from Theorem 9 that
the subset Ron the limit surface, corresponding to~, can be described by

6
R = {f[: F(a) = L aj(f[)j = ku and F(f[, ku) = 0}. (5.89)
j=l

From Eqn. (5.78), it follows that

6 6 N N 6
L: a;(f[); = L ai[ L V;i(f[i)i] = L V;i[ L a;(f[i)i] · (5.90)
j=l j=l i=l i=l j=l

Now if f[i E ~, t hen by Eqns. (5.88) and (5.90)

6 N
L aj (f[) j = L V;i ki = ku.
j=l i=l

Following from (5.89), we conclude that f[ lies in R, which is on the limit surface.

[Corollary]: T he principle of normality holds for any state of a DEM on the limit
surface.

The existence of equilibrium points has been assured by employing the concept
of bounded elastic strain energy. Now we are in a position to address the p roblem
of uniqueness of an equilibrium point associated with a reference strain path. T his
is given as the following theorem.

[Theorem 13]: An equilibrium point assocated with a reference strain path is

uniquely defined (regardless of past response history) if and only if the admissible
stress region bounded by the limit surface is strictly convex

The proof of Theorem 13 can be done simply by considering the schematic diagram
shown in Figure 5.16, where the yield surfaces (and the associated limit surface) are
not strictly convex. Given a uni-directional strain path following different previous
histories, we may end up with different equilibrium points as points 1 and 2 shown
in the figure. If, instead, the admissible stress region is strictly convex, then corre-
sponding to a reference strain path, there is only one point on the limit surface that
 108

has the outward normal in that direction. Thus, following the flow rule based on
the principle of normality, the equilibrium point is uniquely defined.

With the theorems presented above, we may now investigate in detail the prop-
erty of erasure-of-memory that is exhibited by real materials (30]. It may b e deduced
that the existence and uniqueness of equilibrium points associated with different ref-
erence paths are the necessary and sufficient conditions for a DEM to exhibit the
property of erasure-of-memory, since then every time a "big" smooth strain cycle
is prescribed, the system will be brought back to the particular equilibrium states
associated with that strain cycle, regardless of what the previous history is. This
leads to the following important theorem.

(Theorem 14]: A DEM possesses the property of erasure-of-memory if and only

if its admissible stress region bounded by the associated limit surface is bounded
and strictly convex, from which the existence and uniqueness of equilibrium points
follow.

In summary, if the yield functions used in the definition of aDEM is homoge-

neous and strictly quasi-convex* so that the limit surface exists and forms a strictly
convex region, then the DEM can exhibit the property of erasure-of-memory. Ac-
tually, as can b e deduced, the conditions stated in Theorem 14 serve as the general
criteria for any plasticity model to demonstate the property of erasure-of-memory
that real materials have. FUrthermore, establishment of the above theorems provides
us with clear insight into the elastic-plastic response mechanisms of real materials
under complicated cyclic loading conditions, which surely helps further studies on
the related subjects of general plasticity.

* Mathematically, it can be shown [12] t hat if a yield function is strictly quasi-

convex, then the associated yield surface forms a strictly convex region. A scalar
function F(Q.) is strictly quasi-convex at Q.1 if

V Q. =a Q.1 + (1- a) Q.2 , 0 < a < 1.

 109

a*
1

{1 - - - -

~i~
.::
~~ •
·~ •
•
);\ •
·>
{;,.,J - - - J •

X = f. L
Ai=A/N

(a) Parallel-Series Model

EN, LN

......

Li=UN

(b) Series-Parallel Model

Figure 5.1 Two different one-dimensional Distributed-Element Models

 110

Region
(I)
E
I
I
II

(I) Region enclosed by the invariant yield surface in the stress space

(II) Region enclosed by the initial yield surface in the strain space

(III) Region enclosed by the subsequent yield surface in the strain space

Figure 5.2 illustration of the space-dependent yielding behavior of ideal plasticity

 111

Figure 5.3 Invariant yield surfaces nested in the element stress space

<b( k)

k2 k3 k

Figure 5.4 Selection of yield constants for a fmite number of elements

according to the specified scrength discriburion function
 112

Initialize all vanables

Read parameters, NSTEP= 1

Read stratn increment .6emn

Compute F ( k i) Vi = 1. .. ...V

.Y
(JJ I = L (J jt ( k, )/.V
t=l

Stop

Figure 5.5 A flow diagram showing numerical procedure for obtaining srress
response of an N-element DEM
 113

II) Symbol a
..... l All Paths Start l Vchd lot Nut 4 F u~ura }
....... """'"'\
Path S'(ft'bol

O l
.... l
0

..
!11188
8 lUI Q

~~~··· ~ ~

:ci~
c
a 3
!::o
en
'-
0
a>
.
~
en~ !

••
~

t
0: i
0
.....
I Polh
~

(l'l
.....
Lo.25 0 0.25 O.!iO 0. 75
Axtol Stratn, e. %
(a)

~
....
1,6

0
....
5
~
U")
e-:cs
c
·a
!::o
en
a /,

~~ I

4 2
0
....;
I

3,7
~
....
!..o.?S -o..so -<J.E!S 0 0.25 0.50 0. 75
AXIDI Strcrn, €, %
(b)

Figure 5.6 Prescribed strain loading paths for response studies of the proposed
multi-dimensional DEMs (from [30])
 114
~l Start

iii
-...
~0

t-

V'l
..n
OJO

-
l-

en 4

l-
00
OJ..,

~·
Pottl ~ClefS- 0

~I
-4() -co 0 eo 40
AXIOI Str~ c, ksi
( a) biaxial stress space
0
"'t

V'l
~ o Stott-

-L-
en
.~@

~·
0
~ ~------r-----~------~. ----~
~-25 0 0.25 0.50 0. 75
Axial Strain, e. %
( b) axial stress- strain space

6 """"
~- o

~:·+~---~~------------~.----~~·
-1..!5 -LO ~. 5 0
9'm- Stroin, 7, %
0.5 LO !.5

( c) shear stress-strain space

Figure 5.7 Experimentally-observed stress response of copper to the prescribed

strain path given in Figure 5.6(a) (from [30])
 115

·c:n
~0
....

@+---------~--~--~
·~ -eo 0 20
Axial Stress, CT. kst
( a) biaxial stress space

~2
tj
vi
~()
~

Cii
.e2
~ I
4

()

1a~.-~----
~-.so------o~.B!----~o-----a~~-----
a-so----~
a~
Axial Stram, e, %
( b) axial stress-strain space

·v;
,X()
...

2+----r--~----~--~--~---.
1
-L5 -!.0 -o.5 0 0.5 !.0 !.5
9"lecr Strain, 7, %
( c) shear stress-strain space

Figure 5.8 Experimentally-observed stress response of copper to the prescribed

strain path given in Figure 5.6(b) (from [30])
 116

:o r---------~---------r--------~--------~

-1 0

-I S

Axial stress (ksi)

( a) biaxial stress space

""'
(J)
~
v
(J)
(J)
v
""
w

-
(J)

-1 0
.~
)(

<

.8
Axial strain(%)
(b) axial stress-strain space
20

1$

""'
:/l
~
v
:/l
(J)
Q
I.
....
:1)

I. -S
:-:
;:! -1 0
:1)

-IS

-20
· I.
Shear strain (o/o)
(c) shear stress-strain space

Figure 5.9 Stress response predicted by a new OEM subject to the prescribed strain
path given in Figure 5.6(a) (Tresca - , von Mises - - - -)
 117
~0

1'

10
...,
..f.
'"'I: '
>J)
v
'-
~
'- -~
~
v
.I:
il) -1 0

-U

-20

Axial stress (ksi)

( a) biaxial stress space

~ r-----------~----------~------------r---------~

~ .r---------~~----------~----------~~--------~
Axial strain ( o/o)
( b) axial stress-strain space
2D
.. ······ ·······~:::::·· ·- ~
::.:..::::::··.·:::::::.::·.~~::;
;;,;:.
"""
(I)
~
........
'~
:1)
u
I.,
.J
:1)

I., -~
~
~
{/)
-··

Shear strain(%)
(c) shear stress-strain space

Figure 5.10 Stress response predicted by a new DEM subject to the prescribed strain
path given in Figure 5.6(b) (Tresca - , von Mises --- -)
 118

7
2+----r----------~
' -40 --a) 0 20
Axial Stress. a;. ks1
( a)

I -

~+---~-----------
' -40 -20 0 20
Axtal Stress. a, ksl
(b)

'

2+---~------------
1 -40 --a) 0 20
AxJal stress, a: l<sl
( c)

Figure 5.11 Response predicted by different models of plasticity to the strain ~ath given
in Figure 5.6(b): (a) von Mises' yield surface with Prager·s hardemng rule (b)
Tresca's yield surface with Ziegler's hardening rule (c) Tresca's yield surface
and limit surface with Mroz' hardening rule (from [30])
 119

0
E

(a) proponional strain path

(b) stress response behavior

Figure 5.12 Response behavior of ideal plasticity under proponional

strain loading path
 120

Figure 5.13 Definition of the plastic-relaxation stress increment

in the uniaxial case
 121

(a) a big strain cycle

(b) the corresponding srress response

Figure 5.14 Illustration of existence of the equilibrium points

associated with a big strain cycle
 122

tangent plane

Figure 5.15 A diagram showing the rotation of coordinate axes which makes the
Xl axiS perpendicular tO the tangent plane tO the yield SUrface 8f2o at ![a

yield surfaces

Figure 5.16 An illustrative diagram showing non-strict convexity of yield surfaces

 123

CHAPTER 6
GENERALIZED MASING RULES FOR CYCLIC PLASTICITY

6.1 Introduction

In the previous chapter , we have extended the one-dimensional Distributed-

Element Models (DEMs) to a multi-dimensional representation, so that they can
be used for constitutive modeling within the context of general plasticity. While
the formulation of the multi-dimensional DEMs provides a useful and realistic way
for analysis of general multi-axial cyclic response behavior, efficient numerical im-
plementation of the theoretical formulation requires that only a limited number of
elements be introduced. FUrthermore, the model response has to be found by keep-
ing track of each element's behavior throughout the response history. Recall that in
Chapters 3 and 4, we gave the extended Masing rules for response of one-dimensional
DEMs with or without deterioration. As a result , the response of a DEM can be
found without the need to trace each element 's behavior. An interesting question
can then be raised: can we possibly find some mathematical rules similar to those
previously used in the one-dimensional Masing models so that general multi-axial
response of t he models based on the distributed-element formulation can be obtained
without keeping track of ea.ch element's behavior. If such mathematical rules gov-
erning multi-axial cyclic response exist, then based on these rules we may be able
to come up with numerical schemes t hat are more efficient and more accurate than
those based on a finite number of distributed elemen ts.
In the following sections, it will be shown that by introducing a response for-
mula good for initial loading, further unloading and reloading response to a general
loading can then be found by applying a composition of proper transformations
to the state variables involved in the initial loading formula. This method can be
shown to be conceptually equivalent to t he classical multi-yield-surface theory using
the Mroz kinematic hardening rule. However, the idea proposed here can actually
work out the response of a model with a collection of an ( uncountably) infinite num-
ber of yield surfaces and with proper kinematic hardening rules taken into account.
Even though the proposed formulation based on the transformation rules is not ex-
actly mathematically equivalent to that of the generalized three-dimensional DEM
presented in the previous chapter, it still provides us with an alternative way of
 124

obtaining complicated cyclic response in general plasticity. Furthermore, the trans-

formation rules that govern the response behavior corresponding to different loading
branches also give instructive insight into the physics of material behavior in cyclic
plasticity.

6.2 Extension of 1-D Response Formulas to Higher Dimensions

In the one-dimensional case, there have been many response formulas or math-
ematical models proposed for describing nonlinear, hysteretic response behavior of
structural or material systems, such as those presented in Chapters 3 and 4. The
problem of extending such one-dimensional models to the much more involved multi-
dimensional case has been a task of great challenge among researchers in the related
fields, and very little success has been attained on this subject.

6.2.1 Two-Dimensional Bouc-Wen Model

The theory of plasticity provides the theoretical background for analyzing gen-
eral multi-axial hysteretic response of mechanical and structural systems. However,
such an entirely theoretical approach would be usually computationally impracti-
cal for studying structural systems. Park, Wen, and Ang [37) proposed a two-
dimensional hysteretic model for random vibrations of structures subject to biaxial
excitations, which is an extension to the well-known Bouc-Wen model introduced in
Chapter 3. The one-dimensional non-deteriorating Bouc-Wen model, described by

(6.1)

was extended to the biaxial case so as to account for the interaction of the restoring
forces in two different directions. It was proposed [37] that for structural systems
with "isotropic" restoring force behavior, the forces rx and ry in the two directions
are described by the following coupled differential equations:

(6.2)

(6.3)
 125

where x and iJ are the velocities in the x andy directions respectively. The hysteretic
behavior given by Eqns. (6.2) and (6.3) can be illustrated by considering a simple
uni-directional displacement path for which it is assumed that

rx = r cos (), ry = r sin (), x = u cos (), y = u sin (), (6.4)

where rand u are, respectively, the uni-directional force and displacement, and() is
held constant on loading or unloading. Substituting (6.4) into Eqns. (6.2) and (6.3)
we can show that each of (6.2) and (6.3) reduces to Eqn. (6. 1) with n = 2. This
illustrative situation is sketched in Fig. 6.1 for easier understanding. Note that the
ultimate strength r u of the one-dimensional model with n = 2 can be found to be

(6.5)

which may serve as a guideline for choosing the model parameters.

For an orthotropic system, whose stiffness and strength in two orthogonal direc-
tions are different (Ax vs. Ay, (rx)u vs. (ry)u), Eqns. (6.2) and (6. 3) can be replaced
by

. A . Ax I . I {3 Ax . 2 Ay I . I {3 Ay . (6 .6)
rx = xX - a -( )2 xrx rx - -( ) 2 xrx - a - )(
2 yry rx - -( ) yrxry ,
rx u rx u ry u ry u2

ry Ay)( Iyry
. = A yY. - a - . Iry- {3 -Ay( . 2 a Ax
) 2 yry- -)( I. I {3 Ax) xrxry
. (6 .7)
2 2 xrx ry- - (2 ,
Ty u Ty u Tx u Tx u
where we require a+ (3 = 1. For example, in the biaxial tension-torsion loading
case, Eqns. (6.6), (6.7) can be written in terms of stress and strain components as:

where CJ u and Tu represent respectively the ultimate axial stress and the ultimate
shear stress of the system being modeled. An important feature of Eqns. (6.6) and
(6.7) is that under the transformation

(6.8)
 126

Eqns. (6.6), (6.7) will reduce to Eqns. (6.2), (6.3) of the isotropic case.
The foregoing formulation of the two-dimensional Bouc-Wen model is phe-
nomenological in nature; however, the new models exhibit reasonable biaxial hys-
teretic behavior as justified by some experimental results [37, 53]. This may be
attributed to an implicit assumption of the model behavior on the "yield" condi-
tion. From Eqns. (6.4) (using fy = rsinO and treating r and 0 as general polar
coordinates) and (6.8) , it follows that

(f= -r)( ). (6.9)

Tx u

Thus, since rx ~ (rx)u and ry ~ (ry)u we may conclude that the two-dimensional
Bouc-Wen model actually employs the concept of a limit surface of elliptic shape in
the biaxial "stress" space.
An important advantage of the two-dimensional Bouc-Wen model is that it is
versatile and amenable to analytical treatment, and thus can be applied to systems of
considerable complexity and under random excitation. However, the biaxial model
inherits the disadvantage of the one-dimensional model in that it exhibits unstable
drift under small cyclic excitations, as explained in Chapter 3. Moreover, there is
another unrealistic response feature inherent in the model due to the "empirical"
formulation given above. That is, under proportional (displacement) loading, the
biaxial restoring force response is also proportional at all times, even if the response
is in plastic state. This can be easily deduced by noting that when :i; = cfJ in
Eqns. (6.2) and (6.3) , we obtain rx = cry . This behavior is not consistent with the
theory of p lasticity or experimental observations. Another major disadvantage of
the model is that it is difficult to extend it to higher dimensions, due to a lack of
sound theoretical basis, to allow for a general analysis of cyclic plasticity.

6 .2.2 A R ecen t Procedure for G ener alizing 1-D Hysteretic M o d e ls

Recently, Graesser and Cozzarelli (14] presented a systematic procedure for ex-
tending a one-dimensional model of hysteresis to a multi-dimensional tensorial repre-
sentation provided that the model behavior is governed by some simple power laws.
In particular, they considered the generalization of the one-dimensional Ozdemir
model which was proposed originally for hysteretic behavior of yielding structures
(36]. However, due to some unrealistic characteristics associated with the original
 127

one-dimensional Ozdemir model, as discussed in Chapter 3, the extended multi-

dimensional model will inherently exhibit some nonphysical behavior, which may
lead to violation of Drucker 's postulates of stability. The inconsistency of these
generalized models (Bouc-Wen and Ozdemir models) with real behavior can be at-
tributed to their simplified mathematical formulation and lack of physical motiva-
tion. Nevertheless, the approach used in the derivation of the generalized Ozdernir
model still provides useful guidelines to the generalization of one-dimensional hys-
teretic models to the general multi-dimensional case, and is discussed in the follow-
ing.
The one-dimensional Ozdemir model with zero backstress can be described by
[36]
(6. 10)

and
(6. 11)

where E denotes Young's modulus (i.e., initial stiffness), and ay represents the
simple-tension yield stress of the model. The corresponding multi-dimensional model
can be found to be [14]:
(6.12)

and
(6.13)

where we define

and Sij denotes the deviatoric stress components. We may observe the close rela-
tionship between the sets of Eqns. (6.10), (6.11) and (6. 12), (6.13), and note that
(6.12) and (6.13) reduce to (6.10) and (6.11 ) in the uniaxial (one-dimensional) load-
ing case. In the derivation of Eqns. (6.12) and (6. 13), incompressibility of plastic
deformations has been assumed and the term 3J2 j a~ can be seen to be analogous
to the von Mises yield criterion in t he theory of plasticity.
Our aim here is to find multi-dimensional response formulas that can adequately
predict general, physically consistent elastic-plastic response behavior. In the one-
dimensional case, !\lasing's hypothesis and the extended rules given by Jayakumar
[23] provide the theoretical basis for the Masing models, which have been shown
 128

to be equivalent to the associated Distributed-Element Models (DEMs). A special

class of Masing models, proposed by Jayakumar (24] for modeling hysteretic behavior
of structures, uses the following one-dimensional response formulas in addition to
Eqn. (6.10):
(6.14)

for initial loading, and

(6.15)

for other response branches, where n is a parameter that controls the smoothness
of transition from elastic to plastic state, a 0 the stress state corresponding to the
latest unloading point, and au. is the ultimate stress (strength) of the model. Based
on the generalization rules used in Eqns. (6.10) to (6.13), one can find the following
general formula corresponding to Eqn. (6.14) for initial loading:

(6.16)

We remark that Eqns. (6.11) and (6. 14) are identical for initial loading, yet their
generalizations to multiple dimensions (Eqns. (6.13) and (6.16)) are quite different,
even in the case of initial loading. This is because that when applying those gener-
alization rules, which were developed for general response of plasticity, we treated
Eqns. (6.11) and (6.14) as general formulas valid for all response branches. The
formula given by Eqn. (6.16) is mathematically simple; however, it predicts that
the plastic strain increment is always proportional to the total strain increment
at a given stress state, which is obviously incorrect, as suggested by the behavior
of the multi-dimensional DEMs, or demonstrated by the classical theory of plas-
ticity. Fig. 6.2 shows a comparison example where biaxial stress responses under
proport ional strain loading were simulated using both a multi-dimensional DEM
and Eqns. (6.12), (6.16). A deficiency of the model behavior based on Eqns. (6.12)
and (6. 16) is that the stress response never decreases in each of the two components
under the prescribed proportional strain path. This is inconsistent with the result
using the DEM or models based on the classical theory of plasticity. In addition to
the aforementioned problem, there are some other difficulties with the generalization
of Masing rules based on Eqns. (6. 14) and (6.15), which are stated as follows:
 129

1) In the general multi-axial case, it is very difficult to specify appropriately the

virgin loading curve (or the "skeleton" curve) due to the path-dependent chara-
teristic of hysteretic response. We may note that the path-dependent property
actually invalidates the use of a scalar quantity (3h/ k~)n1 2 in Eqn. (6.16), since
otherwise two response histories with identical histories of (3h/ k~)nf 2 will have
the same stress response according to the formula (6.16).
2) To extend the Masing rules given in Eqns. (6.14), (6.15) to higher dimensions,
one has to characterize the response history into different branches including
initial loading, unloading and reloading cases. Based on the classical t heory of
plasticity, characterization of different loading cases can be done by introducing
a loading function F, such as the "yield" function F(q_) = 3J2 - k~ implicitly
used in Eqn. (6. 16). Thus, in t he case of neutral loading (F = 0 and dF = 0)
where 3J2 jk~ stays invariantly at the value of one, the stress state should still
vary with the change of the prescribed strain p ath. But t his is again not true
if Eqn. (6. 16) is used for describing the response to initial loading, which may
possibly include the neutral loading case.

Considering the response behavior of the multi-dimensional DEM proposed in

the previous chapter and the difficulties mentioned above, one can see t hat t he
task of finding simple mathematical formulas for describing general multi-axial hys-
teretic response is formidably challenging. Before we present an innovative idea of
generalizing response formulas for multi-axial plasticity, let us compare two similar
formulations based on different concepts of multiple yield surfaces. The first one is
associated with the new class of mult i-dimensional DEMs in which a collection of
yield surfaces is defined in the element stress space so that the yield surfaces stay
invariant in the space regardless of the change of the model response. Another for-
mulation with multiple yield surfaces is based on the classical theory of plasticity
using t he Mroz kinematic hardening rule [35]. In the second formulation, the yield
surfaces are defined in the model stress space and they move around with the model
response so that the current stress state of the model never lies outside any of the
yield surfaces. The Mroz kinematic hardening rule specifies t hat the "active" yield
surfaces on which the current stress state lies will translate in the same direction
as the line joining the current stress point to the point on the outer yield surface
corresponding to the same direction of outward normal [35]. This rule is sketched
schematically in Fig. 6.3, where the point Pis the current stress state on t he active
 130

yield surface Fm and Q is the point on the outer surface Fm+l corresponding to
the same direction of outward normal. The translation of the surface Fm (as well
as the inner surfaces, such as Fm-l) will follow the direction given by the line PQ.
The Mroz hardening rule ensures that the inscribed yield surfaces have a common
tangent at the current stress point. Based on the theorems given in Chapter 5, it
can be shown that models based on the Mroz hardening rule exhibit correct response
behavior in the sense that the properties associated with equilibrium points and a
limit surface can be adequately demonstrated. Other often-used kinematic hard-
ening rules, such as the Prager and the Ziegler kinematic hardening rules, fail to
exhibit such physical properties of response behavior shown by real materials (30] .
While the Mroz rule leads to good response predictions, its numerical implementa-
tion has been thought to be too involved and inefficient for complicated structural
analysis [30, 32].
It is noted that the response of a multi-dimensional DEM is governed by the
response of its elements, while response behavior of a model based on the classical
formulation of plasticity is determined mainly by t ranslation of yield surfaces. A
comparison of the detailed response behavior between the two multi-yield-surfa.ce
models reveals th at mathematically it is easier to deal with the classical model than
with the DEM in generalizing 1-D response formulas for general plasticity. Effective
response formulas for plasticity based on the classical multi-yield-surface theory will
be proposed in the following sections. It will be shown that these well-proposed
response formulas provide a very efficient way of implementing the classical multi-
yield-surface theory using the Mroz kinematic hardening rule.

6.3 A Class of Generalized Masing Models for Multi-axial Plasticity

6.3.1 A Response Formula for Initial Loading

To find mathematical rules similar to Masing's hypothesis for govening response

behavior in cyclic plasticity, one has to first define a formula good for response
of initial loading, and then find corresponding formulas for describing subsequent
response behavior. By initial loading here, we mean that no unloading defined
according to the classical theory of plasticity* has ever occurred.

* In the classical theory of plasticity, the characterization of different loading cases

for strain hardening materials is done by introducing a loading function (which is
 131

As noted above, the initial loading formula (6.16), derived from extending a
special class of 1-D Masing models, or the formulas (6.14) and (6.15) based on
Ozdemir's model (which was proposed for complete history of response) do not
yield proper response behavior for cyclic plasticity. This kind of deficiency may be
attributed to lack of physical insight in formulating the general response behavior,
so that the formulas may lead to unrealistic or even unstable response behavior.
A new response formula for initial loading is proposed here based on the response
behavior of ideal plasticity and the introduction of a "damage" function. The theory
of classical plasticity is formulated mainly based on experimental observations, and
hence a model showing the behavior of ideal plasticity can be thought of as physically
motivated. Based on the insight obtained in developing the multi-axial DEM which
consists of a collection of elements of ideal plasticity, we extend the "ideal plasticity"
model to account for strain hardening behavior as follows.
Recall that by Eqn. (5.43), the plastic strain increment in a yielding state of
perfectly plastic behavior with a corresponding yield function F can be expressed
in a vector form as
dfP = A (~, df) df_,

= { ~o(~) df if dF = 0;
if dF < 0,

usually the same as the yield function) defined by F(aij) = k so that when

(1) F = k, dF
aF dCTij > 0,
=-a
CTij

it is called loading. When

(2) F = k, dF
aF
= -a daij = 0,
CTij

it is called neutral loading. When

(3) F = k, dF =
aF
-a daij < 0,
CTij

it is called unloading.
 132

where dF is a function of both the current state and the load increment, and A is
the plastic modulus reduction matrix and

(6.17)

where Q is defined in (5.49). Note that in the above, dF is never greater than zero
and A = 0 corresponds to the case where response is purely elastic.
For our purpose of developing a general response formula for init ial loading, a
modulus-reduction function which signifies the "degree of yielding" can be introduced
as
D (f!_) = ( 3:z2 ) ~ , (6. 18)
u

where h = !smn Smn is the second invariant of the deviatoric stress tensor and
the parameter n is introduced to control the smoothness of yielding. The response
formula (6.17) for ideal plasticity in the case of initial loading is then modified by
including the modulus-reduction function D as

(6.19)

By Eqn. (6.19) , when the response is small, 3J2 « k~, and we have dEP ~ Q; when
3h = k~, the response state reaches the limit surface associated with the model
and the response behavior becomes perfectly plastic as loading is continued. In the
biaxial tension-torsion case where von Mises yield criterion is used, Eqn. (6.19) gives

dcf1 = C [Esi 1dcu + 4Gsus12dE12J

(6.20)
dcf2 = C [Esus12dcu + 4Gsi 2dEI2]'

wh ere C =- (3h ( k,..)nf 2

E 8 2 +4 G 8 2 and . the d ev1a
Sij IS
. t one
. stress t ensor.
u 12
A simulation study of the response behavior based on the formulas (6.19) and
(6.13) , respectively, was conducted to get an idea of how the formulas perform in
complicated loading conditions. The prescribed loading path under consideration is
the strain loop 0-1-2-...-8 as shown in Fig. 6.4. In the simulation, the strain loop was
traced twice, while only the response corresponding to the second loop is considered
in order for comparison with the experimental result shown in Fig. 6.5, where only
cyclically stablized behavior is demonstrated. The corresponding response curves
 133

are presented in Figures 6.6 and 6.7, where we can see the superiority of Eqn. (6.19)
in response prediction of initial loading to Eqn. (6.13), as justified by the exper-
imental result and the response predicted by a DEM. The response predicted by
Eqn. (6.19) shows correct behavior in terms of the positions of equilibrium points
and the associated limit surface. We remark that all the loading branches in the
above example are treated as initial loading so t hat only Eqn. (6.19) (together with
Eqn. (6.12)), which is derived for response of initial loading only, is used throughout
the response history.
In summary, we note that there are three important features of the response
formula (6.19) derived for initial loading:
(i) The model based on Eqn. (6. 19) preserves all the equilibrium points of a per-
fectly plastic model so that it always leads to stable and physically consistent
response behavior.
(ii) Analogous to the multi-dimensional DEM, there is no special restriction on the
yield condition (or modulus-reduction function) needed in the formula for initial
loading. In general, we should replace Eqn. (6.18) by

(6.21)

where F(q_) = ku is the equation of the limit surface associated with the model
and can assume any appropriate form. Note that F(q_) is also the "loading"
function controlling the cases of loading or unloading. FUrthermore, the power
law employed in Eqn. (6.21) may be replaced by some other function forms so
as to provide more general, flexible response behavior.
(iii) The key point in the new formula for initial loading is that the modulus-
reduction function D(q_) repla.ces a conventional yield condition and subsequent
hardening rules, so that continuous yielding behavior on initial loading is ade-
quately modeled.

6.3.2 Response Formulas for Unloading and Reloading Branches

With the initial loading formula defined in Eqn. (6.19) , we then want to find
corresponding mathematical rules that can govern appropriately the response be-
havior of unloading and reloading branches, so that complete response history to
 134

any multi-axial loading path can be calculated accordingly. Recall that loading and
unloading correspond to dF;::::: 0 and dF < 0 respectively, where F(g:) is the loading
function as employed in Eqn. (6.21).
It can be recognized that Masing's hypothesis implies mathematically that the
behavior of unloading response can be found from that of virgin response by intro-
ducing a proper transformation on t he state variables describing the response. Mo-
tivated by this concept and t he behavior of the classical multi-yield-surface model
using Mroz' kinematic hardening rule (cf. Fig. 6.3), we propose t he idea of in-
troducing a composition of t ransformations on the state variables involved in the
initial loading formula (6. 19), so that unloading response can be found based on
the same formula as for the response of initial loading. Based on this idea, we have
t he following formula , corresponding to Eqn. (6. 19), for unloading and reloading
branches:
(6.22)

where q_1 denotes t he vector of transformed stress state, which is a function of not
only the current response state, but also the past history.
To determine the effective transformation required for our purpose, we remark
that for the classical multi-yield-surface model, the yield surfaces reached by the
current stress state must be carried along with t he response state instead of stay-
ing invariant, so that they all have the current stress point as a common tangent
point. The movement of the yield surfaces along with the current state is illus-
trated schematically in Fig. 6.8, where the circles represent yield "surfaces" in a
two-dimensional stress space and points A, B denote two instantaneous stress states.
Thus, the response behavior corresponding to the unloading branch from a point B
can be found by transforming the geometrical configuration in Fig. 6.8(c) back into
that in 6.8(a), so that Eqn. (6.22) can be used effectively for response calulation
of any unloading (or reloading) branches. Care must be taken in performing the
transformations so that not only the transformation of geometrical configurations is
appropriately done, but also t he normality rule for determining increments of plastic
strain is preserved (cf. Appendix B). In the following, we will be concerned only with
t he two-dimensional loading case so that we need only deal with transformations of
planar configurations. An effective transformation formula for "steady-state" cyclic
 135

response (i.e., loading between points symmetric to the origin) that is derived from
a composition of a series of proper transformations can be found as follows:

-r1 ·o
a I + tT
· I _
- . (} e
-l 4
, (6.23)
2 sm 2

where (a', T 1 ) denotes the transformed stress state, and

(} 1 =tan- 1 (T- TO) , (}0 = tan-

1
(;~) , (6.24)
a- ao

In (6.24), (a, T) is the current actual stress state and (a0 , To) is the actual stress
state corresponding to the latest unloading point. The detailed derivation of the
transformation rules (6.23), (6.24) is given in Appendix B. Note that in the derivation
of the transformation rules we assumed that the yield surfaces in the 2-D space can be
represented by circles which are initially concentric. This, however, does not put any
limitation on applications using the above idea, since according to the well-known
Riemann 's Mapping Theorem [55], a simply-connected region of arbitrary shape can
always be mapped onto a circle through a conformal mapping.* Therefore, the yield
surfaces in the 2-D space can be of any shape and the transformation rules (6.23),
(6.24) will still work, as long as we can find the transformation whose existence is
guaranteed by Riemann's theorem so that the yield surfaces of arbitrary shape can
be transformed into circles.
The transformation approach mentioned above works, in principle, only for the
2-D case. However, it is also applicable to the general multi-axial plasticity provided
that isotropic materials are considered and the plastic deformation can be treated
as independent, or as some simple function, of the hydrostatic stress state. Many
real materials exhibit approximately these kinds of behavior, such as metals and
soils. In this case, we can always convert a stress state (which is a symmetric two

* The Riemann Mapping Theorem has also been extended to the case where a
region bounded by two simple closed curves, one inside the other, is mapped into a
region bounded by two concentric circles [41}.
 136

tensor) into a corresponding principal stress state (a diagonal two tensor) for which
the shear stress components vanish, and apply the 2-D transformations to the stress
state projected on a shifted 1r plane which is perpendicular to the hydrostatic axis
o- 1 = o-2 = o-3 in the principal stress space. This is shown in Fig. 6.9(a), and 6.9(b),
in which the circular cylinder and cone represent the yield surfaces corresponding
to the von Mises and Drucker-Prager yield conditions [1], respectively. A schematic
diagram illustrating this idea of transformation on the 1r plane is shown in Fig. 6.10.

6.3.3 Rules for Transient Response

With the initial response formula (6.19) and the transformation rules (6.22) to
(6.24) for other response branches, we are able to determine the steady-state cyclic
response of a system characterized by multiple yield surfaces without the need to
calculate the response of elements or to trace the motions of yield surfaces. However,
in the general cyclic loading case, we still need to extend the foregoing formulas to
account for transient behavior of cyclic response, as Jayakumar [23] did in the one-
dimensional case for extending the Masing's hypothesis for hysteresis.
Recall that in the one-dimensional case, we had the rules for incomplete loops
and completed loops of transient response (cf. Sec. 3.2.3). In the multi-dimensional
case, however, the cyclic loops between fixed strain points may not be "strictly
closed" in general. Here, by "strictly closed" we mean that a stress-strain loop is
closed at a load reversal point so that this point is both the starting and the ending
point of the loop. Based on geometrical considerations of multiple yield surfaces and
the Mroz hardening rule, we can modify the definitions for incomplete and completed
loops and deduce corresponding rules for them in the general multi-axial case.
Firstly, we define a completed loop as a "loop" along which the outermost yield
surface that contains the latest point of load reversal is reached again during the
loading process. Otherwise, the loop is said to be incomplete. For example, in
Fig. 6.11, the stress response "loop" ABC is incomplete, while the "loop" BCE is a
completed one since the outermost yield surface (level curve 2) containing the latest
point of load reversal Cis reached at E. With these definitions we can propose the
following two rules for transient response of cyclic plasticity:

Rule 1 : Incomplete loops

 137

The equation of any response curve can be obtained simply be using Eqn. (6.19)
and applying the q_-q_' transformation, as given by Eqns. (6.23) , (6.24) to the latest
point of load reversal (a-0, To) and the outermost yield surface on which (a-0, To) lies.
For example, the response curve CE in Fig. 6.11 can be found by applying the
transformation rule to point C and "level curve" 2.

Rule 2: Completed loops

Once the stress state reaches the outermost yield surface on which the latest
load reversal occcurs, the transformation rule is applied to the previous point of load
reversal and the corresponding outermost yield surface.

For example, in Fig. 6.11, as the loop BCE is completed atE, the transformation
rule is then applied to point B and level curve 4 for the response that follows. Note
that Rule 2 for completed loops is different from that in the one-dimensional case
where two points of load reversal are erased at a time when an interior curve crosses
a curve from a previous load cycle. This rule for one-dimensional hysteresis can
be shown to be actually a special case of the two-dimensional rule, in which only
proportional loading is taken into account.
We remark that in the case of transient response, the geometrical configuration
of yield surfaces is different from that of a steady-state case, not only in the position
of the active point of load reversal, but also in that 80 in (6.24) is measured by
reference to a new center point, which may be different from the origin of the stress
space. Therefore, the 00 in (6.24) should be replaced by the more general formula

B0 = tan - l (To- TC) , (6.25)

a-o- a-c
where (a-c, Tc) represents the coordinates ofthe center of the current reference circle
(the outermost yield surface that the latest unloading point is on). This situation
is illustrated in Fig. 6.12. Fig. 6.12(a) shows the process of initial loading from
the origin 0 to point A and then unloading from there. Fig. 6.12(b) shows t he
geometrical configuration corresponding to the transient response of unloading from
point B , where point C represents the center of the current reference circle. The
coordinates of the new center point C can be generally found as

(6.26)
 138

where r A and r B denote the radii of the outermost "active" circles on which points
A and B lie respectively, and ( CTC', TC') represents the coordinates of the center of
the previous reference circle (C' coincides with the origin 0 for the case in Fig. 6.12).
Based on the preceding rules, numerical implementation of the foregoing algorithm
requires for each point of load reversal to store in a list both the strength constant
(radius) and the coordinates of the center point of the outermost yield surface on
which the reversal point lies. Every time the yield surface with the smallest strength
on the memory list is reached, its corresponding point of load reversal and center
of reference is erased from memory. This phenomenon may be viewed as an addi-
tional attribute of the property of erasure-of-memory, and is the counterpart to the
property of unraveling of interior loops in the one-dimensional case.
While the mathematical manipulation involved in the above approach based on
transformation formulas is simple and effective, a major problem of implementing
the above rules for multi-axial transient response exists. This problem is associated
with the numerical ill-conditioning which occurs when the response formulas are
applied to states near the points of load reversal, which are singular points of the
corresponding transformations as can be deduced from the derivation of the trans-
formation formulas. To illustrate this, we consider the following example. When
unloading occurs from a point, say the point B or C in Fig. 6.11, the "yielding
value," 3J2 (.Q:'), at any point on curve BC or CE is computed by reference to the
corresponding unloading point B or C. After a transient loop is completed, such as
the loop BCD or BCE in Fig. 6.11, the yielding value at D orE should be calcu-
lated by reference to the previous unloading point B , according to the Rule 2 stated
above. However, when the point at which a transient loop is completed is very close
to the previous unloading point (such as point Din Fig. 6.11 which is close to point
B), due to the singular behavior at an unloading point ,the yielding value cannot be
found accurately (as we can see in Fig. 6.11, all the level curves signifying different
yielding values pass through point B). In other words, the calculation of Q:1 from the
transformation formulas is numerically ill-conditioned. In the one-dimensional case,
we can get around this problem by always erasing two points of load reversal when-
ever a loop is completed; but in the more general case, special care must be taken in
doing so in order not to introduce significant error. A remedy for dealing with the
problem is that two latest points of load reversal, instead of just one (Rule 2), will be
 139

erased every time a loop is completed if the current yielding value with reference to
the previous point of load reversal is found to b e considerably less than the yielding
constant of the active yield surface on which the latest point of load reversal lies.
For example, in Fig. 6.11, t he points D and E, which b oth lie on the level curve 2,
should always have the same yielding value, say 2, no matter which unloading point
is referenced. But due to the singular behavior around the unloading point B , the
yielding value at D may be found numerically as much less than 2, e.g., if point D
coincides with (or very close to) the unloading point B, the yielding value will be
found zero there. In this case, we may erase two latest points of load reversal, i.e., B
and C in Fig. 6.11, so that the active reversal point becomes A and then the yielding
value at D with reference to A will be found to be about 4 (corresponding to the
level curve 4), which is correct for continued response from D. On the other hand,
if the response curve goes from C to E at which the yielding value with reference
to point B is close to the yield constant of the active yield surface (level curve 2) ,
then only one point of load reversal (point C) will be removed from the memory.

6.3.4 Simulation Studies

Thus far , based on the classsical formulation of ideal plasticity and multi-yield-
surface t heory, we have derived a class of "generalized Masing models" based on
the response formula (6. 19) for initial loading, together with the transformation
formulas, (6.22) to (6.24), and the rules governing the rest of a response history
for general multi-axial cyclic loading. It is of interest to examine t he p erformance
of such a model that is actually comp osed of an infinite number of yield surfaces
moving in t he stress space according to the Mroz kinematic hardening rule. In t he
following, the model performance will be evaluated under the same biaxial tension-
torsion loading conditions as before.
The results of response predictions using a generalized Masing model for dif-
ferent prescribed strain paths, given in Figures 5.6(a) and (b), are shown in Fig-
ures 6.13 and 6.14 respectively, where t he response predicted by a DEM is also
included for comparison . Recall that t he loading sequence in F ig. 5.6(a) is 0-1-
0-2-0-3-0-... , so as to demonstrate the property of erasure-of-memory exhibited by
real materials. In t he simulations, Tresca's yield criterion has been adopted for the
 140

modulus-reduction function defined in Eqn. (6.21), and the model parameters used
are E = 16, 700 ksi, v = 0.33, o-0 = 30 ksi, and n = 2.5. A comparison between the
model predictions and t he experimentally-observed results, given in Figures 5.7 and
5.8, l~ads to the following remarks. It is immediately recognized that the response
behavior described by the initial loading formula (6.19) (or (6.20) for the biaxial
case) and the transformation rules (6.22) to (6.24) is in good agreement with the
experimental results in almost every aspect. Transition from elastic to plastic regime
is smooth and well-behaved, while the complicated biaxial Bauschinger effect is also
well accounted for. Moreover, t he model behavior clearly shows the existence of
equilibrium points and a limit surface, as well as the property of erasure-of mem-
ory. One may thus conclude that the behavior of the generalized Masing model is
physically consistent. In addition, the computational effort in making response pre-
diction based on the above response-formula approach is even less than that using
a ten-element multi-dimensional DEM, which is already computationally efficient
compared with models based on the classical theory of plasticity. The excellent
accuracy of the model in response prediction may be attributed to the formulation
based on a collection of an uncountably infinite number of yield surfaces and the well-
formulated Mroz kinematic hardening rule.* The numerical efficien cy of the model
is due to the proposed transformation method which avoids costly bookkeeping of
the movement of multiple yield surfaces involved in the model.

6 .4 Comparison of the Generalized Masing models with the DEMs

In the above, we have proposed two constitutive models for plasticity based on
different multi-yield-surface theories. As shown in the simulation studies, both mod-
els are physically consistent and parsimonious in parameters. It is thus of interest
to see whether these two models can become equivalent under general situations.
The generalized Masing models utilize yield surfaces defined in the model stress
space, together with the Mroz hardening rule to account for the Bauschinger effect
in cyclic plasticity, yet the DEMs use invariant yield surfaces defined in the element

* Response behavior based on the Mroz kinematic hardening rule has been shown
to be consistent with t hose observed from real metals [30]. Besides, it has been shown
mathematically that the rule never results in intersection of two yield surfaces [32].
 141

stress space. Even though it may be possible to choose the distribution of the
yield surfaces in a generalized Masing model to match that in aDEM so that they
exhibit the same kinematic hardening behavior, for the two models to be completely
equivalent, we need also take into a.ccount the flow rules adopted.
The flow rule used in the formulation of the multi-dimensional DEMs specifies
that the current plastic strain increment of the model is given by the average of the
corresponding plastic strain increments of the elements, and is not a simple function
of the model state itself. As for the generalized Masing models, the current plastic
strain increment is solely determined by the corresponding model stress state (using
the principle of normality). Thus, we may conclude that in general, the two models
with different formulations cannot be made exactly equivalent. Also, we remark that
for response analysis based on the proposed multi-dimensional DEMs, we need not
distinguish model response into different loading cases, such as loading or unloading;
however, for the generalized Masing models based on the proposed plane-geometry
transformation method, we have to keep track of all the unloading points throughout
reponse histories.
As a final remark, we mention that although the numerical implementation
of the generalized Masing models avoids costly bookkeeping of the movement of
multiple yield surfaces or the response behavior of distributed elements, it requires
in practice smaller load increments than those needed for the implementation of the
proposed DEMs, in order to achieve satisfactory accuracy.
 142

(a) proportional displacement path

(b) hysteretic restoring force behavior

Figure 6.1 Hysteretic behavior of the two-dimensional Bouc-Wen model

 143

12

10

"'"'fJl
~
.
,'.
......, 8 ' '
fJl
,,'
fJl

..e
..
fJl

~ 4
~
..c
en

0
0
Axial stress (ksi)
(a) biaxial stress space
30

..."'"' -------- ------·-----

'Jl
~
......,
fJl
'Jl

...~
"'
~ 10

~ s

Axial s train ( o/o)

(b) axial stress-strain s pace
12

""'
'Jl
~
......,
......... .. ...........
:1)
Ill

....8
..
II)

~ 4
~
..c:
Cl)
2

Shear strain (%)

(c) s hear stress-strain s pace

Figure 6.2 Comparison of the initial response predicted by Eqn. (6.16) and by aDEM
(Eqn. (6. 16) - , OEM----)
 144

Figure 6.3 The Mroz kinematic hardening rule

0. 6
3
- ?

-o.5 ... 4Jl ..

5 8 0.5
€ ( %)
1

6 ....._--1~--=-
7 ;- 0 .6

Figure 6.4 Prescribed strain path for response of initial loading

 145

1
1

·XJO '--_ _ _ _......__ _ _ _.....

-lOO 0 )00
CJ . . . .

( a) biaxial stress space

·l a D ' - - - - - -.........----~
·1
.....0 1

(b) axial stress-strain space

• )00 '------'--------J
·1 0 1
y ....
(c) shear stress-strain space

Figure 6.5 Experimentally-observed stress response of copper to the prescribed strain

path given in Figure 6.4 (from [17])
 146

Axial stress (ksi)

( a) biaxial stress space

...-><
(U

<

Axial strain(%)
(b) axial stress-strain space

:::-<11
~
'-'
<II
<II

..e
<II
... - 10
~
~
.c -20
rl)
-JO

Shear strain (%)

(c) shear stress-strain space

Figure 6.6 Stress response predicted by Eqn. (6.19) with the prescribed strain path
given in Figure 6.4 (Eqn. (6.19) - . DEM--- -)
 147
40

30

......
:n
~
\,../

:n
:n
()

-...
lo
:n
~ -10
()
..::
(/) -20

Axial stress (ksi)

( a) biaxial stress space

-~
-~
<

Axial strain (%)

( b) axial stress-strain space

Shear strain (o/o)

(c) shear stress-strain space

Figure 6.7 Stress response predicted by Eqn. (6.13) with the prescribed strain path
given in Figure 6.4 (Eqn. (6.13) - , DEM--- -)
 148

(a) initial configuration

(b) current stress state A

cr

(c) current stress state B

Figure 6.8 Movement of yield surfaces with current stress

state moving from A to B
 149

shifted 1r plane

(a) von Mises yield condition

/
0"1 = 0"2 = O"J

shifted 1r plane

(b) Drucker-Prager yield condition

Figure 6.9 Different yield surfaces and shifted 1r planes

in the principal stress space
 150
,
(52

mapped
iT plane

;; plane

Figure 6.10 Transformation on the 1r plane where A' is image of A

 151

Figure 6. 11 Illusrrarion of completed loops and numerical difficulty

associated with the rransformation approach
 152

(a) unloading from point A

(b) unloading from point B with new center C

Figure 6.12 Geomerrical consideration of transient response

 153

20 ~--------~---------r--------~---------,

Axial stress (ksi)

(a) biaxial stress space

,-...
'Vi
cVl
Vl
Q)
...""
Vl
-10
';
....
K
< -20

-30

-40
-0
Axial strain (o/o)
(b) axial stress-strain space
20

:::;-
Vl
~
'-'
Vl
Vl
d)
bVl
1-, -S
~
d)
.c -10
Cf)

- IS

-20

Shear strain (o/o )

(c) shear stress-strain space

Fi gure 6_13 Stress response predicted by a generalized Masing model subject to the
prescribed strain path gi ven in Figure 5_6(a)
(Masing - - , DEM- - - - , both using Tresca's yield criterion)
 154

IS

::-
(/)
~
'-" s
(/)
(/)

g
(/)

I. -S
(1
~
.c -10
Cl)

- IS

Axial str ess (ksi)

(a) biaxial stress space
~ r-----------,------------r------------r-----------,

..."""
(/)
~
'-"
VJ
(/)
<.)
...
I.

-
(/)
- 10
r,l
·~
><
< -20

-30

-4()

Axial strain (%)

(b) axial stress-strain space
20

::-
(/)
~
'-"
(/)
(/)

~
w
(/)

1.. -5
(13
Q.)
.c -10
Cl)

- 15

-20

Shear strain (o/o)

(c) shear stress-strain space

Figure 6.14 Stress response predicted by a generalized Masing model subject to the
prescribed strain path given in Figure 5.6(b)
(Masing - - , DEM - - - - , both using Tresca's yield criterion )
 155

CHAPTER 7
SUM MARY AND CONCLU SIONS

Analytical modeling of one-dimensional hysteresis and general multi-axial plas-

ticity is studied in this research, with particular emphasis on the parsimony of model
parameters and the physical consistency of model behavior. Based on the previous
chapters of this thesis, a summary and some conclusions drawn from the results are
presented as follows:

(1) Practical considerations of system identification and its implications for system
modeling are studied, so that criteria of good models for mechanical systems
can be used to provide guidelines for general modeling. A model which is good
for forward (response) analysis is not necessarily good for identification studies,
unless it is parsimonious in its parameters and robust to model error as well as
measurement noise.

(2) Various one-dimensional hysteretic models are examined in detail, including

both deteriorating and non-deteriorating models. Several models described
solely by differential equations have been shown to exhibit unrealistic behav-
ior which violates Drucker's postulates of stability. A general formulation for
modeling of degrading systems is presented based on the formulation of the
Distributed-Element Model (DEM) and the introduction of a damage index
function. A new class of deteriorating Masing models, whose behavior can
be completely described by a few simple mathematical rules and the extended
Masing rules, is also proposed to substitute for a special class of maximum-
displacement-controlled deteriorating DEMs. The modeling procedure and nu-
merical implementation of this class of Masing models is much easier than that
of the DEMs, so that its applicability to system identification studies is im-
proved.
 156

(3) The endochronic theory, which provides a unifying approach for plasticity with-
out the need to introduce yield conditions, is studied and implemented for re-
sponse simulation. A very efficient modeling technique based on the endochronic
theory is proposed to make the model more suitable for identification applica-
tions. Besides, inspired from the study of cyclic hardening behavior, the ex-
tension of a DEM to account for cyclic hardening (or softening) behavior can
be effectively done by simply making the ultimate strength of the model an
appropriate function of plastic deformation.

( 4) The one-dimensional DEMs are generalized so that they can be applied to the
case where multi-dimensional loading conditions are considered. In the gener-
alization, an invariant-yield-surface theory is proposed, in which no kinematic
hardening rule is needed to account for the subsequent yielding and strain hard-
ening behavior. The numerical implementation of the new DEMs is simple and
efficient, and the model behavior is physically consistent, as justified by compar-
ison of predictions with experimental results from the literature. An important
advantage of this new DEM for plasticity is that for an isotropic material, if
the yield condition has been appropriately chosen, then the general model needs
only two parameters, in addition to Young's modulus and Poisson's ratio, which
can be identified simply from a uni-axial virgin loading curve of the material.
In addition, we may choose some constants involved in the yield condition as
parameters to be identified, so that through system identification techniques,
the "best" yield condition for a complex structural or material system can be
found.

(5) Important properties of material behavior in cyclic plasticity are discussed, and
a new theory is presented to elucidate the properties based on the behavior of the
proposed multi-dimensional DEMs. The establishment of the theory provides
us with instructive insight into the elastic-plastic response mechanisms of real
materials under complicated loading conditions.
 157

(6) Generalized Masing rules for cyclic plasticity are proposed based on a plane-
geometry transformation method. When combined with a response formula
valid for initial loading, they provide an alternative model for cyclic plasticity
to the multi-dimensional DEMs, which require the introduction of a substan-
tial number of elements in response calculation in order for sufficient accuracy.
This new approach actually provides a highly efficient way of implementing the
classical multi-yield-surface theory with Mroz' kinematic hardening rule, which
is otherwise computationally impractical. The proposed transformation rules
governing general multi-axial cyclic response again give better insight into the
physical mechanisms of response in cyclic plasticity.

The original motivation of this research was to develop general classes of inelas-
tic models that could be used in system identification of structural systems from their
response data. However, the models proposed in this study are of interest themselves
in that they are simple, parsimonious models which give remarkably good results
of response predictions for copper and presumably for other materials or structural
systems (some minor modification or extension may be needed though). Further-
more, the proposed models involve only a few physically-based parameters so that
in general, no special identification technique is needed for determining appropriate
parameter values for a particular system of interest, although better result might be
obtained if system identification procedures were used.

In light of the above summary, a few suggestions for future exploration in related
subjects may be made as follows:

(I) In the study of degrading hysteretic systems, new damage index functions,
which should be physically consistent in nature and mathematically tractable,
could be proposed to improve the modeling of various effects of degradation,
such as the pinching behavior exhibited by reinforced-concrete structures.

(II) Further tests of the validity of the generalized multi-dimensional Masing models,
which employ the proposed plane-geometry transformation method, could be

I
 158

performed in multi-axial plasticity problems. This additional verification could

be done by comparing with experimental observations obtained from higher-
dimensional loading tests , when available, or by comparing with some other
well-behaved models, such as the new multi-dimensional DEMs and the modi-
fied endochronic models.

(III) Further study could be made of the interactive effects of multi-dimensional

loading on the response behavior of structural systems, such as their ductil-
ity, ultimate strength, and hysteretic energy dissipation, using the proposed
physically-consistent models.

(IV) The multi-dimensional DEMs or Masing models could be applied to identifica-

tion studies using response data from structural systems subject to multi-axial
real or laboratory-simulated earthquake excitations. By treating some constants
in the yield conditions required for the models as parameters to be identified,
we may also find the "best" yield conditions for different complex structural
systems.

(V) An investigation could be made of possible extensions of the proposed theory

of multi-dimensional DEMs and Masing models for rate-dependent plasticity,
creep, relaxation, deterioration and other such material behavior.
 159

REFERENCES

[1) Bathe, K.J. (1982), Finite Element Procedures in Engineering Analysis, Pren-
tice--Hall, New J ersey.
[2) Bazant, Z.P. , and Bhat , P.D. (1976) , "Endochronic theory of inelasticity and
failure of concrete," Journal of the Engineering Mechanics Division, ASCE,
Vol. 102, EM4, PP. 701-722.
[3) Beck, J.L. (1978), "Determining Models of Structures from Earthquake Records,"
Report No. EERL 78-01, California Institute of Technology, Pasadena.
[4) Beck, J .L. and J ayakumar, P. (1986), "Application of system identification
to Pseudo-dynamic test data from a full-scale six-story steel structure," Proc.
Interational Conference on Vibration Problems in Engineering, Xian, China.
[5) Beck, J.L. and Dowling, M.J. (1988), "Quick algorithms for computing ei-
ther displa~ement, velocity or acceleration of an oscillator," Int. J. Earthquake
Eng. and Struct. Dynam., Vol. 16, pp. 245-253.
[6] Beck, J.L. (1990), " Stat istical system identification of structures," Structural
Safety and Reliability, ASCE, II, pp. 1395-1402.
[7] Beck, J .L. and Katafygiotis, L.S. (1991) , "Updating of a model and its uncer-
tainties utilizing dynamic test data," Proc. 1st International Conference on
Computational Stochastic Mechanics, Greece, Computational Mechanics Pub-
lications, Southampton, pp. 125-136.
[8] Chaboche, J.L. (1986) , "Time-independent constitutive theory for cyclic plas-
ticity," Int. J. Plasticity, Vol. 2, pp. 149-188.
[9) Cifuentes, A.O. and Iwan, W.D. (1989), "Nonlinear system identification based
on modelling of restoring force behavior," Soil Dynamics and Earthquake En-
gineering, Vol. 8, No. 1, pp. 2-8.
(10] Clough, R.\V . (1966), "Effect of Stiffness Degradation on Earthquake Ductility
Requirements", Technical Report No. 66-16, Department of Civil Engineering,
University of California, Berkeley.
[ll) Fliigge, W. (1975), Viscoelasticity, 2nd ed. , Springer-Verlag, New York.
 160

[12] Franchi, A., Genna, F. , and P aterlini F. (1990), "Research note on quasi-
convexity of t he yield function and its relation to Drucker's postulate," Int.
J . Plasticity, Vol. 6, pp. 369-375.
[13] Gates, N.C. (1977), "The Earthquake Response of Deteriorating Systems,"
Report No. EERL 77-03, California Institute of Technology.
[14] Graesser E.J. and Cozzarelli , F.A. (1991), "A multi-dimensional hysteretic mo-
del for plastically deforming metals in energy absorbing devices," Technical
Report NCEER-91-0006, State University of New York, Buffalo.
[15] Greenberg, M.D. (1978), Foundations of Applied Mathematics, Prentice-Hall,
Inc., New Jersey.
[16] Hoshiya, M. and Maruyama, 0. (1987), "Identification of nonlinear structural
systems," Proc. 5th International Conference on Applications of Statistics and
Probability in Soil and Structural Engineering, Vancouver, Canada, Vol. 1, pp.
182-189.
[17] Hsu, S.Y., Jain, S.K. , and Griffin O.H. (1991), "Verification of endochronic
theory for nonproportional loading paths," Journal of Engineering Mechanics,
ASCE, Vol. 117, No. 1, pp. 110-131.
[18] Iemura, H. and J ennings, P.C. (1974), "Hysteretic response of a nine-story
reinforced concrete building," Int. J. Earthquake Eng. and Struct. Dynam.,
Vol. 13, pp. 183-201.
[19] lwan, W .D. (1966), "A distributed element model for hysteresis and its steady-
state dynamic response," Journal of Applied Mechanics, ASME, Vol. 33, No. 4,
pp. 893-900.
[20] lwan, W.D. (1967), "On a class of models for the yielding behavior of continuous
and composite systems," Journal of Applied Mechanics, ASME, Vol. 34, No. 3,
pp. 612-617.
(21] lwan , W.D. (1973), "A model for the dynamic analysis of deteriorating struc-
tures," Proc. 5th World Conference on Earthquake Engineering, Rome, Italy,
pp. 1782-1791.
[22] lwan , W.D. and Cifuentes, A.O. (1986), "A model for system identification of
degrading structures," Int. J. Earthquake Eng. and Struct. Dynam., Vol. 14,
pp. 877-890.
 161

[23) Jayakumar, P. (1987) , "Modeling and Identification in Structural Dynamics,"

Report No. EERL 87-01, California Institute of Technology.
[24) Jayakumar, P. and Beck, J.L. (1988), "System identification using nonlinear
structural models," Structural Safety Evaluation Based on System Identification
Approaches, Proc. Workshop at Lambrecht/ Pfalz, pp. 82-102.
[25) Jazwinski, A.H. (1970) , Stochastic Process and Filtering Theory, Academic
Press.
[26) Kachanov, L.M. (1971), Foundations of the Theory of Plasticity, North-Holland
Co., Amsterdam.
[27) Kadison, R.V. and Ringrose, J.R. (1983) , Fundamentals of the Theory of Op-
erator Algebras, Prentice-Hall.
[28) Kalman, R.E. (1960), "A new approach to linear filtering and prediction prob-
lems," Journal of Basic Engineering, ASME, Vol. 83, pp. 35-45.
[29) Katafygiotis, L.S. and Beck, J.L. (1991) , "An efficient treatment of model un-
certainties for the dynamic response of structures," Proc. 1st International
Conference on Computational Stochastic Mechanics, Greece, Computational
Mechanics Publications, Southampton.
[30) Lamba, H.S. and Sidebottom, O.M. (1978) , " Cyclic plasticity for nonpropor-
tional paths: part I - cyclic hardening, erasure of memory, and subsequent
strain hardening experiments; part II - comparison with predictions of three
incremental Plasticity models," Journal of Engineering Materials and Technol-
ogy, Vol. 100, pp. 96-111.
[31) Masing, G. (1926), "Eigenspannungen Und Verfestigung Beim Messing (Self
Stretching and Hardening for Brass)," Proc. 2nd International Congress for
Applied Mechanics, Zurich, Switzerland, pp. 332-335 (in German).
[32) McDowell, D.L. (1989), "Evaluation of intersection conditions for two-surface
plasticity theory," Int. J. Plasticity, Vol. 5, pp. 29-50.
[33) McVerry, G.H. and Beck, J.L. (1983), "Structural Identification of JPL Building
180 Using Optimally Synchronized Earthquake Records," Report No. EERL
83-01, California Institute of Technology.
[34) Mendelson, A. (1968) , Plasticity: Theory and Application, MacMillan, New
York.
 162

[35] Mroz, Z. (1967), "On the description of anisotropic workhardening," Journal

of the Mechanics and Physics of Solids, Vol. 15, pp. 163-175.
[36] Ozdemir , H. (1976) , "Nonlinear Thansient Dynamic Analysis of Yielding Struc-
tures," Ph.D. Dissertation, Division of Structural Engineering and Structural
Mechanics, Department of Civil Engineering, University of California, Berkeley.
[37] Park, Y.J. , Wen, Y.K., and Ang, A.H-S. (1986), "Random vibration of hys-
teretic systems under bi-directional ground motions," Int. J. Earthquake Eng.
and Struct. Dynam., Vol. 14, pp. 543-557.
[38] Saiidi, M. and Sozen, M.A. (1979), "Simple and complex models for nonlinear
seismic response of reinforced concrete structures," Structural Research Series
No. 465, Civil Engineering Studies, University of illinois, Urbana.
[39] Saiidi, M. (1982), "Hysteresis models for reinforced concrete," Journal of the
stuctural Division, ASCE, Vol. 108, ST5, pp. 1077-1087.
[40] Sandler, I.S. (1978), "On the uniqueness and stability of endochronic theories
of material behavior," Journal of Applied Mechanics, ASME, , Vol. 45, No. 2,
pp. 263-266.
[41] Spiegel, M.R. (1964), Theory and Problems of Complex Variables, Schaum's
Outline Series in Mathematics, McGraw-Hill.
[42] Sues, R.H., Mau, S.T. and Wen, Y-K. (1988), "Systems identification of de-
grading hysteretic restoring forces ," Journal of Engineering Mechanics, ASCE,
Vol. 114, No. 5, pp. 833-846.
[43] Sugiura, K., Lee, G.C. and Chang, K.C. (1987), "Endochronic theory for struc-
tural steel under nonproportionalloading," Journal of Engineering Mechanics,
ASCE, Vol. 113, No. 12, pp. 1901-1917.
[44] Takeda, T., Sozen,M.A., and Nielsen, N.N. (1970) , "Reinforced concrete re-
sponse to simulated earthquakes," Journal of the stuctural Division, ASCE ,
Vol. 96, ST12, pp. 2557-2573.
[45] Thyagarajan, R.S. (1989), "Modeling and Analysis of Hysteretic Structural
Behavior," Report No. EERL sg-03, California Institute of Technology.
[46] Valanis, K.C. (1971), "A theory of viscoplasticity without a yield surface, part 1:
general theory; part 2: application to mechanical behavior of metals," Archive
of Mechanics, Vol. 23, No. 4, pp. 517-551.

\
 163

[47] Valanis, K.C. (1980), "Fundamental consequences of a new intrinsic time mea-
sure : plasticity as a limit of the endochronic theory," Archive of Mechanics,
Vol. 23, No. 2, pp. 171-191.
[48] Valanis, K.C. and Read, H. (1982), "New endochronic plasticity model for
soils," Soil Mechanics - Transient and Cyclic Loads, Edited by Pande, G.N.
and Zienkiewicz, O.C., Wiley, New York.
[49] Valanis, K .C. and Fan, J. (1984), "A numerical algorithm for endochronic plas-
t icity and comparison with experiment," Computers and Structures, Vol. 19,
No. 516, pp. 717-724.
[50] Watanbe, 0., and Atluri, S.N. (1986), "Internal t ime, general internal vari-
able, and multi-yield-surface theories of plasticity and creep: a unification of
concepts," Int. J. Plasticity, Vol. 2, No. 1, pp. 37-57.
[51] Wen , Y.K. (1976), "Method for random vibration of hysteretic systems," Jour-
nal of the Engineering Mechanics Division, ASCE, Vol. 102, EM2, pp. 249-263.
[52] Wen , Y.K. (1980), "Equivalent linearization for hysteretic system under random
excitations," Journal of Applied Mechanics, ASME, Vol. 47, No. 1, pp. 150-154.
(53] Wen, Y.K. and Ang , A. H-S. (1988), "Inelastic modeling and system identifica-
t ion," Structural Safety Evaluation Based on System Identification Approaches,
Proc. Workshop at Lambrecht/ Pfalz, pp. 142-160.
[54] Whiteman, I.R. (1959), "A mathematical model depicting the stress-strain di-
agram and the hysteresis loop," Journal of Applied Mechanics, ASME, pp. 95-
100.
[55] Wylie, C.R. (1975), Advanced Engineering Mathematics, 4th ed., McGraw-Hill.
[56] Yoder, P.J. (1980) , "A Strain-Space Plasticity T heory and Numerical Imple-
mentation," Report No. EERL 80-07, California Institute of Technology.
 164

APPENDIX A
Operator Theory on Convex Sets

Some results from the operator t heory on convex sets are summarized here to
provide a theoretical basis for the derivation of important properties of the multi-
dimensional Distributed-Element Models given in Chapter 5. A more complete pre-
sentation of op erator theory can be found in Reference [27].

We define in the following some basic operations of point sets.

[Definition Al] : (Addition and Multiplication)

If X {x: x EX} andY- {y: y E Y} , then

X+ Y = {x + y: x E X,y E Y} , aX= {ax: x EX}. (A. 1)

[Definition A2] : (Convex Combination)

A vector q_ C Rn is said to be a convex combination of N elements, q_k , k =
1, 2, ... ,N, if
N N

q_ = L akq_k , ak > 0, L ak = 1 (A.2)

k= l k=l

[Definition A3]: (Convex Set)

A set n c Rn is said to be convex if it contains every convex combination of
two elements inn, i.e. ,

Based on the above definitions, we have the following important theorems re-
garding convex sets.

[Theorem Al]:
n
A convex set n contains every convex combination L ak q_k of elements q_1 , q_2 ,
./ k= l
... , q_n of n , for all positive integers n.
 165

[Proof]: By mathematical induction on n, i.e. , assuming

n-1 n-1

2::: a~ Q~ En, for any La~ = 1, a~ > 0, Q~ E n v k = 1, 2, ... 'n- 1,

k=1 k=l

then given 2.:":~= 1 ak = 1 with each ak > 0,

n-1

= a1Q 1 + (a2 + a3 + ... + an) L a~ Q~

k=l

since b1 + b2 = 1 and ;£, '}!_ E f!, since l:~::i a~ = 1 by choice of the a~.
[Theorem A2):
A set D is convex if and only if

(A.4)

[Proof]:

a 1D + a 2D = D, (all a2 2: 0)
a1 +a2
aD+ (1 - a)O = n, (0 < a < 1)
{=::} a;£+ (1- a)'}!_ ED, V;£,'}!_ E 0,

{=::} n is convex.
[Corollary]:
n n
Dis convex {=::} L(akf!) = ( L ak)D, ak 2: 0 Vk = 1, ... , n, n positive integer.
k=l k=1

This corollary has been used in the derivation of Eqn. (5.80).

 166

APPENDIX B
Derivation of Transformation Formulas for Generalized Masing Rules
for Multi-Axial Cyclic Response Behavior

In the following, we will derive the transformation formulas (6.23), (6.24) based
on the classical multi-yield-surface theory and the Mroz kinematic hardening rule.
In the derivation, yield surfaces associated with a model are treated as circles in
the two-dimensional u - r stress space. The validity of this simplification has been
discussed in Sec. 6.3.2.
In order to employ the same response formula (e.g., Eqn. (6.19)) for all re-
sponse branches in the multi-axial cyclic (strain) loading case, the stress state vari-
ables Q. involved in the response formula should be modified by a suitable transfor-
mation, as suggested by Masing's hypothesis for cyclic hysteretic response in the
one-dimensional case. The transformation must be able to characterize the change
of situations among different response branches so as to appropriately reflect var-
ious behavior corresponding to different loading conditions. Recall that Masing's
hypothesis for one-dimensional hysteresis implies that the response of an unloading
or reloading branch corresponding to some particular response of initial loading can
be obtained by introducing the transformation:

( x , ,r') = ( x-2 xo ' r - ro)

2 (B. 1)

to the state variables (x, r) involved, where x 0 and r0 represent the displacement
and the restoring force corresponding to a load reversal point for that particular
response branch. Based on the ideas behind Masing's hypothesis, we want to find a
transformation that can adequately reflect the difference between response of initial
loading and that of subsequent load reversals. Consider t he two response situations
I

in Fig. B.1(a) and (b), in which the multi-axial yielding behavior is accounted for by
the classical multi-yield-surface theory with Mroz' kinematic hardening rule. The
 167

idea proposed here is that if we can find a transformation formula that maps the ge-
ometrical configuration in Fig. B.l(a) to that in B .l(b), then we can use a response
formula, which is good for initial loading, to describe the response corresponding
to subsequent unloading or reloading branches. To transform the geometrical con-
figuration in Fig. B.l(a) to that in B.l(b), or vice versa, we introduce a series of
mappings as follows:

(I) w1 = z- (ao +iTo) :

This mapping is a translation of ao +iTo, as defined in Fig. B.l(a), so that the

unloading point A gets mapped to the origin in the w 1 plane.

The second mapping is a counterclockwise rotation of ~- ()0 , where ()0 is defined

in Fig. B.l(a). After the two transformations w 1 , w 2 , the geometrical configuration
in the z plane (Fig. B.l(a)) is mapped into that in the w 2 plane, as shown in
Fig. B.2(a).
2
(III) W3 = - :
W2

This mapping maps the circles in the w2 plane into horizontal lines in the w 3
plane, as shown in Fig. B.2(b).

This mapping maps the horizontal lines in the w 3 plane into concentric cir-
cles in the w 4 plane, as shown in Fig. B.2(c). Note that there may be some other
mappings that can do the same job as W4 does for transforming the overall geo-
metrical configurations, but care must be taken in choosing the mapping so that
.
the direction of a plastic strain increment, which is determined using the normality
principle, is preserved after transformation. To make the idea clearer, let us look
at the two graphs in Fig. B.3(a) and (b), which show respectively the geometrical
configurations after and before transformation. In order to meet the normality rule,
 168

we need that the points A, B, G, and Din Fig. B.3(b) be mapped to A', B', G', and
D' in Fig. B.3(a) so that they have exactly the same outward normal direction.
Thus, by geometry we require that

1
Lx = 2 Ly, Ly = (}0 - B. (B.2)

It can be easily shown that the transformation formula w 4 given in (IV) satisfies
the conditions required by (B.2).
1
(V) w = -
W4
: (w = CJ
1
+iT')

The geometrical configuration after the transformation can be found to be just

the one shown in Fig. B.l(b), in which the coordinates of the unloading point A' is
the same as those of point A in Fig. B.l(a).

With the above transformations, the overall transformation which maps the
configuration in Fig. B.l(a) to that in Fig. B.l(b) can be found by composition rule
as
_ -r1 -i94
w- . e,
2 Sill 02
which is the formula given in Eqn. (6.23), and where

(} 1 = tan - l ( T- To) ,
CJ- CJo

as given in Eqn. (6.24).

 169

z-p lane

(a) configuration before transformation

a'

w-p l ane

(b) configuration after transformation

Figure B.l Geometrical configurations before and after transformation

 170

w2 - plane
(a)

IB
I

/ D
liE
-..
w3-plane
(b)

w4-plane

(c)

Figure B.2 Configurations at differem transformation stages

 171

,..,..,
I

a'

w-p I ane

(a)

z-plane

(b)

Figure B.3 Conditions of the principle of nonnality on the proposed transfonnation