Mathematics in Industry 36

The European Consortium for Mathematics in Industry

Martijn van Beurden
Neil Budko
Wil Schilders   Editors

Scientific
Computing
in Electrical
Engineering
SCEE 2020, Eindhoven, The Netherlands,
February 2020
MATHEMATICS IN INDUSTRY 36
Series Editors
Hans-Georg Bock, Interdisciplinary Center for Scientific Computing IWR,
Heidelberg University, Heidelberg, Germany
Frank de Hoog, CSIRO, Canberra, Australia
Avner Friedman, Ohio State University, Columbus, OH, USA
Arvind Gupta, University of British Columbia, Vancouver, BC, Canada
André Nachbin, IMPA, Rio de Janeiro, RJ, Brazil
Tohru Ozawa,Waseda University, Tokyo, Japan
William R. Pulleyblank, United States Military Academy, West Point, NY,
USA
Torgeir Rusten, Det Norske Veritas, Hoevik, Norway
Fadil Santosa, University of Minnesota, Minneapolis, MN, USA
Jin Keun Seo, Yonsei University, Seoul, Korea (Republic of)
Anna-Karin, Tornberg, Royal Institute of Technology (KTH), Stockholm,
Sweden

THE EUROPEAN CONSORTIUM

FOR MATHEMATICS IN INDUSTRY

SUBSERIES

Managing Editor
Michael Günther, University of Wuppertal, Wuppertal, Germany
Series Editors
Luis L. Bonilla, University Carlos III Madrid, Escuela, Leganes, Spain
Otmar Scherzer, University of Vienna, Vienna, Austria
Wil Schilders, Eindhoven University of Technology, Eindhoven,
The Netherlands
The ECMI subseries of the Mathematics in Industry series is a project of The
European Consortium for Mathematics in Industry. Mathematics in Industry
focuses on the research and educational aspects of mathematics used in industry and
other business enterprises. Books for Mathematics in Industry are in the following
categories: research monographs, problem-oriented multi-author collections,
textbooks with a problem-oriented approach, conference proceedings. Relevance to
the actual practical use of mathematics in industry is the distinguishing feature of
the books in the Mathematics in Industry series.

More information about this series at http://www.springer.com/series/4651

Martijn van Beurden • Neil Budko • Wil Schilders
Editors

Scientific Computing
in Electrical Engineering
SCEE 2020, Eindhoven, The Netherlands,
February 2020
Editors
Martijn van Beurden Neil Budko
Electrical Engineering Department of Applied Mathematics
Eindhoven University of Technology Delft University of Technology
Eindhoven, The Netherlands Delft, The Netherlands

Wil Schilders
Eindhoven University of Technology
Eindhoven, The Netherlands

ISSN 1612-3956 ISSN 2198-3283 (electronic)

Mathematics in Industry
The European Consortium for Mathematics in Industry
ISBN 978-3-030-84237-6 ISBN 978-3-030-84238-3 (eBook)
https://doi.org/10.1007/978-3-030-84238-3

Mathematics Subject Classification: 65Lxx, 65-06, 65Mxx, 65Nxx, 65Zxx, 78-06

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland
AG 2021
This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether
the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse
of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and
transmission or information storage and retrieval, electronic adaptation, computer software, or by similar
or dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors, and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, expressed or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Switzerland AG.
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface

From February 16 until February 20, 2020, the 13th International Conference on
“Scientific Computing in Electrical Engineering” (SCEE) was held in Eindhoven,
The Netherlands. It was jointly organized by the Centre for Analysis, Scientific
Computing and Analysis (CASA) and the Electromagnetics group of Eindhoven
University of Technology, and the group Numerical Analysis of Delft University of
Technology.
Even though 13 is a number often associated with bad luck, this edition was
actually very fortunate. Already prior to and during the conference, the world was
discussing the SARS-CoV-2virus and associated problems and measures, and not
long after the conference ended, there was a lockdown in many countries. We are
very happy that SCEE 2020 took place, not in a virtual way, but with many face-
to-face contacts, meeting our esteemed colleagues once again, having lunches and
dinner together in an excellent location, the “Academisch Genootschap Eindhoven.”
Participants enjoyed the setting and the surroundings, as well as the opportunity to
sit in the garden and discuss.
The thirteenth edition of the SCEE conference brought together some 85 partici-
pants from the fields of applied mathematics, electrical and electronic engineering,
and the computer sciences as well as scientists from industry. Again, it created an
excellent working atmosphere, especially due to its unique workshop character,
where all talks and poster introductions were presented in plenary sessions. In
addition, we had very clear and high-quality talks and poster presentations, lively
and fruitful discussions, and a great deal of personal networking.
The Scientific Program Committee invited four experts to give keynote presen-
tations on the main topics in the regular program. Keynote speakers at SCEE 2020
were (in alphabetical order):
• Liliane Borcea (University of Michigan—USA), “Reduced order model
approach for inverse scattering”
• Romanus Dyczij-Edlinger (Universty of Saarland—Germany), “Reduced-order
finite-element modeling and optimization of antennas”

v
vi Preface

Participants of SCEE 2020 in the garden of Academisch Genootschap Eindhoven

• Slawomir Koziel (Reykjavik University—Iceland), “Forward and Inverse Surro-

gate Modeling for Accelerated Design Optimization of High-Frequency Struc-
tures”
• Jasmin Smajic (ETH Zürich—Switzerland), “Numerical Analysis of Electro-
magnetic Transients in Power Devices”
We also re-installed a tradition, a Sunday evening speaker: Albert Ruehli
(IEEE life fellow, 50 years IBM, currently at Missouri University of Science
and Technology—USA) entertained us with the very nice overview talk entitled
“Retrospective: 50 years of circuit and electromagnetic solutions.” Another feature
of this conference was the Industry Morning, where three renowned speakers
from industry gave very nice presentations on urgent topics within the electronics
industry:
• Rick Janssen (NXP Semiconductors), “Electromagnetic Simulation Challenges
in the Semiconductor Industry”
• Frank Buijnsters (ASML), “Efficient Maxwell Solvers for Optical Semiconduc-
tor Metrology”
• Stefan Kurz (Bosch), “Newton-Kepler-Bosch: Towards the Next Level of Scien-
tific Computing in Engineering”
The topics of the invited and industry speakers were representative of the con-
ference’s range. In addition, from Monday to Thursday, we had a total of 30
oral presentations and 30 poster presentations. The dense conference program was
completed with two special sessions: a meeting of the European project (Marie-
Preface vii

Skłodowska-Curie EID) ROMSOC, and a meeting of the recently established

ECMI Special Interest Group MSOEE. On Wednesday evening, the SCEE Standing
Committee, the Program Committee, and the Local Organizing Committee also
had a meeting, followed by a lavish dinner with the invited speakers. A special
highlight of the SCEE 2020 conference was our conference excursion to either
the Philips museum or the DAF museum. Both companies have been instrumental
in building up the city of Eindhoven, and the collections of these museums were
enjoyed by the participants. After this excursion, the conference dinner took place
in the Academisch Genootschap Eindhoven, where many ideas and new research
directions were discussed in parallel to the enjoyment of good food and wine.
The present book collects the conference outcomes as proceedings papers.
All these papers have successfully passed a standard peer-review process. The
contributions are divided into four parts, which reflect the main focus areas of
SCEE 2020 (“coupled problems,” also a traditional focus area, has been put under
“Mathematical and Computational Methods”):
• Circuit Simulation and Design
• Device Simulation
• Computational Electromagnetics
• Mathematical and Computational Methods
In retrospect, we feel we have compiled a very successful and interesting collection.
We wish to thank all the participants for their valued contributions to the SCEE 2020
conference and to this book, and we hope we will meet each other at future SCEE
conferences!

Eindhoven, The Netherlands Martijn van Beurden

Delft, The Netherlands Neil Budko
Eindhoven, The Netherlands Wil Schilders
October 2020
Acknowledgment

We would like the thank Eindhoven University of Technology, viz. the Centre for
Analysis, Scientific Computing and Applications (CASA) within the Department of
Mathematics and Computer Science and the Electromagnetics (EM) group within
the Department of Electrical Engineering, and Delft University of Technology,
Department of Applied Mathematics (DIAM), for their help and support in the
organization of the SCEE 2020 Conference.
We are also grateful for the financial support by the Applied Mathematics
Institute of the four Universities of Technology in The Netherlands (4TU.AMI), the
mathematics cluster NDNS+ (Nonlinear Dynamics of Natural Systems), the Dutch
National Organisation for Research (NWO), the European Marie-Curie-Skłodowska
Industrial Doctorate Project ROMSOC (Reduced Order Modelling, Simulation and
Optimization of Coupled Systems), and CST—Computer Simulation Technology
AG in Darmstadt, part of Dassault Systèmes.
Last but not least, we would like to thank all the members of the Local Organizing
Committee and the Scientific Committee who helped us very much in preparing and
running the conference. The careful reviewing process was only possible with the
help of the members of the Scientific Committee who were handling the reviewing
process. The anonymous referees did a wonderful job that helped the authors to
improve the quality of their contributions.
Finally, we express our gratitude to our colleagues from Springer Heidelberg for
continued support and patience during the preparation of this volume.

ix
Organization

Local Organizing Committee

Wil Schilders (Chair, TU Eindhoven, The Netherlands)

Martijn van Beurden (TU Eindhoven, The Netherlands)
Neil Budko (TU Delft, The Netherlands)
Enna van Dijk (TU Eindhoven, The Netherlands)

Program Committee

Wil Schilders (Chair, TU Eindhoven, The Netherlands)

Martijn van Beurden (TU Eindhoven, The Netherlands)
Neil Budko (TU Delft, The Netherlands)
Gabriela Ciuprina (Politehnica University of Bucharest, Romania)
Georg Denk (Infineon, Germany)
Herbert de Gersem (TU Darmstadt, Germany)
Michael Günther (University of Wuppertal, Germany)
Stefan Kurz (Bosch, Germany)
Ulrich Langer (Johannes Kepler University Linz, Austria)
Jan ter Maten (University of Wuppertal, Germany)
Jörg Ostrowski (ABB, Switzerland)
Ursula van Rienen (University of Rostock, Germany)
Vittorio Romano (University of Catania, Italy)
Ruth Vazquez Sabariego (KU Leuven, Belgium)
Sebastian Schöps (TU Darmstadt, Germany)
Caren Tischendorf (Humboldt University of Berlin, Germany)

xi
xii Organization

Standing Committee

Ursula van Rienen (Chair, University of Rostock, Germany)

Gabriela Ciuprina (Secretary, Politehnica University of Bucharest, Romania)
Michael Günther (Treasurer, University of Wuppertal, Germany)
Jörg Ostrowski (ABB, Switzerland)
Wil Schilders (TU Eindhoven, The Netherlands)

Sponsors

TU Eindhoven
TU Delft
NWO
CST
4TU.AMI
NDNS+
ECMI
Contents

Part I Circuit Simulation and Design

Efficient Model Reduction of Myelinated Compartments
as Port-Hamiltonian Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3
Ruxandra Barbulescu, Gabriela Ciuprina, Tudor Ionescu, Daniel Ioan,
and Luis Miguel Silveira
1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3
2 Port-Hamiltonian Formulation and Reduction .. . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5
3 System Reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 6
4 Synthesis of Equivalent Reduced Circuit . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 7
5 Results . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 9
6 Conclusions .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 11
References . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 11
Towards a Parallel-in-Time Calculation of Time-Periodic
Solutions with Unknown Period .. . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 13
Iryna Kulchytska-Ruchka and Sebastian Schöps
1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 13
2 Multiple Shooting with Unknown Period . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 14
3 Periodic Time-Parallelization with Coarse Grid Correction . . . . . . . . . . . . . . . . 16
4 Numerical Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 18
5 Conclusions .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 20
References . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 21
On the Exactness of Rational Polynomial Chaos Formulation
for the Uncertainty Quantification of Linear Circuits
in the Frequency Domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 23
Paolo Manfredi and Stefano Grivet-Talocia
1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 23
2 Rational Polynomial Chaos Expansion . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 24

xiii
xiv Contents

3 Transfer Functions of Linear Lumped Circuits . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 24

3.1 Basic MNA Formulation for RGLC Circuits . . . . . .. . . . . . . . . . . . . . . . . . . . 24
3.2 Parameterization for Uncertainty Quantification . .. . . . . . . . . . . . . . . . . . . . 26
4 An Illustrative Example .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 30
5 Conclusions .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 31
References . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 31
Parallel-in-Time Simulation of Power Converters
Using Multirate PDEs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 33
Andreas Pels, Iryna Kulchytska-Ruchka, and Sebastian Schöps
1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 33
2 Power Converter Model .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 34
3 Parareal Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 35
4 Multirate PDEs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 36
5 Numerical Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 38
6 Conclusions .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 40
References . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 40

Part II Device Simulation

A Maximum Principle for Drift-Diffusion Equations
and the Scharfetter-Gummel Discretization . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 45
Kai Bittner, Hans Georg Brachtendorf, Tobias Linn,
and Christoph Jungemann
1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 45
2 A Maximum Principle for the Drift Diffusion Equation .. . . . . . . . . . . . . . . . . . . 46
3 Discretization by Scharfetter-Gummel and Finite Volumes .. . . . . . . . . . . . . . . . 50
References . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 52
Numerical Calculation of Electronic Properties of Transition
Metal-Doped mWS2 via DFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 53
Chieh-Yang Chen and Yiming Li
1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 53
2 The Computational Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 54
3 Results and Discussion .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 59
4 Conclusions .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 61
References . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 61
Numerical Simulation of Thermal Conductivity of Silicon Nanowires . . . . 63
Min-Hui Chuang and Yiming Li
1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 63
2 Computational Structure and Models . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 64
3 Simulation Techniques .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 65
4 Results and Discussion .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 67
5 Conclusions .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 69
References . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 70
Contents xv

A Novel Surface Mesh Simplification Method for Flux-Dependent

Topography Simulations of Semiconductor Fabrication Processes.. . . . . . . . 73
Christoph Lenz, Alexander Scharinger, Paul Manstetten,
Andreas Hössinger, and Josef Weinbub
1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 73
2 Surface Mesh Simplification .. . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 74
2.1 Feature Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 75
2.2 Mesh Partitioning and Movement of Regions . . . . .. . . . . . . . . . . . . . . . . . . . 75
3 Results . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 76
3.1 Distance to Original Geometry .. . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 77
3.2 Time Spent on Simplification.. . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 78
3.3 Flux Calculation and Monte Carlo Ray Tracing .. .. . . . . . . . . . . . . . . . . . . . 78
4 Summary.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 80
References . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 80
Simulations of a Novel DG-GFET .. . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 83
Giovanni Nastasi and Vittorio Romano
1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 83
2 Mathematical Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 84
3 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 86
4 Conclusions .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 90
References . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 90

Part III Computational Electromagnetics

Electric Circuit Element Boundary Conditions in the Finite
Element Method for Full-Wave Frequency Domain Passive Devices.. . . . . . 95
Gabriela Ciuprina, Daniel Ioan, Mihai Popescu, and Sorin Lup
1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 95
2 ECE Boundary Conditions .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 97
3 ECE in FEM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 98
3.1 Numerical Results. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 103
4 Conclusions .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 105
References . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 106
A Convolution Quadrature Method for Maxwell’s Equations
in Dispersive Media . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 107
Jürgen Dölz, Herbert Egger, and Vsevolod Shashkov
1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 107
2 Structure Preserving Discretization . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 109
3 A Convolution Quadrature Approach . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 111
4 Numerical Illustration .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 112
5 Summary.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 114
References . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 114
xvi Contents

On the Stability of Harmonic Coupling Methods with Application

to Electric Machines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 117
H. Egger, M. Harutyunyan, M. Merkel, and S. Schöps
1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 117
2 Model Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 118
3 Harmonic Stator-Rotor Coupling .. . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 121
4 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 123
References . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 125
Multifidelity Uncertainty Quantification for Optical Structures . . . . . . . . . . . 127
Niklas Georg, Christian Lehmann, Ulrich Römer, and Rolf Schuhmann
1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 127
2 Decoupled Uncertainty Propagation with Scattering Matrices.. . . . . . . . . . . . . 128
3 Multifidelity Monte Carlo. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 130
4 Numerical Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 131
5 Conclusions .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 135
References . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 135
Dielectric Breakdown Prediction with GPU-Accelerated BEM . . . . . . . . . . . . 137
Cedric Münger, Steffen Börm, and Jörg Ostrowski
1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 137
2 BEM Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 139
3 Discretization .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 141
4 GPGPU Quadrature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 141
5 Numerical Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 143
5.1 Validation .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 143
5.2 GPU-Acceleration.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 144
6 Conclusions and Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 146
References . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 146
Empirical Analysis of a Coaxial Microwave Structure with Finite
Transmission Zero . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 149
K. Papke, F. Gerigk, and U. van Rienen
1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 149
2 Equivalent Circuit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 151
3 Analyses .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 153
4 Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 154
5 Conclusions .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 156
References . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 157
Frequency-Domain Non-intrusive Greedy Model Order
Reduction Based on Minimal Rational Approximation... . . . . . . . . . . . . . . . . . . . 159
Davide Pradovera and Fabio Nobile
1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 159
2 Available MOR Strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 160
2.1 Greedy Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 161
2.2 A Posteriori Indicators .. . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 161
Contents xvii

3 Numerical Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 163

3.1 An Eigenproblem in Magneto-Hydrodynamics . . .. . . . . . . . . . . . . . . . . . . . 163
3.2 Frequency Response of a Waveguide Diplexer . . . .. . . . . . . . . . . . . . . . . . . . 165
4 Conclusions .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 166
References . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 167
A Comparison Between Different Formulations for Solving
Axisymmetric Time-Harmonic Electromagnetic Wave Problems . . . . . . . . . . 169
Erik Schnaubelt, Nicolas Marsic, and Herbert De Gersem
1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 169
2 Well-Posed Variational Formulation . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 170
2.1 Non-classical Conditions Along the Symmetry Axis . . . . . . . . . . . . . . . . . 170
2.2 Direct Construction of a Subspace of S n (Ω) . . . . .. . . . . . . . . . . . . . . . . . . . 171
3 Comparison and Discussion of the Quasi-3D Methods .. . . . . . . . . . . . . . . . . . . . 172
3.1 Spurious Modes and High-Order FE Discretizations.. . . . . . . . . . . . . . . . . 172
3.2 Convergence Results for Higher Order Finite Elements . . . . . . . . . . . . . . 172
3.3 Influence of α and β on the Convergence Behavior .. . . . . . . . . . . . . . . . . . 174
4 Conclusion .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 176
References . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 176
The Magnetization Analysis of Motor Magnet and Its Influence
on Cogging Torque . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 179
Chenxi Wang, Matthias Willig, Stefan Kurz, and Kevin Gutmann
1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 179
2 BLDC Motor and Cogging Torque .. . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 180
3 Definition of the Magnet Material in Analysis. . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 181
3.1 Magnetization Curve of the Magnet . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 181
3.2 Approximation Method to Describe the Hysteresis Effect .. . . . . . . . . . . 182
4 Modelling of the Magnetization Device and Motor . . . . .. . . . . . . . . . . . . . . . . . . . 183
5 Result of Analysis .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 184
6 Summary.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 187
References . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 188

Part IV Mathematical and Computational Methods

A Combination of Model Order Reduction and Multirate
Techniques for Coupled Dynamical Systems . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 191
M. W. F. M. Bannenberg, A. Ciccazzo, and M. Günther
1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 191
2 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 192
2.1 Mathematical Modelling .. . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 192
2.2 Multirate .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 193
2.3 Model Order Reduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 194
2.4 Combining MR and MOR . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 196
xviii Contents

3 Results . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 196
3.1 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 197
4 Conclusion .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 198
References . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 198
Waveform Relaxation for Low Frequency Coupled Field/Circuit
Differential-Algebraic Models of Index 2.. . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 201
Idoia Cortes Garcia, Jonas Pade, Sebastian Schöps, and Caren Tischendorf
1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 201
2 Field/Circuit Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 202
3 Waveform Relaxation and Convergence . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 204
4 Numerical Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 207
5 Conclusions .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 208
References . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 209
Splitting Methods for Linear Circuit DAEs of Index 1 in
port-Hamiltonian Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 211
Malak Diab and Caren Tischendorf
1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 211
2 Circuit Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 212
3 Operator Splitting for Index-1 Circuit DAEs . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 215
3.1 Subsystem Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 215
3.2 Convergence Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 216
4 Numerical Simulation .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 217
5 Conclusions and Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 218
References . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 218
Reduced Order Modelling for Wafer Heating with the Method of
Freezing . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 221
E. J. I. Hoeijmakers, H. Bansal, T. M. van Opstal, and P. A. Bobbert
1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 221
2 Theory . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 222
2.1 Model Introduction.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 223
2.2 Model Reformulation: Method of Freezing.. . . . . . .. . . . . . . . . . . . . . . . . . . . 224
3 Reduced Order Modelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 225
3.1 Standard Reduced Order Modelling Approach .. . .. . . . . . . . . . . . . . . . . . . . 225
3.2 Proposed Reduced Order Modelling Approach . . .. . . . . . . . . . . . . . . . . . . . 226
4 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 226
5 Conclusion and Future Outlook . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 229
References . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 229
Multirate DAE-Simulation and Its Application in System
Simulation Software for the Development of Electric Vehicles . . . . . . . . . . . . . 231
Michael Kolmbauer, Günter Offner, Ralf Uwe Pfau,
and Bernhard Pöchtrager
1 Background and Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 232
2 Problem Formulation .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 232
Contents xix

3 Multirate Integration for Coupled Network DAEs. . . . . . .. . . . . . . . . . . . . . . . . . . . 236

4 Simulation of a BEV with Cooling System .. . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 237
5 Simulation of a Three Phase Inverter with Cooling System.. . . . . . . . . . . . . . . . 238
6 Conclusion .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 240
References . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 240
A Hysteresis Loss Model for Tellinen’s Scalar Hysteresis Model .. . . . . . . . . . 241
Jan Kühn, Andreas Bartel, and Piotr Putek
1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 241
2 Tellinen’s Scalar Hysteresis Model .. . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 242
3 Adapted Hysteresis Loss Model .. . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 244
4 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 249
5 Conclusion and Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 249
References . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 250
Hybrid Modeling: Towards the Next Level of Scientific
Computing in Engineering .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 251
Stefan Kurz
1 Introduction: What Is Hybrid Modeling? . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 251
2 Blending Data with Physics [16] . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 253
3 Data-Driven Field Simulation [18] . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 255
4 Bayesian Free-Shape Optimization [23] . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 257
5 Summary and Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 260
References . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 262
Machine Learning for Initial Value Problems of
Parameter-Dependent Dynamical Systems . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 265
Roland Pulch and Maha Youssef
1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 265
2 Parameter-Dependent Dynamical Systems. . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 266
3 Time Discretisation.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 267
4 Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 267
5 Numerical Results for Test Example . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 269
6 Conclusions .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 273
References . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 273
Modeling of Thermoelectric Generator via Parametric Model
Order Reduction Based on Modified Matrix Interpolation . . . . . . . . . . . . . . . . . 275
Ananya Roy, Gunasheela Sadashivaiah, Chengdong Yuan, M. Nabi,
and Tamara Bechtold
1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 275
2 Model Description.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 276
3 Parametric Model Order Reduction with Matrix Interpolation . . . . . . . . . . . . . 279
4 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 280
5 Conclusions and Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 282
References . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 282
xx Contents

Nonlinear Model Order Reduction of a Thermal Human Torso Model . . . 285

Gunasheela Sadashivaiah, Chengdong Yuan, and Tamara Bechtold
1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 285
2 Case Study .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 286
3 Model Order Reduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 289
3.1 Proper Orthogonal Decomposition .. . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 289
3.2 Dynamic Mode Decomposition . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 289
4 Numerical Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 290
5 Conclusion and Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 291
References . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 292
Multi-Level Iterations for Microgrid Control with Automatic
Level Choice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 293
Robert Scholz, Armin Nurkanović, Amer Mešanović, Jürgen Gutekunst,
Andreas Potschka, Hans Georg Bock, and Ekaterina Kostina
1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 293
2 Nonlinear Model Predictive Control . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 294
3 Multi-Level Iterations .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 295
3.1 Automatic Level Choice . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 296
4 Dynamic Microgrid Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 297
4.1 Scenario and Model Description . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 297
5 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 299
6 Conclusion .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 300
References . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 301
Multi-Level Inversion Based on Mesh Decoupling . . . . . . .. . . . . . . . . . . . . . . . . . . . 303
Benny Shachor, Hadi Hajibeygi, and Domenico Lahaye
1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 303
2 Problem Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 304
3 Nested Iteration Based on Mesh Decoupling . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 306
4 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 307
5 Conclusions .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 310
References . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 311
 Part I
Circuit Simulation and Design
Efficient Model Reduction of Myelinated
Compartments as Port-Hamiltonian
Systems

Ruxandra Barbulescu, Gabriela Ciuprina, Tudor Ionescu, Daniel Ioan,

and Luis Miguel Silveira

Abstract The information is transmitted in neurons through axons, many of whom

have myelin-covered sections, whose main purpose is to increase the speed of
electrical signal transmission. Modeling the myelinated axons in a realistic way,
by maintaining the physical meaning of components may lead to complex systems,
described by high-dimensional systems of PDEs, whose solution is computationally
demanding. Analysis of larger neuronal circuits including multiple myelinated
axons therefore requires the generation of equivalent low-order models to control
complexity. Such models must preserve the physical interpretation and properties
of the original system including its passivity and stability. The axons’ port-based
structure makes them suitable to be modeled as port-Hamiltonian systems. This
paper uses a structure-preserving reduction method for port-Hamiltonian systems to
reduce the description of a myelinated compartment into a model with comparable
accuracy with the previously used vector fitting technique. The reduced system is
synthesized into an equivalent passive circuit with no controlled sources and only
positive elements, amenable for inclusion in standard neuronal simulators.

1 Introduction

A myelinated axon (Fig. 1) consists of myelinated sections through which the signal
dissipates, which alternate with Ranvier nodes where the signal is regenerated
(“saltatory conduction”). To model the transmission of signals through this chain,
the phenomena occurring in the Ranvier nodes have to be coupled with those in the
myelinated sections (internodes). The underlying mechanisms of a Ranvier node are
well described by the Hodgkin & Huxley model [10], or its reduced versions [19].

R. Barbulescu () · L. M. Silveira

INESC ID/IST Tecnico Lisboa, Universidade de Lisboa, Lisboa, Portugal
e-mail: [email protected]; [email protected]
G. Ciuprina · T. Ionescu · D. Ioan
Politehnica University of Bucharest, Bucharest, Romania
e-mail: [email protected]; [email protected]; [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 3

M. van Beurden et al. (eds.), Scientific Computing in Electrical Engineering,
Mathematics in Industry 36, https://doi.org/10.1007/978-3-030-84238-3_1
4 R. Barbulescu et al.

Fig. 1 Simplified
geometrical model of a
myelinated axon, as a chain
of internode myelinated
compartments and Ranvier
nodes

The most popular approach to model the internodes (myelinated compartments)

is the “cable model”, described by parabolic 1D PDEs [13], i.e., the RC transmission
line equation. In a previous work [11], the authors reduced the internode model
with different methods, resulting in a hierarchical series of models of three spatial
geometry classes: 2.5D, 1D and 0D and three categories of models: analytical,
numerical and reduced order models. The analytical 1D model reduced with the
vector fitting (VF) technique proved to be the most accurate according to our
weighted error metric, as shown in the Results section. In [11] the error is computed
using a weighted norm, where the weights associated to frequencies are extracted
from the spectrum of the standard neuronal signal. This error is suitable to estimate
the global accuracy of neuronal signals, since in the typical neuronal spectrum the
low frequency components (up to a few hundred Hz) are much more significant than
the high frequency ones. The accuracy of the current reduction method is compared
with the results in [11].
For the simulation of the saltatory conduction in a whole axon, the internodes
were replaced in [12] with the differential equation macromodel extracted from
VF, so the equivalent circuit had many controlled sources. This is acceptable when
there is no constraint on the reduced circuit, but in some environments dedicated
to neuronal simulations, such as NEURON [9], one can only create a circuit with
no controlled sources (or a small amount of controlled sources, modeled using Op-
Amps) and with positive parameters. In this work we synthesize the reduced system
into an equivalent circuit with only positive RCs and no controlled sources (we call
this circuit ECi+). We start from the large discretized transmission line lumped RC
model as a substitute of the PDEs (Fig. 2). All elements in this model are linear and
frequency independent. The long network of RC sections has resistive parameters
describing longitudinal electrical conduction phenomena through axoplasm, and
capacitive and transverse conductive effects through the cell membrane.
This particular model of a myelinated compartment, as a chain of RC cells,
is suitable for port-based network modeling, as in the port-Hamiltonian (pH)
framework. The pH systems are widely used in modeling, analysis and control of
(multi-)physical systems [5, 21]. Extensive research has been done on model order
reduction targeting preservation of relevant properties and/or port-Hamiltonian
structure for linear [2, 8, 15] and nonlinear systems [4, 17, 20]. Among these
techniques, the moment-matching procedure is an efficient tool [1, 2, 16]. The
reduced model is obtained by constructing a lower degree rational function that
approximates the given transfer function and matches it at various interpolation
Efficient Model Reduction of Myelinated Compartments as Port-Hamiltonian Systems 5

iP1 Rk Rk Rk +1
1 Vk 1 (t) Vk (t) Vk +1 (t) iP2

uP1 Gk 1
Ck 1 Gk Ck Gk +1 Ck +1
uP2

PORT 1 PORT 2

Fig. 2 The companion circuit of an internode. The network of RC (in our case identical) cells is
generated by the spatial discretization with centered differences of the transmission line equation.
The outer dotted line outlines the system and the inner dashed line denotes one individual RC cell

points in the complex plane. This formulation is preferred to the direct reduction of
the number of cells of the segmented numerical model, which already is a reduced
model.

2 Port-Hamiltonian Formulation and Reduction

The pH representation is based on the energy state space, which represents a

natural state space for the equations composing the mathematical models of physical
systems. The Hamiltonian gives the total stored energy of the system, whereas the
system has boundary ports to interact with the environment, through the exchange
of energy. The mathematical representation of a pH system is

ẋ = (J − R)∇x H(x) + Bu(t)
(1)
y = BT ∇x H(x)

where x ∈ Rn is the state vector; H : Rn → [0, ∞] is continuously differentiable

– the Hamiltonian, describing the internal energy of the system as a function of
state; J = −JT ∈ Rn×n is the structure matrix (skew-symmetric) describing the
interconnection of energy storage elements in the system; R = RT ≥ 0 is the
dissipation matrix describing the energy loss in the system; and B ∈ Rn×m is
the port matrix describing how energy enters and exits the system through the m
terminals/ports (here m = 2).
Our approach is based on describing the myelinated compartment in Fig. 2 as a
pH system (1) and reducing the overall model with structure-preserving moment-
matching. We start from the circuit description of the original model (a SPICE
netlist) and generate the pH form of this system. Next, the system is reduced by
moment-matching. Finally, the equivalent reduced circuit is synthesized from the
state-space representation of the reduced system.
6 R. Barbulescu et al.

We consider the network in Fig. 2 as a 2x2 system with input u =

 T
uP1 (t )
R1 iP2 (t) and output y = [V1 (t) uP2 (t)]T . The state space vector
consists of the charges of the capacitors x = [q1 , q2 , . . . , qn ]T , thus its
 T
derivative ẋ = iC1 , iC2 , . . . , iCn is composed of the currents through the

capacitors. The Hamiltonian is defined as H(x) = 12 nk=1 C1k qk2 = 12 xT Qx
and its derivative with respect to the state variables is a vector of voltages:
 T
∇x H(x) = uC1 , uC2 , . . . , uCn = Qx. 
In this formulation Q is a diagonal matrix, Q = diag C1k , the structure matrix
J = 0, and the dissipative matrix R is tridiagonal, having on line k the elements

T
1 0 ... 0
− R1k , R1k + Rk+1
1
+ Gk and − Rk+1
1
. The port matrix B = . Because
0 ... 0 1
R is positive definite and the capacitors are identical, the product RQ is positive
definite.

3 System Reduction

The reduction is based on a moment-matching technique, part of the family of

interpolatory methods [1]. A linear SISO system described by the state matrices
( , ,  ) has the transfer function K(s) =  (sI −  )−1 , K : C → C.
Consider a point in the complex plane that is not in the spectrum of  . The k-order
moment of the system with the transfer function K at s ∗ ∈ C − σ ( ) is formally
defined as:


∗ (−1)k dk K(s)
ηk (s ) = . (2)
k! ds k s=s ∗

For a fixed point s ∗ , a reduced-order system described by the transfer function

K̂ with the corresponding moments η̂k (s ∗ ) matches the first n∗ moments of K
if ηk (s ∗ ) = η̂k (s ∗ ), k = 1, n∗ , which in fact means it matches the coefficients
of n∗ terms of the Taylor expansion of K [14]. One can either choose one point
s ∗ and match the first moments of the two transfer functions, or choose a set of
points and match the 0-order moment in all the points in the set. The selection
of the interpolation points is important. Whereas selecting n∗ moments at a fixed
s ∗ may improve the approximation accuracy locally, selecting s1 , . . . , sr points
for a reduced order r and matching the 0-order moments at these points better
preserves input-output behaviours (here, the ∗ notation was dropped for readability).
Customary, s1 = 0 is chosen to preserve the step response of the given system.
For SISO systems the interpolation conditions are enforced pointwise, but in
the MIMO case – where K(s) is a m × m matrix-valued rational function – full
matrix interpolation would translate into m × m conditions at every interpolation
point. This would result in an actual larger order of the reduced system than the
Efficient Model Reduction of Myelinated Compartments as Port-Hamiltonian Systems 7

initially imposed r. Instead we only interpolate along certain directions bk (“right

tangential interpolation”). This relaxed notion of interpolation is adequate for an
optimal approximation [7]. For an imposed reduced order r we compute the matrix
 
 = (s1 I −  )−1 b1 (s2 I −  )−1 b2 . . . (sr I −  )−1 br ,

where the vectors b1 , b2 , . . . , br represent the tangential directions of interpolation.

Since the interpolation points sk and the tangential directions bk are dependent
on the reduced model, we use an iterative process to correct the interpolation points
and tangential directions until the interpolation conditions are met [4] .
The reduced matrices are computed as in [2]:
 −1
Jr = T QJQ Qr = T Q

Rr = T QRQ Br = T QB

and they are used to construct the reduced system in the port-Hamiltonian form:

ẋr = (Jr − Rr )Qr xr + Br u(t)
. (3)
y = BTr Qr xr

Such reduced order system matches the 0-order moments of the original sys-
tem at the chosen interpolation points [3]. The reduction procedure is structure-
preserving, in the sense that the reduced system is still in the port-Hamiltonian form,
but the matrices have lost some of their properties, for instance Qr is not diagonal
anymore, Rr is not tridiagonal, but is still symmetric and positive definite, Br is now
likely full.

4 Synthesis of Equivalent Reduced Circuit

There is extensive research on circuit realization of systems, either by direct

interpretation of the mathematical model or from the state-space form or the
system’s transfer function [18]. However, in most approaches the resulting circuit is
not guaranteed to contain only physically-meaningful elements, due to the presence
of negative R, L or C elements or it has a large number of controlled sources.
The transfer function of the reduced system is actually a 2 × 2 symmetrical
matrix of impedances. A possible circuit realization for this is a star (Fig. 3), where
the impedances of the subcircuits result directly from either the transfer function
components or the state-space matrices of the reduced system.
8 R. Barbulescu et al.

Fig. 3 The circuit realization

T scheme for a 2 × 2 system

Each impedance ZT0 , ZT1 and ZT2 can be realized through a pole-residue decom-
position as the sum of the impedances of r cells connected
 in series, each composed
of a capacitor in parallel with a conductance: Zxx = rk=1 1/(Ck s + Gk ).
The reduced system (3) can be viewed in the form of a standard description of an
RC circuit

Cẋr = −Gxr + Bu(t)
y = Exr

where C, G, B and E are defined accordingly.

The reduced states are the capacitors’ voltages in the reduced circuit. To simplify
realization, each state should be involved in only one equation. To that end,
matrices C and G are diagonalized to allow the equations to be separated. Their
diagonalization impacts the matrices B and E, which become full (likely already the
case here). In the reduced system this would translate into the circuit as controlled
sources. To avoid that, the two matrices are scaled so that all their values are either
1 or −1 and consequently the outputs will be algebraic sums of all the states.
The computations lead to the following relations for the components Zxx , where
ckk and gkk (k = 1, r) are the diagonal values of C and G (after diagonalization)
and the denominator actually represents the scaling of B and E:
  
ckk ckk ckk
Ck = Ck = e1k (bk1 −bk2 ) Ck = e2k (bk2 −bk1 )
ZT0 : e1k bk2
gkk
ZT1 : gkk
ZT2 : gkk
Gk = e1k bk2 Gk = e1k (bk1 −bk2 ) Gk = e2k (bk2 −bk1 )

In theory the capacitance Ck and the conductance Gk of a cell may have any sign.
But the CG pair signs differ only by the signs of the diagonal values of the matrices
C and G. Here C is the identity matrix, so clearly positive definite. G is a diagonal
matrix that comes from the original system matrix RQ, which is positive definite.
Since the reduction procedure guarantees passivity, it will preserve the definiteness
of the system matrix. Hence G has only positive values on the diagonal. This means
that for every cell, Ck and Gk are either both positive or both negative.
Consider the synthesized circuit of Zxx as in Fig. 4 (left), where the first two cells
have positive values and the third has negative values. In Fig. 4 (right) the circuit
is split into the “positive” and the “negative” contributions [18]. For the negative
subcircuit the signs for both Ck and Gk are reversed and the same excitation is used
Efficient Model Reduction of Myelinated Compartments as Port-Hamiltonian Systems 9

Fig. 4 Synthesis of a component Zxx of order 3. (Left): The circuit with positive and negative
CG pairs, y = V(1). (Right): The circuit split into the “positive” and “negative” subcircuits, y =
V(1+) − V(1−) extracted as the voltage of a null current source that connects the two subcircuits

for both subcircuits. The initial circuit has the same output y = V(1) as the circuit
after splitting, computed as the difference of two voltages y = V(1+) − V(1−).

5 Results

Figure 5 shows the frequency responses of the two components of the original (50
cells) and reduced (order 5) systems. The response of the transfer component (1,2)
is very far from the original system’s, but the values are so small that this graph is
in fact not relevant accuracy-wise, because the reduction procedure has an implicit
minimization of the H2 norm. This is proved by almost identical step responses.
The relative error is under 2% even for order 1 and is comparable with the one
obtained with vector fitting (Fig. 6 left) with the adaptive frequency sampling (AFS)
procedure described in [6]. In the interest of fairness, the errors are computed for

MIMO 2x2, comp. (1,1) MIMO 2x2, comp. (1,2)

100 100
original model (order 50)
reduced model (order 5) - iteration1
-1
10 reduced model (order 5) - iteration2

10-2 10-50

10-3
original model (order 50)
reduced model (order 5) - iteration1
-4 -100
10 10 reduced model (order 5) - iteration2

10-5

10-6 10-150
100 102 104 106 108 100 102 104 106 108
Frequency [Hz] Frequency [Hz]

Fig. 5 The frequency responses of the original (50 cells) and reduced (order 5) systems (note the
diminutive vertical scale on the graph to the right)
10 R. Barbulescu et al.

10 -1 35

30
10 -2

Electric potential [mV]

25
Relative error

10 -3
20

15
10 -4

10
10 -5
pH moment-matching - s in {0,[1e1,1e3]} (lin) 5
vector fitting AFS - input [1e0,1e7] Hz (log)
10 -6 0
1 2 3 4 5 6 7 8 9 0 5 10 15 20 25 30
Order of reduced model Time [ms]

Fig. 6 (Left): The relative error vs. the order of the reduced model for pH and VF methods;
f
errrel = fmM w(f ) Zorig (f ) − Zred (f ) 2 df/Z0 , frequency f ∈ [fm , fM ] = [100 , 107 ] Hz,
logarithmically spaced, w(f ) is the weight function, Z0 is the d.c. impedance of the line [11].
(Right): The reproduced output of the reduced circuit (order 3) built in NEURON

Fig. 7 The ECi+ circuit extracted from the reduced system of order 3 and reproduced in
NEURON. The output is [V(l1p) − V(l1m); V(r3p) − V(r2m)]

the best set of conditions for each of the two methods. The new method is not meant
to improve accuracy but to generate a model without controlled sources and with
positive elements, which is a requirement for the inclusion in NEURON.
The reduced circuit of order 3 is built in NEURON (Fig. 7) and the output is
reproduced in Fig. 6 (right). The input is a rectangular pulse in the left (i1 = I1p =
Efficient Model Reduction of Myelinated Compartments as Port-Hamiltonian Systems 11

I1m) and open-circuit in the right (i2 = I2p = I2m = 0). The corresponding outputs
copy the shape of the input and the relative difference between the corresponding
peaks of the original circuit (50 cells) and the reduced one is between 2% and 3%.

6 Conclusions

This paper uses a structure-preserving reduction method for pH systems to reduce a

myelinated compartment in the model of a neuron. The automatic procedure starts
from the netlist of the original model and generates its port-Hamiltonian form. The
pH system is reduced using an interpolatory method through moment-matching,
resulting in a reduced system that is still port-Hamiltonian, therefore preserving the
passivity and the stability of the original model. The relative error is acceptable even
for order 1 (less than 2%).
This procedure allows for a trade-off between a good approximation error and
the desired structure preservation. The choice of interpolation points is a degree
of freedom to be used for potentially improved accuracy in the moment matching
reduction.
The state-space representation of the reduced system is subsequently synthesized
into an equivalent circuit with no controlled sources and only positive RLCs (a
ECi+ circuit). This circuit can be used in neuronal simulators such as NEURON
and further integrated into larger models. The current method will further prove
beneficial for the reduction of the entire myelinated axon, with the nonlinear HH
model of a Ranvier node included.

Acknowledgments This work was partially supported by Portuguese national funds through FCT,
Fundação para a Ciência e a Tecnologia, under project UIDB/50021/2020 as well as project
PTDC/EEI-EEE/31140/2017.

References

1. A.C. Antoulas, Approximation of large-scale dynamical systems, in SIAM, vol. 6 (2005)

2. A. Astolfi, T.C. Ionescu, Moment matching for linear port hamiltonian systems, in 50th IEEE
Conference on Decision and Control and European Control Conference, 2011, pp. 7164–7169
3. A. Astolfi, T.C. Ionescu, Families of moment matching based, structure preserving approxi-
mations for linear port Hamiltonian systems. Automatica 49(8), 2424–2434 (2013)
4. C. Beattie, S. Gugercin, S. Chaturantabut, Structure-preserving model reduction for nonlinear
port-Hamiltonian systems. SIAM J. Sci. Comput. 38(5), B837–B865 (2016)
5. C. Beattie, V. Mehrmann, H. Xu, H. Zwart, Linear port-Hamiltonian descriptor systems. Math.
Control Signals Syst. 30(4), 17 (2018)
6. G. Ciuprina et al., Vector fitting based adaptive frequency sampling for compact model
extraction on HPC systems. IEEE Trans. Magn. 48(2), 431–434 (2012)
7. S. Gugercin, C. Beattie, Model reduction by rational interpolation, in Model Reduction and
Approximation: Theory and Algorithms (SIAM, New York, 2017), pp. 297–334
12 R. Barbulescu et al.

8. J.S. Hesthaven, B.M. Afkham, Structure-preserving model-reduction of dissipative Hamilto-

nian systems. J. Sci. Comput. 1–19 (2018). https://doi.org/10.1007/s10915-018-0653-6
9. M.L. Hines, N.T. Carnevale, The NEURON book (Cambridge University Press, Cambridge,
2006). https://neuron.yale.edu/neuron/
10. A.F. Huxley, A.L. Hodgkin, A quantitative description of membrane current and its application
to conduction and excitation in nerve. J. Physiol. 117(4), 500–544 (1952)
11. D. Ioan, R. Bărbulescu, L.M. Silveira, G. Ciuprina, Reduced order models of myelinated
axonal compartments. J. Comput. Neurosci. 47(2–3), 141–166 (2019)
12. D. Ioan, G. Ciuprina, R. Barbulescu, Coupled macromodels for the simulation of the saltatory
conduction. UPB Sci. Bull. Ser. C 18(3) (2019). ISSN:2286-3540
13. K.A. Lindsay et al., An introduction to the principles of neuronal modelling, in Modern
Techniques in Neuroscience Research (Springer, New York, 1999), pp. 213–306
14. D.D. Ling, I.M. Elfadel, A block rational Arnoldi algorithm for multipoint passive model-order
reduction of multiport RLC networks. ICCAD 97, 66–71 (1997)
15. R.V. Polyuyga, Model reduction of port-Hamiltonian systems. PhD thesis, University of
Groningen, 2010
16. R.V. Polyuyga, A. van der Schaft, Structure preserving model reduction of port-Hamiltonian
systems by moment matching at infinity. Automatica 46(4), 665-672 (2010)
17. R.V. Polyuyga, A. van der Schaft, Effort-and flow-constraint reduction methods for structure
preserving model reduction of port-Hamiltonian systems. Syst. Control Lett. 61(3), 412-421
(2012)
18. L.M. Silveira, J.F. Villena, Circuit synthesis for guaranteed positive sparse realization of
passive state-space models. IEEE Trans. Circ. Syst. I 64(6), 1576–1587 (2017)
19. K.K. Sriperumbudur, U. van Rienen, R. Appali, 3d axonal network coupled to microelectrode
arrays: a simulation model to study neuronal dynamics, in 37th Annual International Confer-
ence of the IEEE Engineering in Medicine and Biology Society (EMBC) (IEEE, New York,
2015), pp. 4700–4704
20. A.J. van der Schaft, L2-gain and Passivity Techniques in Nonlinear Control (Springer, New
York, 2000)
21. A. van der Schaft, Port-Hamiltonian systems: an introductory survey. Proceedings of the
International Congress of Mathematicians, vol. 3, pp. 1339–1365. Citeseer, 2006
Towards a Parallel-in-Time Calculation
of Time-Periodic Solutions
with Unknown Period

Iryna Kulchytska-Ruchka and Sebastian Schöps

Abstract This paper presents a novel parallel-in-time algorithm able to compute

time-periodic solutions of problems where the period is not given. Exploiting
the idea of the multiple shooting method, the proposed approach calculates the
initial values at each subinterval as well as the corresponding period iteratively.
As in the Parareal method, parallelization in the time domain is performed using
discretization on a two-level grid. A special linearization of the time-periodic
system on the coarse grid is introduced to speed up the computations. The iterative
algorithm is verified via its application to the Colpitts oscillator model.

1 Introduction

Steady-state analysis is a common task in electrical engineering, for example, during

the initial design stages of, e.g., electric circuits or motors. Classical sequential time
stepping may lead to lengthy transient computation particularly when the underlying
dynamical system possesses a large time constant. Various approaches for efficient
steady-state calculation are known from the literature. For instance, clever methods
to choose the starting value [1] or an explicit error correction [15] could accelerate
the time-domain calculation considerably.
A powerful tool for speeding up the classical time stepping is the class of
parallel-in-time methods, such as the multigrid reduction in time [5] or Parareal
[11]. Originating from the multiple shooting method [13], they are based on
the splitting of the considered time interval into several windows and updating
the solution at synchronization points iteratively. The use of coarse and fine
discretizations propagates quickly low-frequency information of the solution using a
cheap sequential solver followed by a very accurate result with a precise fine solver
applied in parallel.

I. Kulchytska-Ruchka () · S. Schöps

Computational Electromagnetics Group, Technical University of Darmstadt, Darmstadt, Germany
e-mail: [email protected]; [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 13

M. van Beurden et al. (eds.), Scientific Computing in Electrical Engineering,
Mathematics in Industry 36, https://doi.org/10.1007/978-3-030-84238-3_2
14 I. Kulchytska-Ruchka and S. Schöps

Another direction of obtaining the steady state is based on the solution of the
joint space- and time-discrete time-periodic system formulated on the whole period
[8]. There the initial and final values are coupled through the prescribed periodicity
condition. An obstacle within the solution of the periodic problem in the time
domain becomes the large size of the system matrix as well as its special block
structure due to the interdependence of the solution vectors over the period. To deal
with this difficulty a frequency domain approach was proposed in [2]. In case of
linear problems, the method takes advantage of the block-cyclic matrix structure
by applying the discrete Fourier transform. It fully decouples the variables, thereby
allowing for the separate solution of each harmonic coefficient. This approach was
further extended and incorporated into the Parareal framework by the authors in
[10]. There, a simplified Newton-based iterative algorithm was presented together
with its convergence analysis for the efficient treatment of nonlinear problems.
Solutions of time-periodic problems become much more challenging when the
period is not given. Such situation occurs, e.g., when dealing with an autonomous
system [3]. In contrast to a non-autonomous problem, the periodicity cannot be
determined from the applied excitation. This paper proposes a numerical algorithm
capable of determining an appropriate period automatically using parallelization in
the time domain. Extending the idea of the multiple shooting method we include the
unknown period together with multiple initial values as the sought parameters into
the iterative procedure. Verification of the presented approach is illustrated through
its application to the Colpitts oscillator model [9].
The paper is organized as follows. Section 2 describes the basis of the multiple
shooting approach including the unknown period as an additional variable. This is
further expanded to the family of the Parareal-based methods in Sect. 3. Section 4
applies the proposed parallel-in-time approach to the Colpitts oscillator model using
a particular linerization on the coarse level. The paper is finally summarized in
Sect. 5.

2 Multiple Shooting with Unknown Period

We consider the following time-periodic problem for a system of ordinary differen-

tial equations (ODEs)

Mũ (t) = f(ũ(t)), t ∈ (0, T )

(1)
ũ(0) = ũ(T ),

where the period T > 0 and the vector ũ : [0, T ] → Rd , d ≥ 1 are sought. M is a
given non-singular mass matrix, f is a bounded and Lipschitz continous right-hand
side (RHS) function. Following [3] we incorporate the period T as an unknown
parameter by performing the change of variables

[0, T ] t
→ τ := t/T ∈ [0, 1]. (2)
Parallel-in-Time Calculation of Time-Periodic Solutions with Unknown Period 15

The problem (1) is thereby transformed into the equivalent one: find T > 0 and
u : [0, 1] → Rd such that
Mu (τ ) = T f(u(τ )), τ ∈ (0, 1)
(3)
u(0) = u(1).

The unit interval [0, 1] is then partitioned into N windows by the nodes 0 = τ0 <
τ1 < · · · < τN = 1. The n-th subinterval has length Δτn = τn − τn−1 , for n =
1, . . . , N.
For a given discrete variable Un−1 , we consider an initial value problem (IVP)
on the window (τn−1 , τn ]

Mun (τ ) = T f(un (τ )), τ ∈ (τn−1 , τn ]

(4)
un (τn−1 ) = Un−1

and let F (τn , τn−1 , Un−1 , T ) denote the solution operator of (4) for n = 1, . . . , N.
A sketch of the piecewise-defined solution due to the interval splitting is shown
in Fig. 1. In order to eliminate the jumps at the synchronization points τn , n =
1, . . . , N − 1 as well as the difference between the initial value at τ0 and the final
one at τN the matching conditions:

F (τN , τN−1 , UN−1 , T ) − U0 = 0,
(z) := (5)
F (τn , τn−1 , Un−1 , T ) − Un = 0, n = 1, . . . , N − 1
 
have to be satisfied, where z = U 
0 , . . . , UN−1 , T . System (5) represents the
root-finding problem for the mapping : RNd+1 → RNd . The Jacobian of is
given by
⎡ ⎤
−I GN gN
⎢G1 −I g1 ⎥
⎢ ⎥
J (z) = ⎢ .. .. .. ⎥ , (6)
⎣ . . . ⎦
GN−1 −I gN−1

Fig. 1 Example of the interval splitting within the multiple shooting for N = 5. The mismatches
at the synchronization points τn , n = 1, . . . , N − 1 together with the periodicity jump between
the solution at τ0 and τN are eliminated (up to a prescribed tolerance) by solving the root-finding
problem
16 I. Kulchytska-Ruchka and S. Schöps

where

∂F ∂F
Gn = (τn , τn−1 , Un−1 , T ), gn = (τn , τn−1 , Un−1 , T ), (7)
∂Un−1 ∂T

n = 1, . . . , N and I is the identity matrix. The root z of (5) can then be calculated
using the Newton method, i.e., for a given z(0) and k = 0, 1, . . . solution z(k+1) at
the iteration k + 1 is updated through
   
J z(k) Δz(k) = − z(k) , (8)
z(k+1) = Δz(k) + z(k). (9)

Note that due to the introduction of the additional variable T the system of
Eq. (8) is underdetermined and can be solved, e.g., by calculating the Moore-
Penrose pseudoinverse. A generalized eigenvalue-based gauging as well as the
corresponding theory for the Moore-Penrose pseudoinversion was presented in [12].
We note that in case when the size of (8) is large it can be condensed to a d-
dimensional system with d + 1 unknowns by block Gaussian elimination [3].

3 Periodic Time-Parallelization with Coarse Grid Correction

Inheriting the idea of the Parareal algorithm [7, 11] we approximate the derivative in
Gn (7) in a finite difference way using a coarse propagator G , i.e., for the iteration
k and n = 1, . . . , N

(k) ∂F  (k)  (k+1) (k) 

n ΔUn−1 =
G(k) τn , τn−1 , Un−1 , T (k) Un−1 − Un−1
∂Un−1 (10)
 (k+1)   (k) 
≈ G τn , τn−1 , Un−1 , T (k) − G τn , τn−1 , Un−1 , T (k) .

Similar to the fine propagator F , the operator G solves the IVP (4) on each
time window. However, in contrast to the fine solver the coarse propagator has a
considerably lower precision, e.g., it uses a lower-order time integrator or bigger
time steps. Substituting (10) into (8) we obtain the periodic Parareal with periodic
coarse problem [6] with unknown period (PP-PC-UP):
⎡ ⎤
⎡   (k)
⎤ U(k+1) ⎡ (k) ⎤
−I GN ·, T (k) gN 0
⎢ (k+1) ⎥ b
⎢   (k) ⎥ ⎢ U1 ⎥ ⎢ N(k) ⎥
⎢G1 ·, T (k) −I g1 ⎥ ⎢ . ⎥ ⎢ b1 ⎥
⎢ .. ⎥ ⎢ .. ⎥ = ⎢ .. ⎥
⎥ ⎢ ⎥ ⎢
⎢ .. .. ⎥,
⎣ . . . ⎦ ⎢ (k+1) ⎥ ⎣ . ⎦
  ⎣UN−1 ⎦
GN−1 ·, T (k) −I g(k)
N−1 b(k)
T (k+1) N−1

(11)
Parallel-in-Time Calculation of Time-Periodic Solutions with Unknown Period 17

(k) (k)  (k)   (k)

where we denote bn := gn T (k) + G τn , τn−1 , Un−1 , T (k) − F τn , τn−1 , Un−1 ,
    
T (k) and Gn ·, T (k) := G τn , τn−1 , ·, T (k) for n = 1, . . . , N. Following [3] we
have
τn
∂F  (k)  ∂  (k) −1

g(k)
n = τn , τn−1 , Un−1 , T (k)
= Un−1 + M T (k) f(u(τ ))dτ
∂T ∂T
τn−1

τn
−1
  
f(u(τ ))dτ ≈ M−1 Δτn f F τn , τn−1 , Un−1 , T (k) ,
(k)
=M
τn−1
(12)
for n = 1, . . . , N. In the general case, the system of Eq. (11) is nonlinear and
implicit, which requires an additional linearization.
Building upon the ideas presented in [10], which dealt with the time-periodic
problem for a known given period T , we encorporate an additive splitting of the
system matrix in (11). For this let us introduce a modified coarse propagator G¯,
which instead of (4) solves an approximate model with a linearized function f̄(u) =
Au + c on the RHS, i.e.,

Mun (τ ) = T f̄(un (τ )) = T [An un (τ ) + cn ], τ ∈ (τn−1 , τn ]

(13)
un (τn−1 ) = Un−1

with a given Jacobi-matrix An and a vector c. Having the linear coarse model we
construct a fixed point iteration: for s = 0, 1, . . .
⎡ ⎤
⎡   (k) ⎤ U(k+1,s+1) ⎡ (k+1,s) ⎤
−I ¯
GN ·, T (k) gN 0
⎢U(k+1,s+1) ⎥ hN
⎢ ¯   (k) ⎥ ⎢ ⎥ ⎢ (k+1,s) ⎥
⎢G1 ·, T (k) −I g1 ⎥ ⎢ 1 . ⎥ ⎢h1 ⎥
⎢ .. ⎥ ⎢ .. ⎥=⎢ . ⎥
⎢ .. .. ⎥⎢ ⎥ ⎢ . ⎥
⎣ . . . ⎦ ⎢ (k+1,s+1) ⎥ ⎣ . ⎦
  ⎣UN−1 ⎦
¯
GN−1 ·, T (k) −I
(k)
gN−1
(k+1,s)
hN−1
T (k+1)

(14)
   
where h(k+1,s) := b(k) ¯
n + G τn , τn−1 , Un−1
(k+1,s)
, T (k) − G τn , τn−1 , U(k+1,s) , T (k)
n (k)    n−1
and G¯n ·, T := G¯ τn , τn−1 , ·, T (k) for n = 1, . . . , N. Assuming that G¯
solves (13) with the implicit Euler method using a single step on (τn−1 , τn ] and
that all the windows have the same length Δτ , we have an explicit representation
for the coarse solution
   
1/Δτ · M − T (k) A G¯ τn , τn−1 , Un−1 , T (k) = 1/Δτ · MUn−1 + T (k) cn ,
(k+1,s) (k+1,s)

(15)
18 I. Kulchytska-Ruchka and S. Schöps

for n = 1, . . . , N. Denoting by C := 1/Δτ ·M and Q(k) := C−T (k) A and plugging

this into the system (13) we obtain
⎡ ⎤
⎡ ⎤ U(k+1,s+1) ⎡ (k) (k+1,s) ⎤
−Q(k) C Q(k) gkN 0
⎢U(k+1,s+1) ⎥ Q hN − T (k) cN
⎢ ⎥⎢ ⎥ ⎢ (k) (k+1,s) ⎥
⎢ C −Q(k) Q(k) gk1 ⎥ ⎢ 1 . ⎥ ⎢ Q h1 − T (k) c1 ⎥
⎢ ⎥ ⎢ . ⎥ = ⎢ ⎥.
⎢ .. .. .. ⎥⎢ . ⎥ ⎢ .. ⎥
⎣ . . . ⎦ ⎢ (k+1,s+1) ⎥ ⎣ . ⎦
⎣ UN −1 ⎦
C −Q(k) Q(k) gkN −1 Q(k) h(k+1,s) − T (k) c
N −1 N −1
T (k+1)

Remark 1 We note that when the period T is given within the problem setting (1),
the corresponding block-cyclic matrix (system matrix of (3) without the last
column) can be transformed into a block-diagonal using the frequency domain
transformation [2]. A detailed description of the approach as well as a Newton-like
linearization of the periodic system within the parallel-in-time setting is presented
in [10].

4 Numerical Example

We now consider the Colpitts oscillator model presented in [9]. It is described by

the circuit illustrated in Fig. 2, which consists of an inductance, a bipolar transistor,
as well as of four capacitances and four resistances. The Colpitts oscillator model
was exploited in the multi-rate context in [14].

R2

4
R1
L C1 = 50 pF, C2 = 1 nF,
C1 C4 C3 = 50 nF, C4 = 100 nF,
2 Uop
R1 = 12 kΩ, R2 = 3 Ω,
R3 = 8.2 kΩ, R4 = 1.5 kΩ,
1 C3
C2 L = 10 mH, Uop = 10 V.
3
R3
R4

Fig. 2 Circuit of the Colpitts oscillator model [9]

Parallel-in-Time Calculation of Time-Periodic Solutions with Unknown Period 19

The mathematical model of the circuit is given by an implicit system of ODEs

[9], namely, we search for the four node voltages U = [U1 , U2 , U3 , U4 ] s.t.
⎡ ⎤⎡ ⎤
1 0 0 0 U̇1
⎢0 C1 + C3 −C −C ⎥ ⎢
1 ⎥ ⎢U̇2 ⎥
⎥
⎢ 3
⎣0 −C3 C2 + C3 + C4 −C2 ⎦ ⎣U̇3 ⎦
0 −C1 −C2 C1 + C2 U̇4
⎡ ⎤ (16)
(U2 − U1 )R2 /L
⎢ (Uop − U1 )/R2 + xC h(U4 − U2 ) − IS h(U4 − U3 ) ⎥
=⎢⎣
⎥,
⎦
−U3 /R4 + xE h(U4 − U3 ) − IS h(U4 − U2 )
−U4 /R3 + (Uop − U4 )/R1 − yE h(U4 − U3 ) − yC h(U4 − U2 )

with the parameters yE = 10 µA, xE = 1.01 mA, IS = 1 mA yC = 20 µA, xC =

1.02 mA, and the nonlinear function h(x) = exp(x/UT )−1, UT = 2.585 V, coming
from the applied transistor model. Compared to the model introduced in [9], the
value of UT is chosen bigger to ease the convergence of PP-PC using the function h.
In practice, one may need appropriate homotopy or damping strategies, see [4]. The
transient behavior of the oscillator on [0, 1.125] ms is shown in Fig. 3 on the left.
The time step δT = 0.1125 µs and the initial value at t = 0 is u0 = [9.75, 1, 1, 1]
are chosen.
To find the periodic steady-state solution and the corresponding period T we
apply the iteration (14). Linearization of the nonlinear periodic system on the coarse

60
U1 U2
Rel. error ||U(k) − U(k− 1) ||/ ||U(k) ||

PP-PC
U3 U4 10− 1 PP-PC-UP
40
Voltage / V

20
10− 2

10− 3
− 20
0 0.2 0.4 0.6 0.8 1 1 3 5 7 9 11
Time / s ·10− 3 Iteration k

Fig. 3 Left: Transient behavior of the Colpitts oscillator until the steady state. Right: Convergence
of the PP-PC approach with a linearized coarse grid problem for the case when the period T is given
[10] and of PP-PC-UP when T is unknown (3)
20 I. Kulchytska-Ruchka and S. Schöps

level is performed using a surrogate linear model, i.e., G¯ solves the problem (13)
with f̄(U) = AU + c given by
⎡ ⎤
−R2 /L R2 /L
⎢ −1/R2 −xC /ŪT IS /ŪT (xC − IS )/ŪT ⎥
A=⎢⎣ 0
⎥,
⎦
IS /ŪT −1/R4 − xE /ŪT (xE − IS )/ŪT
0 yC /ŪT yE /ŪT −1/R3 − 1/R1 − yE /ŪT − yC /ŪT
(17)
 
c = 0, Uop /R2 , 0, Uop /R1 , (18)

with ŪT = 0.2585 V. The unit interval [0, 1] is split into N = 10 windows. The
coarse time step Δτ = 0.1 and the fine step δτ = 10−4 were chosen within the
time integration. The calculated period with the fixed point iteration (14) is T =
0.1125 ms. The right-hand part of Fig. 3 shows convergence of the PP-PC iteration
with the linearization from [10] for a given period T as well as for an unknown
period (PP-PC-UP). Both results are obtained up to the relative tolerance of 10−3 .
One can see that in case when T is known the method required less iterations, as
one would expect. When comparing the computational cost of the computations in
terms of the number of linear systems solves, PP-PC and PP-PC-UP delivered the
periodic solution effectively 4 and 3 times faster than the sequential time stepping,
respectively.

5 Conclusions

An iterative parallel-in-time method for solving time-periodic problems where the

period is not initially given is proposed in this paper. It complements the system of
equations originating from the Parareal-like algorithm with an additional variable
T and gives an underdetermined system of nonlinear equations. A linearization
using the fixed point iteration is applied on the coarse grid. The algorithm is verified
through its application to a Colpitts oscillator model.

Acknowledgments The authors thank Roland Pulch from Universität Greifswald for his assis-
tance with implementation and for the fruitful discussions on the Colpitts oscillator model.
This research was supported by the Graduate School CE within the Centre for Computational
Engineering at Technische Universität Darmstadt, as well as by DFG grant SCHO1562/1-2 and
BMBF grant 05M2018RDA (PASIROM).
Parallel-in-Time Calculation of Time-Periodic Solutions with Unknown Period 21

References

1. A. Bermúdez, D. Gómez, M. Piñeiro, P. Salgado, A novel numerical method for accelerating

the computation of the steady-state in induction machines. Comput. Math. Appl. 79(2), 274–
292 (2019)
2. O. Bíró, K. Preis, An efficient time domain method for nonlinear periodic eddy current
problems. IEEE Trans. Magn. 42(4), 695–698 (2006)
3. P. Deuflhard, Computation of periodic solutions of nonlinear ODEs. BIT 24, 456–466 (1984)
4. P. Deuflhard, Newton Methods for Nonlinear Problems: Affine Invariance and Adaptive
Algorithms (Springer, Berlin, 2004)
5. R.D. Falgout, S. Friedhoff, T.V. Kolev, S.P. MacLachlan, J.B. Schroder, Parallel time
integration with multigrid. SIAM J. Sci. Comput. 36(6), C635–C661 (2014)
6. M.J. Gander, Y.-L. Jiang, B. Song, H. Zhang, Analysis of two parareal algorithms for time-
periodic problems. SIAM J. Sci. Comput. 35(5), A2393–A2415 (2013)
7. M.J. Gander, S. Vandewalle, Analysis of the parareal time-parallel time-integration method.
SIAM J. Sci. Comput. 29(2), 556–578 (2007)
8. T. Hara, T. Naito, J. Umoto, Time-periodic finite element method for nonlinear diffusion
equations. IEEE Trans. Magn. 21(6), 2261–2264 (1985)
9. W. Kampowsky, P. Rentrop, W. Schmidt, Classification and numerical simulation of electric
circuits. Surv. Math. Ind. 2(1), 23–65 (1992)
10. I. Kulchytska-Ruchka, S. Schöps, Efficient parallel-in-time solution of time-periodic problems
using a multi-harmonic coarse grid correction (2019). arXiv: 1908.05245
11. J.L. Lions, Y. Maday, G. Turinici, A parareal in time discretization of PDEs. Comp. Rend. de
l’Académie des Sci. – Ser. I – Math. 332(7), 661–668 (2001)
12. R. Mirzavand Boroujeni, E.J.W. ter Maten, T.G.J. Beelen, W.H.A. Schilders, A. Abdipour,
Robust periodic steady state analysis of autonomous oscillators based on generalized eigen-
values, in Scientific Computing in Electrical Engineering SCEE 2010, ed. by B. Michielsen,
J.-R. Poirier. Mathematics in Industry, vol. 16 (2012), pp. 293–302
13. D.D. Morrison, J.D. Riley, J.F. Zancanaro, Multiple shooting method for two-point boundary
value problems. Commun. ACM 5(12), 613–614 (1962)
14. R. Pulch, Multi time scale differential equations for simulating frequency modulated signals.
APNUM 53(2–4), 421–436 (2005)
15. Y. Takahashi, T. Tokumasu, A. Kameari, H. Kaimori, M. Fujita, T. Iwashita, S. Wakao,
Convergence acceleration of time-periodic electromagnetic field analysis by singularity
decomposition-explicit error correction method. IEEE Trans. Magn. 46(8), 2947–2950 (2010)
On the Exactness of Rational Polynomial
Chaos Formulation for the Uncertainty
Quantification of Linear Circuits
in the Frequency Domain

Paolo Manfredi and Stefano Grivet-Talocia

Abstract We discuss the general form of the transfer functions of linear lumped
circuits. We show that an arbitrary transfer function defined on such circuits has a
functional dependence on individual circuit parameters that is rational, with multi-
linear numerator and denominator. This result demonstrates that rational polynomial
chaos expansions provide more suitable models than standard polynomial chaos for
the uncertainty quantification of this class of circuits.

1 Introduction

The polynomial chaos expansion (PCE) method [6] has emerged in the macromod-
eling and model-order reduction communities because of the remarkable accuracy
and efficiency in the uncertainty quantification by stochastic systems, including
electric and electronic circuits [3]. Stochastic output variables of interest are
approximated with a suitable polynomial model w.r.t. random input parameters,
from which statistical information is inexpensively extracted. While the method was
demonstrated to provide very high accuracy with a very limited expansion order
in many application scenarios, the modeling of resonant and/or distributed circuits
may require large orders and the accuracy of the calculated PCE coefficients may
be deteriorated by the large variability of the outputs.
A rational polynomial chaos (RPC) model with tensor-product truncation was
recently introduced [4] and was shown to provide better performance, compared
to the conventional single PCE with total-degree truncation that is used in most
engineering applications, specifically in electrical engineering [3]. In this work,
we show that the general form of any transfer function defined for a linear
lumped circuit is rational w.r.t. both frequency and element values. Specifically,
both numerator and denominator are multi-linear functions of element values.

P. Manfredi () · S. Grivet-Talocia

Politecnico di Torino, Torino, Italy
e-mail: [email protected]; [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 23

M. van Beurden et al. (eds.), Scientific Computing in Electrical Engineering,
Mathematics in Industry 36, https://doi.org/10.1007/978-3-030-84238-3_3
24 P. Manfredi and S. Grivet-Talocia

We provide a rigorous and formal proof of this fundamental theoretical result

that is somewhat well-known in electrical engineering [5], but unavailable in an
unambiguous and explicit form. Thanks to the theoretical findings herein presented,
we are able to show that the RPC model is exact and should be the method of choice
for lumped circuits.

2 Rational Polynomial Chaos Expansion

Given an arbitrary transfer function, generically denoted with Z and defined on a

linear lumped electrical circuit with d uncertain elements collected into vector ξ =
(ξ1 , . . . , ξd ), its RPC model reads [4]
L
=1 N (s)ϕ (ξ )
Z(s, ξ ) ≈ L (1)
1+ =2 D (s)ϕ (ξ )

where s is the Laplace variable (complex frequency). In (1), the basis functions
ϕ are multivariate orthogonal polynomials in the uncertain variables ξ , and the
coefficients N and D are computed using a linearized and iteratively re-weighted
regression. It was empirically shown [4] that, for the uncertainty quantification
of electric circuits, the RPC (1) is more accurate than the standard PCE [3]. The
purpose of this work is to provide a rigorous justification.

3 Transfer Functions of Linear Lumped Circuits

We review the basic modified nodal analysis (MNA) formulation [2] of lumped
linear time-invariant (LTI) circuits with RGLC components. The main objective
of this derivation is to reveal in explicit form the functional dependence on the
individual circuit parameters of any transfer function that can be defined on such
circuits.

3.1 Basic MNA Formulation for RGLC Circuits

Let us consider a lumped LTI P -port circuit with n nodes and b branches (one-
port elements). The branches are split into bR resistors with resistance Rk , bG
resistors with conductance Gk , bL inductors with inductance Lk , and bC capacitors
with capacitance Ck , where k is an index identifying individual components. We
distinguish between resistance-defined and conductance-defined resistors to allow
additive variations of either parameter. In addition, the last bJ = P branches are
Exactness of Rational Polynomial Chaos Formulation for the Uncertainty Quantification 25

assumed to represent the P ports of the structure. We place ideal current sources Jk
providing an excitation to the circuit, with the objective of characterizing the P × P
impedance matrix Z(s) in the Laplace domain by computing the corresponding port
voltages as outputs.
The branch voltage and current vectors v, i ∈ Rb are split according to element
types as
 T  T
v = v TR , v TG , v TL , v TC , v TJ , i = i TR , i TG , i TL , i TC , i TJ ,

where v ν , i ν ∈ Rbν for ν ∈ {R, G, L, C, J }, and where the passive sign convention
is used for each branch, including sources. The branch characteristic equations are
collectively written for each class of components as

vR = R i R R = diag(R1 , . . . , RbR ) (2a)

i G = G vG G = diag(G1 , . . . , GbG ) (2b)
d
vL = L iL L = diag(L1 , . . . , LbL ) (2c)
dt
d
iC = C vC C = diag(C1 , . . . , CbC ) (2d)
dt
i J = −J J = (J1 , . . . , JbJ )T . (2e)

Note that the current Jk of each source is incident into its positive node.
Circuit connectivity is described by the (reduced) incidence matrix A ∈ Rn−1,b ,
with the n-th node serving as reference for the definition of the set of nodal voltages
e ∈ Rn−1 . The incidence matrix columns are partitioned according to the branch
classes as
 
A = AR , AG , AL , AC , AJ . (3)

Combining Kirchhoff’s current law (KCL) equations Ai = 0 and Kirchhoff’s

voltage law (KVL) equations v ν = ATν e for ν ∈ {R, G, L, C, J } with the
characteristics (2), leads to the system of linear differential-algebraic equations

d
Gx + C x = Bu (4a)
dt
y = BT x, (4b)

which represents the standard MNA formulation. In (4), u = J denotes the port
currents, considered as inputs, y = v J denotes the corresponding port voltages,
26 P. Manfredi and S. Grivet-Talocia

considered as outputs, vector x ∈ Rm , with m = n − 1 + bR + bL , collects the MNA

variables e, i R , i L , and
⎛ ⎞ ⎛ ⎞ ⎛ ⎞
AG GATG AR AL AC CATC 0 0 AJ
G = ⎝ −ATR R 0 ⎠ , C=⎝ 0 0 0⎠ , B = ⎝ 0 ⎠. (5)
−ATL 0 0 0 0L 0

Throughout this work, we denote with 0 an all-zero matrix or vector, whose size is
inferred from the context.
The so-called stamps of the individual circuit elements in the MNA system (4)
are now easily characterized. A straightforward derivation shows that


ba 
bd
G(θ ) = G0 + (pk p Tk ) θk , C(ζ ) = (q k q Tk ) ζk , (6)
k=1 k=1

where
• ba = bR + bG is the number of adynamic components with values collected in
vector θ ∈ Rba , having elements {θk }bk=1
a
= {Rk }bk=1
R
∪ {Gk }bk=1
G
;
• bd = bL + bC is the number of dynamic components with values collected in
vector ζ ∈ Rbd , having elements {ζk }bk=1d
= {Lk }bk=1
L
∪ {Ck }bk=1
C
;
bG
• the constant vectors p k ∈ R collect the sets {p k }k=1 = {r k }bk=1
m ba R
∪ {gk }k=1
 T  T
individually defined as r k = 0, 1TbR ,k , 0 and g k = a TG,k , 0, 0 , where 1bν ,k
denotes the Euclidean basis vector in Rbν with all vanishing elements except the
k-th component equal to 1, and a G,k is the k-th column of AG ;
• the constant vectors q k ∈ Rm collect the sets {q k }bk=1
d
= {l k }bk=1
L
∪ {ck }bk=1
C
 T  T
individually defined as l k = 0, 0, 1TbL ,k and ck = a TC,k , 0, 0 , where a C,k
is the k-th column of AC ;
• the constant matrix G0 is defined as
⎛ ⎞
0 AR AL
G0 = ⎝−ATR 0 0 ⎠ . (7)
−ATL 0 0

3.2 Parameterization for Uncertainty Quantification

For the uncertainty quantification problem to be well posed, we assume that the
circuit is well defined and uniquely solvable for all parameter configurations, i.e.,
∃s ∈ C for which det(G(θ ) + sC(ζ )) = 0. Equivalently, the pencil (G, C) is regular
for any θ, ζ . We further consider a nominal parameter configuration θ = θ̄ and
Exactness of Rational Polynomial Chaos Formulation for the Uncertainty Quantification 27

ζ = ζ̄ . For instance, this nominal configuration can be considered as the set of

expected values of the circuit element values, assumed to be stochastic variables.
The initial hypothesis also implies unique solvability for this nominal parameter
configuration, which is the only assumption required by the following derivations.
We introduce the variable transformation

θ = θ̄ + ε, ζ = ζ̄ + δ, (8)

where each element of vectors ε and δ is a zero-mean stochastic variable, and we

denote Ḡ = G(θ̄ ) and C̄ = C(ζ̄ ). Due to linearity, (6) can now be written as


ba 
bd
G = G(ε) = Ḡ + (pk p Tk ) εk , C = C(δ) = C̄ + (q k q Tk ) δk . (9)
k=1 k=1

We see that both the static (G) and the dynamic (C) MNA matrices are expressed
as a finite sum of rank-one updates with respect to the nominal circuit formulation.
Each rank-one update pertains to a single individual stochastic circuit element. The
corresponding constant rank-one matrices pk p Tk and q k q Tk are recognized as the
standard MNA stamps of the various circuit elements.
Let us now consider the Laplace-domain solution of (4), which in the present
case corresponds to the impedance matrix of the considered P -port element and
reads

N(s; ξ )
Z(s; ξ ) = BT [G(ε) + s C(δ)]−1 B = , (10)
D(s; ξ )

where we have collected all stochastic parameters in a single vector ξ having

elements {ξk }dk=1 = {εk }bk=1a
∪ {δk }bk=1
d
, with d = ba + bd being the total
number of uncertain circuit elements, as previously defined in Sect. 2. In (10),
the scalar denominator D(s; ξ ) coincides with the determinant of the MNA matrix
Y(s; ξ ) = G(ε) + s C(δ), whereas each element of the numerator N(s; ξ ) is a linear
combination of the determinants of the submatrices (minors) obtained from Y(s; ξ )
by deleting one row and one column.
We now provide an explicit characterization of the numerator and denominator
of (10). To this end, we collect all stochastic parameters in a diagonal matrix

Ξ = diag(ξ1 , . . . , ξd ), (11)

which we use to cast the MNA matrix in the compact form, by restating (9) as

Y(s; ξ ) = Ȳ(s) + U Ξ S(s). (12)

28 P. Manfredi and S. Grivet-Talocia

The matrix Ȳ(s) = Ḡ + s C̄ corresponds to the nominal configuration, and

   T
U= P Q , S(s) = P sQ , (13)

where the constant matrices P and Q collect as columns all the vectors pk and q k ,
respectively. From now on, we will omit the dependence on the Laplace variable s,
since we are interested in the dependence on the stochastic variables ξ .
We introduce two useful lemmas:
Lemma 1 Given a square invertible matrix X and two matrices U , V of compatible
size, we have

det(X + U V T ) = det(I + V T X −1 U ) · det(X).

The above Lemma 1 is known as matrix determinant lemma, see [1] for a proof.
Lemma 2 Let a matrix W ∈ Rn,n have elements in the form Wij = Fij + ξi Bij ,
where Fij , Bij are constants for i, j = 1, . . . , n, and ξi are independent parameters.
Then,

 
n
det(W ) = βk ξ αk , (14)
k =1

where αk ∈ {0, 1} ∀k, , and βk are real constants.

Proof We use an induction argument, noting that the statement is trivially verified
for n = 1. Assuming that the statement holds for size n − 1, we evaluate det(W ) for
size n, for which W reads
⎛ ⎞
F11 + ξ1 B11 F12 + ξ1 F12 · · · F1n + ξ1 B1n
⎜ F21 + ξ2 B21 F22 + ξ2 F22 · · · F2n + ξ2 B2n ⎟
⎜ ⎟
W =⎜ .. .. .. .. ⎟.
⎝ . . . . ⎠
Fn1 + ξn Bn1 Fn2 + ξn Fn2 · · · Fnn + ξn Bnn

Expanding det(W ) using Laplace’s formula along the first row, we get


n
det(W ) = (−1)1+j (F1j + ξ1 B1j )M1j , (15)
j =1

where M1j is the determinant of the submatrix of size n − 1 obtained by deleting

row 1 and column j from W . By the induction ansatz, we have

 
n
α
M1j = βk ξ k , αk ∈ {0, 1} ∀k, . (16)
k =2
Exactness of Rational Polynomial Chaos Formulation for the Uncertainty Quantification 29

Inserting (16) into (15) leads to


n  
n
α
det(W ) = (−1)1+j (F1j + ξ1 B1j ) βk ξ k
j =1 k =2


n 
n
α

n
α
 
n
α
= (−1)1+j F1j βk ξ k + B1j βk ξ1 ξ k = β̂k ξ k ,
k j =1 =2 =2 k =1

where αk ∈ {0, 1} for = 1, . . . , n and ∀k, and β̂k are constants. 


We are now ready to calculate the denominator D(s; ξ ) in (10) as
 
D = det Ȳ + U Ξ S . (17)

Applying Lemma 1 with V T = Ξ S and X = Ȳ, we have

 
D = det (I + Ξ B) · det Ȳ , (18)

−1  
where both B = S Ȳ U and det Ȳ depend only on s and are thus constant with
respect to the stochastic parameters ξ . We have
⎛ ⎞
1 + ξ1 B11 ξ1 B12 ··· ξ1 B1n
⎜ ξ2 B21 1 + ξ2 B22 ··· ξ2 B2n ⎟
⎜ ⎟
I + ΞB = ⎜ .. .. .. .. ⎟.
⎝ . . . . ⎠
ξn Bn1 ξn Bn2 · · · 1 + ξn Bnn

This matrix verifies the conditions of Lemma 2 with F = I . Therefore

 
n
α
det (I + Ξ B) = βk ξ k (19)
k =1

with αk ∈ {0, 1} for all k, , which in turn implies that

 
n
α
D(s; ξ ) = dk (s) ξ k , αk ∈ {0, 1} ∀k, . (20)
k =1

Due to the lumped nature of the system under consideration, the coefficients dk (s)
are polynomials in s of degree up to the dynamic order N of the circuit.
The same arguments used for the denominator D(s; ξ ) can be seamlessly adopted
to show that also the elements of the numerator matrix N(s; ξ ) in (10) have the same
30 P. Manfredi and S. Grivet-Talocia

structural dependence on frequency s and parameters ξ . Therefore, we conclude that

any element (i, j ) of the impedance matrix Z(s; ξ ) has the following structure
Nij
ak;ij (ξ )s k
Zij (s; ξ ) = k=0
N
, (21)
k=0 bk (ξ )s k

where all numerator and denominator coefficients ak;ij (ξ ) and bk (ξ ) have a

multi-linear dependence in the stochastic parameters, i.e., they are multivariate
polynomials in which each element of ξ appears with up to order one. In conclusion,
any impedance element is a rational function of any stochastic parameter ξi with
both numerator and denominator degrees that cannot exceed one.
Based on the above result, the RPC model (1) is exact for linear lumped circuits,
provided that the polynomial basis functions ϕ are multi-linear. This is readily
achieved by adopting a tensor-product truncation of order one [4]. By extension,
the model turns out to be more accurate also for distributed circuits and electro-
magnetic systems, albeit with higher-degree approximations, as was effectively and
empirically demonstrated based on a number of application examples in [4].

4 An Illustrative Example

We consider the filter of Fig. 1 (left), which is designed to exhibit both a band-
and a high-pass behavior. All 9 circuit elements are uncertain, with inductances and
capacitances having independent Gaussian variations with a 20% standard deviation
around the nominal values indicated in the schematic.
The right panel in Fig. 1 shows the variability of the insertion loss of the filter.
The gray lines are a subset of random samples from a reference Monte Carlo (MC)
simulation with 10,000 runs, whereas the solid blue line is the standard deviation
of the MC samples. The dashed red and green lines are the standard deviations
obtained with a conventional PCE having a maximum total degree of three, and

92 nH 1
port 1 116.6 pF port 2
0.8
250 pF

0.6
73.6 pF 73.6 pF
146 nH 146 nH 0.4

312.5 nH 312.5 nH 0.2

0
0 20 40 60 80 100

Fig. 1 Left: filter schematic. Right: variability of the insertion loss of the filter. Gray lines: MC
samples; solid blue, dashed red, and dashed green lines: standard deviation obtained with MC,
conventional PCE, and proposed RPC methods, respectively
Exactness of Rational Polynomial Chaos Formulation for the Uncertainty Quantification 31

with a tensor-product RPC model having a maximum degree of one, respectively.

The conventional model has 220 terms in its single PCE, and the corresponding
coefficients are calculated by means of an ordinary least square regression [3].
The RPC model has a total of 1023 terms (512 in the numerator and 511 in the
denominator), and the coefficients are calculated with an iterative linearized least-
square regression [4]. In both cases, we use a number of regression samples that is
twice the number of unknowns.
As expected, the RPC provides an exact model, and the result is therefore
consistent with the reference MC curve. This is further confirmed by the mean
squared deviation of the two models from the MC samples, which is 3.8574 × 10−2
and 2.6264 × 10−10 for the conventional PCE and the RPC model, respectively.

5 Conclusions

This work presented a formal derivation that any frequency-domain transfer func-
tion defined on linear lumped circuits is a rational function with multi-linear
dependence on the circuit element values. This results provides a rigorous moti-
vation for using a Rational Polynomial Chaos (RPC) model for the uncertainty
quantification of the frequency-domain responses of electrical circuits, and more
generally of electromagnetic systems. Our findings are illustrated based on a lumped
filter example.
While a first-order tensor-product truncation provides an exact model for lumped
circuits, a more compact total-degree truncation (possibly of higher order) can be
used to improve the efficiency, especially for applications in which the exactness
no longer holds. This is the case, for example, of distributed, electromagnetic,
and/or photonic systems. We are also currently investigating a compression strategy,
based on principal component analysis, that avoids having to optimize the model
coefficients separately for each frequency.

References

1. D.A. Harville, Matrix Algebra From a Statistician’s Perspective (Springer, New York, 1997)
2. C.-W. Ho, A. Ruehli, P. Brennan, The modified nodal approach to network analysis. IEEE Trans.
Circ. Syst. 22(6), 504–509 (1975)
3. A. Kaintura, T. Dhaene, D. Spina, Review of polynomial chaos-based methods for uncertainty
quantification in modern integrated circuits. Electronics 7(3), 1–21 (2018)
4. P. Manfredi, S. Grivet-Talocia, Rational polynomial chaos expansions for the stochastic
macromodeling of network responses. IEEE Trans. Circ. Syst. I Reg. Papers 67(1), 225–234
(2020)
5. J. Vlach, K. Singhal, Computer Methods for Circuit Analysis and Design (Wiley, New York,
1983)
6. D. Xiu, Fast numerical methods for stochastic computations: a review. Commun. Comput. Phys.
5(2–4), 242–272 (2009)
Parallel-in-Time Simulation of Power
Converters Using Multirate PDEs

Andreas Pels, Iryna Kulchytska-Ruchka, and Sebastian Schöps

Abstract This paper presents a numerical algorithm for the simulation of pulse-
width modulated power converters via parallelization in time domain. The method
applies the multirate partial differential equation approach on the coarse grid of
the (two-grid) parallel-in-time algorithm Parareal. Performance of the proposed
approach is illustrated via its application to a DC-DC converter.

1 Introduction

Switch-mode power converters are devices which convert electric voltages or

currents between different levels. For this purpose they use transistors to switch on
and off the input voltage or current to obtain the desired average voltage or current
at the output of the converter. A technique called pulse-width modulation (PWM)
is often utilized to control the transistors, i.e., to generate the pulsed voltage from
a given carrier and reference. An exemplary circuit of a buck converter (DC-DC
converter) is depicted in Fig. 1a along with its solution in Fig. 2. It consists of fast
periodically varying ripples and a slowly varying envelope. The simulation of these
power converters with conventional time stepping is computationally expensive
since a high number of time steps is necessary to resolve the fast variations induced
by the transistor switching.
This paper proposes the simulation of power converters using a combination of
two methods, namely the parallel-in-time algorithm Parareal [7] and a multirate
approach based on Multirate Partial Differential Equations (MPDEs) [8]. This is
accomplished via the application of the MPDE approach on the coarse grid of
Parareal. It allows the coarse propagator to obtain a more precise solution given
the PWM input signal, in contrast to the standard coarse propagator when using a
large time step on the original system of equations.

A. Pels · I. Kulchytska-Ruchka () · S. Schöps

Computational Electromagnetics Group, Technical University of Darmstadt, Darmstadt, Germany
e-mail: [email protected]; [email protected];
[email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 33

M. van Beurden et al. (eds.), Scientific Computing in Electrical Engineering,
Mathematics in Industry 36, https://doi.org/10.1007/978-3-030-84238-3_4
34 A. Pels et al.

L RL 100
iL

voltage vi (V)
50
vi vC C R th
0 th
Ts D= Ts

0 2 4 6
time (ms)
(a) (b)

Fig. 1 Power converter model with pulsed voltage source: (a) Circuit of a simplified buck
converter. Transistor switching is modeled as pulsed voltage source. (b) PWM generated pulsed
voltage
Current iL (A) and voltage vC (V)

100

50

iL (A)
0 vC (V)
0 2 4 6 8 10
time (ms)

Fig. 2 Exemplary solution of the buck converter depicted in Fig. 1a. Switching frequency fs =
1/Ts = 5 kHz

The paper is organized as follows: first we introduce our model problem with
pulsed excitation in Sect. 2, then in Sect. 3 the Parareal method is summarized,
Sect. 4 proposes the usage of MPDEs as coarse propagators for Parareal that can
deal with pulsed right-hand sides and finally Sect. 5 discusses a numerical example
before concluding the paper.

2 Power Converter Model

Switch-mode power converters, which convert AC to DC, DC to AC, AC to AC, or

DC to DC voltages, are frequently used devices. They use power electronic switches
to periodically switch the input voltage on and off to regulate the output voltage.
For example a buck converter (DC-DC converter) transforms a given voltage to a
Parallel-in-Time Simulation of Power Converters Using Multirate PDEs 35

lower output voltage. It consists of a part that generates a pulsed voltage vi and
a filter circuit. The latter is shown in Fig. 1a. The pulsed voltage, see Fig. 1b, is
often generated using PWM. Important quantities defining the pulsed signal are the
switching period Ts and the duty cycle D which is the relation between the “on”-
time and the switching period. Given a reference signal r(t) and a carrier signal s(t)
the pulsed voltage is generated by

Vi  
vi (t) = sgn (r(t) − s(t)) + 1 , (1)
2
where sgn denotes the sign function and Vi is the amplitude. The converter circuit
is mathematically described by a system of ordinary differential or differential-
algebraic equations, e.g.,

d
A x(t) + B x(t) = c(t), t ∈ (t0 , T ], (2)
dt

with given initial value x(t0 ) = x0 , where x(t) ∈ RNs is the unknown solution
vector consisting for example of currents and voltages, A, B ∈ RNs ×Ns are matrices,
and c(t) ∈ RNs is the right-hand side containing current and voltage sources,
e.g., the pulsed voltage vi (t). The system may be assembled from lumped element
descriptions based on loop or (modified) nodal analysis as described in [2]. Please
note, that we focus on the linear case but the approach can be straight-forwardly
generalized, e.g., considering B = B(x).
The solution of power converters, as for example the one shown in Fig. 2,
exhibits the multirate phenomenon: slow variations in the solution require large time
intervals, i.e., a large end time point T , while the fast dynamics due to the switching
enforce small time steps. This is the motivation to turn to (parallel) methods that can
exploit this multirate behavior. In the following, we focus on the settling process
until the steady state is reached. If one is interested only in the latter, then other
methods may also be used, for example the application of Parareal for time-periodic
problems is a natural generalization of this work, see, e.g., [5].

3 Parareal Algorithm

Parareal is an iterative algorithm which is able to accelerate the solution of (2)

via parallelization in time. The method originates from [7] and its superlinear
convergence is proven in [3]. The two main ingredients of Parareal are the fine
and the coarse propagators. We denote by F (t, t0 , x0 ) and G (t, t0 , x0 ) the solutions
of the initial value problem (IVP) (2) at t ∈ (t0 , T ] obtained with sequential time
stepping using fine and coarse time steps, respectively.
36 A. Pels et al.

Partitioning the time interval t0 = T0 < T1 < · · · < TN = T we write the

Parareal iteration: for k = 0, 1, . . . and n = 1, . . . , N solve
(k+1)
x0 = x0 , (3)
 (k)   (k+1)   (k) 
x(k+1)
n = F Tn , Tn−1 , xn−1 + G Tn , Tn−1 , xn−1 − G Tn , Tn−1 , xn−1 . (4)

The solution operator F is assumed to deliver a very accurate solution (e.g., using a
numerical time-integration method with small time steps δT ) and can be executed in
parallel, while G gives rough information about the solution using a cheap method
(e.g., using a numerical method with large time steps ΔTi = Ti+1 − Ti ) and has to
be calculated sequentially, cf. (4).
A difficulty in applying Parareal to solve problems with PWM input is that a
naive implementation of a coarse propagator using a time-integrator with large
time steps will not capture the high-frequency dynamics and may also fail to
propagate low-frequency components. A modified Parareal algorithm which still
approximately captures the high-frequency behavior was introduced in [4]. The
idea is to separate the high-frequency (pulsed) components from the low-frequency
components, i.e.,
d
A x(t) + B x(t) = c̄(t) + c̃(t), (5)
dt ! "# $
=c(t )

where c̄ can be given as the sum of a few low-frequency sinusoids from a (fast)
Fourier transform and c̃(t) := c(t) − c̄(t) is the remainder. This allows to define a
reduced coarse propagator G¯fft which solves
d
A x(t) + B x(t) = c̄(t) (6)
dt

and gives rise to a modified Parareal update formula with coarse propagator G¯fft
in (3)–(4). This modified method converges reliably but possibly with reduced order
[4]. In this paper we propose an alternative method to perform time integration by
using the MPDE approach as the coarse propagator.

4 Multirate PDEs

The MPDE approach, which is used for obtaining the coarse solution in Parareal
uses the MPDE concept [1]. For the given problem the solution can be conveniently
decomposed into a slowly varying envelope and fast periodically varying ripples
using the solution expansion [8]

Np
. 
%
xj (t1 , t2 ) = yj,k (t1 )wk (τ (t2 )) = w(τ (t2 ))yj (t1 ), (7)
k=1
Parallel-in-Time Simulation of Power Converters Using Multirate PDEs 37

where yj,k (t1 ) are slowly varying coefficients and wk (τ (t2 )) are a finite set of basis
functions (k = 1, . . . , Np ) whose periodicity is accounted for by the relative time
τ (t2 ) = Tt2s mod 1. Its application to (2) yields
& '
x(t1 , t2 ) ∂%
∂% x(t1 , t2 )
A + x(t1 , t2 ) = %
+ B% c(t1 , t2 ), (8)
∂t1 ∂t2

where the relation between the original (2) and the MPDE (8) solution and right-
hand side are given by

%
x(t, t) = x(t), %
c(t, t) = c(t). (9)

This implies that if a solution to (8) is found, the solution of (2) can be extracted
from it. The specification of a suitable multitime source %c(t1 , t2 ) has to be supplied
by the user. However, the method converges to the correct solution for any choice
that fulfills (9) but it may not be more efficient than conventional time stepping.
A suitable choice for PWM excitations is discussed in Sect. 5. Now, applying a
Galerkin approach along the fast time scale t2 leads to the enlarged equation system

dy
A + B y(t1 ) = C(t1 ) , (10)
dt1

where the matrices are given by [8]

1
A = A ⊗ J, with J = Ts w(τ ) w(τ ) dτ,
0

1
∂w(τ ) 
B = B ⊗ J + A ⊗ Q, with Q=− w (τ ) dτ,
∂τ
0
 Ts
C= %
c(t1 , t2 ) ⊗ w(τ (t2 )) dt2 .
0

Suitable basis functions, which can well represent the ripples in the power converter
solution, are, e.g., B-Splines with suitable continuity or the PWM basis functions
[6]. The latter are global polynomial ansatz functions with w1 (τ, D) = 1, w2 (τ, D)
piecewise linear and wk (τ, D) is obtained recursively by integrating wk−1 (τ ) and
orthonormalizing for 3 ≤ k ≤ Np , see Fig. 3. It has been shown in [8] that they are
capable of very effectively representing the ripples in linear problems.
Finally, Eq. (10) can be time-stepped along t1 by using much larger time steps
than are needed to solve (2) since the fast variations are taken into account by the
basis functions. The accuracy of the solution (reconstructed using (7)) increases with
Np . However increasing Np also makes each time step of an implicit method more
38 A. Pels et al.

2 w1 Vi
w2
1 w3
w4
0

−1

−2 0
0 D 1 0 D 1
relative time τ relative time τ
(a) (b)

Fig. 3 Construction of basis functions with cusp at relative switching time D: (a) PWM basis
functions on relative time interval and (b) right-hand side

costly since an enlarged linear equation system has to be solved. Nevertheless, even
with very few basis functions the reconstructed solution can be expected to capture
the main features of the exact solution. This motivates the introduction of another
coarse propagator G¯mpde in Parareal which solves (10) and extracts afterwards the
single-time solution according to (7).

5 Numerical Experiments

The proposed approach is applied to the academic example of the buck converter
(see Fig. 1a). Its circuit is described by the IVP (2) given by




L 0 RL 1 vi (t)
A= ; B= and c(t) = , (11)
0C −1 1/R 0

with inductance L = 10−3 H, capacitance C = 10−4 F, resistances RL = 10−2 

and R = 0.8 . The PWM input vi (t) has the amplitude of Vi = 100 V and is
generated by a sawtooth carrier signal s(t) = tfs mod 1 with switching frequency
of fs = 5 kHz and the reference signal r(t) = 0.7 according to (1). The considered
time interval [0, 12] ms is partitioned into N = 40 windows for all Parareal variants.
The coarse time step size is ΔT = T /N = 3 × 10−4 s and the fine propagator
uses the time step δT = 10−6 s. All solutions are obtained with the implicit Euler
method.
First, the classical Parareal method (3)–(4) is applied. It solves the original
system (2) with the PWM input in both propagators, i.e., G and F . It is compared
to two variants where G is changed to: 1. G¯fft which solves system (6) containing
only the DC component instead of the PWM signal on the right-hand side (modified
Parallel-in-Time Simulation of Power Converters Using Multirate PDEs 39

PWM
maximal relative l2 error
10− 1 DC
MPDE 1
10− 2 MPDE 3
10− 3
10− 4
10− 5
10− 6
10− 7
1 2 3 4 5 6 7 8 9
number of iterations k
Fig. 4 Convergence of Parareal towards the sequential (reference) solution using different coarse
propagators for the given example

Parareal [4]); 2. G¯mpde which solves (2) using the MPDE approach with Np = 1 and
Np = 3 with the right-hand side% c(t1 , t2 ) = c(t2 ).
The maximal relative l2 error w.r.t. the sequential (reference) solution xseq (t)

x,k (ti ) − xseq (ti )2

max with t0 ≤ ti := t0 + iδT ≤ T
i∈N xseq (ti )2

is depicted in Fig. 4 for all the considered approaches  ∈ {PWM, DC, MPDE 1,
MPDE 3} at iteration k. The conventional Parareal converges for our test case (11)
up to a relative error of 10−6 in 9 iterations which is remarkable since the time step
of the coarse propagator does not resolve the dynamics of the PWM input and also
violates smoothness assumption, see [4] for details.
This method requires 2 700 and 360 sequential solutions of linear algebraic
systems of size Ns = 2 on the fine and the coarse levels, respectively, or 3 060
linear systems in total. By the number of sequential solves we mean the number
of solver calls which cannot be carried out in parallel (communication costs are
neglected). The approaches using the DC component and the MPDE approach with
Np = 1 both required 8 iterations (2 400 fine and 320 coarse solves, or in total
2 720 solutions of linear systems in 2 variables). Finally, the MPDE approach with
Np = 3 basis functions on the coarse level converged after 7 iterations, thereby
solving 2 100 linear systems of size Ns = 2 on the fine level and 280 linear systems
of size Ns × Np = 6 on the coarse level. One observes that the conventional coarse
propagator requires always roughly 1-2 Parareal iterations more than the MPDE
40 A. Pels et al.

approach with Np = 3 to obtain the same accuracy, e.g., MPDE 3 needs 3 instead
of 5 iterations for an error of 4 × 10−4 .
From Fig. 4 we see that Parareal with coarse propagator G¯mpde using a constant
basis function, i.e., Np = 1 and the modified Parareal with G¯fft using only the
DC excitation perform very similarly (if not identically). This resemblance is not
surprising since the MPDE 1 approach with Np = 1 computes only the envelope of
the solution, which is conceptually similar to the modified Parareal with a smooth
(in this case constant) coarse input. Finally, further tests show that the exploitation
of more basis functions (Np > 3) does not improve the convergence of Parareal,
they are similar to the case Np = 3.

6 Conclusions

In this paper we introduced a novel parallel-in-time algorithm, able to treat systems

excited by pulse-width modulated signals. The method extends the two-grid Parareal
algorithm by exploiting the MPDE solution approach on the coarse grid. It was
applied to the time-domain simulation of a buck converter supplied by a PWM
voltage source. Future research will further investigate the similarity of Parareal
with the MPDE coarse propagator and the modified Parareal as well as higher order
MPDE approaches as coarse propagators.

Acknowledgments The authors thank Ruth Vazquez Sabariego from KU Leuven for many fruitful
discussions on the MPDE approach. This research was supported by the Graduate School CE
within the Centre for Computational Engineering at Technische Universität Darmstadt, as well as
by DFG grant SCHO1562/1-2 and BMBF grant 05M2018RDA.

References

1. H.G. Brachtendorf, G. Welsch, R. Laur, A. Bunse-Gerstner, Numerical steady state analysis

of electronic circuits driven by multi-tone signals. Electr. Eng. (Arch. für Elektrotech.) 79(2),
103–112 (1996)
2. D. Estévez Schwarz, C. Tischendorf, Structural analysis of electric circuits and consequences
for MNA. Int. J. Circ. Theory Appl. 28(2), 131–162 (2000)
3. M.J. Gander, E. Hairer, Nonlinear convergence analysis for the parareal algorithm, in Domain
Decomposition Methods in Science and Engineering XVII, ed. by U. Langer, M. Discacciati,
D.E. Keyes, O.B. Widlund, W. Zulehner (Springer, New York, 2008), pp. 45–56
4. M.J. Gander, I. Kulchytska-Ruchka, I. Niyonzima, S. Schöps, A new parareal algorithm for
problems with discontinuous sources. SIAM J. Sci. Comput. 41(2), B375–B395 (2019)
5. M.J. Gander, I. Kulchytska-Ruchka, S. Schöps, A new parareal algorithm for time-periodic
problems with discontinuous inputs, in Domain Decomposition Methods in Science and
Engineering XXV. Lecture Notes in Computational Science and Engineering (Springer, New
York, 2019)
Parallel-in-Time Simulation of Power Converters Using Multirate PDEs 41

6. J. Gyselinck, C. Martis, R.V. Sabariego, Using dedicated time-domain basis functions for the
simulation of pulse-width-modulation controlled devices – application to the steady-state regime
of a buck converter, in Electromotion 2013, 2013
7. J.L. Lions, Y. Maday, G. Turinici, A parareal in time discretization of PDEs. Comp. Rend. de
l’Académie des Sci. – Ser. I – Math. 332(7), 661–668 (2001)
8. A. Pels, J. Gyselinck, R.V. Sabariego, S. Schöps, Efficient simulation of DC-DC switch-
mode power converters by multirate partial differential equations. IEEE J. Multiscale Multiphys.
Comput. Tech. 4(1), 64–75 (2019)
 Part II
Device Simulation
A Maximum Principle for Drift-Diffusion
Equations and the Scharfetter-Gummel
Discretization

Kai Bittner, Hans Georg Brachtendorf, Tobias Linn,

and Christoph Jungemann

Abstract The solution of the drift-diffusion equation does not satisfy a maximum
principle in general. Here it is shown that a maximum principle can be established
for the so called Slotboom variable, which permits statements on uniqueness,
stability, and positivity. This maximum principle is preserved for the discretized
system obtained by the Scharfetter-Gummel scheme.

1 Introduction

A maximum principle states that the solution of certain partial differential equations
attains its maximum on the boundary of the solution domain. Usually, if a
maximum principle holds, there follows also a corresponding minimum principle
by straightforward arguments (as changing signs). The maximum principle implies
several properties as uniqueness and stability of solutions. Furthermore, it often
ensures positiveness of physical quantities, as e.g. electron and hole densities in
semiconductors, where negative values would be non-physical.
A drift-diffusion equation, used e.g. to model the transport of electrons and
holes in a semiconductor, does not exhibit a maximum principle for the densities in
general. The densities in semiconductor devices may vary by orders of magnitude
due to huge differences in doping concentrations. On the over hand the drift-
diffusion equations ensures the positivity of densities. In a numerical scheme this
positivity shall still be guaranteed, which is indeed the case for the Scharfetter-
Gummel discretization [1]. However, the situation looks different if one considers
the Slotboom variable instead of the density.

K. Bittner () · H. G. Brachtendorf

University of Applied Sciences of Upper Austria, Hagenberg, Austria
e-mail: [email protected]; [email protected]
T. Linn · C. Jungemann
Institute of Electromagnetic Theory, RWTH Aachen University, Aachen, Germany
e-mail: [email protected]; [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 45

M. van Beurden et al. (eds.), Scientific Computing in Electrical Engineering,
Mathematics in Industry 36, https://doi.org/10.1007/978-3-030-84238-3_5
46 K. Bittner et al.

A maximum principle for Fermi potentials, which implies the maximum prin-
ciple for the Slotboom variables was shown in [2] for static drift diffusion
equations without recombination term. An estimate for the Slotboom variables in
the stationary semiconductor equations (van Roosbroeck system) was established by
Markowich [3, Theorem 3.2.1]. Here, the bound depends on the Dirichlet boundary
values, but do not yield a Maximum principle as presented by us in Theorem 1. A
discretized version of this estimate for a Scharfetter-Gummel finite volume scheme
can be found in [4]. An estimate for the time dependent van Roosbroeck system
is given in [5], where it is shown that the solution of a Scharfetter-Gummel finite
volume scheme is bounded by constants depending only on boundary and initial
data.
The results in this paper were motivated by the investigation of extended drift-
diffusion equations (see e.g. [6]) containing a time derivative of the flux and the
convective derivative. This equations are hyperbolic and do therefore not satisfy
a maximum principle. Thus it is a challenge to ensure positivity of densities for
the original equation as well as for discretizations. However, the study of the drift
diffusion equation and the Scharfetter-Gummel discretization lead to a maximum
principle with improved bounds for continuous solutions. This maximum principle
and its implications are presented in Sect. 2. In Sect. 3 we show how this maximum
principle is preserved by the Scharfetter-Gummel scheme.

2 A Maximum Principle for the Drift Diffusion Equation

We consider the drift-diffusion equation for the steady state:

 
∇ T ∇n(x) − n(x)∇φ(x) = d(n(x), x), (1)

where n is the unknown particle density, φ is a potential, and d is a source term (e.g.
for carrier generation and recombination in semiconductors). Here we have used
without loss of generality a scaling of quantities and equations which simplifies the
formulation. In the sequel Ω ⊂ Rn will be an open domain, Ω denotes its closure
and ∂Ω = Ω \ Ω is the boundary. We split the boundary into the closed Dirichlet
boundary ΓD ⊂ ∂Ω, with ΓD = ∅, and the Neumann boundary ΓN = ∂Ω \ ΓD .
We consider the Dirichlet boundary conditions

n(x) = g(x), x ∈ ΓD (2)

for the density and Neumann boundary conditions

 
ν T ∇n(x) − n(x)∇φ(x) = h(x), x ∈ ΓN (3)

for the flux, where ν denotes the outer normal vector.

Maximum Principle for Drift-Diffusion and Scharfetter-Gummel Discretization 47

The densities n(x) do not satisfy a maximum principle in general. In particular,

rapid changes of the potential (caused e.g. by differences of the doping concentra-
tion in semiconductors) result in changes of the density n(x) such that a maximum
on the boundary is not guaranteed. However, the Slotboom variable [7]

ñ(x) := e−φ(x) n(x) (4)

takes into account changes of the potential. For the equilibrium one has e.g. a
constant Slotboom variable. In particular, in terms of the Slotboom variable the flux
is written as

∇n − n∇φ = eφ ∇ ñ, (5)

such that (1) becomes an elliptic problem in terms of ñ(x), which allows us to
formulate a maximum principle for the Slotboom variable. One formulation could
be obtained from [8, Theorem A.1], however we will present sharper bounds here.
Theorem 1 Let n : Ω → R be a continuous solution of (1) on the open domain
Ω ⊂ Rn . For the Neumann boundary conditions (3) we require h(x) ≤ 0, x ∈ ΓN .
If d(n, x) ≥ 0, n ≥ n0 (x), x ∈ Ω, then the Slotboom variable ñ(x) satisfies
& '
−φ(x)
ñ(x) ≤ max n0 (x)e , max ñ(y) , x ∈ Ω. (6)
y∈ΓD

Proof Let us first assume that the maximum is not attained in ΓD but in x ∗ ∈ Ω
and n(x ∗ ) > n0 (x). For any sufficiently small ε > 0 0ne has n(x) > n0 (x) for
x ∈ Bε (x ∗ ) ⊂ Ω, where Bε (x ∗ ) := {y ∈ Rn : y − x ∗  < ε} is the ε-ball around
x ∗ . Furthermore, from ñ(x) ≤ ñ(x ∗ ) it follows that ν T ∇ ñ(x) ≤ 0 for x ∈ ∂Bε (x ∗ )
with ε > 0 sufficiently small. Without loss of generality we can further assume
that ν T ∇ ñ(x) < 0 on a subdomain of ∂Bε (x ∗ ) ⊂ Ω.1 Then we obtain from the
divergence theorem that
 
 
0≤ d(n(x), x) dx = ∇ T eφ(x) ∇ ñ(x) dx
Bε (x ∗ ) Bε (x ∗ )
(
= eφ(x) ν T ∇ ñ(x) dS < 0,
∂Bε (x ∗ )

which is a contradiction.

1 If ν T ∇ ñ(x) = 0, then ñ(x) is constant in a neighborhood of x ∗ and one replaces x ∗ by a point

from the boundary of that neighborhood. In the particular case that ñ(x) is constant on the entire
domain (6) follows immediately.
48 K. Bittner et al.

If we assume that the maximum is attained in x ∗ ∈ ΓN , we choose ε > 0 such

that Bε (x ∗ ) ∩ ΓD = ∅ is satisfied in addition to the conditions above. Now we
conclude analogously
 
 
0≤ d(n(x), x) dx = ∇ T eφ(x) ∇ ñ(x) dx
Bε (x ∗ )∩Ω Bε (x ∗ )∩Ω
( (
= e ν ∇ ñ(x) dS +
φ(x) T
eφ(x) ν T ∇ ñ(x) dS < 0,
∂Bε (x ∗ )∩Ω Bε (x ∗ )∩ΓN
! "# $ ! "# $
<0 ≤0

which is again a contradiction. 


Remark 1 An analogous statement holds for the minimum if we have d(n, x) ≤ 0
for n < n0 (x) and h(x) ≥ 0. For h(x) = 0, x ∈ ΓN , we obtain upper and lower
bounds. As a particular result we obtain the positivity of densities for positive g(x)
and non-negative h(x).
Remark 2 The above assumptions reflect the simulation of semiconductors, where
the drift-diffusion equation is used to describe the movement of electrons (or
holes) in an electric field generated by the potential φ (see e.g. [9] for a detailed
treatment). Dirichlet boundary conditions occur at semiconductor-metal interfaces
with positive boundary values, while Neumann boundary conditions correspond
to semiconductor-insulator interfaces where the normal of the current density is
h(x) = 0.
The right hand side it then the generation-recombination rate, typically given as

d(n, x) = r(n, p) (np − Nintr

2
),

where p is the hole density, Nintr the intrinsic density, and r(n, p) is a model
dependent positive factor. That is, the assumptions of Theorem1 are satisfied with
N2
n0 (x) = p(x)
intr
.
In the complete semiconductor model n, p, and φ are solutions of the Rosen-
broeck system, consisting of two drift diffusion equations for electrons and holes as
well as the Poisson equation for the potential. However the assumption of arbitrary
p and φ might be justified in a numerical scheme, as e.g. the Gummel iteration.
From the above theorem follows immediately a stability result
Corollary 1 Let ni : Ω → R, i = 1, 2 be continuous solutions of (1), which both
fulfill Neumann boundary conditions (3) for the same arbitrary h(x). If d(n, x), is
monotonically increasing with respect to n, then the Slotboom variables satisfy
) ) ) )
)ñ1 (x) − ñ2 (x)) ≤ max )ñ1 (y) − ñ2 (y)), x ∈ Ω.
y∈ΓD
Maximum Principle for Drift-Diffusion and Scharfetter-Gummel Discretization 49

Proof Let us first assume that the maximum is attained in x ∗ ∈ Ω. Without loss of
generality we assume that ñ1 (x ∗ ) > ñ2 (x ∗ ). Analogously to the proof of Theorem 1
we conclude
 (
 
0≤ d(n1 (x), x)−d(n2 (x), x) dx = eφ(x) ν T ∇ ñ1 (x)−ñ2 (x) dS < 0,
Bε (x ∗ ) ∂Bε (x ∗ )

which is a contradiction.
If we assume that the maximum is attained in x ∗ ∈ ΓN the proof is analogous
using that
 
ν T ∇ ñ1 (x) − ñ2 (x) = 0, x ∈ ΓN .



The above corollary states that small distortions of the Dirichlet boundary
conditions will result only in small distortions of the solution. Although it is a
statement for the Slotboom variable it provides also a stability statement for the
densities based on relation (4). As a particular result we obtain here the uniqueness
of the boundary value problem for the drift-diffusion equation.
For the time dependent drift-diffusion equation
   
∂
∂t n(x, t) = ∇ T ∇n(x, t) − n(x, t)∇φ(x, t) − d n(x, t), x, t , (7)

the situation is more involved. However, we can still show the positivity of the
densities, under suitable assumptions.
Theorem 2 Let n : Ω × (0, T ) → R be a continuous solution of (7) on the open
domain Ω × (0, T ). We require positive Dirichlet (2) and nonnegative Neumann (3)
boundary conditions, i.e., g(x, t) > 0 and h(x, t) ≥ 0, as well as positive initial
conditions n(x, 0) > 0, x ∈ Ω. If there is a δ > 0 with d(n, x, t) ≤ 0, x ∈ Ω ∪ ΓN ,
t ∈ (0, T ), n ≤ δ, then

n(x, t) > 0, x ∈ Ω ∪ ΓN , t ∈ (0, T ).

Proof We assume there is a (x ∗ , t ∗ ) such that n(x ∗ , t ∗ ) = 0, while n(x, t) > 0 for
t < t ∗ , x ∈ Ω. As in the proof of Theorem 1 we conclude

 ∗ 
∇ T eφ(x,t ) ∇ ñ(x, t ∗ ) − d(n(x, t ∗ ), x, t ∗ ) dx > 0
Bε (x ∗ )∩Ω

for any sufficient small ε > 0. Here we have assumed without loss of generality
that ñ(x, t ∗ ) attains positive values in any neighborhood of x ∗ , which is admissible
due to g(x, t ∗ ) > 0. This implies that ∂t∂ n(x ∗ , t ∗ ) > 0, i.e., there is a t < t ∗ , with
n(x ∗ , t) < 0, which is a contradiction. 

50 K. Bittner et al.

3 Discretization by Scharfetter-Gummel and Finite Volumes

The Scharfetter-Gummel scheme was introduced to avoid non-physical behavior

resulting from conventional discretization techniques, e.g., central differences. The
approach is based on the integration of the flux along the interval between two
discretization nodes under the assumption of constant current density and electric
field. Here, we present a different derivation which is based on midpoint finite
differences and equality (5) for the Slotboom variable.
Consider the adjacent nodes xi and xj in a discetization grid. We denote by hij =
x −x
xj − xi  their distance, the unit vector νij = jhij i gives us the direction of the
xi +xj
link and zij = 2 is the midpoint. The flux in the direction of the link is given by
 
∇ν n e−φ ∇ν φ  
νijT J = ∇ν n − n ∇ν φ = −φ
=− −φ
∇ν n e−φ ,
e ∇ν e

where ∇ν := νijT ∇ denotes the directional derivative. Using midpoint finite

differences for all derivatives we obtain
 −φ(xj ) − n(x ) e −φ(xi )

1 φij n(xj ) e i
νij J (zij ) + O(hij ) = − hij
T 2
e−φ(xj ) − e−φ(xi )
 n(x ) n(xi ) 
φ j
= − hijij − = 1
n(xj )B(φ ij ) − n(xi )B(−φ ij ) =: Jij ,
1 − eφij e−φij − 1 hij

where φij = φ(xj ) − φ(xi ) and B(x) := exx−1 is the Bernoulli function. That is, we
have indeed obtained the Scharfetter-Gummel discretization (with an error of order
h2ij ). In terms of the Slotboom variables this becomes
   
Jij = 1
hij B(−φij ) n(xj ) e−φij − n(xi ) = 1
B(−φij ) eφ(xi ) ñ(xj ) − ñ(xi ) ,
hij
! "# $
wij

with
φ(xj ) − φ(xi )
wij = wj i = 1
−φ(x
>0
hij
e j ) − e −φ(xi )

(see also [10] for a different derivation).

To discretize (1) one uses a finite volume approach. Consider the grid nodes
X = {xk : k ∈ I} ⊂ Ω. We distinguish between the interior nodes X := Ω ∩ X =
{xk : k ∈ II } and Dirichlet and Neumann boundary nodes ∂XD := ΓD ∩ X = {xk :
k ∈ ID } = ∅ and ∂XN := ΓN ∩ X = {xk : k ∈ IN }, respectively. Here I, II , ID ,
and IN are suitable finite index sets. The set of links L = {xij := (xi , xj )} ⊂ X2
Maximum Principle for Drift-Diffusion and Scharfetter-Gummel Discretization 51

contains all pairs of adjacent nodes, with intersecting Voronoi cells ωi and ωj . The
finite volume approach leads to
 ( 

∇ T J (x) dx = ν T J (x) dS ≈ Aij Jij + AN
i h(xi ), i ∈ II ∪ IN .
ωi ∂ωi j :xij ∈L

where Aij and AN i are the size of ωi ∩ ωj and ωi ∩ ΓN . respectively. Note that for
interior nodes Ai = 0. That is, we obtain the equations
N

  
Aij wij ñj − ñi = d(ni , xi ) − AN
i h(xi ), i ∈ II ∪ IN , (8)
! "# $
j :xij ∈L
Wij

to determine the approximative solution ñi ≈ ñ(xi ) for i ∈ II ∪ IN . Note that

Wij = Wj i > 0.
Writing the left hand side of (8) in matrix vector notation M ñ one sees easily that
M = (mij ) is symmetric, weakly diagonal dominant with mii > 0 and mij ≤ 0,
i = j , as noted e.g. in [4]. From this relation we finally obtain a maximum principle
for the discretized equations.
Theorem 3 Let {ni : i ∈ II ∪ IN } be a solution of (8). If h(xi ) ≤ 0, i ∈ IN , and
d(n, xi ) ≥ 0, n ≥ n0 (xi ), i ∈ II ∪ IN , then
* +
−φ(xi )
ñi ≤ max n0 (xi )e , max ñj ) , i ∈ I.
j ∈ID

Proof We use a similar argument as for the proof of Theorem 1. Let us assume that
the maximum is attained for i ∈ II ∪ IN and ni > n0 (xi ). Then
  
0 ≤ d(ni , xi ) − AN
i h(xi ) = Wij ñj − ñi < 0,
! "# $
j :xij ∈L
<0

which is a contradiction. 

Using analogous arguments we obtain also discrete versions of Corollary 1.
Corollary 2 Let {ni : i ∈ II ∪ IN } and {mi : i ∈ II ∪ IN } be solutions of (8),
with identical h(x). If d(n, x) is monotonically increasing with respect to n, then
the Slotboom variables satisfy
) ) ) )
)ñi − m̃i ) ≤ max )ñj − m̃j ), i ∈ I.
j ∈ID

For the semi-discretized version of (7) we obtain also a discrete version of

Theorem 2 for the positivity of the densities.
52 K. Bittner et al.

Theorem 4 Let {ni (t) : i ∈ II ∪ IN } be a solution of

 
∂
∂t ni (t) = Aij nj (t)B(φij (t)) − ni (t)B(−φij (t)) + AN
i h(xi ) − d(ni (t), xi , t),
j :xij ∈L

i ∈ II ∪ IN ,

where ni (t) > 0, i ∈ ID ; h(xi ) ≥ 0, i ∈ IN , and ni (0) > 0, i ∈ II ∪ IN .

If there is a δ > 0 with d(n, xi , t) ≤ 0, i ∈ II ∪ IN , t ∈ (0, T ), n ≤ δ, then

ni (t) > 0, i ∈ I, t > 0.

Acknowledgments This project has been funded by Deutsche Forschungsgemeinschaft (DFG,

German Research Foundation) Ref. No.: JU406/14-1 and the Austrian Science Fund (FWF): I3130-
N30.

References

1. D.L. Scharfetter, H.K. Gummel, Large-signal analysis of a silicon read diode oscillator. IEEE
Trans. Electron Devices 16(3), 391–415 (1969)
2. J. Jerome, Analysis of Charge Transport: A Mathematical Study of Semiconductor Devices
(Springer, Berlin, Heidelberg, 1996)
3. P. Markowich, The Stationary Semiconductor Device Equations. Computational Microelec-
tronics (Springer, New York, 1986)
4. K. Gärtner, Existence of bounded discrete steady-state solutions of the van Roosbroeck system
on boundary conforming Delaunay grids. SIAM J. Sci. Comput. 31(2), 1347–1362 (2009)
5. M. Bessemoulin-Chatard, C. Chainais-Hillairet, A. Jüngel, Uniform L∞ estimates for
approximate solutions of the bipolar drift-diffusion system. in Finite Volumes for Complex
Applications VIII - Methods and Theoretical Aspects (Springer International Publishing,
Cham, 2017), pp. 381–389
6. Z. Kargar, T. Linn, D. Ruić, C. Jungemann, Investigation of transport modeling for plasma
waves in THz devices. IEEE Trans. Electron Dev. 63(11), 4402–4408 (2016)
7. J.W. Slotboom, Computer-aided two-dimensional analysis of bipolar transistors. IEEE Trans.
Electron Dev. 20(8), 669–679 (1973)
8. J. Jerome, T. Kerkhoven, L∞ stability of finite element approximations to elliptic gradient
equations. Numer. Math. 57(6–7), 561–576 (1990)
9. P. Farrell, N. Rotundo, D.H. Doan, M. Kantner, J. Fuhrmann, T. Koprucki, Drift-diffusion
models, in Handbook of Optoelectronic Device Modeling and Simulation, ed. by J. Piprek.
Series in Optics and Optoelectronics, vol. 2 (CRC Press, Boca Raton, 2017), pp. 733–771
10. R.E. Bank, D.J. Rose, W. Fichtner, Numerical methods for semiconductor device simulation.
IEEE Trans. Electron Dev. 30(9), 1031–1041 (1983)
Numerical Calculation of Electronic
Properties of Transition Metal-Doped
mWS2 via DFT

Chieh-Yang Chen and Yiming Li

Abstract In this work, we use the spin-polarized density functional theory (DFT)
to study the atomic structures of transition metal-doped monolayer WS2 (mWS2 ).
The structures of doped mWS2 are simulated via atomic relaxation which moves the
ions according to the interactive force between electrons and the ions until converge
condition is reached, where the Kohn-Sham equation is solved numerically. We do
reveal not only simulation flow but also the accuracy examination for the explored
mWS2 . The estimated physical properties are further described and discussed.

1 Introduction

Two-dimensional materials, the monolayer transition metal dichalcogenide disulfide

(TMD), feature a high on/off ratio, low power consumption, and thermal stability,
especially the direct energy band gap of monolayer structure becoming eye-catching
study issues. Our recent study revealed the key steps for the stability of doping sites
for discussing electronic properties of TMD materials [1–3]. For WS2 , although
some doping techniques on monolayer tungsten disulfide have been reported, they
only focused on the certain doping material [4, 5]; thus, in this work, we analyze
the doping sites, formation energy, work function, and charge transfer of mWS2
with 3d transition metals doped mWS2 . The considered doping materials consist
of scandium (Sc), titanium (Ti), vanadium (V), chromium (Cr), manganese (Mn),
iron (Fe), cobalt (Co), nickel (Ni), copper (Cu), and zinc (Zn). We use symbols
4 × 4 and 2 × 2 as the dimension of the supercell containing a dopant, and the
aforementioned effective doping concentrations are 2.04% and 7.69% in the atomic
percentage, corresponding to 7.13 × 1013 and 2.85 × 1014 cm−2 , respectively.
The doped mWS2 structure can be obtained via atomic relaxation process which
is an iteration of ion moving steps according to electron charge density. The ions are

C.-Y. Chen · Y. Li ()

Parallel and Scientific Computing Laboratory, Institute of Communications Engineering, National
Yang Ming Chiao Tung University, Hsinchu City, Taiwan
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 53

M. van Beurden et al. (eds.), Scientific Computing in Electrical Engineering,
Mathematics in Industry 36, https://doi.org/10.1007/978-3-030-84238-3_6
54 C.-Y. Chen and Y. Li

moved by conjugate gradient algorithm, where the direction decided by calculated

force and stress tensor. Then the electron distribution is updated according to new
ion locations. By solving the Kohn-Sham equation constructed from electrons’ and
ions’ relation, the total energy, force, and stress tensor are updated. These steps
are repeated until the force and stress tensor reach the convergence condition. The
exchange-potential term in the Kohn-Sham equation [6] is critical in DFT approach,
and it can include other corrections such as van der Waals correction. Deriving
from an approximated Schrödinger equation, the Kohn-Sham equation is solved
by blocked Davidson algorithm [7]. Notably, the suitable correlation-exchange
function is first examined by experimental values and the sampling k-points are
tested.

2 The Computational Model

We are interested in the electronic properties of advanced material such as mWS2 .

To achieve the correct doped mWS2 , the spin-polarized DFT is used for atomic
level structure relaxation process and studying electronic properties. The first
step of simulation flow is to construct the initial atomic geometry based on the
periodic boundary condition of a given system. In the second step, we calculate the
wavefunction and electron density in the ground state by self-consistent electronic
structure calculation. However, these electronic properties are not in the equilibrium
condition because of the Coulomb interaction between ions. We relax the ions
configuration to its equilibrium state based on the atomic forces calculated by
Hellmann-Feynman theorem using the ground-state wavefunction. Once the atomic
forces between arbitrary two atoms are smaller than a tolerance, we recalculate the
electronic properties (wavefunction distribution, electron density, band structure,
and density of states) based on the relaxed atomic structure and given wavevector.
Starting from the electronic Schrödinger equation and the Hartree–Fock approx-
imation [6] based on the molecular orbital theory, the Kohn-Sham equation in (1) is
a simplified form which mainly constructed from electron density. The Hohenberg-
Kohn theorem [8] states that the ground-state energy of a given potential distribu-
tion, i.e., atomic configuration, is a unique functional of electron density.

(Te + Vnuclear + VH at ree + Vxc )φi

 −ZI e2 
p2 n(r  ) 
=( i + ) ) + e2 ) ) d r + Vxc )φi
2mi )  ) )  ) (1)
I )ri − RI ) )r − r )

=εi φi , i = 1, 2, . . . , N,

where Te , Vnuclear , VH at ree , and Vxc are the kinetic energy operator of elec-
tron i, nucleus-electron potential energy, Hartree potential energy, and exchange-
correlation potential energy, respectively. pi , mi , I , ZI , e, ri , RI , εi , φi (r ), and
Numerical Calculation of Electronic Properties of Transition Metal-Doped. . . 55

N are the momentum operator, electron mass, the index of nucleus, the charges
of nucleus, the charge of electron, the position vector of electron, the position
vector of nucleus, the orbital energy, the Kohn-Sham orbital, and N non-interacting
electrons, respectively. The number of these equations depends on the number of
valence electrons which come all of the simulated atoms. The Vxc (r ) term can be
obtained by different approaches; for example, Perdew-Burke-Ernzerhof (PBE) [9]
is used as an exchange-correlation function after our intensive accuracy test. The
aforementioned equations are construct for single k-point. For sampling of Brillouin
zone, Monkhorst-pack [10] k-points centered at the Γ point (0, 0, 0) is generated:


3
ni + 1/2
k = bi , ni = 0, . . . , Ni − 1, Ni is even, (2)
Ni
i=1

where k,  bi , and Ni are the vector in k-space, the basis of reciprocal lattice, and the
mesh numbers for b1 , b2 , and b3 directions.
To solve the Kohn-Sham equation, the blocked Davidson algorithm[7] has been
considered in the numerical calculation. It consists of five steps: basis initialization,
subspace construction, residual vector calculation, correction vector calculation,
and subspace expansion. In the basis initialization, a set of orthonormal basis
ψi , i = 1, 2, . . . , m, ∀m ≥ n for the lowest n states are guessed and built. In the
subspace construction, the full-size Hamiltonian matrix H is projected on a set
of sub-matrices {H˜ ij = ψiT H ψi } and solved for the eigenpairs in the subspace
(H˜ ϕ k = εk ϕ k , ∀k = 1, 2, . . . , n). Next, the calculated eigenpairs are used for
residual vector calculation. The residual vector is defined as ri = (H − εi I)ϕi ,
while εi and ϕi are the eigenvalue and eigenvector of sub-matrices. We check the
individual element in the residual vector for the convergence. If the elements are
larger than tolerance, we calculate the correction vector based on the residual vectors
and the eigenpair of sub-matrices. However, several ways evolve to calculate the
correction vector and result in different branches of Davidson algorithm.
The correction vectors {g k , k = 1, 2, . . . , n} are given by gIk = (εk −
HI I )−1 rIk , I = 1, 2, . . . , N and normalized, while N is the number of determinants
m k
of H and r k = 
i=1 ϕi (H − ε )ψi . In the final step, the correction vectors
k

{g , k = 1, 2, . . . , n} orthonormalized against the set {ψi , i = 1, 2, . . . , m} using

k

Gram-Schmidt process and appended in the set {ψi } if the orthonormalized norm
value is larger than a threshold said 10−3 . The resulting number of basis might
increase by a, while 1 ≤ a ≤ n. The whole process returns to subspace construction
using the updated orthonormal basis {ψi , i = 1, 2, . . . , m , m = m + a} until the
residual vectors reach convergence.
Under the Kohn-Sham formalism, the DFT was developed based on the local
and semi-local functionals, such as local density and generalized-gradient approx-
imation. However, it may not work well in describing the long-range charge
dynamics such as van der Waals interaction. This may cause a significant inac-
curacy in delineating the system energy, lattice constant, and electronic prop-
56 C.-Y. Chen and Y. Li

Fig. 1 The comparison between experimental and simulation values from different van der Waals
models. (a) The displacement of lattice parameters, a and c. (b) The displacement of bandgap, Eg

erty of two-dimensional materials because it is the van der Waals interaction

that forms stacked structure in two-dimensional materials. Thus, for different
materials, the examination and comparison to experimental value are necessary.
For bulk WS2 , the comparisons between experimental[11] and simulation values
with different exchange-correlation functions are shown in Fig. 1. We examine
the exchange-correlation functions combining local-density approximation (LDA),
Perdew-Burke-Ernzerhof (PBE), and van der Waals correlation. The discussed
van der Waals model includes DFT-D2, DFT-D3, optB86b-vdW, optB88b-vdW,
optPBE-vdW[12], vdW-DF[13], and vdW-DF2[14]. Due to abundant exchange-
correlation functions have been established there for different atoms, Eq. (1)
with different van der Waals correlations are solved by using Vienna ab initio
Simulation Package (VASP)[15].
We normalize the displacement between simulation and experimental value:
Displacement(%) = (Xsim − Xexp )/Xexp × 100%, where Xsim and Xexp are values
from simulation and experiment, respectively. Figure 1a shows that LDA-optB86b-
vdW, PBE-DFT-D3, and PBE-optB86b-vdW can achieve very small displacement
of lattice parameters compared with the experimental data[11]. Among all lattice
parameters, PBE has the largest displacement of the lattice parameter c, but it
is merely about 7.89% overestimation. Notably, a relatively small displacement
along the lattice parameter a implies that the strain along horizontal direction
can be properly described. Figure 1b shows that these functions all underestimate
the bandgap compared with the experimental data[11]. PBE function shows the
smallest displacement of bandgap which only has 1.34% underestimation while
other exchange-correlation functions have large underestimation. Note that the
vdW-DF and vdW-DF2 corrections result in incorrect indirect bandgap position
for the bulk WS2 . Our calculation indicates that the bandgap of bulk WS2 mainly

depends on the strain of a - and b-direction instead of it of c-direction.
The previous examination of the different exchange-correlation functions is with
a large number of Brillouin zone sampling which is performed by Monkhorst-pack
k-points. To find proper numbers of k-points which can shorten computation time
Numerical Calculation of Electronic Properties of Transition Metal-Doped. . . 57

Fig. 2 (a) The ground state energies of bulk WS2 with different examined numbers of k-points.
 the numbers of kx and ky
Because bulk WS2 have the same length of lattice parameters a and b,
are set equal and it of kz varies. (b) A zoom-in plot of (a)

and preserve good accuracy, we fix the bulk WS2 structure cell and alter the number
of k-points. Since the lattice parameters a and b of bulk WS2 have the same length,
the numbers of kx and ky are set to be equal. Figure 2 shows the ground state
energies of bulk WS2 versus the different numbers of k-points. The k-points of
kx and ky became stable when they are larger than 8. Notably, they are sensitive
due to the nature of the two-dimensional materials, where the kz of 12 is the most
insensitive to kx and ky numbers. The numbers of suitable k-points is related to
the lattice parameter; lengths of a and c are 3.1532 and 12.323 angstrom from
the experiment [11], respectively. The length of c is nearly 4 times to a , thus the
suitable k-points of kz will be around one-fourth to it of kx . Since the monolayer
WS2 has an additional vacuum layer and similar lattice parameter a and c compared
to bulk WS2 , the numbers of kx and ky are the same as bulk WS2 but less for kz .
The number of kz of monolayer WS2 is set to 1 due to its long lattice parameter c.
Finally, 12 × 12 × 4 and 12 × 12 × 1 k-points for bulk and monolayer WS2 are
concluded, respectively.
Our structure relaxation flow of monolayer WS2 is shown in Fig. 3a. First,
we obtain the reliable bulk WS2 from previous examined settings, and the band
structure is shown in Fig. 3b. From the results of exchange-correlation function and
k-points examinations, they indicate the importance of correct structure relaxation
 thus, the atomic force along these two directions should be preserved.
along a and b;
For this reason, monolayer WS2 is built from relaxed bulk WS2 which adds a 10
angstrom thick vacuum layer along c and fixes the simulation cell. Figure 3c shows
verified calculated band structure of monolayer WS2 . Confidently, the calculated
energy bandgaps from this method are in agreement with the experiment [11]. Then,
we continuously analyze the TM-doped mWS2 . The discussed characteristics of
TM-doped mWS2 include formation energy, work function, and band structure. As
58 C.-Y. Chen and Y. Li

Fig. 3 (a) The structure relaxation flow of monolayer WS2 . The monolayer structure relaxation
is complete under fixed unit cell for maintaining same atomic force along a and b directions. The
cutoff kinetic energy is 500 eV; the force acting on each atom of relaxed structure is smaller than
0.01 eV/angstrom; the energy difference is less than 10−6 eV per atom. From the simulation, the
atom-projected band structure of (b) bulk WS2 and (c) monolayer WS2 are obtained. The color
bars indicate the weighting of band dominated by tungsten atoms [1]

shown in Fig. 4a, because the new structure is built from different materials, we
consider the formation energy formula:

Ef orm = Edoped,mW S2 − EmW S2 + ni μi , (3)

where Edoped,mW S2 and EmW S2 are the total energies of the doped mWS2 system
and the pristine mWS2 , ni and μi are the number of atom i added (-1) or removed
(+1) and the corresponding chemical potential, respectively. The four possible
doping sites are discussed, as illustrated in Fig. 4b. To study the effect result from
two different doping concentrations, 4 × 4 and 2 × 2 supercells with one doping
atom were built from pristine mWS2 . The definition of work function is the external
energy exciting an electron from the surface of solid material into the vacuum space.
It can be calculated from the energy difference between the simulated vacuum
energy and the Fermi level, i.e., Evacuum − Ef ermi .
Numerical Calculation of Electronic Properties of Transition Metal-Doped. . . 59

Fig. 4 (a) The formula of the formation energy calculation. The Edoped,mW S2 and EmW S2 are
total energies of relaxed doped- and undoped-monolayer WS2 , respectively. For the calculation of
chemical potential, μi , an atom is located in a large cell. The potential can be obtained until the
value is stable by keeping enlarging the cell size. (b) The structures of four possible doping sites.
The “TM” means the atom of doping material. The notations “I-” and “S-” mean interstitial and
substitutional sites, respectively. The structures are built from repeatedly extended monolayer cell
and add doping atom for different doping concentrations

3 Results and Discussion

From the atomic relaxed structures, we can plot the band structure according to the
solved eigen energies, as shown in Fig. 5a and b. Both two plots are shifted so that
the simulated Fermi energy is located at zero. Since our simulated Fermi energy
is located at the band which is occupied by the last valence electron, i.e., the band
contributed by the doping atom. For example, comparing to Fig. 5b and a show
that the Sc doping contributes additional bands between the original conduction
and valence band. The calculated formation energy of discussed doping sites with
two different concentration are summarized in Fig. 5c. The formation energies of
discussed interstitial sites are lower than that of the substitutional sites, it indicates
the structure stability. Figure 5d plots the work function of pristine and TM-doped
mWS2 with respect to two concentrations. For simplicity, only the results of I-T
doping site are shown. The titanium (Ti)-doped mWS2 has the lowest work function
with higher concentration while zinc (Zn)-doped has the highest one with both
concentrations. The Sc possesses the largest range of modulation of work function,
1.63 eV, among discussed doping species and concentrations. It implies that there is
high flexibility in tuning work function of mWS2 which is promising for the design
and fabrication of future advanced nano-devices.
 60

Fig. 5 The band structure of (a) a Sc substitutional doping which a tungsten atom is replaced and (b) the undoped monolayer WS2 . The solid red lines and
dotted orange lines are of spin-up and spin-down states, respectively. (c) The calculated formation energy of discussed doping sites. (d) The work functions of
TM-doped mWS2 with respect to different TM materials and concentrations. Here only shows the results of the doping site with the lowest formation energy.
The arrows indicate how the work function changes as the doping concentration increases
C.-Y. Chen and Y. Li
Numerical Calculation of Electronic Properties of Transition Metal-Doped. . . 61

4 Conclusions

In this work, the numerical method and simulation flow for studying doped
monolayer tungsten disulfide have been described. The studies are completed with
the examined exchange potential model and k-points sampling. The key simulated
results indicate the values of work function of Sc- and Cr-doped mWS2 have
relatively large flexibility for work function modulation via doping technology.

Acknowledgments This work was supported in part by the Ministry of Science and Technology,
Taiwan, under Grant MOST 108-2221-E-009-008, Grant MOST 108-3017-F-009-001, Grant
MOST 109-2221-E-009-033, Grant MOST-109-2634-F-009-030, and Grant MOST 110-2221-E-
A49-139, and in part by the “Center for mmWave Smart Radar Systems and Technologies”İ under
the Featured Areas Research Center Program within the framework of the Higher Education Sprout
Project by the Ministry of Education in Taiwan.

References

1. C.-Y. Chen, Y. Li, On electronic structure of monolayer tungsten disulfide doped by 3d

transition metals, in Proceedings of International Conference on Solid State Devices and
Materials SSDM 2019, Nagoya, September 2–5 (2019), pp. 161–162
2. Y.-C. Tsai, Y. Li, On electronic structure and geometry of MoX2 (X = S, Se, Te) and black
phosphorus by ab initio Simulation with various Van der Waals corrections, in Proceedings of
IEEE Int. Conf. Simul. Semicond. Process. Devices SISPAD 2017, Kamakura, September 7–9,
2017, pp. 169–172
3. Y.-C. Tsai, Y. Li, Impact of doping concentration on electronic properties of transition metal-
doped monolayer molybdenum disulfide. IEEE Trans. Electron Devices 65, 733–738 (2018)
4. S. Sasaki et al., Growth and optical properties of Nb-doped WS2 monolayers. Appl. Phys.
Express 9, 71201 (2016)
5. C. Sun et al., N-doped WS2 nanosheets: a high-performance electrocatalyst for the hydrogen
evolution reaction. J. Mater. Chem. A 4, 11234–11238 (2016)
6. W. Kohn, L.J. Sham, Self-consistent equations including exchange and correlation effects.
Phys. Rev. 140, A1133–A1138 (1965)
7. E.R. Davidson, Matrix eigenvector methods, in Methods in Computational Molecular Physics,
ed. by G.H.F. Diercksen, S. Wilson (Springer, Dordrecht, 1983), pp. 95–113
8. P. Hohenberg, W. Kohn, Inhomogeneous electron gas. Phys. Rev. 136, B864–B871 (1964)
9. J.P. Perdew, K. Burke, M. Ernzerhof, Generalized gradient approximation made simple. Phys.
Rev. Lett. 77, 3865–3868 (1996)
10. H.J. Monkhorst, J.D. Pack, Special points for Brillouin-zone integrations. Phys. Rev. B 13,
5188–5192 (1976)
11. R. Lv et al., Transition metal dichalcogenides and beyond: synthesis, properties, and applica-
tions of single- and few-layer nanosheets. Acc. Chem. Res. 48, 56–64 (2015)
12. J. Klimeš, D.R. Bowler, A. Michaelides, Chemical accuracy for the van der Waals density
functional. J. Phys. Condens. Matter 22, 022201 (2010)
13. G. Román-Pérez, J.M. Soler, Efficient implementation of a van der Waals density functional:
application to double-wall carbon nanotubes. Phys. Rev. Lett. 103, 96102 (2009)
14. K. Lee, Ë.D. Murray, L. Kong, B.I. Lundqvist, D.C. Langreth, Higher-accuracy van der Waals
density functional. Phys. Rev. B 82, 022201 (2010)
15. G. Kresse, J. Hafner, Ab initio molecular dynamics for liquid metals. Phys. Rev. B 47, 558–561
(1993)
Numerical Simulation of Thermal
Conductivity of Silicon Nanowires

Min-Hui Chuang and Yiming Li

Abstract To provide the sufficient power of trillion sensors in the era of internet-
of-things, the thermoelectric materials and devices have been of great interest
recently. In this paper, we construct a model for the periodic silicon nanowires
(SiNWs) embedded in Si0.7 Ge0.3 (SiNWs-Si0.7 Ge0.3 composite) and propose a
simulation flow for the calculation of its thermoelectric properties. The electron
band structure and phonon energy dispersion of SiNWs-Si0.7 Ge0.3 composite are
simulated by using the effective mass Schrödinger equation formulated by the Bloch
theorem and the elastodynamic wave equation, respectively. The aforementioned
equations are discretized by using the finite element method and the corresponding
eigenvalue problems are solved by the implicitly restarted Arnoldi method. Then,
the thermoelectric properties of SiNWs-Si0.7 Ge0.3 composite are estimated by
Landauer approach.

1 Introduction

For thermoelectric (TE) energy conversion materials in solid-state power generation,

the dimensionless figure of merit (FOM) given by

S2σ T
ZT = (1)
κph + κel

is used to indicate the TE performance, where S is the Seebeck coefficient, σ is the

electrical conductivity, κph is the thermal conductivity from phonon, and κel is the
thermal conductivity from electron. To increase the value of ZT , researchers are
either focus on reducing the thermal conductivity or enhancing the power factor,

M.-H. Chuang · Y. Li ()

Parallel and Scientific Computing Laboratory, Institute of Communications Engineering,
Department of Electrical and Computer Engineering, National Yang Ming Chiao Tung
University, Hsinchu, Taiwan
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 63

M. van Beurden et al. (eds.), Scientific Computing in Electrical Engineering,
Mathematics in Industry 36, https://doi.org/10.1007/978-3-030-84238-3_7
64 M.-H. Chuang and Y. Li

S 2 σ . The Landauer approach can be used to calculate the TE properties; however,

these properties are determined by the electron and phonon energy dispersions
but not independently controlled. Recent experiments have shown that the lattice
thermal conductivity would be significantly reduced without suffering from the
loss of power factor by using nanostructures [1–3]. With these nanostructures,
the lattice thermal conductivity can be suppressed due to the phonon boundary
scattering. Furthermore, alloys shows a better TE performance with respect to
nonalloys because of the reduction of the lattice thermal conductivity [4]. Studies
on nanocomposite TEs show that proper nanostructures in SiGe nanocomposite
materials can lead to a reduction in the thermal conductivity [5].
In this paper, the conductive matrix material of Si0.7 Ge0.3 is studied [6]. This
composite film consists of the SiNWs embedded in Si0.7 Ge0.3 and is modelled as a
periodic superlattice for the following simulation. The electron and phonon energy
dispersions are solved from two eigenvalue problems. To solve the Schrödinger
equation, researches give several approaches in which the finite-element discrete
variable representation plays an important role to get an efficient and highly accurate
result [7]. This paper is organized as follows. In Sect. 2, we show the simulation
structure and the physical setting. In Sect. 3, we show the numerical approach for
the calculation of the band profile and TE properties. A part of the simulation results
and discussions are illustrated in Sect. 4. Finally, we draw the conclusions in Sect. 5.

2 Computational Structure and Models

As shown in Fig. 1, the direction of the carrier and phonon transports is parallel to
the x-y plane so that the nanowires can play as interface for the phonon transport and
reduce the thermal conductivity. To simplify the simulation structure, we assume
that the nanowires are periodic with the radius r, the space s between the closest two
nanowires, and the height h, as shown in Fig. 1b. For the band profile calculation,

Fig. 1 (a) The three-dimensional (3D) schematic structure of the periodic nanowires. (b) The
geometry parameter of the simulation structure. (c) The CB and VB of SiNWs-Si0.7 Ge0.3
composite [10]. (d) The definition of irreducible Brillouin zone (IBZ), where , X, and M are
in the x-y plane of the k-space.  is the original point
Numerical Simulation of Thermal Conductivity of Silicon Nanowires 65

Table 1 The adopted physical parameters. m∗l and m∗t are the effective masses of electrons in the
longitudinal and transverse directions, respectively. They are used for the quantized energy band
calculation in the CB. Similarly, m∗hh and m∗lh are the effective masses of heavy holes and light
holes, respectively. They are used for quantized energy band calculation in the VB [10, 13, 14]
Electron mass Hole mass Elastic constant
Material Bandgap m∗l m∗t m∗hh m∗lh C11 C12 C44
Si 1.12 0.98 0.19 0.49 0.16 165.8 63.9 79.6
Si0.7 Ge0.3 1.00 1.14 0.12 0.41 0.10 154.6 59.2 75.8

we simplily consider the undoped situation, where the conduction and valence
bands are plotted in Fig. 1c. For an electron or a hole in a periodic potential, the
Bloch theorem [8] is used to describe the phase change; thus, the corresponding
Schrödinger equation with effective mass approximation is given by [9]

−2 i2  2 k 2
∇[ ∇ 
u k ] − k · ∇ 
u k (
r ) + [V (
r ) + ]
uk (r ) = En,k uk (r ), (2)
2m∗ m∗ 2m∗

where , m∗ , V (r ), En,k , and uk (r ) are the reduced Plank’s constant, the effective
mass, the position-dependent potential energy, quantum energy levels, and the
corresponding wave function, respectively. Notably, V (r ) is equal to the conduction
band (CB) or valence band (VB) for electrons and holes [10], respectively. In
addition, the phononic band structure is calculated by the elastodynamic wave
equation [11]

 u(r )] = ρω2 u(r ),

∇ · [C∇ (3)

where C is the elastic constant matrix, u is the Fourier transform of the displacement
vector [12], ρ is the mass density, and ω is the eigenfrequency, respectively. The
elastic constant matrix C describes second-order strain energy density. Since Si has
cubic symmetry, the number of independent elastic constants can be reduced to
C11 , C12 , and C44 [13]. The elastic constants of Si0.7 Ge0.3 are decided by the linear
interpolation of the values from Si and Ge, as listed in Table 1.

3 Simulation Techniques

To estimate the FOM in (1), by considering the physical transparency and the
computational efficiency of the solution method, the Landauer approach has
been implemented on the TE region [15–17]. In situations close to equilibrium,
the Landauer approach is mathematically equivalent to the Boltzmann transport
66 M.-H. Chuang and Y. Li

equation under the relaxation time approximation if the mean-free-path (MFP) for
backscattering is

2vx2 τ 
λ(E) = , (4)
|vx |

where v is the group velocity, τ is the momentum relaxation time, and the
subscription x represents the transport direction of the carriers or phonons [16].
Within the Landauer approach, the number of modes and the MFP for backscattering
are two important physically parameters. The calculation flows for the TE properties
related to the electrons and phonons are listed in Algorithms 1 and 2, respectively.
In Algorithm 1, the differential conductivity is given by [15]

2q 2 Me (E) ∂f0
σ  (E) = λe (E) (− ), (5)
h A ∂E

where 2q 2/ h is the quantum of the conductance, λe is the MFP of the electron

transport, Me /A is the number of modes per area A, and ∂f0 /∂E is the window
function. f0 is the equilibrium Fermi-Dirac function which is related to the band
structure in the CB. The differential lattice thermal conductivity in Algorithm 2 is
given by [15]

 π 2 kB2 T Mph (ω) 3 E 2 ∂n0

κph (ω) = λph (ω) ( 2 )( ) (− ), (6)
3h A π kB T ∂ω

where π 2 kB2 T /3h is the quantum of the thermal conductance, λph is the MFP of the
∂n0
phonon transport, Mph /A is the number of modes per area, and ( π32 )( kBET )2 (− ∂ω )
is the window function. n0 is the Bose-Einstein distribution which is related to the
phonon dispersion. The carriers are under the diffusion transport with the MFP for
back-scattering calculated by the Matthiessen rule, where the average MFP for back-
scattering without nanowire structure is extracted by setting the thermal conductivity
of 150 W/m-K from the measured data of bulk silicon [18].
For the calculation of the electronic and phononic band structures, the boundary
conditions of (2) and (3) are set periodically owing to highly periodical array of
SiNWs [6]. We solve these two discretized eigenvalue problems by the implicitly
restarted Arnoldi method [19]. The finite element method with Lagrange elements
is implemented to discretize the Schrödinger and elastodynamic equations. A finite
element is a triple including a geometry domain, a space function in this domain, and
a set of linear functionals (so-called the degree of freedom) [20]. The band structure
is calculated by sampling in k-space, more specifically, in the irreducible Brillouin
zone (IBZ) [21], as shown in Fig. 1d. The calculation flow to get the band diagrams
is listed in Algorithm 3.
Numerical Simulation of Thermal Conductivity of Silicon Nanowires 67

Algorithm 1: The Landauer approach for electronic TE proprieties

Require: E-k relationship and the energy upper limitation Em
Ensure: Calculate S, σ , and κe
1: while energy E < Em do
2: Calculate the differential electrical conductivity σ 
3: Conductivity σ + = σ 
4: κ0 + = (E − EF )2 σ 
5: Seebeck coefficient S+ = (E − EF )σ 
6: end while
7: κ0 = κ0 /q 2 T
8: S = −S/qT σ
9: κe = κ0 − T σ S

Algorithm 2: The Landauer approach for phononic TE proprieties

Require: ω-q relationship and the frequency upper limitation ωm
Ensure: Calculate κph
1: while frequency ω < ωm do
2: 
Calculate the differential lattice thermal conductivity κph
3: κph + = κph

4: end while

Algorithm 3: The band diagram calculation by sampling some specific points

in IBZ
Require: effective masses, the geometry structure, and the sampling points
Ensure: eigenvalues or eigenfrequencies
1: while There is at least one point in IBZ which has not been solved. do
2: Assign the value of k
3: Solve the eigenvalue or eigenfrequency problem at k
4: Record the eigenvalues or eigenfrequencies at k
5: end while
6: Connect the i th eigenvalue or eigenfrequency at all sampling points
7: Output the band structure

4 Results and Discussion

Because h is relatively ten times larger than r, we forcus on the 2D simulation.

The first ten lowest energies of electrons and light holes calculated from (2) are
shown in Fig. 2. When s = 2 nm, the barrier of the CB (Si0.7 Ge0.3 ) is with a small
width and the electrons are easy to tunneling. Thus, the ground state (E0 ) is with the
smallest energy compared with Figs. 2b and c. As the space between each nanowire
increases, the electrons are more localized in the finite energy well, which leads to
an increased ground state energy. For example, in -valley, the ground state energies
are 0.565, 0.569, and 0.570 eV with s =2, 15, and 50 nm, respectively. Similarly, the
band structure of light holes is clustered and the ground state energy of light holes
in the VB rise as the barrier becomes widers. As the space decreases, as shown in
68 M.-H. Chuang and Y. Li

Fig. 2 (a)–(c) The first ten energies of electrons in the CB. (d)–(f) The first ten energies of light
holes in the VB. Ei represents the ith energy state. For electrons, E0 is the ground state and E1 is
the first excited state

Fig. 3 (a)–(c) The energy dispersion of phonons in the SiNWs-Si0.7 Ge0.3 composite with s = 2,
15, and 50 nm, respectively

Fig. 3, the phonons meet more interfaces and the scattering rate is huge. In room
temperature, the low energy phonon plays an important role to carry heat. Thus, the
thermal conductivity can be expected to be decreased as s decreases.
Notably, the number of eigenvalues will influence the accuracy of the results.
Thus, to find an optimal sampling number with a minimal time cost, Fig. 4a shows
Numerical Simulation of Thermal Conductivity of Silicon Nanowires 69

Fig. 4 (a) The computational time versus the number of sampling points and (b) the relative error
versus the numbers of iterations with respect to different s, where the results are solved from the
Schrödinger equation. The lines are the results with s =2, 15, and 50 nm, respectively. The tested
PC is with an Intel® CoreTM i7–7500 CPU and the RAM is 16 GB

the computational time when solving (2) versus the sampling numbers with respect
to different s. For each s, we solve the first ten eigenvalues for electrons in the
CB. There are 894, 1126, and 1344 elements in our simulation with s=2, 15, and
50 nm, respectively; the sizes of corresponding matrices are 1856, 2329, and 2765,
respectively. The computational time will increase as the sampling points increase
linearly. Figure 4b shows the relative error between iterations of the implicitly
restarted Arnoldi method with respect to different s. The stopping criterion is
the relative error <10−6 . Notably, the accuracy of the computed eigenenergy is
almost the same when the matrix size increases from several thousands to ten
thousands; however, the time cost increases significantly. Not shown here, we have
the similar numerical tests when solving (3), where both the computational time and
convergence behavior are faster than that of (2).
By considering the doping effect, TE properties calculated via the Landauer
approach will vary as the Fermi level. The calculated lattice thermal conductivity
is about 2.2 W/mK in Algorithm 2, which is close to the experimentally measured
data of 3.5 W/mK [6] when the density of SiNWs is 1.6 × 1011 cm−3 (r = 5 nm and
s = 15 nm). For the SiNWs-Si0.7 Ge0.3 composite with p-type doping of 1.16×1015
cm−3 , ZT calculated from (1) is about 1.5 × 10−4 at room temperature.

5 Conclusions

In this paper, we have applied the numerical method to calculate the electronic and
phononic band structures of the silicon nanowires embedded in Si0.7 Ge0.3 by solv-
ing the Schrödinger equation and the elastodynamic wave equation. The Landauer
approach is used for the calcultion of thermoelectric properties, respectively. The
simulated thermal conductivity is close to the measurement.
70 M.-H. Chuang and Y. Li

Acknowledgments This work was supported in part by the Ministry of Science and Technology,
Taiwan, under Grant MOST 108-2221-E-009-008, Grant MOST 108-3017-F-009-001, Grant
MOST 109-2221-E-009-033, Grant MOST-109-2634-F-009-030, and Grant MOST 110-2221-E-
A49-139, and in part by the “Center for mmWave Smart Radar Systems and Technologies” under
the Featured Areas Research Center Program within the framework of the Higher Education Sprout
Project by the Ministry of Education in Taiwan.

References

1. Y.-M. Lin, X. Sun, M.S. Dresselhaus, Theoretical investigation of thermoelectric transport

properties of cylindrical Bi nanowires. Phys. Rev. B 62, 4610 (2000)
2. A.I. Hochbaum, R. Chen, R.D. Delgado, W. Liang, E.C. Garnett, M. Najarian, A. Majumdar, P.
Yang, Enhanced thermoelectric performance of rough silicon nanowires. Nature 451, 163–167
(2008)
3. P. Martin, Z. Aksamija, E. Pop, U. Ravaioli, Impact of phonon-surface roughness scattering on
thermal conductivity of thin Si nanowires. Phys. Rev. Lett. 102, 125503 (2009)
4. C. Bera, N. Mingo, S. Volz, Marked effects of alloying on the thermal conductivity of
nanoporous materials. PRL 104, 115502 (2010)
5. M.S. Dresselhaus, G. Chen, M.Y. Tang, R. Yang, H. Lee, D. Wang, Z. Ren, J.-P. Fleurial, P.
Gogna, New directions for low-dimensional thermoelectric materials. Adv. Mater. 19, 1043–
1053 (2007)
6. A. Kikuchi, A. Yao, I. Mori, T. Ono, S. Samukawa, Composite films of highly ordered Si
nanowires embedded in SiGe0.3 for thermoelectric applications. J. Appl. Phys. 122, 165302
(2017)
7. B.I. Schneider, Parallel solver for the time-dependent linear and nonlinear Schrödinger
equation. Phys. Rev. E 73, 036708 (2006)
8. P. Kratzer, J. Neugebauer, The basics of electronic structure theory for periodic systems. Front.
Chem. 7, 1–18 (2019)
9. M.-Y. Lee, Y. Li, S. Samukawa, Miniband calculation of 3-D nanostructure array for solar cell
applications. IEEE Trans. Electron Dev. 62, 3709–3714 (2015)
10. F. Schaffler, Silion-germanium, in Properties of Advanced Semiconductor Materials: GaN,
AlN, InN, BN, SiC, SiGe, ed. by M.E. Levinshtein, S.L. Rumyantsev, M.S. Shur (Wiley, New
York, 2001), pp. 149–188
11. R. Anufriev, M. Nomura, Thermal conductance boost in phononic crystal nanostructures. Phys.
Rev. B, 91, 245417 (2015)
12. G. Mascali, V. Romano, A hierarchy of macroscopic models for phonon transport in graphene.
Physica A 548, 124489 (2020)
13. W.-W. Zhang, H. Yu, S.-Y. Lei, Q.-A. Huang, Modelling of the elastic properties of crystalline
silicon using lattice dynamics. J. Phys. D 44, 335401 (2011)
14. Y. Shiraki, N. Usami, Silicon–Germanium (SiGe) Nanostructures: Production, Properties and
Applications in Electronics (Woodhead Publishing, Cambridge, 2011)
15. J. Maassen, M. Lundstrom, The Landauer approach to electron and phonon transport. ECS
Trans. 69, 23–36 (2015)
16. C. Jeong, R. Kim, M. Luisier, S. Datta, M. Lundstrom, On Landauer versus Boltzmann and full
Band versus effective mass evaluation of thermoelectric transport coefficients. J. Appl. Phys.
107, 023707 (2010)
17. L. Musl, E. Flage-Larsen, Thermoelectric transport calculations using the Landauer approach,
ballistic quantum transport simulations, and the Buttiker approximation. Comput. Mater. Sci.
132, 146–157 (2017)
Numerical Simulation of Thermal Conductivity of Silicon Nanowires 71

18. A.I. Hochbaum, R. Chen, R.D. Delgado, W. Liang, E.C. Garnett, M. Najarian, A. Majumdar, P.
Yang, Enhanced thermoelectric performance of rough silicon nanowires. Nature 451, 163–167
(2008)
19. R.B. Lehoucq, D.C. Sorensen, C. Yang, ARPACK Users’ Guide: Solution of Large-Scale
Eigenvalue Problems with Implicitly Restarted Arnoldi Methods (Society for Industrial and
Applied Mathematics, Philadelphia, 1998)
20. J. Sun, A. Zhou, Finite Element Methods for Eigenvalue Problems (CRC Press, New York,
2017)
21. M.S. Kushwaha, P. Halevi, L. Dobrzynski, B. Djafari-Rouhani, Acoustic band structure of
periodic elastic composites. Phys. Rev. Lett. 71, 2022 (1993)
A Novel Surface Mesh Simplification
Method for Flux-Dependent Topography
Simulations of Semiconductor
Fabrication Processes

Christoph Lenz, Alexander Scharinger, Paul Manstetten, Andreas Hössinger,

and Josef Weinbub

Abstract In etching and deposition simulations of a semiconductor fabrication

process the calculation of the surface rates of particles is an essential but also
the computationally most demanding step. A promising approach is to preprocess
the simulation domain by simplifying the surface. We thus propose a new surface
mesh simplification method that takes advantage of geometric domain-specific
surface properties that are prevalent in topography simulations. We compare our
method to a suitable reference algorithm and show that our method maintains
higher geometric accuracy and accordingly maintains the original geometry in great
detail. Furthermore, the evaluation of the simplified meshes show an enhanced
performance of the particle surface rate calculation.

1 Introduction

Process technology computer-aided design (TCAD) tools are used to simulate

fabrication processes of semiconductor devices. One important branch of process
TCAD is the evolution of the topography during etching and deposition processes.
In each time step the three essential computational tasks are: (a) the calculation
of the particle flux on the surface, which is used to (b) calculate the surface
velocity according to a surface model and (c) the calculation of the new position
of the surface using the surface velocities [1]. In Process TCAD the surface can be
represented implicitly using the level-set method where the domain is discretized on

C. Lenz () · A. Scharinger · P. Manstetten · J. Weinbub

Christian Doppler Laboratory for High Performance TCAD, Institute for Microelectronics, TU
Wien, Wien, Austria
e-mail: [email protected]; [email protected]; [email protected];
[email protected]
A. Hössinger
Silvaco Europe Ltd., St. Ives, UK
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 73

M. van Beurden et al. (eds.), Scientific Computing in Electrical Engineering,
Mathematics in Industry 36, https://doi.org/10.1007/978-3-030-84238-3_8
74 C. Lenz et al.

a regular grid. This approach is attractive due to the robust handling of topographical
changes in a level set framework [2]. The particle flux on the semiconductor surface
denotes the number of particles interacting on the surface. One possible numerical
method for calculating the surface flux is Monte Carlo ray tracing [3]. At practically
relevant surface resolutions the flux calculation dominates the overall execution
time of an etching or deposition simulation [1]. It is thus useful to investigate
approaches that speed up the flux calculation. One promising approach is to use
temporary explicit surface meshes as there exists a large body of knowledge about
ray tracing on explicit surfaces. The marching cubes algorithm [4] is commonly
used to extract an explicit surface from the level set. However, the resulting surface
meshes typically contain very narrow and long triangles (needles) or small triangles
in flat regions that contain no geometric variation. Therefore eliminating those
surface elements reduces the total surface element count which speeds up the ray
tracing tasks, further underlining the attractiveness of an explicit surface mesh
approach.
There exist several algorithms that reduce the resolution of surface meshes with
respect to a given metric; several metrics have been proposed in literature [5–8].
However, some of these algorithms try to simplify the geometry homogeneously
[5, 6] or use computationally expensive metrics [7, 8]. The latter is particularly
relevant when considering the entire etching or deposition workflow where the mesh
simplification has to be conducted at every single time step. Mesh simplification, or
more general domain simplification, is a commonly used approach in process TCAD
simulations [9, 10]. In particular, in [11] the authors evaluate the flux on a mesh by
sampling only a sparse set of surface elements to accelerate the simulation.
In this paper we introduce a flexible and computationally lightweight simplifi-
cation method based on the local surface curvature. We evaluate the impact of our
mesh simplification method on typical process TCAD topography simulations by
using the high performance ray tracing library Embree [12] by conducting a ray
tracing performance analysis. Specifically we compare the flux calculation time for
surfaces obtained with the presented method, with the flux calculation time obtained
for surfaces generated by the reference Lindstrom-Turk algorithm [5], by comparing
the execution time of the simplification process and the performance of the flux
calculation using Monte Carlo ray tracing.

2 Surface Mesh Simplification

The simplification method presented in this work is based on the Lindstrom-Turk

algorithm [5]. This algorithm uses an Edge Collapse procedure to simplify the
surface mesh. It offers a relatively low computational complexity and takes the
quality of triangles into account: The latter is particularly important for process
TCAD simulations, as the mesh quality directly influences subsequent procedures.
Our method uses the mean curvature of each vertex to partition the mesh into
regions. This allows us to adjust the amount of simplification according to the local
A Novel Mesh Simplification Method for Flux-Dependent Topography Simulations 75

geometric properties in each region. This simplification method has been designed
to simplify regions of the mesh offering negligible geometric variation (e.g. flat
areas) to a higher degree, thus allowing to maintain a higher resolution in regions of
the mesh with high geometric variation. Furthermore, our method is not limited to
the Lindstrom-Turk simplification algorithm, hence other simplification algorithms
[6] can be used in combination with our method.

2.1 Feature Detection

The first step in our simplification method is the detection of geometric features in
the mesh: We use the absolute mean curvature of each vertex and calculate it via a
discrete approximation of the Laplace-Beltrami operator [13] in the vertex xi

 j ∈N1 (i) (cot αij − cot βij )(xi − xj )
|H (xi )| = , (1)
4Aavg

where H (xi ) denotes the mean curvature in the vertex xi and N1 (i) is the set of
all vertices adjacent to xi . The angles αij , βij are the angles of the triangles that
share the edge between xi and xj , which are opposite to this edge and Aavg is
the average area of the triangles surrounding the vertex xi . The mean curvature
is used to categorize each vertex to be either a flat or a feature vertex. In particular,
an empirical threshold is used to identify vertices with small curvature (numerical
artifacts), which are considered to be flat.

2.2 Mesh Partitioning and Movement of Regions

The Mesh is partitioned into the feature regions and the transition regions according
to the metrics above. The feature region encompasses the triangles of the mesh
with significant geometric variation. The transition region contains the triangles
that do not hold information about the geometric variation. This partition of the
mesh allows to simplify the transition region to a greater extent, which reduces the
overall number of mesh elements without loosing information about the geometric
variation. Furthermore, this approach allows to keep a high resolution in regions of
the mesh with high geometric variation by simultaneously limiting the overall mesh
size in terms of number of triangles. However, simplifying the flat region to a higher
degree than the feature region leads to low quality triangles (e.g. needles).
To prevent the formation of low quality elements the transition region is
simplified with linearly increasing parameters, thus creating a reasonable mesh
grading. Figure 1 schematically depicts two steps of the discussed process. At first
the whole mesh, including the feature region, is simplified until the smallest edge
has an edge length of l0 . If the feature region should not be simplified l0 is set
76 C. Lenz et al.

transition
region
→ → →
feature
region

(1) (2) (3) (4)

Fig. 1 Example of the simplification process: (1) shows the mesh after it has been divided into
regions. (2) shows the simplification of the feature region. (3) shows the extension of the feature
region. (4) shows again the simplification of the transition region with an increased edge length

to 0. After this initial simplification step the transition region is simplified until
the smallest edge has an edge length of l1 = l0 + sl, where sl denotes the step
length. Next, the feature region is expanded into the transition region. Afterwards
the now smaller transition region is simplified until the smallest edge has an edge
length of li+1 = li + sl with i ∈ {0, 1, . . . , n ∈ N}. These last two steps continue
until the feature region cannot move any further into the transition region, and
thus terminates the simplification process. To avoid unwanted side effects of the
potentially large edge lengths produced by our iterative scheme, another parameter
lmax is used to terminate the refinement once the edge length li in the transition
region has reached lmax .
The parameter for the simplification of the feature region l0 , when using the level
set method, can be connected to the level-set and is chosen in concordance with
the minimal grid size Δt . When using meshes not originating from a level-set, this
parameter can be chosen by averaging the edge length of all feature vertices. We
have empirically determined that the step length sl should be approximately the
edge length of the feature region after the simplification with the parameter l0 stops.
A bigger step size increases the amount of edges that are removed. However, the
bigger the difference between the edge length of the feature region and the step
length, the worse the triangle quality of the mesh.

3 Results

The simplification method has been evaluated in the context of process TCAD
in three ways: geometric distance to the original geometry, execution time of
the simplification method, and the execution time of a subsequent surface flux
calculation by ray tracing. In this study two example geometries have been analyzed
and each example geometry has been simplified applying eight different degrees of
simplification, resulting in a reduction of vertices from 20–90%. Figure 2 shows the
A Novel Mesh Simplification Method for Flux-Dependent Topography Simulations 77

(a) (b)

Fig. 2 Process TCAD surface meshes simplified with our method. (a) Surface 1 with 78% of the
vertices of the original mesh removed by our simplification method. (b) Surface 2 with 52% of the
vertices of the original mesh removed by our simplification method

two surface meshes after they have been simplified with our method. The original
surface meshes of Surface 1 and Surface 2 have 70,831 and 175,550 vertices,
respectively. The performance benchmarks presented in the following are based on a
serial C++ implementation of our method executed on a 64bit GNU/Linux platform
equipped with an Intel Devil’s Canyon CPU.

3.1 Distance to Original Geometry

Surface mesh simplification introduces geometric distortions into the simplified

mesh. To measure the error introduced by the simplification process we use
the Hausdorff distance [14] between the original and the simplified mesh. The
Hausdorff distance is measured from each vertex of the original mesh to the
simplified mesh. Figure 3 shows the results for one test case of our analysis. The
distance to the original mesh is smaller when using our simplification method.
On average our simplification method has 20–40% lower Hausdorff distance to
the original geometry than using the Lindstrom-Turk algorithm. The reason for
the significantly improved Hausdorff distance is our method which allows to use
more vertices in areas of high geometric variation, allowing to represent the overall
geometry better.
78 C. Lenz et al.

Fig. 3 Hausdorff distance from each vertex of the original mesh to a mesh simplified with our
method and a mesh simplified using the Lindstrom-Turk algorithm. (a) Surface 1 with 78% of the
vertices of the original mesh removed by our simplification method. (b) Surface 2 with 52% of the
vertices of the original mesh removed by our simplification method

3.2 Time Spent on Simplification

The simplification method presented in this work introduces an overhead to the

simplification process. This overhead consists primarily of the feature detection, at
the start of the simplification process, and the movement of the feature regions. As
can be seen in Fig. 4 our simplification method takes on average 17% longer than
the Lindstrom-Turk algorithm.

3.3 Flux Calculation and Monte Carlo Ray Tracing

A common approach to compute the surface flux in a Process TCAD application is

to use a Monte Carlo simulation [15]. This is a randomized procedure and the results
of the Monte Carlo method are of stochastic nature. To compute the trajectories
Simplification time [sec]

39 42 47 52 67 73 78 81
Amount of simplification in [%]

(a) (b)

Fig. 4 Average simplification time of our method and the Lindstrom-Turk algorithm. The amount
of simplification denotes the number of vertices which have been removed from the original mesh.
(a) Surface 1. (b) Surface 2
A Novel Mesh Simplification Method for Flux-Dependent Topography Simulations 79

Execution Time [%]

314 sec =
100%
95%

90%

85%

80%

39 42 47 52 67 73 78 81
Amount of simplification [%]

(a) (b)

Fig. 5 Execution time of Monte Carlo ray tracing using 108 rays. The amount of simplification
denotes the number of vertices which have been removed from the original mesh. (a) Surface 1.
(b) Surface 2

on which particles move through the simulation domain (modeling the surface
flux) we use the Embree ray tracing library [12]. In Embree a bounding volume
hierarchy data structure [3, 12] is used to efficiently compute the paths on which
the particles move through space. The internal structure of the bounding volume
hierarchy depends eminently on the structure and the coarseness of the surface mesh.
Figure 5 shows the execution times measured to perform the Monte Carlo ray
tracing on the meshes with different degrees of simplification. As the simplified
meshes contain less triangles the bounding volume hierarchy data structure used for
ray tracing will have less elements than the data structure for the original mesh. As
the size of the data structure is decreased the memory footprint is reduced and this
leads to faster flux calculations because less data has to be processed and the caches
of the processor are used more effectively. Figure 5a and b show that the empirical
speedup in flux calculation depends on the shape of the surface mesh. When tracing
Surface 1, the meshes of both simplification methods perform approximately the
same and are faster than the original mesh. When tracing Surface 2, the meshes
generated by our simplification method clearly outperform the meshes simplified
with the Lindstrom-Turk algorithm and the original geometry. Surface 2 contains
deep trenches and the rays of the tracing algorithm need to travel towards the
bottom of these trenches. As the walls of the trenches do not have high curvature
the bounding volume hierarchy data structure created from the mesh simplified
with our method will be less complex within the deep trenches and hence, the
traces of the rays down the trench can be computed by performing less operations.
Also, the rays which travel towards the bottom of the trench usually reflect off the
surface many times, which makes the difference in computational effort for using
a bounding volume hierarchy from a mesh simplified with our method even more
evident. Figure 5b for Surface 2 shows a speedup of about 12% compared to the
Lindstrom-Turk algorithm for simplification levels of 52 and 67%.
80 C. Lenz et al.

4 Summary

We introduce a new surface mesh simplification method that uses the curvature of
the surface mesh to identify regions which can be simplified with different sets of
parameters depending on the local surface properties. Our approach is well suited for
meshes that are common in flux-dependent process TCAD simulations since such
meshes often contain large flat regions with high resolutions from the originating
regular grid. We have evaluated our method with respect to geometric distances and
execution times for simplification and subsequent computations of flux estimates.
The geometric distances in the experiments have improved in comparison to the
reference algorithm. In particular, the average Hausdorff distance of the investigated
geometries has improved by 20–40%. The ray tracing time in all our experiments has
been improved on average by 15%, furthermore, demanding real world geometries
from process TCAD have shown a compelling improvement of 12% of time spent
on ray tracing. The execution time of our simplification method is on average
17% slower than the reference algorithm. When considering entire topography
simulations, the accelerated ray tracing significantly exceeds the additional time
spent on our simplification method.

Acknowledgments The financial support by the Austrian Federal Ministry for Digital and
Economic Affairs and the National Foundation for Research, Technology and Development is
gratefully acknowledged.

References

1. P. Manstetten, Efficient Flux Calculations for Topography Simulation. Doctoral Dissertation,

TU Wien (2018). http://www.iue.tuwien.ac.at/phd/manstetten/
2. J.A. Sethian, Level Set Methods and Fast Marching Methods: Evolving Interfaces in Com-
putational Geometry, Fluid Mechanics, Computer Vision, and Materials Science (Cambridge
University Press, 1999)
3. A.S. Glassner, An Introduction to Ray Tracing (Elsevier, 1989)
4. W.E. Lorensen, H.E. Cline Marching cubes: A high resolution 3D surface construction
algorithm. ACM SIGGRAPH Comput. Graph. 21, 163–169 (1987)
5. P. Lindstrom, G. Turk, Fast and memory efficient polygonal simplification, in Proceedings of
the IEEE Visualization Conference, pp. 279–286 (1998)
6. M. Garland, P.S. Heckbert, Surface simplification using quadric error metrics, in Proceedings
of the Conference on Computer Graphics and Interactive Techniques, pp. 209–216 (1997)
7. H. Borouchaki, P.J. Frey, Simplification of surface mesh using Hausdorff envelope. Comput.
Methods Appl. Mech. Eng. 194, 4864–4884 (2005)
8. S.J. Kim, C.H. Kim, D. Levin, Surface simplification using a discrete curvature norm. Comput.
Graph. 26, 657–663 (2002)
9. F. Rudolf, Symmetry- and Similarity-Aware Volumetric Meshing. Doctoral Dissertation, TU
Wien (2016). https://www.iue.tuwien.ac.at/phd/rudolf/
10. L. Gnam, High Performance Mesh Adaptation for Technology Computer-Aided Design.
Doctoral Dissertation, TU Wien (2019). https://www.iue.tuwien.ac.at/phd/gnam/
A Novel Mesh Simplification Method for Flux-Dependent Topography Simulations 81

11. L. Gnam, P. Manstetten, A. Hössinger, S. Selberherr, J. Weinbub, Accelerating flux calculations

using sparse sampling. Micromachines 9, 1–17 (2018)
12. I. Wald, S. Woop, C. Benthin, G.S. Johnson, M. Ernst, Embree: A kernel framework for
efficient CPU ray tracing. ACM Trans. Graph. 33, 143:1–143:8 (2014)
13. M. Meyer, M. Desbrun, P. Schröder, A.H. Barr, Discrete differential-geometry operators for
triangulated 2-manifolds, in Visualization and Mathematics III, pp. 35–57 (2003)
14. R. Straub, Exact computation of the Hausdorff distance between triangular meshes, in
Proceedings of the Conference of the European Association for Computer Graphics, pp. 17–20
(2007)
15. R.Y. Rubinstein, D.P. Kroese, Simulation and the Monte Carlo Method (Wiley, 2016)
Simulations of a Novel DG-GFET

Giovanni Nastasi and Vittorio Romano

Abstract A peculiar geometry for a graphene double gate field effect transistor
is proposed. It allows us to overcome the problems encountered for a standard
MOSFET geometry due to the zero gap in monolayer graphene. It is found that for
a wide range of the gate voltage the current is in an off state with a ratio current-on
over current-off of about 104 .

1 Introduction

As quoted in [1] “Graphene has changed from being the exclusive domain of
condensed-matter physicists to being explored by those in the electron-device
community. In particular, graphene-based transistors have developed rapidly and
are now considered an option for post-silicon electronics. However, many details
about the potential performance of graphene transistors in real applications remain
unclear.”
Device engineers devote considerable effort for developing transistor designs in
which short-channel effects are suppressed and series resistances are minimized.
Scaling theory predicts that a FET with a thin barrier and a thin gate-controlled
region will be robust against short-channel effects down to very short gate lengths.
The possibility of having channels that are just one atomic layer thick is perhaps
the most attractive feature of graphene for its use in transistors. Main drawback of
a large-area monolayer graphene is the zero gap. This has the consequence that the
current versus the gate voltage is no longer a monotone function and the off region
is very narrow (see [2]), making graphene not usable in a straightforward way for
transistors. Moreover, graphene on substrate suffers also from the degradation of the
mobility because of the additional interaction with the phonons of the oxide.

G. Nastasi () · V. Romano

Department of Mathematics and Computer Science, University of Catania, Catania, Italy
e-mail: [email protected]; [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 83

M. van Beurden et al. (eds.), Scientific Computing in Electrical Engineering,
Mathematics in Industry 36, https://doi.org/10.1007/978-3-030-84238-3_9
84 G. Nastasi and V. Romano

Fig. 1 Schematic representation of the investigated DG-GFET

Here we propose a special geometry for a double gate graphene FET (DG-GFET)
which overcomes the problem related to the zero gap as will be shown by the
numerical simulations. The devices is depicted in Fig. 1. The active area is made
of just one graphene layer.
Usually the GFETs are investigated by adopting reduced one dimensional models
of the Poisson equation with some averaging procedure [3, 4]. Here a full two-
dimensional simulation is presented based on a drift-diffusion-Poisson system with
the mobilities proposed in [5].
Other approaches are based on hydrodynamical models, e.g. those deduced with
the maximum entropy principle [6–9], or the direct solution of the Boltzmann
equation [10–14] or Monte Carlo methods [15]. Thermal effects can also be
included [16–21]. Here the crystal lattice will be kept at a constant temperature
and considered as a thermal bath. For the inclusion of quantum effects the interest
reader is referred to [22, 23].

2 Mathematical Model

The mathematical model we adopt to simulate the charge transport in the graphene
layer of the DG-GFET is the bipolar drift-diffusion in 1D case,
& '
∂n 1 ∂ ∂n ∂φ
− μn k B T L − enμn = 0,
∂t e ∂x ∂x ∂x
& '
∂p 1 ∂ ∂p ∂φ
+ −μp kB TL − epμp = 0,
∂t e ∂x ∂x ∂x
Simulations of a Novel DG-GFET 85

where n(t, x), p(t, x) are the graphene electron density and hole density respec-
tively, e is the positive elementary charge, kB is the Boltzmann constant, TL is the
lattice temperature (kept constant), μn (x) and μp (x) are the mobility models for
electrons and holes respectively and φ(x, y) is the electric potential. We adopt the
mobility model proposed in [5] (for other models the interested reader is referred to
[2, 24, 25]) given by
νs
μs (x) = ,
[1 + (νs E/vsat )γ ]1/γ

where E = |∂φ/∂x| is the absolute value of the x-component of the electric field,
vsat is the saturation velocity (we take the value 0.2 µm/ps), γ ≈ 2 and
μ0
νs (x) = ,
(1 + s/nref )α

where μ0 = 0.4650 µm2 /V ps is the low field mobility, nref = 1.1 × 105 µm−2
and α = 2.2. The symbol s indicates the carrier density: s = n for electrons and
s = p for holes.
In order to determine the electric potential a 2D Poisson equation is coupled to
the drift-diffusion system

∇ · (∇φ) = h(x, y),

where

e(n(x) − p(x) − Nimp )/tgr if y = ygr
h(x, y) =
0 if y = ygr

being ygr the y-coordinate (see Fig. 1), Nimp = 3.5×103µm−2 the impurity density
due to the SiO2 , tgr the distance between the two layers of oxide which is assumed
to be equal to 1 nm. We remark that the charge in the graphene layer is considered
distributed in the volume enclosed by the parallelepiped of base the area of the
graphene and height tgr . Recall that n and p are areal densities. Moreover  is given
by

gr if y = ygr
(x, y) =
ox if y = ygr

where gr and ox are the dielectric constants of the graphene and oxide respectively.
The source and drain contacts are assumed to be thermal bath charge reservoirs.
86 G. Nastasi and V. Romano

3 Numerical Results

Here some numerical results are presented in order to show that the proposed DG-
GFET is able to perform as a transistor. The length is 100 nm. The width of both the
oxide layers (SiO2 ) is 10 nm. The source and drain contacts are positioned in the
direction transversal with respect the graphene sheet and they occupy all the device
height (21 nm). The two gate potentials are set as equal. At the metallic contacts the
total voltage including the work function is considered equal to 0.25 V plus the bias
voltage, which is zero at source. Indeed the work function depends on the specific
material the contacts are made of.
A full 2D discretization of the Poisson equation is adopted in the whole device
by standard central differencing enforced with a Gummel iteration, while the drift-
diffusion equation is solved only in the graphene sheet as a 1D problem with
a Scharfetter-Gummel method (indeed only one row of grid points is used by
considering a kind of average in the y direction). The interested reader is referred to
[2] for the details. By numerical experiments a good resolution is already obtained
with 41×23 grid points.
In Figs. 2, 3 the shape of the electrical potential is plotted when the source-drain-
potential is 0.3 V and the gate-source potential is −1 V and 1 V respectively. In
the first case the device is off while in the second case is on. The Fig. 4 shows the
characteristic curve current versus gate voltage with source drain voltage equal to
0.2 V while Fig. 5 shows the same but in a logarithmic scale.

Fig. 2 Electrostatic potential when the gate-source potential is −1 V and the source-drain-
potential is 0.3 V
Simulations of a Novel DG-GFET 87

Fig. 3 Electrostatic potential when the gate-source potential is 1 V and the source-drain-potential
is 0.3 V

4
Vsd = 0.1V
Vsd = 0.2V
Vsd = 0.3V
Current density (A/cm)

3 Vsd = 0.4V

0
-1 -0.5 0 0.5 1
Gate voltage (V)

Fig. 4 Current versus gate voltage for several values of the bias voltage
88 G. Nastasi and V. Romano

102
Vsd = 0.1V
Vsd = 0.2V
100 Vsd = 0.3V
Current density (A/cm)

Vsd = 0.4V

10-2

10-4

10-6

10-8
-1 -0.5 0 0.5 1
Gate voltage (V)

Fig. 5 Current versus gate voltage for several values of the bias voltage in logarithmic scale

It is evident that the peculiar geometry of the DG-FET we have investigated

leads to characteristic curves that are appropriate for field effect transistors mainly
because there exists a clear and wide off region similar to the case of tradition
semiconductors like Si or GaAs. With the particular set up of gate-source and
source-drain voltage the current of the minor charges is not triggered and transport
remain mainly unipolar (Figs. 6 and 7). This at variance with the case of standard
MOSFET configurations where minority charges are triggered for sufficiently gate-
drain voltage limiting the current-off zone to a short range [2, 26].
The current-on over current-off ratio is about of four orders of magnitude which
is acceptable for electrical engineering purposes. This is clearly shown by the figure
in logarithmic scale.
Simulations of a Novel DG-GFET 89

2.5
Current density (A/cm)

2
VG = 0.1V
1.5 VG = 0.3V
VG = -0.1V
VG = -0.3V
1

0.5

0
0 0.2 0.4 0.6 0.8 1
Gate voltage (V)
Fig. 6 Current versus bias voltage for several values of the gate voltage

100
Current density (A/cm)

10-5

VG = 0.1V
VG = 0.3V
VG = -0.1V
VG = -0.3V
10-10
0 0.2 0.4 0.6 0.8 1
Gate voltage (V)
Fig. 7 Current versus bias voltage for several values of the gate voltage in logarithmic scale
90 G. Nastasi and V. Romano

4 Conclusions

A novel geometry for a double gate FET with active area made of a single layer of
graphene has been proposed and simulated with a drift-diffusion model by solving
a full 2D Poisson equation for the electrostatic potential. The results are rather
encouraging because a good transistor effect is obtained at variance with of other
GFETs proposed in the literature. The simulation based on more sophisticated
models is currently under investigation by the authors in order to get a further
validation of the devised device.

Acknowledgments The authors acknowledge the support from INdAM (GNFM). The author
G.N. acknowledges the support from Progetto Giovani GNFM 2019 Modelli matematici, numerici
e simulazione del trasporto di cariche e fononi nel grafene. This work has been also supported by
the Università degli Studi di Catania, Piano della Ricerca 2018/2020 Linea di intervento 2.

References

1. F. Schwierz, Graphene transistors. Nat. Nanotechnol. 5, 487–496 (2010)

2. G. Nastasi, V. Romano, A full coupled drift-diffusion-Poisson simulation of a GFET. Commun.
Nonlinear Sci. Numer. Simul. 87, 105300 (2020)
3. D. Jiménez, O. Moldovan, Explicit drain-current model of graphene field effect transistors
targeting analog and radio-frequency applications. IEEE Trans. Electron. Dev. 58(11), 4049–
4052 (2011)
4. A.K. Upadhyay, A.K. Kushwaha, S.K. Vishvakarma, A unified scalable quasi-ballistic trans-
port model of GFET for circuit simulations. IEEE Trans. Electron. Dev. 65(2), 739–746 (2018)
5. V.E. Dorgan, M.-H. Bae, E. Pop, Mobility and saturation velocity in graphene on SiO2 . Appl.
Phys. Lett. 97(8), 082112 (2010)
6. V.D. Camiola, V. Romano, Hydrodynamical model for charge transport in graphene. J. Stat.
Phys. 157, 1114–1137 (2014)
7. L. Luca, V. Romano, Hydrodynamical models for charge transport in graphene based on the
maximum entropy principle: the case of moments based on energy powers. Atti Accad. Pelorit.
Pericol. Cl. Sci. Fis. Mat. Nat. 96(S1), A5 (2018)
8. V.D. Camiola, G. Mascali, V. Romano, Charge Transport in Low Dimensional Semiconductor
Structures - The Maximum Entropy Approach. Springer International Publishing (2020)
9. L. Luca, V. Romano, Comparing linear and nonlinear hydrodynamical models for charge
transport in graphene based on the Maximum Entropy Principle. Int. J. Nonlin. Mech. 104,
39–58 (2018)
10. A. Majorana, G. Nastasi, V. Romano, Simulation of bipolar charge transport in graphene by
using a discontinuous Galerkin method. Comm. Comp. Phys., 26(1), 114–134 (2019)
11. P. Lichtenberger, O. Morandi, F. Schürrer, High-field transport and optical phonon scattering
in graphene. Phys. Rev. B 84(4), 045406 (2011)
12. V. Romano, A. Majorana, M. Coco, DSMC method consistent with the Pauli exclusion
principle and comparison with deterministic solutions for charge transport in graphene. J.
Comput. Phys. 302, 267–284 (2015)
13. M. Coco, A. Majorana, G. Nastasi, V. Romano, High-field mobility in graphene on substrate
with a proper inclusion of the Pauli exclusion principle. Atti Accad. Pelorit. Pericol. Cl. Sci.
Fis. Mat. Nat., 97(S1), A6 (2019)
Simulations of a Novel DG-GFET 91

14. M. Coco, G. Nastasi, Simulation of bipolar charge transport in graphene on h-BN. COMPEL
39(2), 449–465 (2020)
15. M. Coco, A. Majorana, V. Romano, Cross validation of discontinuous Galerkin method and
Monte Carlo simulations of charge transport in graphene on substrate. Ricerche mat. 66, 201–
220 (2017)
16. M. Coco, G. Mascali, V. Romano, Monte Carlo analysis of thermal effects in monolayer
graphene. J. Comput. Theor. Trans. 45(7), 540–553 (2016)
17. M. Coco, V. Romano, Simulation of electron–phonon coupling and heating dynamics in
suspended monolayer graphene including all the phonon branches. J. Heat Transfer. 140(9),
092404 (2018)
18. M. Coco, V. Romano, Assessment of the constant phonon relaxation time approximation in
electron-phonon coupling in graphene. J. Comput. Theor. Trans. 7(1–3), 246–266 (2018)
19. G. Mascali, V. Romano, Charge transport in graphene including thermal effects. SIAM J. Appl.
Math. 77, 593–613 (2017)
20. G. Mascali, V. Romano, Exploitation of the maximum entropy principle in mathematical
modeling of charge transport in semiconductors. Entropy 19(1), 36 (2017)
21. G. Mascali, V. Romano, A hierarchy of macroscopic models for phonon transport in graphene.
Physica A 548, 124489 (2020)
22. O. Morandi, F. Schürrer, Wigner model for quantum transport in graphene. J. Phys. A Math.
Theor. 44(26), 265301 (2011)
23. L. Luca, V. Romano, Quantum corrected hydrodynamic models for charge transport in
graphene. Ann. Phys. 406, 30–53 (2019)
24. A. Majorana, G. Mascali, V. Romano, Charge transport and mobility in monolayer graphene.
J. Math. Industry 7(4), 4 (2016).
25. G. Nastasi, V. Romano, Improved mobility models for charge transport in graphene. Commun.
Appl. Ind. Math. 10(1), 41–52 (2019)
26. G. Nastasi, V. Romano, Simulation of graphene field effect transistors, in Scientific Computing
in Electrical Engineering, SCEE 2018, Taormina, September 23–27, ed. by G. Nicosia, V.
Romano. Mathematics in Industry, vol. 32 (Springer Nature, Switzerland AG, 2020), pp. 171–
178
 Part III
Computational Electromagnetics
Electric Circuit Element Boundary
Conditions in the Finite Element Method
for Full-Wave Frequency Domain Passive
Devices

Gabriela Ciuprina, Daniel Ioan, Mihai Popescu, and Sorin Lup

Abstract A natural coupling of a circuit with an electromagnetic (EM) device

is possible if special boundary conditions, called Electric Circuit Element (ECE),
are used for the EM field formulation. This contribution shows how these ECE
boundary conditions can be implemented into the finite element method for the
solving of coupled full-wave EM field-circuit problems in the frequency domain.
The implementation is based on a weak formulation that uses the electric field
strength strictly inside the domain and a scalar potential defined solely on the
boundary. Edge elements are used inside the three-dimensional domain and nodal
elements are used on its two-dimensional boundary surface. The weak formulation
is given and its discrete form is validated on a 2D example, with known analytic
solution.

1 Motivation

Many EM devices with distributed parameters and field effects specific to full-wave
(FW) or Magneto-Quasi-Static (MQS) EM field regime are connected to circuits
with lumped parameters (e.g. in measuring and control applications). For this, the
EM devices need boundary conditions compatible with external circuits (Fig. 1,left).
By definition, an isolated electric circuit has a finite number of components
connected to common terminals. Each terminal is characterized by its voltage with
respect to the ground. A non-isolated circuit, i.e. a sub-circuit with m terminal nodes
has each of these terminals characterized by a pair of scalar quantities, a current ik

G. Ciuprina () · D. Ioan · M. Popescu · S. Lup

Politehnica University of Bucharest, Bucharest, Romania
e-mail: [email protected]; [email protected]; [email protected]; [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 95

M. van Beurden et al. (eds.), Scientific Computing in Electrical Engineering,
Mathematics in Industry 36, https://doi.org/10.1007/978-3-030-84238-3_10
96 G. Ciuprina et al.

Fig. 1 Left: Coupling of electric circuits and EM device models are naturally ensured by means
of terminals. Right: To ensure the coupling, “node voltages” (potentials) and electric currents of
non-isolated circuits must have a correspondent in the EM device model

entering into the sub-circuit and a “node voltage” (potential) vk (Fig. 1,right). The
power transferred to it is


m 
m−1 
m−1
P = ik vk = ik (vk − vm ) = ik vk (1)
k=1 k=1 k=1

if im is expressed according to Kirchhoff current law for a cutset and the terminal m
is connected to ground. This power expression shows that the state of a m-terminal
circuit is characterized by 2(m−1) independent quantities: m−1 currents and m−1
voltages. The assumption vm = 0 is not a restriction for the purpose of this paper,
which is stated at the end of Sect. 2. A natural coupling of this sub-circuit with
an EM device is possible if some connecting surfaces are defined on the device’s
boundary, for which currents and potentials are defined, in order to satisfy Kirchhoff
relationships and provide the same transmitted power formula (1) as subcircuits do.
The conditions that satisfy these requirements are the ones proposed in [10], used
in [4, 8] and called Electric Circuit Element (ECE) boundary conditions.
The ECE boundary conditions, combined with current excited terminals, are the
“realistic boundary conditions” used in [1] to solve eddy current problems with the
finite element method (FEM) using a formulation in H and an ungauged T − ϕ, ϕ
one in [2]. Similar conditions, although with a different definition for the terminal
voltages are proposed in [5] and used for A,  V eddy current formulations [7].
The use of ECE in MQS problems for inductance extraction with an A, V
formulation is discussed in [9]. Our aim is to use ECE boundary conditions to
solve full-wave (FW) problems with FEM. We have successfully used ECE to model
passive on-chip components such as resistors, inductors, capacitors, interconnects or
RF-MEMS switches in FW [3], with the Finite Integration Technique as numerical
method. According to our knowledge, the ECE conditions are not available in FEM
codes which implement the formulation of microwave ports for FW. Theoretical
studies exists, e.g. in [4], based on an E, V formulation for the whole domain. In
this paper we use E strictly inside the domain and V solely on the boundary. During
the reviewing process of this paper, Hiptmair and Ostrowski released a relevant
report [6], proving the interest for this subject.
Electric Circuit Element Boundary Conditions for Passive Devices 97

2 ECE Boundary Conditions

Assume a simply connected domain Ω, with a Lipschitz boundary ∂Ω that includes

m disjoint parts Sk , k = 1, 2, . . . , m (device’s terminals), so that conditions (ECE1),
(ECE2) and (ECE3) are satisfied (Fig. 2):

• (ECE1) there is no magnetic coupling with the exterior: n · ∂ B(r,t
∂t
)
= 0, ∀r ∈
∂Ω;
• (ECE2) the electric coupling is carried out only through the terminals: n · (∇ ×
H (r, t)) = 0, ∀r ∈ ∂Ω − ∪m k=1 Sk ;
 t) = 0,
• (ECE3) the terminals are equipotential: n × E(r,  ∀r ∈ Sk , k =
1, . . . , m.
According to Faraday’s law, (ECE3) implies (ECE1) for the terminals, the inclusion
of the terminals in (ECE1) is kept only for emphasizing the physical meaning.
By definition, the currents and potentials of any terminal are:
(  * 
+ 
∂D
ik (t) = H · dl
 =− J + · n ds, vk (t) = E · dl,

∂Sk Sk ∂t Ck ⊂∂Ω
(2)
where, in order to ensure conservation, each terminal current is the total current
(conductive and displacement) and the potential is properly defined as the voltage
between this terminal and the reference one, along a path Ck included in the domain
boundary. Due to (ECE1) the voltage between two points placed on the boundary
surface is independent of the path of the integration line connecting these points,
provided that this path is included in the surface. Thus, the potential on the surface
is well defined, although this is not the case in a general time-varying EM field.
Under these conditions, (1) holds for the EM device, where ik and vk are given

Fig. 2 Electric terminals are

disjoint surfaces on the
domain’s boundary. The
non-grounded terminals can
be either voltage excited (its
potential is given) or current
excited (its total current is
given)
98 G. Ciuprina et al.

Fig. 3 Each non-grounded terminal of the EM device with ECE boundary conditions can be either
current excited or voltage excited. Its hybrid transfer matrix is obtained after computing voltages
of the current excited terminals and currents of the voltage excited terminals in linear problems

by (2), and thus the ECE boundary conditions are perfectly compatible with the
power transferred through its terminals by a multipolar circuit [8, 10].
If we assume that the terminals have known potentials, then it can be proved that
the problem of EM field analysis in a linear domain with ECE boundary conditions
has a unique solution. Consequently, the terminal currents are output signals and
are obtained by solving the field problem [10]. As the domain is linear, so are
the equations, hence the device with ECE conditions is a linear system, defining
a multiple input multiple output (MIMO) type dynamic system with m − 1 inputs
and m − 1 outputs (Fig. 3).
In the frequency domain, the input-output relationship is expressed as:


 T ZA  T
V 1 . . . V n I n+1 . . . I m−1 = I 1 . . . I n V n+1 . . . V m−1 . (3)
BY


Z(f ) A(f )
The problem to be solved is: “Find , where f is the frequency in
B(f ) Y(f )
a given frequency range of interest, defined by its minimum and maximum values
fmin and fmax f ∈ [fmin , fmax ], from the EM field solution.” If this hybrid matrix is
known, then the “field” element can be realized with common circuit elements and
included in any circuit simulator.

3 ECE in FEM

It is useful to recall the formulation in E with classical boundary conditions, since

the newly proposed formulation inherits a part of it.
Strong Formulation of PDE for E with Classical Boundary Conditions
The well known FW Maxwell equations in the frequency domain, for linear media
and no internal field sources are: ∇ × E = −jωμH , ∇ × H = σ E + jωεE, 
 
∇ · (μH ) = 0, ∇ · (εE) = ρ, where permittivity ε, permeability μ and conductivity
σ are positive, space dependent material parameters. The reluctivity ν = 1/μ might
be used instead of μ. The solution of these equations is unique if in any point of
Electric Circuit Element Boundary Conditions for Passive Devices 99

∂Ω, either exclusively Et or Ht are known (given). The subscript t indicates the
tangential component of the vector on the surface. It is useful to denote a disjoint
partition of the boundary: ∂Ω = SE ∪ SH , SE ∩ SH = ∅, and thus Et : SE → C2 ,
Ht : SH → C2 . The imposed boundary conditions are: Et (r ) = n × (E(  r ) × n ), for
r ∈ SE and Ht (r ) = n ×(H (r )× n ), for r ∈ SH . In what follows we will name them
classical boundary conditions. The uniqueness of the field solution can be proven
on the basis of the complex form of the Poynting’s theorem that gives the expression
of the transmitted power (assuming a linear field domain, with no moving parts):
(   *  ∗  ·D∗
+
∗ B · H E
− (Et × H∗t ) · n ds = E · J + 2jω − . (4)
∂Ω Ω Ω 2 2

The proof assumes that there exist two such fields that satisfy the same boundary
conditions. This means that the Poynting theorem in complex form is valid for
the difference field, which satisfies Maxwell’s equations (due to linearity) and
zero boundary conditions. This implies that the real part is zero which conduces
to zero difference electric field (conductivity of the domain is assumed non-zero
everywhere) and the imaginary part is zero with conduces to zero difference
magnetic field.
The second order equation is:

 + jω(σ + jωε)E = 0.
∇ × (ν∇ × E)  (5)

Weak Formulation in E with Classical Boundary Conditions

In general, solving of (5) implies a numerical approach, e.g. FEM, which is based on
weak formulations. The needed functionals result by projecting (5) onto a set of test

functions E , then integrating by parts and applying Gauss-Ostrogradski formula:
   (  
 · (∇ × E  ) + jω(σ + jωε)E · E  dx = −
(ν∇ × E)  × E  · n ds
(ν∇ × E)
Ω ∂Ω

Replacing the expression of the magnetic field strength in the right hand side we get
   ( 
  
 · (∇ × E ) + jω(σ + jωε)E · E
(ν∇ × E) dx = jω H × E · n ds.
Ω ∂Ω
(6)
  
With classical boundary conditions, the right hand side is equal to SE E t × n ·
  
H ds + SH n × H t · E ds. E t are essential boundary conditions that is why the
100 G. Ciuprina et al.


test functions are chosen so that E t is zero on SE . Thus, the weak equation for the
trial functions E is:
     
 
 · (∇ × E ) + jω(σ + jωε)E · E
(ν∇ × E) dx = jω n × H t · E ds.
Ω SH
(7)

The boundary conditions H t are natural, they appear in the functional equation.
In conclusion, the weak formulation in E with classical boundary conditions is:
 E  ) = f (E  ), ∀E  ∈ H0 where
Find E in H , such that a(E,
  
 E  ) =
a((E,  · (∇ × E  ) + jω(σ + jωε)E · E  dx,
(ν∇ × E) (8)
Ω
  

f (E ) = jω n × H t · E ds, (9)
SH
, -
H = u ∈ H (curl, Ω)| u × n ) = Et on SE ,
n × ( (10)
, -
H0 = u ∈ H (curl, Ω)| u × n ) = 0 SE .
n × ( (11)

Discrete Formulation in E with Classical Boundary Conditions

Assume a simplicial mesh (tetrahedrons in 3D, triangles in 2D), numerical test
functions N k that correspond to edge elements of order (0,1), and degrees of
freedom that represent the complex representations of voltages U k along the edges.
Ne
The numerical solution is approximated as E = 
j =1 U j Nj , where Ne is the
total number of edges in the domain. For any cell, the sum involves 6 terms in 3D
and 3 terms in 2D. By substituting the approximation of the numerical solution

in (6), choosing the test function E = N i and rearranging the sums we obtain
a relationship that reveals how the matrices assembling has to be done for all
i = 1, . . . , Ne:
Ne . 
  /  
(ν∇ × N j ) · (∇ × N i ) + jω(σ + jωε)Nj · N i dx U j = jω n × H t · N i ds.
j =1 Ω SH

(12)

The initial assembling is carried out for all the edges in the domain. The next step
refers to the boundary conditions. Assume that the edges were numbered in the
following order: first—the inner edges, second—the edges on the boundary SH and
finally, the edges on the boundary SE . This leads to the following partitioning:
⎡ ⎤⎡ ⎤ ⎡ ⎤
Ain−in Ain−SH Ain−SE Uin 0
⎣ ASH−in ASH−SH ASH−SE ⎦ ⎣ U ⎦ = ⎣ bSH ⎦ (13)
SH
ASE−in ASE−SH ASE−SE USE 0
Electric Circuit Element Boundary Conditions for Passive Devices 101

The group of equations that correspond to edges on the SE boundary is deleted and
the essential boundary conditions E t are translated into imposed values of electric
voltages along edges on the SE boundary. The system to be solved is





Ain−in Ain−SH Uin 0 Ain−SE  
= − USE , (14)
ASH−in ASH−SH USH bSH ASH−SE

the coefficient matrix being symmetric and positive defined.

Weak Formulation in E, V with ECE Boundary Conditions
If we use ECE boundary conditions, the unknowns are the electric field inside the
domain and an electric scalar potential solely defined on ∂Ω. That is why the
formulation is still named E, V , but is different from other formulations, such as
the E, V in [4] where V is defined also inside the domain. An E, V interpretation of
the ECE boundary conditions (ECE 1,2,3)) is:
0
• (ECE1b) Γ E · dl  = 0, ∀Γ ∈ ∂Ω;

• (ECE2b) n · E(r) = 0, ∀r ∈ ∂Ω − ∪m k=1 Sk ;
 
• (ECE3b) E t (r) = 0 ∀r ∈ Sk , k = 1, . . . , m.
From (ECE1b) an electric scalar potential V can be defined on the boundary ∂Ω,
such that E t = −∇2 V . Condition (ECE3b) requires that the electric terminals
are equipotential. For uniqueness reasons, one terminal has to be defined by any
value. Without lack of generality we can assume it is grounded in what follows. For
the other terminals the uniqueness implies that, exclusively, either their voltages or
currents are known.

Using (5) we get the weak equation for E:
 
 
(ν∇ × E) · (∇ × E ) + jω(σ + jωε)E · E dx = jω V k I k , (15)
Ω k∈Ic

where Ic is the set of indices of current excited terminals. Similarly, we will denote
by Iv is the set of indices of voltage excited terminals. We need an equation for the
electric potential on the boundary, as well. Let’s denote the normal component of
not
the total current density in any point on the boundary as J n = (∇ × H) · n. We will
project J n onto a set of scalar test functions V  :
( ( m 
 
(ECE2)
(∇ × H) · n V  ds = J n V  ds = J n V  ds = V k I k
∂Ω ∂Ω k=1 Sk k∈Ic

The integrand of the left hand side can be further computed by using the
integration by parts formula that involves the surface differential operators and the
102 G. Ciuprina et al.

substitution of the magnetic field with its expression with respect to the electric
field, as it follows from Faraday’s law:
( ( (
(∇ × H) · n V  ds = V  n · curl H ds = V  div2 (H) ds =
def
∂Ω ∂Ω ∂Ω
 ( (
ν
= V  (n × H) ds − H · grad2 V  ds = curl E · grad2 V  ds
∂(∂Ω) ∂Ω ∂Ω jω

Consequently it follows that the weak form of the equation on the boundary is
( 
(ν∇ × E) · ∇2 V  ds = j ω V k I k (16)
∂Ω k∈Ic

 V with ECE boundary conditions.

Finally, we get the weak formulation in E,
Find E ∈ HE , V ∈ HV , such that

a(E, E ) = f (E ), ∀E ∈ HE,0 ; b(E, V  ) = g(V  ), ∀V  ∈ HV ,0

(
H · dl = I k , k ∈ Ic ; Et = −∇2 V , on ∂Ω,
∂Sk

where
 
 
a(E, E ) = (ν∇ × E) · (∇ × E ) + jω(σ + jωε)E · E dx, f (E ) = jω V k I k ;
Ω k∈Ic
( 
b(E, V  ) = (ν∇ × E) · ∇2 V  ds, g(V  ) = jω V k I k ;
∂Ω k∈Ic

where Et = −∇2 V  .

HE = {u ∈ H (curl, Ω)| n × (u × n) = −∇2 V  on ∂Ω, V  ∈ HV

1
n × (u × n) = 0 on ∪m k=1 Sk
HE,0 = {u ∈ H (curl, Ω)| n × (u × n) = −∇2 V  on ∂Ω, V  ∈ HV ,0
1
n × (u × n) = 0 on ∪m k=1 Sk
HV = {u ∈ H (grad, ∂Ω)| u = V k on Sk , k ∈ Iv ,
u = constant(unkown) on Sk , k ∈ Ic }

HV ,0 = {u ∈ H (grad, ∂Ω)| u = 0 on Sk , k ∈ Iv
u = constant(unkown) on Sk , k ∈ Ic }
Electric Circuit Element Boundary Conditions for Passive Devices 103

Note: We have investigated two other formulations

0 for the boundary equations for
which b(E, V  ) = 0. In one version b(E, V  ) = ∂Ω (σ + jωε)(∇2 V ) · (∇2 V  ) ds +
0    
0
∂Ω ∂n (σ + jωε)E · n V ds and another version is b(E, V ) = ∂Ω (σ + j ωε)n ·
∂

EV  ds. Due to lack of space we will not present them here.

Formulation in E, V with ECE Boundary Conditions—Algorithm in FEM
Step 1 We start with
the discreteform

of classical

 BC, given by (14), written
Au,u Au,ub u 0
for all the edges) = . Only the first block row of
Aub ,u Aub ,ub ub bb
equations, corresponding to the inner edges, is kept.
Step 2 Write the discrete form of the Eq. (16) on the 2D surface boundary mesh.
Ne 0  
j =1 ∂Ω (ν∇ × Nj ) · (∇2 ϕi ) ds Uj = j ωI i , where ϕi is the nodal element
i. This is written for all the nodes on the boundary

and willbe

placed

together

Au,u Au,ub u 0
with the discrete equation obtained at step 1: = .
AVb ,u AVb ,ub ub bb
Step 3 On the boundary, the variables are changed, from electric voltages to
electric potentials,


ub as potential differences. The system becomes

by expressing
Au,u Au,Vb u 0
= .
AVb ,u AVb ,Vb Vb bb
Step 4 Finally, Vb has to be split in three (V-for nodes that are not on terminals,
Vt,c -voltages of current excited terminals, Vt,v -voltages of voltage excited
terminals), in order to impose the rest of the natural conditions (potentials for
voltage excited, or currents for current excited terminals): Finally, the system to
solve is
⎡ ⎤⎡ ⎤ ⎡ ⎤ ⎡ ⎤
Au,u Au,V Au,Vt,c u 0 Au,Vt,v  
⎣ AV ,u AV ,V AV ,Vt,c ⎦ ⎣ V ⎦ = ⎣ 0 ⎦ − ⎣ AV ,Vt,v ⎦ Vt,v .
AVt,c ,u AVt,c ,V AVt,c ,Vt,c Vt,c j ωIt,c AVt,c ,Vt,v

After solving, we get the unknown potentials V and Vt,c . The currents through
the terminals in Iv can be computed as a postprocessing step.

3.1 Numerical Results

Figure 4 shows a quantitative validation for a 2D simple case, with two terminals
and with analytical solution. It is a single input single output (SISO) system, both
current and voltage excitations give accurate results. The domain is a brick that
occupies the space x ∈ [−a, a], y ∈ [0, l] and x ∈ [0, h]. One excited terminal
(in voltage or in current) is on the z = 0 boundary and the grounded terminal is on
the z = h boundary. The material inside is assumed homogeneous with ε, μ, σ .
The analytic solution can be obtained by solving the Helmholtz equations and
considering the current excited terminal (I ). The complex power absorbed by this
104 G. Ciuprina et al.

0.04

0.035

0.03

0.025
R [Ω]

0.02

0.015

0.01

Analytic
0.005
Voltage excitation
Current excitation
0
0 10 20 30 40 50 60 70 80 90 100
Frequency [GHz]

× 10 -13
5.5
Analytic
5 Voltage excitation
Current excitation
4.5

3.5
L [H]

2.5

1.5

0.5
0 10 20 30 40 50 60 70 80 90 100
Frequency [GHz]

Fig. 4 Quantitative validation of the implementation for a 2D case with analytical solution. The
problem is a rectangle with two opposite terminals, consequently the system is SISO. Both voltage
and current excitations lead to relative errors less than 2% for the whole frequency range
Electric Circuit Element Boundary Conditions for Passive Devices 105

Fig. 5 Qualitative validation for a 2D case, MIMO (3 terminals), hybrid excitation (one terminal
grounded, one is voltage excited and one is current excited)

domain is P = 2E y H ∗z lh, where E y = γ /(σ + j ωε) cosh(γ a)/sinh(γ a) I /(2h)

and H z = I /(2h). The extracted complex impedance is Z = P /|I |2 and its
components shown in Fig. 4 are R = real P and L = real P /ω for a = 2.5 µm,
l = 10 µm, h = 10 µm, σ = 6.6 · 107 S/m, μ = μ0 , ε = ε0 , fmin = 0.01 GHz,
fmax = 100 GHz.
Figure 5 shows a qualitative validation for a MIMO test. The rectangular domain
is occupied by a T-shape conductor of high conductivity, having 3 terminals, out of
which the one at the right hand side of the figure is grounded.

4 Conclusions

The advantages of ECE BC for Maxwell equations are that the ports are clearly and
well defined, without ambiguity, fully compatible with the circuit terminals. There
is no restriction on the field regime (full wave, nonlinear). For MIMO systems, the
hybrid excitation is obtained in a natural way. This paper proposed a FEM algorithm
for ECE, which E strictly inside the domain and V on the boundary. The degrees of
freedom are the electric voltages on the inner edges and the potentials of the floating
nodes on the boundary (nodes outside terminals and current excited terminals). Our
next research will compare the 3 mentioned formulations.

Acknowledgments S. Lup acknowledges the support of the Operational Programme Human

Capital of the Ministry of European Funds through 51675/09.07.2019, SMIS code 125125.
106 G. Ciuprina et al.

References

1. A. Bermudez, R. Rodriguez, P. Salgado, Numerical solution of eddy current problems in

bounded domains using realistic boundary conditions. Comput. Methods Appl. Mech. Eng.
194(2), 411–426 (2005)
2. A. Bermudez, M. Pineiro, R. Rodriguez, P. Salgado, Analysis of an ungauged T,ϕ-ϕ
formulation of the eddy current problem with currents and voltage excitations. ESAIM Math.
Modell. Numer. Anal. (ESAIM: M2AN) 51(6), 2487–2509 (2017)
3. G. Ciuprina, A.S. Lup, C.B. Diţă, D. Ioan, A. Stefanescu, Extraction of TL-lumped RF
macromodels MEMS switches. Nr. Elmag. Multiphy. Mod. and Opt., NEMO, Canada (2015)
4. I.F. Hantila, D. Ioan, Voltage-current relation of circuit elements with field effects. Rev.
Roumaine des Sci. Tech. Serie Electrotechnique et Energetique 39(3), 405–416 (1994)
5. R. Hiptmair, O. Sterz, Current and voltage excitations for the eddy current model. Int. J.
Numer. Modell. Electron. Networks Dev. Fields 18(1), 1–21 (2005)
6. R. Hiptmair, J. Ostrowski, Electromagnetic port boundary conditions: Topological and
variational perspective. Technical Report 2020–27, Seminar for Applied Mathematics, ETH
Zürich, 2020. Int. J. Numer. Modell. 34(3), 1–23 (2020)
7. R. Hiptmair, F. Kramer, J. Ostrowski, A robust Maxwell formulation for all frequencies. IEEE
Trans. Magnet. 44(6), 682–685 (2008)
8. D. Ioan, I. Munteanu, Missing link rediscovered: The electromagnetic circuit element concept.
JSAEM Stud. Appl. Electromagn. Mech. 8, 302–320 (1999)
9. S. Kurz, Some remarks about flux linkage and inductance. Adv. Radio Sci. 2, 39–44 (2004)
10. R. Radulet et al., Introduction des parametres transitoires dans l’etude des circuits electrique
lineaires ayant des elements non filiformes et avec pertes suplimen-taires. Rev. Roum. Sci
Techn. Electrotech. et Energ 11(4), 565–639 (1966)
A Convolution Quadrature Method
for Maxwell’s Equations in Dispersive
Media

Jürgen Dölz, Herbert Egger, and Vsevolod Shashkov

Abstract We study the systematic numerical approximation of Maxwell’s equa-

tions in dispersive media. Two discretization strategies are considered, one based on
the traditional leapfrog time integration method and the other based on convolution
quadrature. The two schemes are proven to be equivalent and to preserve the
underlying energy-dissipation structure of the problem. The second approach,
however, is independent of the number of internal states and in principle allows
to handle rather general dispersive materials. Using ideas of fast-and-oblivious
convolution quadrature, the method can be implemented efficiently.

1 Introduction

We consider electromagnetic wave propagation through linear dispersive media. The

underlying physics are described by Maxwell’s equations

∂t d = curl h, ∂t b = − curl e (1)

with e, h and d, b denoting the electric and magnetic fields and fluxes, respectively,
which are mutually related by the constitutive relations

b = μ0 h, d = 0 ∞ e + p. (2)

Here 0 , μ0 are the permittivity and permeability of vacuum, and ∞ = 1 + ∞  is

the high frequency limit of the relative permittivity. Further, p denotes the memory

J. Dölz
University of Twente, Enschede, Netherlands
e-mail: [email protected]
H. Egger () · V. Shashkov
TU Darmstadt, Darmstadt, Germany
e-mail: [email protected]; [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 107
M. van Beurden et al. (eds.), Scientific Computing in Electrical Engineering,
Mathematics in Industry 36, https://doi.org/10.1007/978-3-030-84238-3_11
108 J. Dölz et al.

 e + p, which is described in frequency domain

part of the polarization pt ot = 0 ∞
by

p̂(s) = 0 χ̂(s)ê(s). (3)

The system is complemented by appropriate boundary and initial conditions. For

ease of presentation, we assume that e(0) = p(0) = 0 in the following. By the
convolution theorem for the Laplace-transform, see e.g. [21, Ch 12], the polarization
can then be expressed in time domain by
 t
p(t) = 0 χ(t − r)e(r)dr. (4)
0

We further assume throughout the paper that the susceptibility kernel χ can be
written as a superposition of simple Debye functions [4], i.e.,

 
i,s − i,∞
χ̂(s) = χ̂i (s) with χ̂i (s) = , (5)
i 1 + sτi

where τi denotes the relaxation time and i,s > i,∞  are the static and high-
 
frequency limits of the electric susceptibility of the ith component with i i,∞ =
 . Such multipole Debye models have been used, e.g., for the modeling of the
∞
dielectric response of biological tissue; see [2, 5] and the references given there. In
general, the summation in (5) may be over infinitely many terms.
One of the key features of the multipole Debye model is its provable passivity,
which follows from the energy–dissipation principle [1, 12]

d  2
E =−   ( τ−
i
 ∂t pi 2 , (6)
dt i 0 i,s i,∞ )

valid for any sufficiently smooth solution of (1)–(3) with homogeneous

or periodic
boundary conditions. Here  ·  is the L2 -norm, further p = i pi is the
decomposition of the memory part of the polarization into its components according
to (5), and
&  '
1 √ √
E =  μ0 h2 +  0 ∞ e2 +  √ 1  pi 2 (7)
2 i 0 (i −∞ )

denotes the electromagnetic energy of the system. Due to the rational structure
of the transfer functions χ̂i , the individual polarizations pi can be characterized
equivalently by the differential equations

τi ∂t pi + pi = 0 (i,s − i,∞ )e, (8)
A Convolution Quadrature Method for Maxwell’s Equations in Dispersive Media 109

with initial values pi (0) = 0, which is the basis for various simulation methods.
Corresponding finite difference and finite element schemes have been considered,
for instance, in [1, 6, 9, 10, 12, 13, 17, 20]. Let us note that with increasing number
of internal states pi , all methods become computationally more and more expensive.
In this paper, we consider a different approach for the numerical solution of (1)–
(3), which allows us to compute the time evolution of e, h, and p without explicitly
computing the internal states pi . As indicated in [8], this can be accomplished
through discretization of the integral (4) by means of appropriate convolution
quadratures [14, 16], instead of integrating (8) with time-differencing schemes. The
complexity of every time step is then independent of the number of internal states
pi . Moreover, using ideas of [18, 19], the additional memory cost for storing the
history of the field e can be reduced to the logarithm of the number of time steps.

2 Structure Preserving Discretization

After space discretization by appropriate finite-difference or finite-element methods

and time-discretization by the leapfrog scheme, the system (1)–(2) with polarization
components defined by (8) can be written in matrix–vector notation as

Mh dτ hn + C en = 0, (9)
 n+1/2 n+1/2
Me dτ en+1/2 + dτ pi − CT h = 0, (10)
i
n+1/2 n+1/2
Md,i dτ pi + Mp,i pi = en+1/2 , i ≥ 1. (11)

The equations hold for all n ≥ 0 and are complemented by appropriate initial
conditions. Note that en and hn+1/2 are the approximations for e(t n ) and h(t n+1/2 )
at staggered grid points t r = rτ with τ denoting the time step size. Furthermore,
dτ en+1/2 = τ1 (en+1 − en ) and dτ hn = τ1 (hn+1/2 − hn−1/2 ) are the central
n+1/2
difference quotients, and en+1/2 = 12 (en+1 + en ), h = 12 (hn+1 + hn ) and
pn+1/2 = 2 (pn+1 + pn ) are the averages of two consecutive steps. Further note that
1
 ).
Eq. (11) was obtained from (8) after dividing by 0 (i,s − i,∞
For appropriate space discretization schemes, the mass matrices Mh , Me are
symmetric, positive-definite, and diagonal or block-diagonal [3, 7], such that (9)–
(11) amounts to an explicit time-stepping scheme. Moreover, the method satisfies
the following discrete equivalent of the underlying energy–dissipation identity.
Lemma 1 Set a2M = (a, a)M and (a, b)M = b T Ma, and denote by

1  n+1/2 n−1/2 
En = (h ,h )Mh + en 2Me + pni 2Mp,i
2 i
110 J. Dölz et al.

the discrete energy at time step t n = nτ . Then any solution of (9)–(11) satisfies
 n+1/2 2
dτ E n+1/2 = − dτ pi Md,i , n ≥ 0.
i

Note that a CFL condition is required to ensure E n ≥ 0; see Remark 1 below.

Proof By elementary computations, one can verify that

1
dτ E n+1/2 = (dτ hn+1 + dτ hn , hn+1/2 )Mh + (dτ en+1/2 , en+1/2 )Me
2
 n+1/2 n+1/2
+ (dτ pi , pi )Mp,i .
i

Note that (a, b)M = (Ma, b) = (Mb, a) where (·, ·) denotes the Euclidean scalar
product. We then test equation (10) with en+1/2 and (11) with dτ pn+1/2 . Moreover,
we test the average of Eq. (9) for step n and n+1 with hn+1/2 . This allows to replace
all terms on the right hand side of the above formula and leads to
 n+1/2 n+1/2
dτ E n+1/2 = −(Cen+1/2 , hn+1/2 ) + (CT hn+1/2 − dτ pi ,e )
i
 n+1/2 n+1/2
+ (en+1/2 − Md,i dτ pi , dτ pi ).
i

Using that (Ca, b) = (CT b, a), one can see that most of the terms drop out and we
obtain the assertion of the lemma. 

Remark 1 Method (9)–(11) automatically inherits the energy-dissipation principle
of the continuous problem. We therefore call it a structure-preserving discretization
scheme. The first term in the energy E can be estimated from below by

(hk+1/2 , hk−1/2 )Mh = hk+1/2 2Mh + τ (hk+1/2 , dτ hk )Mh

1 k+1/2 2 τ2
= hk+1/2 2Mh − τ (Cek , hk ) ≥ h Mh − Cek 2 −1 ,
2 2 Mh

and the last term can be further bounded from below under the assumption that

τ 2 Cv2 −1 ≤ v2Me for all vectors v. (12)

Mh

This CFL condition, restricting the time step τ in dependence of the space
discretization, implies stability of the scheme and allows to show that the energy
E n is a positive and symmetric quadratic functional and thus induces a norm on the
space of state vectors (hn+1/2 , en , pn1 , pn2 , . . .). Together with Lemma 1, this is the
basis for the error analysis of method (9)–(11); we refer to [11] for details.
A Convolution Quadrature Method for Maxwell’s Equations in Dispersive Media 111

3 A Convolution Quadrature Approach

The dimension of the state space and hence also the computational cost for
computing one time step of method (9)–(11) obviously increases with  increasing
number of internal states pi . We will now show that e, h, and p = i pi can
be computed without explicit reference to the internal states pi , which results in
an algorithm that is independent of the number of internal states. Instead of using
Eq. (8), we directly discretize the integral (4) by a convolution sum
n
pn = ωn−k ek . (13)
k=0

This is the field of convolution quadrature, and we refer to [14, 16] for details on the
mathematical background. As illustrated in [8], a proper choice of the convolution
weights {ωn }n≥0 allows to obtain the following equivalence statement.
Lemma 2 Let {ωn }n≥0 be the coefficients of the power series

 ∞

2(1−ξ )
0 χ̂ τ (1+ξ ) = ωn ξ n . (14)
n=0

Then the solution {hn+1/2 , en , pn }n≥0 of the scheme (9)–(11) with e0 = p0i = 0
coincides with the solution of the convolution-quadrature method (9)–(10) and (13).
Proof For convenience of the reader, we briefly summarize the basic ideas of the
proof, which closely follows the arguments presented in [8]. We start by multiplying
equations (11) with ξ n and sum over all n ≥ 0 to obtain
  
Md,i ( ξ1 − 1)pni ξ n + 1
Mp,i ( 2ξ + 12 )pni ξ n = (1 + 12 )en ξ n .
n≥0 n≥0 n≥0 2ξ

An appropriate rearrangement of terms then further leads to

  
2(1−ξ )
pni ξ n = χ̂i τ (1+ξ ) en ξ n ,
n≥0 n≥0

with transfer function χ̂i as defined in (5). Summation over all i and using pn =
 n
i pi and the definition of the weights ωn then yields the assertion. 

Remark 2 According to the above lemma, the convolution quadrature (CQ) method
defined by (9)–(10) and (13)–(14) has the same passivity and stability properties as
the underlying difference scheme (9)–(11). Let us note that instead of the internal
states {pni }i≥0 , the CQ approach utilizes the history {ek }k≤n of the electric field
values to compute the memory part pn of the polarization.
Before closing this section, we briefly comment on the practical computation of
the weights {ωn }n≥0 and the efficient realization of the proposed CQ approach.
112 J. Dölz et al.

Remark 3 Following [14, 15], also see [8], the convolution weights {ωn }n≥0 can be
computed with high accuracy using fast Fourier transforms, i.e.,
L−1 
2 1−ρeiφ
ωn ≈ 1
Lρ n χ̂ τ 1+ρeiφ e−inφ , φ = 2π /L,
=0

and the quadrature error can be controlled by appropriate choice of the parameters
L and ρ; see [14–16] for details. The computation of all weights {ωn }N n=0 with
machine precision requires O(N) evaluations of χ̂. If the material parameters are
inhomogeneous, then the weights ωn will also depend on the spatial variable.
Remark 4 A direct implementation of the CQ approach requires the storage of
the complete history {ek }k≤n to compute the polarization pn via (13), and a naive
computation of the N convolution sums {pn }n≤N is of O(N 2 ) complexity; this
can be reduced to O(N log2 N) by FFT [19]. Using fast and oblivious convolution
quadrature (FOCQ), the required storage can be reduced to O(log N) field vectors
[18, 19]. The basic idea of these approaches is to split the sum (13) into subsums
with exponentially growing number of summands
n L B −1 L
ωk en−k = ωk en−k =: Un ,
k=0 =0 k=B −1 =0

where B > 1 is an integer; we set B −1 := 0 and further assumed for simplicity

that n + 1 = B L is a power of the basis B. Under certain regularity assumptions
on χ̂, each subsum Un can be approximated efficiently using interpolation [18]
or contour integration [19]. In comparison to standard CQ, which requires O(N)
historic field values and O(N) evaluations of the transfer function χ̂ to compute all
weights ωn , the FOCQ algorithm only requires O(log N) historic field vectors and
O(log N) evaluations of the transfer function χ̂ , which also improves the setup cost
substantially.

4 Numerical Illustration

In our test problem, we consider the propagation of an electromagnetic pulse across

the interface between air and human tissue. The dielectric response of the tissue
is characterized by a five-pole Debye model which was taken from [6]. Using the
notation of Sect. 1, the total polarization in this model is prescribed in frequency
 + χ̂(s))ê(s) with   = 3.3 and
domain by p̂t ot (s) = 0 (∞ ∞

8.5 · 105 8.19 · 103 1.19 · 103

χ̂(j ω) = + +
1 + j ω/(138π) 1 + j ω/(86π · 103) 1 + j ω/(1.34π · 106 )
32 45.8
+ + .
1 + j ω/(460π · 106 ) 1 + j ω/(40π · 109)
A Convolution Quadrature Method for Maxwell’s Equations in Dispersive Media 113

For our computational tests, we consider a plane wave setting, in which the fields
are of the form e = (ex , 0, 0), h = (0, hy , 0), and pi = (px,i , 0, 0), and only
depend time t and the propagation direction z. Then (1)–(4) leads to a one–
dimensional wave propagation problem for unknown fields ex , px and hy . As
computational domain, we consider the interval (−1, 1) and we impose periodic
boundary conditions for the electric and magnetic field. The initial values are
described by ex,0 (z) = px,i,0 (z) = 0 and hy,0 (z) = 10e−10z . All quantities are
2

given in SI-units.
For the spatial discretization, we utilize piecewise linear finite elements for ex
and px,i , and piecewise constants to represent hy . Numerical integration by the
vertex rule is used for the assembling of the mass matrices Me , Mp,i , and Md,i , which
leads to a diagonal structure, and the matrix Mh is diagonal automatically. In the case
of piecewise constant material properties only one scalar convolution weight ωn has
to be stored per time step n and per subdomain covered by a dispersive material.
In Fig. 1, we display the magnetic field component hy for the two schemes
presented in Sects. 2 and 3 for some selected time steps. As predicted, the numerical
solutions cannot be distinguished by visual inspection; in our computations, the
maximal difference, caused by inexact computation of the weights ωn , was in the
order of 10−12 , and thus much smaller than the discretization errors. We tested both,
the classical CQ with O(N 2 ) complexity and the FOCQ approach with O(N log N)

n = 0.7833 ·10− 5 n = 0.7800 ·10− 5

6 6
4 4
2 2
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
(a) (b)
n = 0.6511 ·10− 5 n = 0.6154 ·10− 5
6 6
4 4
2 2
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
(c) (d)

Fig. 1 Snapshots of the component hy of the numerical solution restricted to the interval [0, 1] at
different time steps. The solution of the leapfrog method (9)–(11) is drawn in red while that of the
convolution-quadrature method (9)–(10) and (13) is depicted in black. The gray area indicates the
location of the dispersive medium. (a) t = 0.977 · 10−9 . (b) t = 2.930 · 10−9 . (c) t = 4.883 · 10−9 .
(d) t = 6.836 · 10−9
114 J. Dölz et al.

cost; see Remark 4. Both approaches lead to almost identical results. The latter was
however substantially faster, in particular for a large number N of time steps.
From the results in Fig. 1, one can also recognize the basic physical behavior: In
the initial phase, the pulse propagates through air and the total energy of the system
is conserved exactly. When impinging on the air-tissue interface, a part of the pulse
gets reflected and the rest penetrates into the dispersive medium. Propagation in
the medium is substantially slower and, moreover, energy is dissipated according to
Lemma 1. We were able to reproduce this energy balance up machine precision.

5 Summary

We presented two discretization strategies for simulating Maxwell’s equation in

dispersive media, which were proven to be equivalent for certain classes of problems
and to comply with the underlying energy–dissipation structure of the problem.
The second scheme, which is based on a convolution quadrature approach, is
independent of the number of internal states or relaxation times, and can be
applied to dispersive media with rather general memory kernels. This might become
particularly useful also in the context of uncertainty quantification.

Acknowledgments The authors are grateful for support by the German Research Foundation
(DFG) via grants TRR 146 project C03, TRR 154, project C04, and Eg-331/1-1 and through grant
Center for Computational Engineering at TU Darmstadt.

References

1. V.A. Bokil, N.L. Gibson, Convergence analysis of Yee schemes for Maxwell’s equations in
Debye and Lorentz dispersive media. Int. J. Numer. Anal. Model. 11, 657–687 (2014)
2. J. Clegg, M. Robinson, A genetic algorithm for optimizing multi-pole Debye models of tissue
dielectric properties. Phys. Med. Biol. 57, 6227–43 (2012)
3. G. Cohen, Higher-Order Numerical Methods for Transient Wave Equations (Springer, Berlin,
2002)
4. P. Debye, Polar Molecules (Chemical Catalogue Company, New York, 1929)
5. S. Gabriel, R.W. Lau, C. Gabriel, The dielectric properties of biological tissues: III. Parametric
models for the dielectric spectrum of tissues. Phys. Med. Biol. 41, 2271–93 (1996)
6. O.P. Gandhi, B.-Q. Gao, J.-Y. Chen, A frequency-dependent finite-difference time-domain
formulation for general dispersive media. IEEE Trans. Microw. Theory Tech. 4144–4, 658–
665 (1993)
7. H. Egger, B. Radu, A mass-lumped mixed finite element method for Maxwell’s equations
(2018). arXiv:1810.06243. to appear in Proceedings of SCEE 2018
8. H. Egger, K. Schmidt, V. Shashkov, Multistep and Runge–Kutta convolution quadrature
methods for coupled dynamical systems. J. Comput. Appl. Math. 387, 112618 (2020)
9. D. Jiao, J.-M. Jin, Time-domain finite-element modeling of dispersive media. IEEE Microw.
Wireless Components Lett. 11, 220–222 (2001)
A Convolution Quadrature Method for Maxwell’s Equations in Dispersive Media 115

10. M.J. Jenkinson, J.W. Banks, High-order accurate FDTD schemes for dispersive Maxwell’s
equations in second-order form using recursive convolutions. J. Comput. Appl. Math. 336,
192–218 (2018)
11. P. Joly, Variational methods for time-dependent wave propagation problems, in Topics in
Computational Wave Propagation. LNCSE, vol. 31 (Springer, Berlin, 2003), pp. 201–264
12. S. Lanteri, C. Scheid, Convergence of a discontinuous Galerkin scheme for the mixed time-
domain Maxwell’s equations in dispersive media. IMANUM 33, 432–459 (2012)
13. J. Li, Error analysis of finite element methods for 3-D Maxwell’s equations in dispersive media.
J. Comput. Appl. Math. 188, 107–120 (2006)
14. C. Lubich, Convolution quadrature and discretized operational calculus. I. Numer. Math. 52,
129–145 (1988)
15. C. Lubich, Convolution quadrature and discretized operational calculus. II. Numer. Math.
52(4), 413–425 (1988)
16. C. Lubich, A. Ostermann, Runge-Kutta methods for parabolic equations and convolution
quadrature. Math. Comp. 60(201), 105–131 (1993)
17. R. Luebbers, F.P. Hunbserger, K.S. Kunz, R.B. Standler, M. Schneider, A frequency-dependent
finite-difference time-domain formulation for dispersive materials. IEEE Trans. Electromag.
Compat. 32, 222–227 (1990)
18. J. Roychowdhury, Reduced-order modeling of time-varying systems. IEEE Trans. Circuits
Syst. II 46, 1273–1288 (1999)
19. A. Schädle, M. López-Fernandez, C. Lubich, Fast and oblivious convolution quadrature. SIAM
J. Sci. Comput. 28, 421–438 (2006)
20. S. Shaw, Finite element approximation of Maxwell’s equations with Debye memory. Adv.
Numer. Anal. 2010, 923832 (2010)
21. W.A. Strauss, Partial Differential Equations. An Introduction (Wiley, New York, 1992)
On the Stability of Harmonic Coupling
Methods with Application to Electric
Machines

H. Egger, M. Harutyunyan, M. Merkel, and S. Schöps

Abstract Harmonic stator-rotor coupling offers a promising approach for the

interconnection of rotating subsystems in the simulation of electric machines. This
paper studies the stability of discretization schemes based on harmonic coupling in
the framework of mortar methods for Poisson-like problems. A general criterion
is derived that allows to ensure the relevant inf-sup stability condition for a
variety of specific discretization approaches, including finite-element methods and
isogeometric analysis with harmonic mortar coupling. The validity and sharpness of
the theoretical results is demonstrated by numerical tests.

1 Introduction

Electric drives naturally consist of different subdomains, i.e. the stator and rotor,
which move relative to each other. The time-varying geometry and nonlinearities
caused by saturation effects formally require a time-domain analysis, which is
often realized by solving a sequence of quasi-stationary problems at different
working points. Several strategies have been proposed for the simulation of the
corresponding equations of magnetostatics and, in particular, for the coupling of the
fields across the air gap between stator and rotor. As it is common practice, see e.g.
[13, 14, 17], we consider a two dimensional regime, in which the unknown fields are
described by the axial component of the magnetic vector potential. The governing
system then consists of two Poisson-like problems for the stator and the rotor, which
can be coupled via Lagrange multipliers. Such domain decompositions of mortar

H. Egger ()
Numerical analysis and Scientific Computing, Technische Universität Darmstadt, Darmstadt,
Germany
e-mail: [email protected]
M. Harutyunyan · M. Merkel · S. Schöps
Computational Electromagnetics, Technische Universität Darmstadt, Darmstadt, Germany
e-mail: [email protected]; [email protected];
[email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 117
M. van Beurden et al. (eds.), Scientific Computing in Electrical Engineering,
Mathematics in Industry 36, https://doi.org/10.1007/978-3-030-84238-3_12
118 H. Egger et al.

methods, which couple subdomains via Lagrange multipliers, have been investi-
gated intensively in the literature [2, 3, 5, 20]; see [8, 13] for results concerning
electric machines. It is well-known that a careful choice of approximation spaces is
required to obtain stable discretization schemes for underlying saddlepoint problems
[6, 18]; appropriate stabilization [15] could be used as an alternative approach.
In this paper, we investigate the stability of mortar discretizations using trigono-
metric functions as Lagrange multipliers, called harmonic coupling methods in
[4, 13]. We discuss in detail the discrete inf-sup condition which is necessary
and sufficient to guarantee the stability of such approximations. We provide a
simple criterion for the maximal number of harmonics used as Lagrange multipliers
depending on the mesh size and polynomial degree of the subdomain discretizations
which guarantees the stability of the scheme. Our analysis applies to the harmonic
coupling of various discretization methods, e.g. obtained by isogeometric analysis
(IGA) [4, 7, 16], and can in principle be extended to other Lagrange multiplier
spaces.
The remainder of this note is organized as follows: In Sect. 2, we introduce the
model problem to be considered and we summarize some well-known results about
its analysis and discretization. In Sect. 3, we then turn to the harmonic stator-rotor
coupling, and we state and prove our main results. Section 4 is concerned with
numerical tests, in which we demonstrate the validity of our stability criterion for
low and high order discretizations based on IGA.

2 Model Problem

We consider a typical geometric setup that consists of two subdomains Ω1 , Ω2

representing, respectively, the stator and rotor, separated by a small air gap which
contains the interface Γ = ∂Ω1 ∩ ∂Ω2 ; see Fig. 1. Let Σ = ∂Ω \ Γ , = 1, 2,
be the remaining parts of the subdomain boundaries and f = f |Ω denote the
restriction of a function f defined on Ω1 ∪ Ω2 to the subdomain Ω . We then
consider the following elliptic interface problem: Inside the two subdomains, we
require

−div(ν ∇u ) = j , in Ω , (1)
u = 0, on Σ , (2)

where u denotes the z-component of the magnetic vector potential, ν the magnetic
reluctivity, and j = js +divm⊥ a generalized current density with js denoting the z-
component of the source currents and m⊥ = (my , −mx ) the rotated magnetization
vector of the permanent magnet. The corresponding in-plane components of the
magnetic flux density and field strength are given by b = (∂y u, −∂x u) = ∇ ⊥ u
Inf-Sup Stability of Harmonic Coupling Methods 119

stator

windings
rotor
permanent
magnets

Fig. 1 Typical structure of a 6-pole permanent magnet synchronous machine (left) and the coarsest
mesh of the stator domain as used in our numerical tests (right)

and h = νb, respectively. The coupling of the fields across the interface Γ is
accomplished by the conditions

u1 = u2 , on Γ, (3)
n · (ν1 ∇u1 ) = n · (ν2 ∇u2 ), on Γ, (4)

which correspond to the conditions for the normal continuity of b and the tangential
continuity of h, respectively; see e.g. [4, 13]. Here n = n2 is the unit normal vector
at Γ pointing from Ω2 to Ω1 . In the context of electric machines, it is natural to
assume that Ω are bounded domains with piecewise smooth boundaries Σ and Γ ,
having non-zero measure. Moreover, we can assume that ν is bounded from above
and below by positive constants ν, ν, i.e. ν ≤ ν(x) ≤ ν for all x ∈ Ω1 ∪ Ω2 .
The weak formulation of the interface problem (1)–(4) then reads as follows:
Find u ∈ V = {v ∈ H 1 (Ω1 ∪ Ω2 ) : v|Σ = 0} and λ ∈ M = H −1/2 (Γ ) such that

(ν∇u, ∇v)Ω1 ∪Ω2 + λ, [v]Γ = j, vΩ1 ∪Ω2 ∀v ∈ V , (5)

[u], μΓ = 0 ∀μ ∈ M. (6)
 
Here (a, b)Ω1 ∪Ω2 = Ω1 a·b dx + Ω2 a·b dx is the usual scalar product of functions
a, b ∈ L2 (Ω1 ∪ Ω2 ), while a, bΩ1 ∪Ω2 , a, bΓ are the duality products on V ×
V  and M × M  , respectively, with V  , M  denoting the dual spaces of V and M.
Furthermore, H 1 (Ω1 ∪ Ω2 ) denotes the space of piecewise smooth functions v with
restrictions v = v|Ω ∈ H 1 (Ω ) for = 1, 2 and [v] = v1 −v2 denotes the jump of
such functions across the interface Γ , and H −1/2(Γ ) can be obtained by mapping
the space of 2π-periodic functions u(φ) = a0 + n≥1 an cos(nφ) + bn sin(nφ)

whose Fourier coefficients satisfy a02 + n≥1 (1 + n2 )−1/2 (an2 + bn2 ) < ∞.
120 H. Egger et al.

Lemma 1 For any js ∈ L2 (Ω) and m ∈ L2 (Ω)2 , the variational problem (5)–(6)
with j = js + divm⊥ has a unique solution (u, λ) ∈ V × M and there holds

uH 1 (Ω1 ∪Ω2 ) + λH −1/2 (Γ ) ≤ C js L2 (Ω1 ∪Ω2 ) + mL2 (Ω1 ∪Ω2 ) )

where the constant C does not depend on u, λ, js or m. Moreover, u is the unique

weak solution of (1)–(4) and λ = n · (ν∇u ) the associated tangential component
of the magnetic field strength h at the interface.
Remark 1 The result is well-known and a similar assertion can already be found in
the work of Babuska [1]. Using Brezzi’s theory for saddlepoint problems [6], the
essential ingredient turns out to be the inf-sup stability condition

μ, [v]Γ
inf sup ≥ β > 0. (7)
μ∈M v∈V μH −1/2 (Γ ) vH 1 (Ω1 ∪Ω2 )

Following [18], condition (7) can be proven as follows: Let z1 ∈ H 1 (Ω1 ) be the
weak solution of the mixed boundary value problem

−Δz1 = 0 in Ω1 with z1 = 0 on Σ1 and ∂n z1 = μ on Γ. (8)

Then by standard arguments for elliptic problems [1, 2], one can show that

z1 H 1 (Ω1 ) ≤ c2 μH −1/2 (Γ ) and μ, z1 Γ ≥ c1 μ2H −1/2 (Γ ) ,

with positive constants c1 , c2 only depending on Ω1 , Σ1 , and Γ .

Now define z ∈ H 1 (Ω1 ∪ Ω2 ) by z = z1 on Ω1 and z = 0 on Ω2 . Then
c1
μ, [z]Γ = μ, z1 Γ ≥ c1 μ2H −1/2 (Γ ) ≥ μH −1/2 (Γ ) z1 H 1 (Ω1 ) .
c2
c1
The result now follows with β = c2 by noting that zH 1 (Ω1 ∪Ω2 ) = z1 H 1 (Ω1 ) .


Remark 2 The solution z1 of the auxiliary problem (8) suffices to prove the inf-sup
stability conditions but does not yield the supremum in (7). The simplicity of the
auxiliary problem however allows us to calculate z1 analytically later on and thus to
obtain a computable bound β > 0 depending only on the geometric setting.
Discretization As a next step, we now consider Galerkin approximations of the
weak formulation (5)–(6): Find uh ∈ Vh ⊂ V and λN ∈ MN ⊂ M such that

(ν∇uh , ∇vh )Ω1 ∪Ω2 + λN , [vh ]Γ = j, vh Ω1 ∪Ω2 ∀vh ∈ Vh , (9)
[uh ], μN Γ = 0 ∀μN ∈ MN . (10)

Following the usual convention, we assume that Vh and MN are finite dimensional.
Inf-Sup Stability of Harmonic Coupling Methods 121

Lemma 2 Let the conditions of Lemma 1 be valid and assume that

μN , [vh ]Γ

inf sup ≥ β > 0 (11)
μN ∈MN vh ∈Vh μN H −1/2 (Γ ) vh H 1 (Ω1 ∪Ω2 )

Then problem (9)–(10) has a unique solution uh ∈ Vh , λN ∈ MN . Furthermore

u − uh H 1 (Ω1 ∪Ω2 ) + λ − λN H −1/2 (Γ ) (12)

 
≤ C inf u − vh H 1 (Ω1 ∪Ω2 ) + inf λ − μN H −1/2 (Γ )
vh ∈Vh μN ∈MN

where (u, λ) is the solution of the continuous variational problem (5)–(6) and the
constant C depends only on β  in (11), the bounds for ν, and the geometry.
Remark 3 All conditions required for the proof of the corresponding result on the
continuous level, except the inf-sup stability condition, are inherited by the Galerkin
approximation. The existence of a unique solution can thus again be deduced from
Brezzi’s saddlepoint theory [6]. The error estimate (12) follows from Galerkin
orthogonality and standard arguments; we refer to [5, 6] for details. Hence any
choice of approximation spaces Vh , MN that allows to prove the discrete inf-sup
stability condition (11) will lead to a well-posed discrete problem.

3 Harmonic Stator-Rotor Coupling

We now consider a particular class of Galerkin approximations (9)–(10) in which

Vh is constructed by piecewise polynomials, while the Lagrange multiplier space
MN is defined by trigonometric polynomials. Our analysis in particular also covers
the harmonic-coupling of the methods considered in [4, 13].
Using polar coordinates, the computational domain Ω = Ω1 ∪ Ω2 can be
represented as the image of a rectangle Ω % under a mapping F : Ω % → Ω; see
Fig. 2. Now let T%h denote a shape-regular partition of Ω
%1 ∪ Ω%2 into triangles and
/or rectangles of size h. The meshes of the two sub-domains are assumed to be
geometrically conforming, but they may be non-matching across the interface. We
denote by Pk (T%h ) the space of piecewise polynomials over T%h of degree ≤ k and
by M%N = span{sin(nπξ ), cos(nπξ ) : 0 ≤ n ≤ N} the spaces of trigonometric
polynomials of degree ≤ N. We then choose the approximation spaces Vh , MN s.t.

%N )
MN = F (M and Vh = Vh |Ω1 ∪ Vh |Ω2 ⊂ F (P k (Th )) ∩ V . (13)

By the condition Vh = Vh |Ω1 ∪ Vh |Ω2 we mean that discrete functions, when

restricted to one of the sub-domains and extended by zero to the other still belong to
the approximation space Vh . The basic assumption for the discrete inf-sup stability
condition (7) of the corresponding Galerkin approximation (9)–(10) is the following.
122 H. Egger et al.

Σ̂1
Ω1 Σ1
Ω̂1

Γ̂
Γ
Ω2
Ω̂2
Σ̂2 Σ2

Fig. 2 Sketch of a subset of the rectangular reference domain Ω̂ and its mesh (left) and the
physical domain Ω = F (Ω̂) and mesh obtained after mapping. The boundaries on the left and
right are only introduced for the illustration but not present in our application

Theorem 1 Assume that there exists a linear operator Πh : V |Ω1 → Vh |Ω1 such
that

Πh v1 H 1 (Ω1 ) ≤ c3 v1 H 1 (Ω1 ) , (14)

Πh v1 = πh v1 on Γ, (15)
v − πh vH −1/2 (Γ ) ≤ c4 hk vH 1/2 (Γ ) , (16)

where πh : L2 (Γ ) → Vh |Ω1 ∩Γ denotes the L2 -projection on Γ . Then there exists

a constant 0 <  < 1, depending only on c3 , c4 in (14)–(16), such that the discrete
inf-sup condition (11) holds with β  = β  () whenever N and h are chosen such
that

Nh/k ≤ 1 − . (17)

Proof As an immediate consequence of the continuous inf-sup condition (7), we

can find for μ = πh λN a function z ∈ V with z = 0 on Ω2 , such that

πh λN , [z]Γ ≥ βzH 1 (Ω1 ∪Ω2 ) πh λN H −1/2 (Γ ) .

We then define zh = Πh z1 on Ω1 and zh = 0 on Ω2 , and observe that

πh λN , [zh ]Γ = πh λN , [Πh z]Γ = πh λN , πh [z]Γ = πh λN , [z]Γ ,

where we used property (15) and the orthogonality of the L2 -projection πh . Together
with the previous estimate and employing condition (14), we thus obtain

β
πh λN , [zh ]Γ ≥ βzH 1 (Ω1 ∪Ω2 ) πh λN H −1/2 (Γ ) ≥ zh H 1 (Ω1 ∪Ω2 ) πh λN H −1/2 (Γ ) .
c3
Inf-Sup Stability of Harmonic Coupling Methods 123

Using the triangle inequality, we can further estimate

πh λN H −1/2 (Γ ) ≥ λN H −1/2 (Γ ) − λN − πh λN H −1/2 (Γ ) ,

and the last term can be bounded with the approximation error estimate (16) by

λN − πh λN H −1/2 (Γ ) ≤ c4 hk λN H 1/2 (Γ ) ≤ C  hk NλN H −1/2 (Γ ) .

In the second estimate, we here used an inverse inequality for the finite dimensional
Lagrange multiplier space MN . In summary, we thus obtain

πh λN , [zh ]Γ ≥ β(1 − C  Nh/k)zh H 1 (Ω1 ∪Ω2 ) λN H −1/2 (Γ ) ,

from which the assertion of the theorem follows immediately.

Remark 4 The conditions of the theorem hold for a variety of discretization
methods, e.g. FEM or IGA. The projection operator Πh can here be constructed
following the ideas of [11, 19] or [9] and the approximation property (16) for πh
is well-known; details will be given in a forthcoming publication. The resulting
harmonic-coupling mortar methods are thus stable, if the number of degrees of
freedom n ∼ k/ h located at the interface exceeds the number of coupling modes N
to some extent, cf. (17). Our main arguments may be applied to other problems and
discretization strategies.

4 Numerical Results

We now illustrate the theoretical results of Theorem 1 by some numerical tests

using an IGA discretization [16] as implemented in GeoPDEs [12]. We here only
report about results concerning the discrete inf-sup condition; numerical results
concerning the magnetic fields can be found e.g. in [4, 13]. The geometry used in
our computations is depicted in Fig. 1. Following the arguments given in Remark 1
and underlying the proof of Theorem 1, we have

μ, vΓ μ, z1 Γ

sup ≥ sup ≥ βμH −1/2 (Γ ) , (18)
v∈V vH 1 (Ω1 ∪Ω2 ) z1 ∈V1 z1 H 1 (Ω1 )

where V = {v ∈ H 1 (Ω1 ∪ Ω2 ) : v|Σ1 ∪Σ2 = 0} and V1 = {v ∈ V : v|Ω2 = 0} ⊂

V . Due to the Dirichlet boundary conditions on Σ1 , we can choose v1 H 1 (Ω1 ) =
∇v1 L2 (Ω1 ) as the norm on V1 . One can then show that the second supremum
in (18) is attained by the solution z1 of the mixed boundary value problem (8). For
our model problem, Ω1 = {x ∈ R2 : R1 < |x| < R2 } is a simple annulus with
radii R1 = 0.0447 and R2 = 0.0675, and the solution of the above problem can be
computed analytically in the form of a Fourier series, and we define S1 μ := v1 |Γ .
124 H. Egger et al.

Table 1 Discrete inf-sup constants obtained for n gridpoints at the interface Γ and harmonic order
N = cn of the Lagrange multipliers for different refinement levels and scaling parameters c
c \ 1 2 3 4
1/4 0.135237 0.135556 0.135676 0.135693
1/3 0.135237 0.135556 0.135661 0.135684
3/8 0.135237 0.135536 0.135611 0.135684
1/2 3.526e–08 2.532e–08 2.401e–08 2.401e–08

Table 2 Discrete inf-sup constants for n spline degrees of freedom on the interface Γ and
harmonic order N = cn of the Lagrange multipliers for polynomial degree k and scaling
parameter c
c \k 2 3 4 5
1/4 0.135721 0.135723 0.135723 0.135723
1/3 0.135721 0.135722 0.135723 0.135723
3/8 0.135720 0.135723 0.135723 0.135723
1/2 3.652e–08 0 8.082e–08 1.825e–08

The largest possible constant β such that the second estimate of (18) remains true
for all μ ∈ H −1/2 (Γ ) can then be characterized by the minimal eigenvalue of

μ, S1 3
μH −1/2 (Γ )×H 1/2 (Γ ) = β 2 (μ, μ̃)2H −1/2 (Γ ) ∀μ̃ ∈ H −1/2(Γ ).

For the problem

√ under consideration, the solution can be computed explicitly which
gives β = R1 ln(R2 /R1 ) ≈ 0.13573. The discrete inf-sup constant is evaluated
by numerically solving the corresponding discretized eigenvalue problem; see e.g.
[10].
In the first series of tests, we utilize the lowest order approximation and consider
a sequence of uniformly refined meshes. The discrete inf-sup constant is computed
as outlined above. The results of these computations are depicted in Table 1. The
coarsest mesh has n = 144 vertices at the interface Γ and is doubled in every
refinement step; see Fig. 1. For c = 1/2, we have dim(MN ) = 2N + 1 = n + 1 > n,
and the discrete inf-sup stability condition is violated. The results of Table 1 thus
perfectly agree with the theoretical predictions of Theorem 1. In a second sequence
of tests, we study the dependence of the inf-sup constant on the polynomial degree k
of the spline approximation on the mesh with refinement level 2. The corresponding
results are summarized in Table 2. For the choice c = 1/2, the number of Lagrange
multipliers 2N + 1 = n + 1 > n again exceeds the number of the spline degrees at
the interface Γ and the discrete inf-sup stability fails. The computational results are
again in perfect agreement with the theoretical predictions.

Acknowledgments This work is supported by the ‘Excellence Initiative’ of the German Federal
and State Governments and by the Graduate School of Computational Engineering at Technische
Universität Darmstadt and the grants TRR 154 project C04 and TRR 146 project C03.
Inf-Sup Stability of Harmonic Coupling Methods 125

References

1. I. Babuška, The finite element method with Lagrangian multipliers. Numer. Math. 20, 179–192
(1973)
2. F. Ben Belgacem, The mortar finite element method with Lagrange multipliers. Numer. Math.
84, 173–197 (1999)
3. C. Bernardi, Y. Maday, A.T. Patera, A new nonconforming approach to domain decomposition:
the mortar element method, in Nonlinear Partial Differential Equations and their Applications.
Pitman Research Notes in Mathematics Series, vol. 299 (1994), pp. 13–51
4. Z. Bontinck, J. Corno, S. Schöps, H. De Gersem, Isogeometric analysis and harmonic stator-
rotor coupling for simulating electric machines. Comput. Meth. Appl. Mech. Engrg. 334,
40–55 (2018)
5. D. Braess, W. Dahmen, C. Wieners, A multigrid algorithm for the mortar finite element
method. SIAM J. Numer. Anal. 37, 48–69 (1999)
6. F. Brezzi, On the existence, uniqueness and approximation of saddle-point problems arising
from Lagrangian multipliers. RAIRO Anal. Numer. 8, 129–151 (1974)
7. E. Brivadis, A. Buffa, B. Wohlmuth, L. Wunderlich, Isogeometric mortar methods. Comput.
Methods Appl. Mech. Engrg. 284, 292–319 (2015)
8. A. Buffa, Y. Maday, F. Rapetti, A sliding mesh-mortar method for a two dimensional currents
model of electric engines. ESAIM Math. Model. Numer. Anal. 35, 191–228 (2001)
9. A. Buffa, E.M. Garau, C. Giannelli, G. Sangalli, On quasi-interpolation operators in spline
spaces, in Building Bridges: Connections and Challenges in Modern Approaches to Numerical
Partial Differential Equations. Lecture Notes in Computational Science and Engineering, vol.
114 (Springer, Berlin, 2016), pp. 73–91
10. D. Chapelle, K.J. Bathe, The inf-sup test. Comput. Struct. 47(4/5), 537–545 (1993)
11. P. Clément, Approximation by finite element functions using local regularization. RAIRO 9,
77–84 (1975)
12. C. de Falco, A. Reali, R. Vázquez, GeoPDEs: a research tool for Isogeometric Analysis of
PDEs. Adv. Eng. Softw. 42, 1020–1034 (2011)
13. H. De Gersem, T. Weiland, Harmonic weighting functions at the sliding interface of a finite-
element machine model incorporating angular displacement. IEEE Trans. Magn. 40, 545–548
(2004)
14. G. Gyselinck, L. Vandevelde, P. Dular, C. Geuzaine, A general method for the frequency
domain FE modeling of rotating electromagnetic devices. IEEE Trans. Magn. 39, 1147–1150
(2003)
15. P. Hansbo, C. Lovadina, I. Perugia, G. Sangalli, A Lagrange multiplier method for the finite
element solution of elliptic interface problems using non-matching meshes. Numer. Math. 100,
91–115 (2005)
16. T.J.R. Hughes J.A. Cottrell, T. Bazilevs, Isogeometric analysis: CAD, finite elements, NURBS,
exact geometry and mesh refinement. Comput. Meth. Appl. Mech. Eng. 194, 4135–4195
(2005)
17. E. Lange, F. Henrotte, K. Hameyer, A variational formulation for nonconforming sliding
interfaces in finite element analysis of electric machines. IEEE Trans. Magn. 46, 2755–2758
(2010)
18. P.A. Raviart, J.M. Thomas, Primal hybrid finite element methods of 2nd order elliptic
equations. Math. Comp. 31, 391–413 (1977)
19. L.R. Scott, S. Zhang, Finite element interpolation of nonsmooth functions satisfying boundary
conditions. Math. Comp. 54, 483–493 (1990)
20. B.I. Wohlmuth, A mortar finite element method using dual spaces for the Lagrange multiplier.
SIAM J. Numer. Anal. 38, 989–1012 (2000)
Multifidelity Uncertainty Quantification
for Optical Structures

Niklas Georg, Christian Lehmann, Ulrich Römer, and Rolf Schuhmann

Abstract This work addresses uncertainty quantification for optical structures. We

decouple the propagation of uncertainties by combining local surrogate models with
a scattering matrix approach, which is then embedded into a multifidelity Monte
Carlo framework. The so obtained multifidelity method provides highly efficient
estimators of statistical quantities jointly using different models of different fidelity
and can handle many uncertain input parameters as well as large uncertainties. We
address quasi-periodic optical structures and propose the efficient construction of
low-fidelity models by polynomial surrogate modeling applied to unit cells. We
recall the main notions of the multifidelity algorithm and illustrate it with a split
ring resonator array simulation, serving as a benchmark for the study of optical
structures. The numerical tests show speedups by orders of magnitude with respect
to the standard Monte Carlo method.

1 Introduction

Manufacturing on the nanometer-scale exhibits strong variability in the finally built

structures which should be addressed in a simulation based design approach. The
field of uncertainty quantification provides suitable tools to model and quantify

N. Georg ()
Institut für Dynamik und Schwingungen, Technische Universität Braunschweig, Braunschweig,
Germany
Centre for Computational Engineering, Technische Universität Darmstadt, Darmstadt, Germany
e-mail: [email protected]
C. Lehmann · R. Schuhmann
Theoretische Elektrotechnik, Technische Universität Berlin, Berlin, Germany
e-mail: [email protected]; [email protected]
U. Römer
Institut für Dynamik und Schwingungen, Technische Universität Braunschweig, Braunschweig,
Germany
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 127
M. van Beurden et al. (eds.), Scientific Computing in Electrical Engineering,
Mathematics in Industry 36, https://doi.org/10.1007/978-3-030-84238-3_13
128 N. Georg et al.

uncertainties in the geometrical and material constitutive parameters. In this con-

tribution, we focus on the benchmark example of a split ring resonator (SRR)
array with random input data. In particular, we model uncertainties in the SRR
geometry with random variables and quantify the implied variation in the frequency
response of the system. Such quasi-periodic optical structures may feature a large
number of uncertain parameters which makes the application of many methods,
such as standard or even sparse Polynomial Chaos difficult, see [1] for instance.
We present a remedy by applying spectral polynomial expansions on the unit cell
level in the framework of the Scattering Matrix Approach (SMA) [2], which yields
a significant reduction of the computational effort. Since the coupled surrogate
model may be biased, we use a Multifidelity Monte Carlo (MFMC) method [3],
which combines different numerical models with different fidelity, to obtain efficient
statistical estimators. In particular, through limited recurrences to a high-fidelity
simulation of the entire structure, the MFMC method then corrects for possible
approximation errors in the low-fidelity data.

2 Decoupled Uncertainty Propagation with Scattering

Matrices

Our benchmark application is a simplified model of an array of coupled SRRs,

motivated by the research on optical metamaterials [4, 5]. It consists of a periodic,
but finite-size array of metallic SRR structures on a nanometer-scale, each of which
can be interpreted as a realization of a resonance circuit, with the ring and the
small gap acting as inductance and capacitance, respectively. More details on the
geometry and setup will be given in Sect. 4. Due to the unavoidable tolerances in
manufacturing of such small structures the geometric properties of each SRR will
slightly vary, and the periodicity of the array of coupled resonator will not be perfect
(see Figs. 54 and 57 in [5] for an illustration). Thus, we introduce a parameter vector
ycell,j ∈ Ξ ⊂ RP , which models variations in the geometry or material of the
structure in cell j . The full input vector is then given as y = (yTcell,1 , . . . , yTcell,N )T ⊂
RN·P , and all results of the forward model depend on this input vector.
The structure is excited by a plane wave and the reflection and transmission
coefficients are evaluated. Translated into the language of dispersion analysis, the
array is expected to feature a number of bandgaps, i.e. intervals on the frequency (or
wavelength) axis where no transmission through the structure is possible. Both the
finite size of the arrays (in our case up to seven unit cells) and the parameter variation
in each SRR will have some influence on the corresponding limit frequencies.
The electromagnetic treatment of this application example requires the solution
of the wave equation with an appropriate excitation at the ports. From the field
solutions the amplitudes ai and bi of properly normalized incoming and outgoing
waves are determined. They are coupled by the scattering matrix S(j ω),

(. . . bi (j ω) . . .)T = S(j ω) (. . . ai (j ω) . . .)T ,

Multifidelity Uncertainty Quantification for Optical Structures 129

with r(j ω) = S11 (j ω) the reflection coefficient at the input port. Note that we omit
the frequency dependency of S in the following to enhance the readability.
For the discretization we apply the efficient Finite Integration Technique (FIT)
time-domain algorithm [6]. It relies on a three-dimensional Cartesian mesh and
allows calculating broadband results with single transient simulation runs (using
Discrete Fourier Transform on the signals). The calculation of scattering parameters
proceeds in two steps: First, the two-dimensional eigenvalue problem of the port
apertures is analyzed to obtain the field patterns and cutoff-frequencies of the so-
called waveguide modes. Note that a lossfree model is considered here, and the array
of SRRs is transversally terminated by perfect electric and magnetic boundary con-
ditions. Second, these mode patterns and their well-known orthogonality properties
are used to both excite the three-dimensional structure and to extract the amplitudes
of the out-going waves at the ports. From one simulation run, one column of the
scattering matrix can be obtained. For further details on the FIT we refer to the
literature.
A technique to reduce the computational cost in the analysis of periodic
structures is to decompose the SRR array into its single unit cells and to calculate
separate scattering matrices S(i) for each of them. The final concatenation of these
single-cell results can be accomplished by switching to the transfer matrices T(i)
which map the wave amplitudes of the right hand side of each cell to the left hand
side (rather than from input to output quantities as with S). For a system with 2
ports:
* + * + * + * + * +
−1 −1
b1 a1 b1 a2 S12 − S11 S21 S22 S11 S21
=S ↔ =T with T = −1 −1 .
b2 a2 a1 b2 −S21 S22 S21

Extended formulas for larger S, T, which take several port-modes into account, can
easily be derived. Using transfer matrices, the total system behavior of N cells is
simply given by a matrix multiplication T = T(1) · . . . · T(N) . This approach has
been used previously in [2, 7] and is referred to as SMA.
This procedure has the intrinsic weakness that the coupling between the unit
cells is not governed by a single waveguide mode alone, but an unknown number
of higher modes may contribute. Of course, the coupling of modes at frequencies
below their cutoff-frequency decreases rapidly with increasing spatial distance of
the single SRRs. However, especially if there are resonances within the frequency
range of interest (which clearly is the case for the SRRs as one of their working
principles), this systematic error may become significant. In theory an extension of
the SMA to an arbitrary number of coupling modes is straight-forward. However, the
required number (and/or selection) of modes is sometimes hard to estimate a-priori,
and the calculation of the extended transfer matrix increases the computational cost.
Our approach removes any possible systematic error introduced in the coupling, by
treating the SMA-based predictions as low-fidelity data and by correcting them with
a couple of time domain solutions of the entire structure.
130 N. Georg et al.

Non-intrusive Uncertainty Quantification (UQ) usually requires the repeated

evaluation of the scattering matrices S(y) for different values of the inputs y. Even
with SMA the computational cost to evaluate a large number of sample points using
the FIT might become prohibitive. Hence, we propose to construct a surrogate model
for a unit cell of the periodic structure. In particular, we use a spectral collocation
method, i.e. an approximation


C
 
S(j ) (ycell,j ) ≈ Suc;C (ycell,j ) := S(j ) y(i)
cell,j Ψi (ycell,j ) (1)
i=1

where uc is short for unit cell and j = 1, . . . , N refers to an arbitrary unit cell of the
structure. Also, {y(i)
cell,j }i=1 ⊂ Ξ denotes a set of collocation points, e.g. Chebyshev
C

nodes, and Ψi denote the corresponding barycentric Lagrange polynomials. We

emphasize that the same surrogate model is employed for all cells. It can be
straightforwardly employed to obtain a surrogate of the full structure based on the
SMA as (after transformation into T matrices)

T(y) ≈ TC (y) := Tuc;C (ycell,1 ) · . . . · Tuc;C (ycell,N ). (2)

We also emphasize that (2) can be evaluated with negligible computational cost. In
order to highlight the efficiency of the proposed combination of SMA and spectral
surrogates for the unit cell, we give a few comments on the alternative approach,
i.e. spectral approximation of the full structure. Due to spectral convergence
properties, global polynomial approximations can be highly efficient, even up to a
moderately large number of parameters (e.g., up to 10–20) using adaptive sparse
approximations, see e.g. [1]. However, these methods still suffer from the so-
called curse-of-dimensionality, i.e. the rapid growth of computational cost w.r.t.
the number of parameters. As the full structure has a significant larger number of
parameters, i.e. by a factor of N, this would quickly result in a very large number
of simulation runs. Additionally, the computational cost for each model evaluation
would also be significantly larger, when the full structure is considered instead of a
single unit cell.

3 Multifidelity Monte Carlo

MFMC generalizes the multilevel Monte Carlo approach, which was recently used
in [8] for a high-frequency application. MFMC simulation combines low-fidelity
models of different kinds, without quantified model errors, into an efficient sampling
framework. By sampling the high-fidelity model at least one time, the MFMC
approach provides an unbiased estimator. Moreover, a low variance and hence, a
low root-mean-square error, is realized through optimal model management and the
resulting estimator is typically much more efficient than the standard Monte Carlo
Multifidelity Uncertainty Quantification for Optical Structures 131

(MC) estimator. The MFMC methodology was introduced in a series of papers [3, 9]
and is now well-established. Hence, in the following we limit ourselves to the key
aspects and refer to the literature for a more complete introduction into the field.
We adopt a probabilistic approach to represent uncertainty, where y represents
a realization of a random vector Y. Let g(Y) denote an output quantity derived
from the simulated frequency response. MC simulation is then based on a sample
{Yi , g(Yi )}K
i=1 , which can be used to estimate for instance the mean value of the
model output. The mean value approximation and its mean-square error read

1 
K
V[g(Y)]
E[g(Y)] ≈ ĝK := g(Yi ), E[|E[g(Y)] − ĝK |2 ] = . (3)
K K
i=1

Following [9], we consider a model family {g (i) }Mi=1 , where g

(1) represents the

high-fidelity model, and g (i) for i ≥ 2 represent low-fidelity models, obtained for
instance by SMA in combination with surrogate modeling. The MFMC estimator
samples all models and combines the results into a single estimator as


M 
(1) (i) (i)
E[g] ≈ ĝMFMC = ĝK (1) + αi ĝK (i) − ĝK (i−1) ,
i=2

(i)
where ĝK (i) denotes the standard MC estimator based on the sample
(i)
{Yj , g (i) (Yj )}K
j =1 and 0 < K
(1) ≤ K (2) ≤ . . . ≤ K (M) .

In place of low-fidelity error control, the model management of MFMC employs

the Pearson correlation coefficient ρ1,i between the high-fidelity model g (1) and the
low-fidelity model g (i) . In particular, low-fidelity models with a high ρ1,j and a low
computational cost wi are sampled extensively. For a given computational budget
B, MFMC minimizes the mean-square error by appropriately choosing αi , K (i) , see
[9] for details. With σi = V[g (i) (Y)]1/2, the resulting estimator is unbiased with a
mean-square-error of

σ12  1 1 2 2
M
E[|ĝMFMC − E[g(Y)]| ] = (1)
2
+ − (αi σi − 2αi ρ1,i σ1 σi ).
K K (i−1) K (i)
i=2
(4)

4 Numerical Examples

We apply the UQ methods presented in the previous section to the benchmark

problem of an SRR array introduced in the beginning of Sect. 2. First, we give some
details on the considered numerical models, before investigating the performance of
the proposed UQ methodology.
 132 N. Georg et al.

70 nm

90 nm

320 nm
L~ (305, 335) nm
(a) (b)

Fig. 1 Numerical model of SRR array. Depicted is only one cell out of seven. (a) Unit cell of size
1 µm × 0.6 µm × 0.6 µm. Red boundaries indicate the ports. (b) Geometry specification. Thickness:
20 nm. Uncertain longitudinal length L of SRR element

0
Magnitude |S11 | [dB]

− 10

− 20

− 30
80 100 120 140 160 180 200 220
Frequency f [THz]

Fig. 2 Broadband scattering parameter for different realizations of SRR array. Dashed vertical
lines indicate respective limit frequencies of considered bandgaps. Dotted line refers to −3 dB line

We consider a SRR array with N = 7 cells. The employed Cartesian grid as well
as the geometric dimensions (taken from [4], except for the enlarged cell size) are
presented in Fig. 1, where we consider an uncertain longitudinal length L(j ) of each
SRR element in the range of 320 nm ± 15 nm. Hence, the random vector Y is given
as (L(1) , . . . , L(N) )T , where L(j ) , j = 1, . . . , N are assumed to be independent
and identically uniformly distributed. Figure 2 presents a broadband scattering
parameter, in particular the fundamental reflection coefficient |S11 |, for different
realizations of the structure. Two bandgaps can be observed, which can be defined
by their limit frequencies, where the scattering parameter drops below −3 dB. The
corresponding bandwidths bi and center frequencies fc,i , where i ∈ {1, 2} refers
to the first or second bandgap, can be computed from S11 in a post-processing
step. For brevity, we restrict ourselves to the computation of the mean value of the
center frequencies E[fc,i ] in the following. However, very similar findings hold for
the bandwidths bi as well. We further note that for some parameter sample points
some additional resonances within the second bandgap appear which are due to
the slightly detuned resonances in the series of SRRs. This effect is ignored in
Multifidelity Uncertainty Quantification for Optical Structures 133

Table 1 Employed numerical models of SRR array for MFMC study. The last two columns show
the estimated correlation coefficients for both bandgaps
Symbol Model Cost wi ρ1,i for fc,1 ρ1,i for fc,2
g (1) Full model (FIT, 2 · 105 time-steps) 197.50 s 1.000000 1.000000
g (2) Full model (FIT, 2 · 104 time-steps) 11.25 s 0.999236 0.998035
g (3) SMA (FIT, 1 port-mode) 9.64 s 0.999943 0.968376
g (4) SMA (FIT, 2 port-modes) 115.47 s 0.999998 0.999998
g (5) SMA + unit cell surrogate (1 port-mode) 0.006 s 0.999943 0.967540
g (6) SMA + unit cell surrogate (2 port-modes) 0.026 s 0.999998 0.999886

the following evaluation of the MLMC algorithm and only the outer limits of this
bandgap are considered.
An overview of the employed numerical models as well as the corresponding
computational costs (measured in computation time for an in-house MATLAB
implementation on a standard workstation) is given in Table 1. For the full FIT
model g (1) we terminate the time stepping procedure if either the energy decays to
−120 dB or a maximum number of 2 · 105 time-steps is reached. The low-fidelity
model g (2) is obtained by restricting the maximum number of time-steps to 2 · 104 .
The low-fidelity models g (3) and g (4) are obtained by the SMA approach. For g (3)
only the propagating fundamental TEM mode is considered, while g (4) additionally
takes the evanescent first TM mode into account. The selection of suitable models is
based on a pilot run (with a small sample) and model selection techniques, see also
[3, 9].
The construction of the respective unit cell surrogate models for g (5) and g (6) in
the offline-phase is based on C = 7 Chebyshev nodes, which are well-established
non-equidistant interpolation nodes. Note that other choices are equally feasible,
Gauss-Legendre nodes for instance. Surrogate modeling requires some additional
computational effort, which, however, only needs to be invested once. Also, in
this case, even a single model evaluation of g (1) requires a larger computational
effort than constructing the surrogate models. Hence, we will neglect this cost
here, for simplicity. We further note that the evaluation times of all models scale
approximately linear w.r.t. to an eventually increased number of cells N, while
the offline-cost for the surrogate models is independent of N. Accordingly, similar
MFMC results, as presented in the following for N = 7, are also expected for SRR
arrays with a different number of cells. Exemplarily, this has been confirmed for
N = 14 numerically. However, we note that for larger models some care has to
be taken regarding the concatenation within the SMA, since the multiplication of
transfer matrices can become numerically unstable.
In order to evaluate the performance of the proposed methodology for the
considered benchmark problem, we draw an input sample {Yi }K̃ i=1 of size K̃ =
(j )
500 and employ each model g to compute the corresponding output samples
i=1 , j = 1, . . . , 6. The correlation coefficients with the high-fidelity
{g (j ) (Yi )}K̃
134 N. Georg et al.

First bandgap Second bandgap

Estimated RMSE [THz]
10− 1 10− 1

10− 2 10− 2

10− 3 10− 3

10− 4 10− 4

10− 5 10− 5
103 104 105 106 107 103 104 105 106 107
Budget B [s] Budget B [s]

MC: g(1) MFMC: all models MFMC: g(1) , g(2) MFMC: g(1) , g(3)
MFMC: g(1) , g(4) MFMC: g(1) , g(5) MFMC: g(1) , g(6) MC: g(6)

Fig. 3 Estimated RMSE for different MC and MFMC variants, see Table 1

model g (1) are then estimated as shown in Table 1. It can be observed that all low-
fidelity models show a strong correlation with the high-fidelity model.
We employ an MFMC implementation which is based on the open-source Matlab
library github.com/pehersto/mfmc, see [9]. In the following, we will compare the
root-mean-square-errors (RMSEs) of MC and MFMC for given computational bud-
gets B, which can both be accurately estimated based on the samples {g (j ) (Yi )}K̃
i=1 ,
as explained in the following. The RMSE of standard MC on the high-fidelity model
g (1) is obtained by (3), where K is given by wB1 and the variance is replaced by the
MC estimate for the variance using {g (1) (Yi )}K̃
i=1 . This is shown in Fig. 3 in blue
color. Similarly, the RMSE of MFMC can be estimated according to (4), as shown
in black color in Fig. 3. We note that the proposed approach yields speedups by
several orders of magnitude w.r.t. standard MC (for a fixed accuracy).
We note that the MFMC algorithm sorts out some models, as, for example, g (2)
and g (3) have a smaller correlation with the high-fidelity model than the surrogate
model g (6) but a higher computational cost. For completeness, we additionally show
the convergence of MFMC using only g (1) and g (j ) , j ∈ {2, . . . , 6} with dashed
lines in Fig. 3. As expected, in all cases this approach performs better than MC
but worse than the combination of models chosen by the MFMC algorithm. It can
be observed that, for both bandgaps, mainly the proposed unit cell surrogate models
lead to the tremendous efficiency gains. While for the first bandgap considering only
one port-mode could also be sufficient, for the second bandgap it is clearly necessary
to consider two port-modes for the SMA. This is expected as the first bandgap is
mainly governed by the fundamental resonance of the SRRs itself, whereas for the
second one the mutual coupling between the cells play a larger role.
Finally, we show that the high-fidelity model evaluations within the MFMC
framework are indeed required to remove the biasing error. If one would apply
a standard MC method on the surrogate model g (6) solely (instead of g (1) )
Multifidelity Uncertainty Quantification for Optical Structures 135

the associated error is represented by the dotted red line in Fig. 3. Both error
contributions, the sampling and the biasing error, are estimated again with a Monte
Carlo sample.

5 Conclusions

We have presented an uncertainty propagation technique for quasi-periodic optical

structures with random influences, which combines surrogate modeling of unit cells,
SMA and MFMC. The resulting multifidelity approach can significantly improve the
efficiency of Monte Carlo sampling. In particular speedups by orders of magnitude
were obtained for a split ring resonator. The proposed method exhaustively samples
unit cell models which are combined through the scattering matrix approach and
hence, can be evaluated efficiently. Only a single unit cell surrogate was required
which significantly reduced the number of uncertain parameters and hence, the
computational complexity. The surrogate-SMA data was then corrected with a few
time domain simulations of the entire structure to obtain unbiased estimates of the
bandgap properties.

Acknowledgments The work of Niklas Georg is supported by the DFG grant RO4937/1-1, the
Excellence Initiative of the German Federal and State Governments and the Graduate School of
Computational Engineering at TU Darmstadt. Christian Lehmann’s work is funded by the DFG
grant SCHU1157/11-1.

References

1. N. Georg, D. Loukrezis, U. Römer, S. Schöps, Enhanced adaptive surrogate models with

applications in uncertainty quantification for nanoplasmonics. Int. J. Uncertain. Quan. 10(2),
165–193 (2020)
2. B. Bandlow, R. Schuhmann, G. Lubkowski, T. Weiland, Analysis of single-cell modeling of
periodic metamaterial structures. IEEE Trans. Magn. 44(6), 1662–1665 (2008)
3. B. Peherstorfer, K. Willcox, M. Gunzburger, Survey of multifidelity methods in uncertainty
propagation, inference, and optimization. SIAM Rev. 60(3), 550–591 (2018)
4. S. Linden, C. Enkrich, M. Wegener, J. Zhou, T. Koschny, C. Soukoulis, Magnetic response of
metamaterials at 100 Terahertz. Science 306, 1351–1353 (2004)
5. K. Busch, G. von Freymann, S. Linden, S. Mingaleev, L. Tkeshelashvili, M. Wegener, Periodic
nanostructures for photonics. Phys. Rep. 444(3–6), 101–202 (2007)
6. T. Weiland, Time domain electromagnetic field computation with finite difference methods. Int.
J. Numer. Model. El. 9(4), 295–319 (1996)
7. H. Glock, K. Rothemund, U. van Rienen, CSC - A procedure for coupled S-parameter
calculations. IEEE Trans. Magn. 38(2), 1173–1176 (2002)
8. A. Litvinenko, A. Yucel, H. Bagci, J. Oppelstrup, E. Michielssen, R. Tempone, Computation
of electromagnetic fields scattered from objects with uncertain shapes using multilevel Monte
Carlo method. IEEE J. Multiscale Multiphys. Comput. Tech. 4, 37–50 (2019)
9. B. Peherstorfer, K. Willcox, M. Gunzburger, Optimal model management for multifidelity
Monte Carlo estimation. SIAM J. Sci. Comput. 38(5), A3163–A3194 (2016)
Dielectric Breakdown Prediction
with GPU-Accelerated BEM

Cedric Münger, Steffen Börm, and Jörg Ostrowski

Abstract The prediction of a dielectric breakdown in a high-voltage device is based

on criteria that evaluate the electric field along possible breakdown paths. For this
purpose it is necessary to efficiently compute the electric field at arbitrary points
in space. A boundary element method (BEM) based on an indirect formulation,
realized with MPI-parallel collocation, has proven to cope very well with this
requirement. It deploys surface meshes only, which are easy to generate even for
complex industrial geometries. The assembly of the large dense BEM-matrix, as
well as the iterative solution of the resulting system, and the evaluation along the
field-lines all require to carry out the same type of calculation many times. Graphical
Processing Units (GPUs) promise to be more efficient than Central Processing Units
(CPUs) when it is possible to do the same type of calculations for large blocks
of data in parallel. In this paper we therefore investigate if GPU acceleration is a
measure to further speed up the established CPU-parallel BEM solver.

1 Introduction

Every high-voltage device has to pass dielectric type tests, in which a large voltage
is applied to the device. The test is passed if no dielectric breakdown occurs. A
breakdown usually starts from an electrode-surface with high dielectric stress, and
then propagates through the volume along a field-line of the electric field E towards
the opposite electrode, see Fig. 1. The propagation stops if the electric field along

C. Münger ()
Seminar for Applied Mathematics, ETH Zürich, Zürich, Switzerland
e-mail: [email protected]
S. Börm
Math. Seminar, Christian-Albrechts-Univ. Kiel, Kiel, Germany
e-mail: [email protected]
J. Ostrowski
ABB Corporate Research, Baden, Switzerland
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 137
M. van Beurden et al. (eds.), Scientific Computing in Electrical Engineering,
Mathematics in Industry 36, https://doi.org/10.1007/978-3-030-84238-3_14
138 C. Münger et al.

Fig. 1 The electric field strength on the surface of a disconnector and possible breakdown paths
along the field lines

this breakdown path γ is not strong enough. For details see [1]. An inception of a
streamer, i.e. the initial state of a breakdown only occurs if the criterion

αeff (|E|)ds > Kstr (1)
γ

is fulfilled. Here αeff is the effective ionization function that depends on the strength
of the electric field |E|, and Kstr is the (gas-specific) streamer constant. The
prediction of a dielectric breakdown during a type test relies on the evaluation of
this criterion along the most probable breakdown paths. It requires the computation
of the electric field at all surface points and along field lines in the volume.
Simulation-based dielectric design became a standard procedure because a user-
friendly, i.e., fast, robust, reliable, and easy-to-use computational method was
developed, see [2]. In the following we will first describe this boundary-element-
based method and then introduce how general-purpose graphics processing units
(GPGPUs) can be used to massively reduce computing times.
Dielectric Breakdown Prediction with GPU-Accelerated BEM 139

2 BEM Formulation

In this section we derive the BEM-formulation as it is in use since many years at

ABB, see [2, 3]. The device consists of the subdomains
4 Ω0 , . . . , Ωm−1 , and the
unbounded exterior subdomain is Ωm = R3 \ k={0..m−1} Ωk , see the example in
Fig. 2. The electric field E = − grad ϕ is calculated by solving the Laplace equation

div  grad ϕ = 0 (2)

for the electric scalar potential ϕ in each of these subdomains. The permittivity is
denoted by . We use an indirect formulation with a single-layer potential

σ (y)
ϕ(x) = ΨSL [σ ](x) = dSy , (3)
∂Ω 4π|x − y|

and search for the unknown scalar virtual surface charge density σ that is related
to the physical surface charge density σs . Each conductor, i.e. each separated
conducting part with electrical conductivity σel > 0, is on a constant electrical
potential. If a conductor is connected to an electric potential V0 , like Ω0 in Fig. 2,
then it holds

ϕ(x) = V0 ∀x ∈ Ω0 . (4)

The electric potential V of floating conductors like Ω1 in Fig. 2 is unknown

ϕ(x) = V ∀x ∈ Ω1 , (5)

and is to be determined by a charge neutrality condition, see [4]. The total charge Q
of the floating conductor can be derived from the Gauss law as
 
Q= σs dS = D · n dS. (6)
∂Ω1 ∂Ω1

Fig. 2 Example of a device

that consists of a conductor
Ω0 that is connected to a
prescribed electric potential
V0 , a floating conductor Ω1 ,
an insulator Ω2 , and the
unbounded exterior domain
Ω3
140 C. Münger et al.

The normal component of the displacement field D can be expressed as

D · n = ε+ E · n = −ε+ grad ϕ · n = −ε+ grad ΨSL [σ ] · n. (7)

Here ε+ denotes the permittivity of the exterior domain. The Neumann trace of the
single layer potential and can be expressed with help of the adjoint double layer K 

1
grad ΨSL [σ ] · n = σ + K  σ , with (8)
2

x−y
K  (σ )(x) = · n(x)σ (y)dSy . (9)
∂Ω 4π|x − y|
3

Combining the Eqs. (6)–(8) with Q = 0 due to charge neutrality yields


1 +
ε σ (y) + ε+ (K  σ )(y)dSy = 0. (10)
∂Ω1 2

We model thin floating conductive sheets only by a single surface. Then the electric
fields from both sides (±) need to be considered for charge neutrality, since

σs = n · (D+ − D− ) $⇒ (11)

1 +
(ε + ε− )σ (y) + (ε+ − ε− )(K  σ )(y)dSy = 0. (12)
∂Ω1 2

There is no surface charge on non-conductors: σs = 0 on ∂Ω2

1 +
(ε + ε− )σ (x) + (ε+ − ε− )(K  σ )(x) = 0 ∀x ∈ ∂Ω2 . (13)
2
So for our simple but quite general example of Fig. 2 we have to solve the following
set of equations:

σ (y)
dSy = V0 ∀x ∈ ∂Ω0 (14)
∂Ω 4π|x − y|

σ (y)
dSy − V = 0 ∀x ∈ ∂Ω1 (15)
∂Ω 4π|x − y|

1 +
ε σ (y) + ε+ (K  σ )(y)dSy = 0 (16)
∂Ω1 2
1 +
(ε + ε− )σ (x) + (ε+ − ε− )(K  σ )(x) = 0 ∀x ∈ ∂Ω2 (17)
2
Dielectric Breakdown Prediction with GPU-Accelerated BEM 141

The solution of the system of Eqs. (14)–(16) yields the virtual surface charge
distribution from which the electric field can be compute at any point in space as

x−y
E(x) = σ (y)dSy ∀x ∈ R3 . (18)
∂Ω 4π|x − y|3

3 Discretization

We use a collocation boundary element approach: the surface ∂Ω is represented by a

collection of triangles τ1 , . . . , τN with vertices x1 , . . . , xn . The unknown function σ


n
σh (y) = uj ψj (y),
j =1

is approximated by suitable basis functions ψ1 , . . . , ψn . This approach is to be

inserted into (14)–(16) and this set of equations is only required to hold in the
collocation points x1 , . . . , xn . This is an n + Nf l -dimensional system of linear
equations, with Nf l being the number of floating conductors. The implementation
of this approach poses a number of challenges:
• High-voltage devices have smooth curved surfaces in order to avoid field
enhancements. We use piecewise quadratic parametrizations with (curved) tri-
angles τ1 , . . . , τN to minimize the geometrical discretization error.
• The entries of the matrix corresponding to the linear system have to be computed.
In the established CPU-based method we employ an MPI-parallel implementa-
tion of suitable quadrature rules.
• The system of linear equations has to be solved. We use Krylov subspace
methods, since these methods only need matrix-vector multiplications, which
can easily provided in the established method by the MPI-based distributed
representation of the matrix.

4 GPGPU Quadrature

Even on modern processors of parallel computers it is time-consuming to compute

the matrix entries vij for complex industrial geometries. It is advantageous to
calculate the surface integral for pairs of collocation points and triangles.

ψj (y)
vij = dSy for all i, j ∈ {1, . . . , n} (19)
∂Ω 4π|xi − y|
142 C. Münger et al.

Matrix assembly is however ideally suited for parallelization on SIMD-type of pro-

cessors, because the integrations require us to carry out mostly identical operations
for all matrix entries. The currently most common processors are general-purpose
graphics processing units (GPGPUs) like the NVidia Tesla™ or AMD Radeon
Instinct™ cards that contain thousands of floating-point arithmetic units and offer
teraflop-level performance. Porting the quadratures to GPGPUs poses challenges:
• The most powerful GPGPU models are equipped with fast local memory. We
have to ensure that geometrical data is efficiently transferred to the local memory.
• We are using piecewise (mapped) linear basis functions and multiple triangles
contribute to the same matrix entry. Thus we have to avoid collisions between
different triangles that may try to simultaneously update the same matrix entry.
For the parallelization it is beneficial that each collocation point corresponds to
one row in the matrix. In our implementation we take advantage of this fact by
assigning each collocation point to a thread. Then we consecutively iterate through
all triangles of the discretization and simultaneously compute the contribution of a
triangle to all collocation points, see Fig. 3. This way collisions can be avoided. For
good performance it is required that all threads execute exactly the same operations.
This can be ensured by grouping the collocation points depending on whether they
are on the surface of a conductor or part of a dielectric interface.
For the quadrature we distinguish between regular, near-singular, and singular
integration of the pairs of triangles and collocation points. If the collocation point
is a node of the triangle then it is a singular pair. When the distance D between the
circumcenter of the triangle and the collocation point is larger than a threshold that
scales with the circumradius R of the triangle like

D >η·R (20)

Fig. 3 Near-singular assembly: The gray triangles are near-singular and are computed after the
regular and singular triangles
Dielectric Breakdown Prediction with GPU-Accelerated BEM 143

then it is a regular pair. We used the scaling parameter η = 1.2. All other pairs are
near-singular. Circumcenters and circumradii of the triangles can be precomputed.
All three types of pairs can be integrated by using the well-known Duffy-
transformation, see [5]. This is straightforward for regular and singular pairs, only
the near-singular pairs need to be integrated adaptively for accuracy. They frequently
occur, e.g. in cases with narrow gaps in the geometry, or during the postprocessing,
when a point near the surface is to be evaluated. In this near-singular case we first
compute the point of the triangle that is closest to the collocation point. Next we
subdivide the triangle into smaller triangles such that this closest point is a corner-
node of a subdividing triangle. Then we again employ the Duffy-transformation to
integrate over all smaller triangles. This yields an adaptive quadrature with increased
accuracy around the closest point. The adaptivity can impact the performance of a
GPU-computation badly if no attention is paid, because then there is divergence in
the control flow on the GPU. In order to minimize this divergence, we first compute
the regular and singular pairs, and deal with the near-singular integrals later.
The categorization into regular, near-singular and singular pairs is carried out
on the fly during the iteration through the triangles. If a pair is marked as near-
singular, then it will be marked as not processed, see Fig. 3. They are computed in
parallel by using the subdivision method after the regular and singular cases have
been completed. This strategy allows that all three types of integrations are carried
out in parallel, without the need to mix operations.
The full matrix may not fit in the memory of one GPU for larger problems.
Therefore, but also to speed up the computation we use multiple GPUs. Due to
the independence of matrix rows we split the matrix into multiple blocks of rows
that can be computed and stored independently on different GPUs.

5 Numerical Experiments

In this chapter we show some examples that were computed with the novel GPU-
implementation that is based on the H2Lib package, see [6]. We first validated our
implementation for an axial-symmetric case. We compared the results of the H2Lib
with the results of the already existing simulation tools Polopt (3D) see [3], and Elfi
(2D) see [2]. Next we compared the performance of the new GPU-parallel H2Lib
implementation with the performance of the existing MPI-parallel Polopt tool.

5.1 Validation

The benchmark problem that is used to validate the GPU-implementation in H2Lib

is a bushing, see Fig. 4. It consists of an insulator that is wrapped around a conductor
on 100 kV high-voltage. Five thin conducting sheets are embedded in this insulator.
They accomplish the field grading. The outermost is grounded, and the potentials
144 C. Münger et al.

Fig. 4 Cross-section of the bushing, including dimensions in mm

Table 1 Potentials of the ELFI(2D) POLOPT H2LIB

conducting sheets of the
bushing Vfoil1 70.8 kV 70.15 kV 70.22 kV
Vfoil2 51.4 kV 50.47 kV 50.50 kV
Vfoil3 35.0 kV 34.00 kV 34.02 kV
Vfoil4 18.9 kV 18.02 kV 18.02 kV

of the other (floating) sheets are unknown and are to be determined. The sheets
are treated as single surfaces according to Eq. (12). Their potentials were computed
with all three solvers. POLOPT and H2Lib use the same mesh with 3’526 nodes.
The results agree very well, see Table 1. The small remaining differences are due to
the use of different quadratures.

5.2 GPU-Acceleration

After the successful validation of the H2Lib implementation, we compared the

computing times for the GPU-parallel H2Lib and the CPU-parallel Polopt. In
both cases we assembled the dense BEM matrix and solved the system with an
iterative GMRES with diagonal preconditioner. So the expected numerical work is
quadratically depending on the degrees of freedom, i.e. here the number of nodes (#
Nodes). Figure 5 shows the times that it took for matrix-assembly, iterative solution,
and computation of the electric field at the surfaces of a realistic high-voltage device
for different mesh-sizes. POLOPT used 180 CPU cores distributed over 5 nodes
with two 18-core CPUs each. The CPU cluster was optimized for these calculations
because it is in use by ABB product designers. For H2Lib we used a total of 12
NVidia GTX 1080Ti, distributed over multiple nodes. For POLOPT we clearly
recognize the quadratic scaling. The H2Lib also scales quadratically, see Fig. 6, the
proportionality constant seems however to be much better than the one for Polopt.
We computed another larger example see Table 2. We used 180 cores for
POLOPT for all meshes except the largest one, where we used 360 cores. The
Dielectric Breakdown Prediction with GPU-Accelerated BEM 145

Fig. 5 Cumulative times for assembly, solving and surface electric field computation for POLOPT
and H2LIB

Fig. 6 Cumulative times for assembly, solving and surface electric field computation for the
H2Lib only

number of GPUs were chosen such that the matrix fitted in the combined memory
of all GPUs in single precision for H2Lib. Again the GPU-parallel implementation
clearly outperforms the CPU-parallel version. We also evaluated the breakdown
146 C. Münger et al.

Table 2 Calculation times #Nodes #GPUs H2Lib POLOPT

for different meshes. H2Lib
used multiple GTX1080Ti, 68’218 2 18 s 33 s
POLOPT is executed on 180 140’183 8 19 s 130 s
CPU-Cores 232’029 20 26 s 371 s
330’706 40 34 s 702 s
432’084 72 41 s 732 sa
a
360 Cores

Table 3 Time for POLOPT H2Lib

computation and fieldline
180 Cores 4xTesla P100
evaluation. Model size:
68’218 Nodes. Polopt in Assembly 11 s 2s
serial for the evaluation Solving 11 s 2s
E surface 11 s 2s
Total 33 s 6s
Evaluation 102 s for 67 s for 9219
17 fieldlines fieldlines

criterion along the most probable breakdown paths (field lines), see Table 3. The
acceleration seems even higher, however the evaluation is only implemented in
serial in Polopt.

6 Conclusions and Outlook

The usage of GPUs drastically accelerates the computation of electrostatic fields, as

well as their evaluation with respect to breakdown inception. This opens the door not
only for the computation of larger problems, but also for the inclusion of additional
physical models (e.g. for surface charging), or for optimization. Most promising is
the combination of the GPU-acceleration with a compression technique, see [7, 8].
Another strong acceleration is to be expected in this case. Moreover the asymptotical
behavior will no longer depend quadratically on the degrees of freedom.

References

1. A. Blaszczyk, J. Ekeberg, S. Pasncheshnyi, M. Saxegaard, Virtual High Voltage Lab , in

Scientific Computing in Electrical Engineering. Mathematics in Industry (Springer Book, Cham,
2016)
2. A. Blaszczyk, H. Steinbigler, Region oriented charge simulation. IEEE Trans. Magn. 30(5),
2924–2927 (1994)
3. N. De Kock, M. Mendik, Z. Andjelic, A. Blaszczyk, Application of 3D boundary element
method in the design of the EHV GIS components. IEEE Magn. Electr. Insul. 14(3), 17–22
(1998)
Dielectric Breakdown Prediction with GPU-Accelerated BEM 147

4. D. Amann, A. Blaszczyk, G. Of, O. Steinbach, Simulation of floating potentials in industrial

applications by boundary element methods. J. Math. Ind. 4(1), 1–15 (2014)
5. M.G. Duffy, Quadrature over a pyramid or cube of integrands with a singularity at a vertex.
SIAM J. Num. Anal. 19(6), 1260–1262 (1982)
6. S. Börm, S. Christophersen, et al., H2Lib, a software library for H - and H 2 -matrices. Open
source, available at http://www.h2lib.org
7. S. Börm, S. Christophersen, Approximation of integral operators by Green quadrature and
nested cross approximation. Num. Math. 133(3), 409–552 (2016)
8. S. Börm, S. Christophersen, GCA-H 2 matrix compression for electrostatic simulations, in
International Conference on Scientific Computing in Electrical Engineering, Taormina (2018)
Empirical Analysis of a Coaxial
Microwave Structure with Finite
Transmission Zero

K. Papke, F. Gerigk, and U. van Rienen

Abstract Empirical studies are presented on a certain radio frequency (RF) struc-
ture that has not yet been well understood. The coaxial structure provides almost
ideal conditions to approximate high-pass filter functions. It has been investigated
by the aid of numerical simulations accompanied by the search for appropriate
equivalent microwave networks. A particular feature is a finite transmission zero
which allows not only for maximally flat and Chebyshev approximations but also
the synthesis of elliptic filter functions. The synthesis is drawn by means of two
examples taking into account the topology of the equivalent circuit.

1 Introduction

Coaxial microwave filters have been applied for decades to damp unwanted resonant
modes in accelerating and deflecting type cavities operating at tens of megahertz up
to few gigahertz while the extracted power may reach the level of 1 kW in particular
cases [1–4]. These so-called higher-order mode couplers are essentially high-pass
or pseudo-high-pass filters consisting of coaxial lines and certain discontinuities
in between. Early design procedures were focused on the implementation of
narrow-band band-pass filters using reactance-coupled λ/2 resonators [5, pp. 528].
However, such semi-analytical approaches provide only rough estimates for the

K. Papke ()
CERN, Geneva, Switzerland
University of Rostock, Rostock, Germany
e-mail: [email protected]; [email protected]
F. Gerigk
CERN, Geneva, Switzerland
e-mail: [email protected]
U. van Rienen
University of Rostock, Rostock, Germany
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 149
M. van Beurden et al. (eds.), Scientific Computing in Electrical Engineering,
Mathematics in Industry 36, https://doi.org/10.1007/978-3-030-84238-3_15
150 K. Papke et al.

geometrical parameters of coaxial microwave filters. Since the 1990s, the filter
design was more and more based on numerical simulations which permit the precise
evaluation of scattering properties associated with arbitrarily shaped microwave
structures and their systematic adaptation according to individual requirements.
Still, the selection of a suitable topology, as chosen prior to the numerical analyses,
is very much in the realm of intuition and experience [1]. Even for a specified
topology, the applied numerical optimization scheme may not be able to converge
against the best solution, given a certain set of requirements, as too many variables
may be involved.
This paper proposes a generally applicable procedure to systematically design
coaxial microwave filters on the basis of filter or transfer functions; this further
implies the most suitable topology for the given problem. The synthesis of a filter
function rests on the idea that scattering properties of discontinuities in coaxial
guides are well described by lumped elements within the interesting frequency
range, i.e. by equivalent circuits. A large variety of microwave structures with
certain filter characteristics and appropriate equivalent circuits has been worked
out already until the 1950s [5, 6]. Still, scattering properties of coaxial microwave
structures with multiple discontinuities being relatively close to each other, so that
evanescent modes may interact, are partially unexplored by means of equivalent
circuits. The structure sketched in Fig. 1 is such an example. In the limit of
vanishing coaxial lines, it may equivalently be described by a canonical network
realization of third-order high-pass filter functions. A particular feature is the
transmission zero at finite, non-vanishing frequency which allows not only for
maximally flat and Chebyshev approximations but also the synthesis of elliptic filter
functions. The equivalent circuit can be considered as surrogate system, significantly
cheaper to evaluate than the three-dimensional field problem, and with excellent
approximation properties within a certain frequency range. This together with the
fact that equivalent circuit parameters allow for large adjustment ranges as shown

2∆1 dgap 2∆2

ro
Z1 , l1 Z0 , l20 Z0 , l20 Z2 , l2
ri

2rfix 1 2rfix 2
α
θ1 θ1 θ2 θ2

Fig. 1 Cross-sectional and side view of a coaxial structure with two cylindrical fixings of radii
rfix1 , rfix2 between the inner and outer conductor of radii ri and ro , respectively. Both fixings are
rotated against each other in the transverse plane by the angle α. The parameters Δ1 and Δ2
represent the electric thicknesses of the corresponding fixings. The inner conductor is interrupted
at the center by a distance dgap . The sections of coaxial guides are described by the lengths lν
and characteristic impedances Zν with ν = 0, 1, 2. Terminal planes are denoted as θμ or θμ with
μ = 1, 2
Empirical Analysis of a Coaxial Microwave Structure with Finite Transmission Zero 151

in the following, makes the microwave structure in Fig. 1 particularly interesting for
the synthesis of high-pass filter functions with finite transmission zeros.
After introducing the equivalent network in Sect. 2, individual circuit elements
are further investigated in Sect. 3, some of which provide unexpected and qualita-
tively new behavior for the microwave circuit theory. Finally, the synthesis is drawn
in Sect. 4 by means of two examples.

2 Equivalent Circuit

The propagation of the transverse electromagnetic mode shall be considered so that

the microwave structure can be represented by a two-port. Figure 2 sketches an
equivalent circuit suitable to approximate the scattering matrix S of the structure
shown in Fig. 1 between the terminal planes θ1 and θ2 over a considerable frequency
range. It is the result of an intense systematic research. A typical fit of simulated
scattering functions is shown in Fig. 3. Remarkable is the transmission zero at finite,
non-vanishing frequency
√ accounted for by the parallel LC resonator. Its resonant
frequency ω0 = 1/ L0 C0 is seen from the transmission power gain in Fig. 3a.
To derive the network parameters of Fig. 2, at first, the scattering matrix S of the
microwave structure between the terminal planes θ1 and θ2 in Fig. 1 is obtained by
numerical simulations.1 Let the matrix S be defined by the elements sij ; i, j = 1, 2
which are functions of the frequency ω. A subsequent transformation provides the
scattering matrix S at the terminal planes of interest, θ1 and θ2 [9, pp. 184]

S = w S w (1)

where the diagonal matrix w = diag{eiβl1 , eiβl2 } invokes the inward phase shifts
along each terminal translation given the propagation constant β and lengths l1 and
l2 which are per se not known due to the finite thickness of the obstacles in the
structure. Consequently, the elements sij ; i, j = 1, 2 of the scattering matrix S are
considered as functions of these lengths and the frequency.
Let l1 + Δ1 be the distance from the terminal plane θ1 to the center of the left
fixing in Fig. 1 which is a priori known. Similar, let l2 + Δ2 be the distance from the
terminal plane θ2 to the center of the right fixing. Since Δ1 and Δ2 define half of
the electric “thickness” of the individual fixing, the sum l0 + dgap + Δ1 + Δ2 must
correspond to the distance between the centers of both fixings. Linear transforms
{Δ1 , Δ2 }
→ {l0 , l1 , l2 } are introduced to reduce the number of length variables as
well as to confine their variation to the vicinity of the corresponding fixing.

1Numerical simulations are mostly carried out using CST STUDIO SUITE® software [7] for the
present work. In part, they are verified with COMSOL Multiphysics [8].
152 K. Papke et al.

iωL0
Z1 , l1 Z0 , l20 Z0 , l20 Z2 , l2

iωL1 1 iωL2
iωC0

θ1′ θ1 θ2 θ2′

Fig. 2 Equivalent circuit model composed of lumped elements and transmission lines

Fig. 3 Approximation of the numerically simulated RF reflection and transmission with respect to
the terminal planes θ1 and θ1 by means of the equivalent circuit model according to Figs. 1 and 2.
The structure is assumed to be symmetric, hence, Z1 = Z2 , l1 = l2 , and rfix1 = rfix2 , with ri = 5 mm,
ro = 22.5 mm, and rfix1 = 3 mm. The cross section of the coaxial guide in between the fixings is
identical to those of the input and output regions. The fixings are separated by a distance of d =
22.5 mm while the inner conductor is separated by a distance of dgap = 0.3 mm. Circuit parameters
L0 , C0 , L1 are derived from the minimization problem (6). (a) Transmission and reflection power
gains |s12 |2 , |s11 |2 . (b) Real and imaginary part of the reflection at the terminal plane θ1 . The R 2
value reveals very good approximation in the considered frequency range f ≤ 2 GHz

The equivalent circuit in Fig. 2 admits a transmission matrix Tmodel between the
terminal planes θ1 and θ2 whose elements tij ; i, j = 1, 2 are given by

t12 1 ω
t11 = + cos βl0 − sin βl0 , (2)
iωL2 2Z0 C0 ω02 − ω2
1 ω βl0
t12 = i cos2 + iZ0 sin βl0 , (3)
C0 ω02 − ω2 2
t12 1 ω
t22 = + cos βl0 − sin βl0 , (4)
iωL1 2Z0 C0 ω02 − ω2
t11 t22 − t12 t21 = 1. (5)
Empirical Analysis of a Coaxial Microwave Structure with Finite Transmission Zero 153

The necessary condition for the frequency response of the microwave structure
being approximated by the equivalent circuit can be formulated as

min ||T(ωk ) − Tmodel (ωk )||2 , (6)
Δ1 ,L1 ,Δ2 ,L2 ,Z0 ,C0
k

where T results from the simulated and phase shifted scattering matrix S sampled
at the frequencies ωk . The relationship between the scattering matrix S and
transmission matrix T can be found in [9, p. 192]. The sufficient condition requires
the residual to become small and, thus, defines the applicable frequency range for
the model. Since the resonant frequency of the LC resonator is directly obtained
from the simulated transmission power gain |s12 |2 only one parameter, either L0 or
C0 is involved in the nonlinear least-square problem (6). It can be solved by iterative
minimization schemes, such as a constrained BFGS algorithm.2

3 Analyses

The dependency of circuit parameters has been studied by successively changing

the geometry. Notable characteristics are related to the LC resonator causing the
transmission zero at finite, non-vanishing frequency. For simplicity, the input and
output waveguide regions as well as cylindrical fixings in Fig. 1 are respectively
chosen to be identically, hence Z1 = Z2 and L1 = L2 . For all geometric variations,
the frequency response of the coaxial structure is well approximated for f ≤ 3 GHz
with residuals in the order of 10−2 by solving (6). The convergence error of the
underlying numerical simulations is significantly smaller. Consistent and smooth
variations of {Δ1 , L1 , Δ2 , L2 , Z0 , C0 , L0 } further justify the circuit in Fig. 2 being
equivalent.
Figure 4 shows the capacitance and inductance of the LC resonator as functions
of the rotating angle α and distance d between the centers of both fixings.
To illustrate the influence of fringe fields, the capacitance C0 is normalized to
C0∗ = ε0 πri2 /dgap , where ε0 is the permittivity constant of vacuum. Moreover,
L0 is normalized to the shunt inductance L1 which only marginally varies with
d. Significant alterations with the angle, in particular, at small distances d provide
evidence for the parallel resonator being related to evanescent higher order multipole
modes in the coaxial guide. It is remarkable that the frequency response of such a
problem is fully characterized by a simple resonator in between two transmission
lines. Furthermore, parameter variations over orders of magnitude in Fig. 4b allow
for any reasonable inductance L0 while the impact of transmission lines can be
reduced as desired.

2 Broyden-Fletcher-Goldfarb-Shanno algorithm.
154 K. Papke et al.

Fig. 4 (a) Capacitance of the parallel LC resonator normalized to the definition C0∗ = ε0 πri2 /dgap
and (b) ratio of series and shunt inductances, both as functions of the distance between the fixings,
d. Furthermore, it is ri =5 mm, ro=22.5 mm, rfix1=rfix2=3 mm, dgap=0.3 mm

4 Application

Given the squared magnitude of a rational transfer function according to

c0
H (iω)H ∗ (iω) = , (7)
1 + εDn (iω)Dn∗ (iω)

where c0 and ε are scalars and Dn (iω) is the filter function, the systematic approach
to derive a microwave circuit being able to approximate this frequency response is
defined as synthesis. In accordance to the equivalent circuit in Fig. 2, any rational
third-order high-pass filter function may be considered. Particularly interesting are
elliptic filter functions as they yield the steepest transition between passband and
stopband given certain attenuation limits in both frequency bands [10, pp. 207]. The
synthesis of elliptic filters is drawn in the following.
Consider the frequency map f : Ω
→ ω. It maps a normalized frequency
space associated with a low-pass to the frequency space of a high-pass according
√
to the definition ω = ωp ωs /Ω, with the passband and stopband edges ωp and ωs ,
respectively. The filter function of a normalized elliptic low-pass of odd order n is
defined as [11]


(n−1)/2
Ω 2 − Ω0ν
2
Dn (iΩ) = c1 iΩ 2 Ω2 − 1
, (8)
ν Ω0ν

where n = 3, 5, 7, · · · . The zeros are calculated by Jacobian elliptic sine functions

as Ω0ν = k sn (2νK/n, k), where K is the complete elliptic integral of the first kind
with the modulus k = Ωs−2 . The factor c1 in (8) normalizes the maximum deviation.
Empirical Analysis of a Coaxial Microwave Structure with Finite Transmission Zero 155

The transfer function H is directly related to |s12 |2 taking into account the
impedance normalization to Z1 and Z2 at the corresponding terminal planes in
Fig. 1 [10, pp. 163]. Lossless two-ports admit unitary scattering matrices, hence,
SH S = I. They further fulfill reciprocity, so that s12 = s21 . Both properties are used
to derive S from |s12 |2 , only. The corresponding impedance matrix is obtained via
1 1
Z = P 2 [I + S] [I − S]−1 P 2 , (9)

where P = diag{Z1 , Z2 } accounts for the impedance normalization, and I is the

identity matrix. The impedances and admittances in the series, or respectively, shunt
arms of the ladder network are derived by continued fraction expansion of z11 and
z22 about Ω = ∞ [10, p. 165]. Finally, the map Ω
→ ω is applied.
The insertion loss method described above does not account for distributed
elements such as transmission lines. The procedure provides an initial set of values
for the lumped elements assuming that the transmission lines are not present. The
structure is simulated repeatedly, with the geometry being gradually changed so
that the fitted circuit parameters in (2)–(4) approximately match the desired ones.
Table 1 lists the parameter values for a third-order elliptic filter with passband and
stopband edges of fp = 1.19 GHz and fs = 0.79 GHz, respectively. For the same
example, the geometry and insertion loss [9, p. 63] are shown in Fig. 5. There is a
notable influence by transmission lines despite a relative short distance between the

Table 1 Parameters of a Parameter Lumped circuit Approximation by coaxial guide

third-order elliptic filter with
fs = 0.79 GHz and L0 18.493 nH 18.412 nH
fp = 1.19 GHz C0 2.739 pF 2.751 pF
L1 , L2 4.201 nH 4.227 nH

Fig. 5 (a) Optimized microwave structure to approximate a third-order elliptic high-pass filter
characteristics. (b) Insertion loss as given by the predefined transfer function H (iω) in black and
the approximation by the optimized structure in blue. Deviations are caused by transmission lines
156 K. Papke et al.

Fig. 6 (a) Optimized microwave structure to approximate a fifth-order elliptic high-pass filter
characteristics. (b) Insertion loss as given by the predefined transfer function H (iω) in black and
the approximation by the optimized structure in blue. Deviations are caused by transmission lines

fixings. Another example using the same passband and stopband edges but higher
order is shown in Fig. 6. It results from a cascade of two structures each adjusted
as a third-order filter. The subsequent connection requires significant changes of the
rotation angles in order to achieve the attenuation curve shown in Fig. 6b.

5 Conclusions

To the authors’ best knowledge, this work contains three new scientific contri-
butions. First, the systematic design of coaxial microwave filters on the basis of
abstract filter or transfer functions was demonstrated. It enables both the design of
coaxial high-order mode couplers under completely new aspects and fundamental
predictions about the topology prior to any computational refinement. The synthesis
is based on equivalent circuit models based on a finite cascade of lumped, lossless
two-ports and transmission lines whose parameters are fitted according to simulated
scattering functions. One structure was elaborated and is particularly suitable for the
synthesis of rational high-pass filter functions. A second important finding is that the
ladder network topology for the equivalent circuit remains valid even in the presence
of evanescent mode coupling between adjacent discontinuities of the coaxial guide.
Finally, the empirical studies on the considered microwave structure, i.e. the nature
of its transmission zero at finite, non-vanishing frequency, open up new research
topics for the microwave circuit theory and await field theoretical analyses.
Empirical Analysis of a Coaxial Microwave Structure with Finite Transmission Zero 157

References

1. E. Haebel, Couplers, tutorial and update. Part. Accel. 40, 141–159 (1992)
2. Q. Wu, B. Sergey, I. Ben-Zvi, et. al., Operation of the 56 MHz superconducting rf cavity in
RHIC with higher order mode damper. APS 22, 102001 (2019)
3. J. Mitchel, Higher Order Modes and Dampers for the LHC Double Quarter Wave Crab Cavity.
Ph.D. thesis, University of Lancaster, 2019
4. K. Papke, F. Gerigk, U. van Rienen, Comparison of coaxial higher order mode couplers for the
CERN Superconducting Proton Linac study. APS 20, 060401 (2017)
5. G.L. Matthaei, L. Young, E.M.T. Jones, Microwave Filters, Impedance-matching Networks,
and Coupling Structures (Artech House, Norwood, 1980)
6. N. Marcuwitz, Waveguide Handbook (McGraw-Hill Book Company, New York, 1951)
7. CST - Computer Simulation Technology Ver. 2016. CST AG, Darmstadt, Germany (2016)
8. COMSOL Multiphysics Ver. 5.3. COMSOL Multiphysics GmbH, Stockholm, Sweden (2017)
9. D.M. Pozar, Microwave Engineering, 4th edn. (Wiley, New York, 2012)
10. O. Wing, Classical Circuit Theory (Springer, New York, 2008)
11. R. Saal, E. Ulbrich, On the design of filters by synthesis. IRE Trans. Circ. Theory 5, 284–327
(1958)
Frequency-Domain Non-intrusive Greedy
Model Order Reduction Based
on Minimal Rational Approximation

Davide Pradovera and Fabio Nobile

Abstract We present a technique for Model Order Reduction (MOR) of frequency-

domain problems relying on rational interpolation of vector-valued functions. The
selection of the sample points is carried out adaptively according to a greedy
procedure. We describe several options for the choice of a posteriori error indicators,
which are used to drive the greedy algorithm and define its termination condition.
Namely, we illustrate a tradeoff between each indicator’s accuracy and its “intru-
siveness”, i.e. how much information on the underlying high-fidelity model needs
to be available. We test numerically the effectiveness of this technique in solving a
non-Hermitian eigenproblem and a microwave frequency response analysis.

1 Introduction

Consider the function u : C μ

→ u(μ) ∈ Cn implicitly defined as the solution of
the linear parametric problem with a single parameter

u(μ) = f(μ),
A(μ) (1)

with A and f smooth functions taking values in Cn×n and Cn , respectively. To

be more specific, we consider here parametric problems (1) arising from spatial
discretization (e.g., by FEM [3]) of frequency domain problems, with the parameter
μ representing the frequency. For most such problems, A(μ) depends at most
quadratically on μ:

A(μ) = A0 + μA1 + μ2 A2 ,

whereas f is usually of the form f(μ) = θ0 (μ)f0 , with θ0 : C → C.

D. Pradovera () · F. Nobile

CSQI, Institute of Mathematics, EPFL, Lausanne, Switzerland
e-mail: [email protected]; [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 159
M. van Beurden et al. (eds.), Scientific Computing in Electrical Engineering,
Mathematics in Industry 36, https://doi.org/10.1007/978-3-030-84238-3_16
160 D. Pradovera and F. Nobile

In applications, it is often computationally unfeasible to solve (1) as many times

as needed for a frequency response analysis. In recent years, this issue has been
solved through MOR, whose main purpose is the construction of a surrogate 3 u(μ)
for u (μ), much cheaper to evaluate at any given μ than solving (1), and with good
approximation properties.

2 Available MOR Strategies

A plethora of MOR techniques have been employed to compute surrogates for

frequency response problems; some notable ones include:
• projective techniques, e.g. the Reduced Basis (RB) and multi-moment-matching
methods [2], are extremely powerful, but require knowledge of, and access to,
the specific structure of A and f;
• strategies based on rational approximation, e.g. the Löwner framework [4] or the
Vector Fitting (VF) algorithm, are non-intrusive, i.e. they rely only on evaluations
of u at few frequencies, which we will refer to as snapshots or samples; in
particular, there is no need for any information on (nor access to) the specific
structure of A and f in (1); the price to pay for this additional flexibility is a
reduced accuracy of the method for a given number of snapshots.
More recently, the minimal rational interpolation (MRI) technique was proposed
[7], trying to achieve non-intrusiveness and optimal snapshot management at the
same time. We summarize here a practical scheme for MRI:
1. fix a set of sample points μ1 , . . . , μS ∈ C, and a polynomial basis {ψi }S−1
i=0 ⊂
P (C) (e.g., one could choose monomials, or Chebyshev polynomials); also,
S−1

let { j }Sj=1 ⊂ PS−1 (C) be the Lagrangian basis associated to the sample points;
2. build the Vandermonde matrix V ∈ CS×S and the diagonal weight matrix D:
&
'
dS−1 1 dS−1 S
(V)ij = ψj (μi ) and D = diag , . . . , ∈ CS×S ;
dμS−1 dμS−1

3. compute the snapshots u(μ1 ), . . . , u(μS ) and assemble a QR decomposition of

the snapshot matrix
 ) ) ) 
) ) )
u(μ1 ))u(μ2 )) · · · )
u(μS ) = WR, with W ∈ Cn×S , R ∈ CS×S ; (2)

4. compute a minimal eigenvector q ∈ CS of the positive semidefinite (Gramian)

matrix (RDV)H RDV, and define the surrogate denominator as


S−1
Q ∈ PS−1 (C), Q(μ) = (
q )i ψi (μ);
i=0
Frequency-Domain Non-intrusive Greedy MOR Based on Minimal Rational. . . 161

5. define the reduced minimal rational approximation ˚

u as

R diag ([Q(μ1 ), . . . , Q(μS )])   −

S−1
C μ
→ ˚
u(μ) = V ψi (μ) ∈ CS ,
Q(μ) :i
i=0

where (A): i denotes the i-th column of matrix A; then the full minimal rational
approximation 3 u(μ) = W˚
u ≈ u can be found as 3 u(μ).

2.1 Greedy Approach

A common feature of all the techniques cited above is that a “sufficiently large”
number of samples is needed to guarantee the accuracy of the surrogate model;
in the particular case of frequency-domain problems, there exist lower bounds [3]
for the number of samples required to achieve reasonable accuracy. Unfortunately,
such number depends on the unknown spectral properties of A, and on the
approximability of f. For RB and MRI, one can identify adaptively the correct
number of samples by relying on the so-called greedy algorithm, which can be
summarized as follows:
1. Initialize a set V = { u1 , . . . , uS0 } with some preliminary snapshots at
μ1 , . . . , μS 0 .
2. Build a surrogate model (e.g., by MRI) based on V .
3. Choose a measure r(μ) of the discrepancy between exact and surrogate solution,
and find its maximal point % μ: r = r(μ) ≤ r(% μ) for all μ.
4. If r(%μ) is smaller than a prescribed tolerance, terminate.
5. Compute a snapshot at % μ, add it to V , and go to 2.
The main difficulty in setting up the greedy algorithm is choosing a good r.
Given the presence of resonances, it is standard [3] to use as a posteriori estimator
the residual of (1), namely, given some suitable norm  ·  ,

r(μ) = A(μ)3
u(μ) − f(μ) . (3)

2.2 A Posteriori Indicators

In an intrusive framework, an efficient way to compute (3) has been known in the
RB literature for quite a while, see e.g. [2], assuming f(μ) to depend affinely on μ,
i.e.
Nf−1

f(μ) = θi (μ)fi ,
i=0
162 D. Pradovera and F. Nobile

with fi ∈ Cn and θi : C → C for all i. Then, as long as the matrices Ai , the vectors
fi , the weights θi (μ), and the reduced surrogate solution ˚
u(μ) are available, we can
evaluate the residual at μ as

Nf−1
⎛ ⎞
 
2
r(μ)2 = θi (μ)θj (μ)fj , fi  + ˚
u(μ)∗ ⎝ μi μj Aj W, Ai W ⎠ ˚
u(μ)
i,j =0 i,j =0
⎛⎛N −1 ⎞ ⎞
f
 
2
−2Re ⎝⎝ θi (μ)μj Aj W, fi  ⎠ ˚
u(μ)⎠
i=0 j =0
(I)

in O((S + Nf)2 ) operations. This idea can be employed in MRI as well, at the cost
of making the procedure intrusive. However, we propose here some alternatives.
In [7] it was observed that, if 3u is the MRI of u with samples at {μj }Sj=1 and
denominator Q, and both A(μ) and f(μ) depend at most linearly on μ (i.e. A2 = 0
and f(μ) = f0 + μf1 ) or μ2 (i.e. A1 = 0 and f(μ) = f0 + μ2 f2 ), then

c 
S
r(μ) = |μ − μj |, (4)
|Q(μ)|
j =1

with c = c(μ1 , . . . , μS , A, f) independent of μ. In particular, since the location

of the maximum of r does not depend on c, see (4), % μ can be found even without
knowing c. In order to determine the value of c non-intrusively, it is enough to
compute r using (3) at a single new point μ (in practice, we take μ = % μ):
) ) S ) )
) Q(μ ) )  ) μ − μj )
r(μ) = r(μ ) )) ) ) )
) μ − μ ) . (R)
Q(μ) ) j
j =1

In certain situations, however, a direct evaluation of the quantity r(μ ) within

each greedy iteration might be impossible (e.g. if the solver used to evaluate u is a
black-box that does not allow residual evaluation) or too computationally expensive.
In order to have a viable greedy loop, we need to design an alternative termination
condition in step 4. In this case, we propose to employ some heuristic indicator
based on snapshot collinearity [3]: more explicitly, let W be the Q-factor in the QR
factorization of the current snapshot matrix (2), and assume that the sample at %
μ has
been computed; we opt to terminate the greedy algorithm if

u(% μ) < tol 
μ) − WW∗ u(% u(%
μ) . (C)

For this last indicator, as long as the greedy iterations continue, the extra snapshot
does not go wasted, since it is precisely the one which gets added to V in step 5: on
Frequency-Domain Non-intrusive Greedy MOR Based on Minimal Rational. . . 163

the whole, this procedure computes only one “extra” snapshot, at the final greedy
iteration, with respect to the two previous versions of the algorithm. Actually, one
can adjust the greedy algorithm so as to employ even the extra snapshot in the final
surrogate model: it suffices to build an updated MRI using all the samples, including
the last one, once the termination condition has been satisfied.
We remark that the last two strategies rely on (4), which is valid only under some
strong assumptions (linear dependence on the parameter) on A and f. However, (4)
can still be used for general parametric problems (1), and will give a reasonable
d2 d2 
estimation of the residual as long as dμ 2 A and dμ2 f are small.

3 Numerical Examples

Here, through two practical examples, we showcase the usefulness of the greedy
MRI procedure, as well as the effectiveness of the three termination strategies based
on (I), (R), and (C).

3.1 An Eigenproblem in Magneto-Hydrodynamics

Take the generalized eigenproblem from [5]: K v = μM v in Cn , with n = 4800; it

stems from modal analysis of a FE discretization of a dissipative problem in MHD.
Here, we restrict our interest to the part of the spectrum with positive imaginary
part: the eigenvalues are located on 3 so-called Alfvén branches around the branch
point μb ≈ −0.082 + 0.613i. Our aim is to approximate the number and location
of eigenvalues around μ0 = −0.175 + 0.5i; more precisely, we focus on the disk
D = {μ ∈ C : |μ − μ0 | ≤ 0.175}.
In order to cast this problem in the form (1), let f ∈ Cn be a (normal Gaussian)
random vector: we define the non-homogeneous problem

find u : C → Cn u(μ) = f,

s.t. (K − μM)

and build a surrogate for u using MRI. Then, our estimates for the eigenvalues will
be the roots of the MRI denominator Q.
As a first MOR method, we apply greedy MRI: in particular, we employ
indicator (I) with relative tolerance 10−2 , and the initial snapshots are at the S0 = 30
shifted roots of unity1 {μ0 − 0.175e2iπj/S0 }Sj =1
0
.

1 The shifted roots of unity are chosen as sample points because they allow for very stable and
efficient interpolation schemes, relying on Fast Fourier Transform. We refer to [1] for a more
detailed discussion of their properties.
164 D. Pradovera and F. Nobile

Fig. 1 Results of standard (left) and greedy (right) MOR. The exact and approximate eigenvalues
are pluses and crosses, respectively, whereas the sample points are full dots. The contour plots show
the logarithm of the greedy residual indicator; the dashed line represents the locus {μ : r(μ) = tol},
i.e. the boundary of the set where the prescribed tolerance is not satisfied

Additionally, we consider a non-greedy approach as a reference: we build an

MRI starting from S = 45 samples at shifted roots of unity {μ0 − 0.175e2iπj/S }Sj=1 .
The number of samples is chosen so that, overall, the two methods employ exactly
the same number of snapshots: the only difference is where the samples are taken.
The results are shown in Fig. 1. At the beginning of the greedy procedure (which
corresponds to a standard MRI with S = 30), the spectrum is approximated quite
poorly; this is correctly identified by the a posteriori indicator, which shows that the
prescribed tolerance is not satisfied over a large portion of D. After 15 iterations of
the greedy procedure, the residual is globally below the tolerance, and the algorithm
ends. We can verify that all the eigenvalues in D are well captured.
In the standard approach with S = 45, most of the eigenvalues are well identified,
but the quality of the approximation deteriorates around μb . In particular, among
the two surrogates obtained with 45 snapshots, the greedy one is clearly superior.
However, the improved accuracy of the greedy approach is accompanied by some
risks:
• The locations of the greedy snapshots are close to the exact eigenvalues, since,
according to the residual indicator (4), sampling there yields “the most informa-
tion” for the surrogate model. However, sampling close to an eigenvalue may
require solving numerically an ill-conditioned or even singular linear system.
• While the Vandermonde matrix for samples at the roots of unity has optimal
condition number, adding new sample points at arbitrary locations is guaranteed
to hinder the well-conditioning of the interpolation problem. Indeed, the residual
indicator at the end of the greedy iterations shows a slightly unstable behavior
near the bottom of D.
Frequency-Domain Non-intrusive Greedy MOR Based on Minimal Rational. . . 165

Table 1 Timing results of greedy MOR (average over 3 simulations with the same parameters for
each method). All simulations were carried out on a single node of the Fidis cluster at EPFL [6]

No. of Average time per iteration

Method snapshots State solve Indicator Surrogate update
RB+(I) 24 97.1 s 15.0 s 3.2 s
MRI+(I) 23 4.31 s 2.9 s
MRI+(R) 23 1.04 s
MRI+(C) 22(+1) 1.08 s

3.2 Frequency Response of a Waveguide Diplexer

We consider a frequency response problem involving the FE discretization (n =

90,258) of a waveguide diplexer [3], for frequencies μ between 9.5 and 11 GHz.
We are interested in approximating the scattering parameters

& 5 '−1
1 − (μc /μ0 )2 ∗ −1
S : C μ
→ I − 2 I + iμ F (K − μ M) F
2
∈ C3×3 , (5)
1 − (μc /μ)2 ! "# $
U(μ)∈C90,258×3

where we set μc = 6.56 GHz, μ0 = 10 GHz, and the state matrix U(μ) has one
column for each port of the waveguide.
We build a surrogate for U using greedy RB and MRI, employing indicators (I)
and (R) with relative tolerance 10−2 , and (C) with tol = 10−4 . The reduced
tolerance for (C) can be justified by the considerably different nature of the indicator.
To obtain an approximation of S, we just replace the exact state with the surrogate
one in (5).
The results are summarized in Table 1 and visually depicted in Fig. 2. We
remark that, by construction, the snapshot “history” of MRI is independent of
the indicator, i.e. the parameter value % μ which is selected at a given iteration is
the same: the only effect of the choice of the indicator, besides timing, is the
number of greedy iterations which are carried out before termination. In this regard,
we observe that MRI+(I) and MRI+(R) yield exactly the same indicator, whereas
MRI+(C) terminates one snapshot sooner,2 causing some slight instability in the
approximations of the scattering parameters for low frequencies, noticeable mostly
in S13 .
A comparison of the surrogate S obtained by MRI+(I) and RB+(I) shows that
the two methods yield very similar approximations, and reconstruct well the exact
values. In fact, the approximated scattering parameters for RB are not included
in Fig. 2, as they are almost indistinguishable from those obtained with MRI+(I).

2 Here we are discarding the final extra snapshot used to check the termination condition (C). If it

had been included, we would have recovered the same surrogate model as MRI+(I)/(R).
166 D. Pradovera and F. Nobile

0
|Si j | [dB]

− 100
|S11 | |S12 |
|S13 | |S22 |
− 200 |S23 | |S33 |

− 50
r(μ ) [dB]

− 100

− 150 RB+(I)
MRI+(I)

9.6 9.8 10 10.2 10.4 10.6 10.8 11

μ [GHz]

Fig. 2 Results of greedy MOR. On top the surrogate scattering parameters: the points are
measurements from the original problem (5), whereas full and dotted lines are the surrogates
obtained with MRI+(I) and MRI+(C), respectively. On the bottom the relative residual at the end
of the greedy iterations for RB+(I) and MRI+(I); the points indicate the snapshot positions

However, RB requires one more snapshot, and the locations of the snapshots (and
the residual profiles) for RB and MRI are quite different.
In terms of computing time, the efficiency of MRI seems quite remarkable,
particularly for the two “least intrusive” indicators: the overhead time needed for
the evaluation of indicators (R) and (C) is just a fraction of the time required for
computation of (I) in RB.

4 Conclusions

We have presented several a posteriori indicators which can be employed in the

greedy MRI algorithm, characterized by different degrees of intrusiveness and
applicability. Good approximation properties, as well as a substantial speed-up in
residual computation with respect to classical methods, have been observed in two
numerical examples.

Acknowledgments This work has been funded by the Swiss National Science Foundation through
FNS Research Project number 182236.
Frequency-Domain Non-intrusive Greedy MOR Based on Minimal Rational. . . 167

References

1. A.P. Austin, P. Kravanja, L.N. Trefethen, Numerical algorithms based on analytic function
values at roots of unity. SIAM J. Numer. Anal. 52, 1795–1821 (2014)
2. P. Benner, S. Gugercin, K. Willcox, A survey of projection-based model reduction methods for
parametric dynamical systems. SIAM Rev. 57, 483–531 (2015)
3. V. De La Rubia, M. Mrozowski, A compact basis for reliable fast frequency sweep via the
reduced-basis method. IEEE Trans. Microw. Theory Tech. 66, 4367–4382 (2018)
4. A.C. Ionita, A.C. Antoulas, Data-driven parametrized model reduction in the loewner frame-
work. SIAM J. Sci. Comput. 36, A984–A1007 (2014)
5. M.N. Kooper, H.A. Van Der Vorst, S. Poedts, J.P. Goedbloed, Application of the implicitly
updated Arnoldi method with a complex shift-and-invert strategy in MHD. J. Comput. Phys.
118, 320–328 (1995)
6. SCITAS: EPFL Fidis cluster webpage. Online document (2020). http://www.epfl.ch/research/
facilities/scitas/hardware/fidis. Cited 11 Feb 2020
7. D. Pradovera, Interpolatory rational model order reduction of parametric problems lacking
uniform inf-sup stability. SIAM J. Numer. Analy. 58(4), 2265–2293 (2020)
A Comparison Between Different
Formulations for Solving Axisymmetric
Time-Harmonic Electromagnetic Wave
Problems

Erik Schnaubelt, Nicolas Marsic, and Herbert De Gersem

Abstract In many time-harmonic electromagnetic wave problems, the considered

geometry exhibits an axial symmetry. In this case, by exploiting a Fourier expansion
along the azimuthal direction, fully three-dimensional (3D) calculations can be
carried out on a two-dimensional (2D) angular cross section of the problem, thus
significantly reducing the computational effort. However, the transition from a full
3D problem to a 2D analysis introduces additional difficulties such as, among others,
a singularity in the variational formulation. In this work, we compare and discuss
different finite element formulations to deal with these obstacles. Particular attention
is paid to spurious modes and to the convergence behavior when using high-order
elements.

1 Introduction

When treating a problem exhibiting axial symmetry, a Fourier expansion along the
azimuthal direction can be exploited in order to restrict the computation to a two-
dimensional (2D) angular cross section of the geometry, while still considering a
fully three-dimensional (3D) solution [1]. Therefore, these methods are also referred
to as quasi-3D or 2.5D methods. Let us consider a cylindrical coordinate system
(r, ϕ, z), and let us expand the electric field e(r, ϕ, z) into a Fourier series along ϕ:
⎡ ⎤ ⎛⎡ m ⎤ ⎡ −m ⎤⎞
er0 (r, z) ∞ er (r, z) cos(mϕ) er (r, z) sin(mϕ)
e(r, ϕ, z) = ⎣ eϕ0 (r, z) ⎦ + ⎝⎣ eϕm (r, z) sin(mϕ) ⎦+⎣ eϕ−m (r, z) cos(mϕ) ⎦⎠ ,
0
ez (r, z) m=1 ezm (r, z) cos(mϕ) ez−m (r, z) sin(mϕ)

E. Schnaubelt () · N. Marsic · H. De Gersem

Technische Universität Darmstadt, Institut für Teilchenbeschleunigung und Elektromagnetische
Felder (TEMF), Darmstadt, Germany
e-mail: [email protected]; [email protected]; [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 169
M. van Beurden et al. (eds.), Scientific Computing in Electrical Engineering,
Mathematics in Industry 36, https://doi.org/10.1007/978-3-030-84238-3_17
170 E. Schnaubelt, N. Marsic, and H. De Gersem

 T
where the Fourier coefficients en (r, z) = ern , eϕn , ezn with n ∈ Z are functions of
the radial and axial coordinates only. Furthermore, by exploiting the orthogonality
of the trigonometric functions, we can write the Maxwell eigenvalue problem for an
axisymmetric cavity V with perfect electric conducting boundaries as [1]:
⎧
⎪
⎨ For a given mode n ∈ Z, find the eigenpairs (en , ω2 ) with en ∈ S n (Ω) :
 
−1 n ω2  
⎪
⎩ μ r curln 
e n
· curln 
e dΩ − 2
εr en · en dΩ = 0 ∀en ∈ S n (Ω),
Ω c0 Ω
(1)

with εr = εr (r, z) and μr = μr (r, z) the scalar relative electric permittivity

and magnetic permeability of the medium, Ω a 2D angular cross section of V ,
dΩ = r dr dz, ω the angular frequency, c0 the speed of light in vacuum, S n (Ω) the
function space of the nth Fourier coefficient and
⎡  ⎤
−r −1 nezn + ∂z (reϕn )
⎢ ⎥
curln en = ⎣ ∂z ern − ∂r ezn ⎦.
 
+r −1 nern + ∂r (reϕn )

2 Well-Posed Variational Formulation

In order to construct an appropriate subspace of S n (Ω) and in order to account for

the singular behavior of curln at r = 0, two strategies have been proposed in the
literature.

2.1 Non-classical Conditions Along the Symmetry Axis

A first approach consists in taking the unknown fields eϕ,n = reϕn ∈ H 1 (Ω) and
 
n = e n , e n T ∈ H (curl, Ω) [2] together with non-classical discrete conditions at
erz r z
the symmetry axis [3, Section 4.4]. By following this strategy, all integrals are well-
posed but exhibit singular integrands, hence requiring either i) a classical Gaussian
quadrature with a large number of quadrature points or ii) specialized quadrature
rules [3, Section 5.1] which differ from element to element, thus preventing the
use of fast assembly techniques [4]. In what follows, this approach will be further
referred to as transformation “TA”.
Comparison of Different Formulations for Axisymmetric EM Wave Problems 171

2.2 Direct Construction of a Subspace of S n (Ω)

Another approach consists in directly constructing an appropriate subspace of

S n (Ω) such that the variational formulation is guaranteed to be always well-posed,
as shown in [1, 5–7] for instance, thus avoiding the need for non-classical conditions
on the symmetry axis. To this end, the unknowns eϕn and erz n are transformed into

un ∈ H 1 (Ω) and U n ∈ H (curl, Ω) by following the methodology shown in

 T
Table 1, with gradrz eϕn = ∂r eϕn , ∂z eϕn and r̂ the unit vector along the r-axis.
The parameters α and β in TC(α, β) must satisfy, according to [6], the constraints
shown in Table 2. Furthermore, some transformations need an additional homoge-
neous Dirichlet condition at r = 0, as shown in Table 3. Finally, for appropriate
choices of α and β, TC(α, β) leads to polynomial integrands (see Sect. 3.2 for more
details). This property is also met by TB for n = 0 and TD for n = ±1.

Table 1 Different transformations for constructing a subspace of S n (Ω)

Mode Transf. TB [7] Transf. TC(α, β) [6, Section 1.3] Transf. TD [5]
n=0 u0 = eϕ0 r β u0 = reϕ0 u0 = eϕ0
U 0 = erz
0 U 0 = erz
0 U 0 = erz
0

n = ±1 u±1 = eϕ±1 r β u±1 = reϕ±1 u±1 = eϕ±1

n ±1 eϕ±1 n ±1 eϕ±1
U ±1 = erz + r̂ r α U ±1 = ±erz
±1 + grad (re±1 )
rz ϕ U ±1 = erz + r̂
r r r r
|n| > 1 u = eϕ
n n r β un = reϕn u = eϕ
n n

n n eϕn n n
U n = erz + r̂ r α U n = nerz
n + grad (ren )
rz ϕ U n = erz
r r r

Table 2 Constraints on α and β for TC(α, β) according to [6, Section 1.5]

n=0 n = ±1 |n| > 1
β ≥ 0.5 α ≥ 0.5 and β = 1 α ≥ 0.5 and β > 0

Table 3 Conditions on the symmetry axis

Mode Transf. TB [7] Transf. TC(α, β) [6, Section 1.5] Transf. TD [5]
n=0 u0 = 0 u0 = 0 if β ∈ [0.5, 1.5[, none otherwise u0 = 0
n = ±1 None None None
|n| > 1 un = 0 un = 0 if β ∈]0, 1], none otherwise un = 0
172 E. Schnaubelt, N. Marsic, and H. De Gersem

3 Comparison and Discussion of the Quasi-3D Methods

As already stated, this work compares the aforementioned transformations to treat

the eigenvalue problem (1). To this end, they were implemented in a homemade
high-order finite element (FE) code.1 All following numerical experiments are
performed on a pillbox cavity with radius r = 150 mm, height h = 294 mm, and
μr = εr = 1, for which closed-form solutions are well-known [8, Sections 1.13 and
1.14]. A structured triangular mesh is used which is refined by uniformly splitting
each triangle into four subtriangles.

3.1 Spurious Modes and High-Order FE Discretizations

Let us start our comparison by determining if the methods discussed previously can
avoid spurious modes. As we search the azimuthal unknown (eϕ∗,n for TA and un
for TB, TC and TD) in a finite subspace2 of H 1 (Ω) of polynomial order q and the
in-plane unknown (erz for TA and U n for TB, TC and TD) in a finite subspace (see
footnote 2) of H (curl, Ω) of polynomial order p, the dimension of each subspace
must be selected with care. In particular, in order to satisfy the exactness of the
discrete de Rham sequence [10], one must impose that q = p + 1 [11].
In order to validate this choice, we ran multiple numerical tests with the different
transformations, different modes n and different values for p and q. As a result, we
observed that, apart from TD, all eigenspectra were free of spurious modes when
q = p + 1. Interestingly, we also observed no spurious modes when q > p + 1. On
the other hand, spurious modes were systematically observed when q < p + 1, and
when transformation TD was used with |n| > 1 (for all possible values of p and q).
For this reason, TD will not be investigated further. As an illustration, Fig. 1 shows
a part of the numerical spectrum of a pillbox cavity for n = 1 and different mesh
densities. It was computed with TB, once for q = 3, p = 2 and once for q = p = 2.
When n = 0, the in-plane and azimuthal unknowns are decoupled from each
other [6, Section 1.6]. Therefore, q and p can be chosen independently.

3.2 Convergence Results for Higher Order Finite Elements

In this subsection, the convergence behavior of the different formulations in

combination with high-order basis functions will be compared. As an example, the

1 See https://gitlab.onelab.info/gmsh/small_fem/blob/master/simulation/Quasi3D.cpp.
2
 
In this paper, a finite subspace of H 1 (Ω) resp. H (curl, Ω) is built using grad-conforming
(resp. curl-conforming) finite elements from [9, Chapter 4.5]. In particular, we consider complete
subspaces of H (curl, Ω) with both irrotational and rotational functions.
Comparison of Different Formulations for Axisymmetric EM Wave Problems 173

TM112 TE113 TE121 TM113 TE122 TE114 TM120 TE123 TM121

TM121
Ang. eigenfreq. (rad/s)

TE123
TM120 1.4 ·1010
TE114
1.3 ·1010
TE122
TM113
1.2 ·1010
TE121
1.1 ·1010
TE113
TM112 1 ·1010
2 4 8 16 32 2 4 8 16 32
# of mesh elements per wavelength of the TM112 mode (− )
(a) (b)

Fig. 1 Part of the spectrum of a pillbox cavity obtained with TB and n = 1. (a) Polynomial order
q = 3, p = 2. (b) Polynomial order q = 3, p = 3

10− 1
Rel. error eigenfreq. (− )

TA TB TC(1, 1)
G=4
10− 3 G=6
G=7
G = 12
10− 5
G = 13
G = 16
10− 7 G = 19
4
1

2 4 8 16 32
# mesh elements/wavelength (− )

Fig. 2 Convergence results when computing the eigenfrequency of the TE111 mode of a pillbox
cavity with TA, TB and TC(1, 1), using q = 3, p = 2 and G Gauss-Legendre quadrature points

evolution of the relative error between the numerically computed eigenfrequency

and its analytical counterpart is shown in Fig. 2 for different mesh densities and
different numbers of Gauss-Legendre quadrature points G. A second-order FE
method with q = 3, p = 2 is used, hence resulting in an expected convergence
slope of 4 [12]. This slope is indeed achieved for TA, TB and TC(1, 1) once the
number of quadrature points passes a certain threshold. Again, this is not an isolated
case but can be observed for all n and for different element orders.
Let us also stress that TB and TC(1, 1) (i) yield a lower relative error than TA
and (ii) depend less on G than TA. This last observation can be easily explained:
as TB and TC(1, 1) lead to polynomial integrands, the final solution is independent
of G, at least for a G sufficiently large to integrate a polynomial of the given order
exactly.
The parameters of TC(α, β) need to meet the following criteria to yield polyno-
mial integrands in the variational formulation (1): (i) α and β must be multiples
174 E. Schnaubelt, N. Marsic, and H. De Gersem

of 0.5, (ii) their sum must be an integer and (iii) β ≥ 1.5 for n = 0 and
α ≥ 0.5, β ≥ 0.5 for n = 0.

3.3 Influence of α and β on the Convergence Behavior

Let us now focus on the transformation TC(α, β), and let us carry out a convergence
test similar to the previous section. However, now, the influence of the parameters
α and β (chosen according to Table 2) on the convergence rate will be investigated.
The results of this numerical experiment are displayed in Fig. 3. As it can be
observed directly, while all choices converge towards the sought eigenvalue, only
particular pairs (α, β) exhibit the expected convergence rate. This behavior has been
observed for other choices of (n, p, q) with q = p + 1 as well.
This behavior can be easily explained if we assume that en ∈ C ∞ in the vicinity
of the symmetry axis. This assumption is of course restrictive, but applies to the
pillbox cavity [8], and gives already a good insight into the underlying numerical
mechanisms. In what follows, only the case n = ±1 will be discussed, but the same
methodology applies to the other cases.

Fig. 3 Convergence rate of TC(α, β) for different values of α and β, as allowed by Table 2.
The symbol “∗” in (a) indicates that the result is independent of p due to the decoupling of the
in-plane and azimuthal unknowns (see Sect. 3.1). (a) TM111 mode with q = 4, p = 3. (b) TE022
mode with q = 4, p = ∗
Comparison of Different Formulations for Axisymmetric EM Wave Problems 175

Let us start by expanding e±1 into a Taylor series in the vicinity of r = 0 and
z = z0 . As ez±1 = 0 at r = 0 (see [1]), we have:
⎧
⎪ ±1 z rr 2 zz 2 rz
⎨ eϕ (r, z) = a0 + a1 r + a1 (z − z0 ) + a2 r + a2 (z − z0 ) + 2a2 r(z − z0 ) + . . . ,
r
±1 2 rz
r rr
ez (r, z) = b1 r + b2 r + 2b2 r(z − z0 ) + . . . ,
⎪
⎩ ±1
er (r, z) = c0 + c1r r + c1z (z − z0 ) + c2rr r 2 + c2zz (z − z0 )2 + 2c2rz r(z − z0 ) + . . . .
(2)

Then, by exploiting the definition of U ±1 (see Table 1), we can write:

.
r α Ur±1 = ±er±1 +eϕ±1 +r(a1r + 2a2rr r + 2a2rz (z − z0 ) + . . . ),
(3)
r α Uz±1 = ±ez±1 + 0 +r(a1z + 2a2zz (z − z0 ) + 2a2rz r + . . . ).

Restricting the further analysis to the axial component, we then have that:

(3)  
Uz±1 = r −α ± ez±1 + r(a1z + 2a2zz (z − z0 ) + 2a2rz r + . . . ) ,

(2)   
= r −α r b1r + b2rr r + 2b2rz (z − z0 ) + · · · + a1z + 2a2zz (z − z0 ) + 2a2rz r + . . . ,

def 1−α
=r f (r, z),

where f is a polynomial function of r and z. As r → 0, we have that

f (r, z) → f (0, z). Thus, the unknown Uz±1 ∼ r 1−α as r → 0. Analogously, using
er±1 ± eϕ±1 = 0 at r = 0 [1], we find that Ur±1 ∼ r 1−α as r → 0. Therefore, for
the sought FE solution to be differentiable at r = 0, and because of the constraint
enforced by Table 2 (α ≥ 0.5), we need to impose that α = 1. Any other choice (in
accordance with Table 2) will lead to a non-differentiable U n , jeopardizing thus the
convergence of the FE scheme. By applying the same strategy to the cases n = ±1,
and by taking into account the restrictions imposed in Table 2, the set of allowed
couples (α, β) must be further narrowed, as shown in Table 4. To the best of our
knowledge, these last results have been derived for the first time.
For the integrands of the variational formulation (1) to be polynomial, we
further need to impose the following restriction: β = 2 when n = 0 (see
Sect. 3.2). Moreover, as the number of quadrature points depends on the order of the
integrands, it is preferable to select the smallest acceptable (α, β) couple. Therefore,
the values given in Table 5 are recommended. Let us finally note that, apart from the
case n = 0, these recommendations are in accordance with [6, Section 1.6], where
the best (α, β) couples were determined on the basis of numerical experiments.

Table 4 Values of α and β for TC(α, β) leading to a high FE convergence rate

n=0 n = ±1 |n| > 1
β ∈ {1, 2} α = 1 and β = 1 α ∈ {1, 2} and β ∈ {1, 2}
176 E. Schnaubelt, N. Marsic, and H. De Gersem

Table 5 Recommended choice of α and β for TC(α, β)

n=0 n = ±1 |n| > 1
β =2 α = 1 and β = 1 α = 1 and β = 1

Concerning the case n = 0, the choice β = 1 leads to slightly more accurate

results for a given mesh density in a numerical experiment carried out in [6, Section
1.6]. However, our numerical experiments do not confirm these results as the choice
β = 2 leads to a slightly lower relative error for the same mesh density (see Fig. 3a).
This behavior is supported by the fact that the integrands are polynomial for the
latter choice β = 2.

4 Conclusion

This paper compared four different transformations to treat three-dimensional time-

harmonic electromagnetic wave problems in axisymmetric geometries proposed in
the literature. We first determined numerically that the transformations TA, TB and
TC(α, β) lead to eigenvalue problems which are free of spurious modes, while
the transformation TD exhibits spurious modes when |n| > 1. We then compared
numerically the accuracy of TA, TB and TC(α, β), and found that TB and TC(α, β)
produce the most accurate results for a given mesh density. Finally, we analyzed
theoretically the convergence rate of TC(α, β) for different values of α and β in a
high-order FE context, and determined new restrictions on α and β.

Acknowledgments The authors would like to express their gratitude to Abele Simona for his
valuable advice and the fruitful discussions on axisymmetric problems.

References

1. P. Lacoste, Solution of Maxwell equation in axisymmetric geometry by Fourier series

decompostion and by use of H(rot) conforming finite element. Numer. Math. 84(4), 577–609
(2000)
2. J.F. Lee, G.M. Wilkins, R. Mitra, Finite-element analysis of axisymmetric cavity resonator
using a hybrid edge element technique. IEEE Trans. Microwave Theory Techn. 41(11), 1981–
1987 (1993)
3. O. Chinellato, The complex-symmetric Jacobi-Davidson algorithm and its application to the
computation of some resonance frequencies of anisotropic lossy axisymmetric cavities. Ph.D.
Thesis, Eidgenössische Technische Hochschule Zürich, 2005
4. N. Marsic, C. Geuzaine, Efficient finite element assembly of high order Whitney forms. IET
Sci. Meas. Technol. 9(2), 204–210 (2015)
5. E.A. Dunn, J.K. Byun, E.D. Branch, J.M. Jin, Numerical simulation of BOR scattering and
radiation using a higher order FEM. IEEE Trans. Antennas Propag. 54(3), 945–952 (2006)
Comparison of Different Formulations for Axisymmetric EM Wave Problems 177

6. S. Cambon, Méthode d’éléments finis d’ordre élevé et d’équations intégrales pour la résolution
de problème de furtivité radar d’objets à symétrie de révolution. Ph.D. Thesis, Institut National
des Sciences Appliquées de Toulouse, 2012
7. M. Oh, de Rham complexes arising from Fourier finite element methods in axisymmetric
domains. Comput. Math. Appl. 70(8), 2063–2073 (2015)
8. T.P. Wangler, RF Linear Accelerators. 2nd, Completely Revised and Enlarged edition. (Wiley-
VCH Verlag GmbH & Co. KGaA, Weinheim, 2008)
9. S. Zaglmayr, High Order Finite Element Methods for Electromagnetic Field Computation.
Ph.D. Thesis, Johannes Kepler Universität Linz, 2006
10. A. Simona, L. Bonaventura, C. de Falco, S. Schöps, Isogeometric approximations for
electromagnetic problems in axisymmetric domains (2019). arXiv preprint arXiv:1912.08570
11. L. Demkowicz, P. Monk, L. Vardapetyan, W. Rachowicz, de Rham diagram for hp finite
element spaces. Comput. Math. Appli. 39(7–8), 29–38 (2000)
12. S. Sauter, hp-finite elements for elliptic eigenvalue problems: error estimates which are explicit
with respect to λ, h, and p. SIAM J. Numer. Analy. 48(1), 95–108 (2010)
The Magnetization Analysis of Motor
Magnet and Its Influence on Cogging
Torque

Chenxi Wang, Matthias Willig, Stefan Kurz, and Kevin Gutmann

Abstract In this article, a simulation procedure of a BLDC motor is described that

includes the magnetization process of the magnet poles. This leads to more realistic
magnetic field distributions in the motor during the cogging torque analysis. The
process of the magnetization calculation as well as the handling of the material
properties is explained and the influence of isotropic and anisotropic material
definition on the results is shown.

1 Introduction

The acoustic performance is one of the most important indicators to evaluate the
comfort of automobile, therefore, the acoustics of electric machines as a common
device in vehicles is a critical point that needs to be considered in the machine design
process.
However, the increasing precision requirements for the accuracy of prediction
of noise-exciting forces of electric machines pose a significant challenge to the
assumptions and idealizations applied in motor design process today. In order
to meet higher precision requirements, it is necessary to adapt influences from
statistical geometry variations, material fluctuations and the manufacturing process
of the machine. As an increasingly used electric machine in automobile, the acoustic
and vibration performance of the BLDC motor (Brushless Direct Current Motor) is
an important quality indicator of the machine. In the no load operation, it is mainly
determined by its cogging torque, which is highly influenced by the motor geometry
and magnetic field in the air gap [1–3]. Hence, instead of using an idealized
magnetic field, a more accurate and realistic description of the magnetic field of

C. Wang · S. Kurz
Technical University Darmstadt, Darmstadt, Hessen, Germany
e-mail: [email protected]
M. Willig · K. Gutmann ()
Robert Bosch GmbH, Buehl, Baden-Wuerttemberg, Germany
e-mail: [email protected]; [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 179
M. van Beurden et al. (eds.), Scientific Computing in Electrical Engineering,
Mathematics in Industry 36, https://doi.org/10.1007/978-3-030-84238-3_18
180 C. Wang et al.

magnet needs to be taken into consideration, in order to achieve a higher accuracy

of the cogging torque calculation [1, 4].

2 BLDC Motor and Cogging Torque

The motor considered in this analysis is a BLDC motor in outer rotor configuration.
The topology of the motor is 12/8 i.e. the stator has 12 slots and the rotor has 8
magnet poles. In the motor used for this analysis, these 8 poles are magnetized on a
magnet ring as indicated by the vectors of the magnetic flux density B in Fig. 1.
The generation of cogging torque in the motor is determined by the interaction
of the magnet poles of the rotor and the slots of the stator. Based on the co-energy
in the system, the calculation of the motor cogging torque is given by the following
equation, which is the fundament of the torque calculation in the FEA-tool [2, 3].
  HB
dW ∂
T = = [ ( B(H )dH )dV ] , (1)
dθ ∂θ V 0

where W is the magnetic coenergy, θ is the rotor position, H the magnetic field,
B is the flux density, HB is the magnetic field in operating point, V the integration
volume and T the calculated torque. For the applied method of virtual work, the
change in the coenergy of the system (and therefore the virtual torque) is given by
the change in the coenergy of the virtually distorted finite elements.

Fig. 1 Model structure of the BLDC motor

The Magnetization Analysis of Motor Magnet and Its Influence on Cogging Torque 181

Fig. 2 The cogging torque of BLDC motor with 12 slots and 8 poles

The fundamental of the cogging torque of a 12/8 motor topology is given by

the least common multiple [2, 5] of 12 and 8 which is the 24th mechanical order.
Figure 2 shows a typical cogging torque curve of the motor. The distribution of
magnetic flux for minimum and maximum of rotor position dependent no-load
torque is also shown.
One can observe that the no load torque at different rotor positions depends
on the flux distribution in the motor when there is a magnetic field imposed
by the permanent magnet poles. This flux distribution is mainly determined by
the magnetization of the permanent magnet poles and the shape of the magnetic
circuit. A more realistic representation of the magnetization leads to a more precise
prediction of cogging torque of the electric machine.

3 Definition of the Magnet Material in Analysis

3.1 Magnetization Curve of the Magnet

By using the 3D FEA-tool to simulate the magnetization process, the input data
required for this method is the magnetization curve in the first quadrant of B − H
coordinate system, as shown in Fig. 3. This curve was measured using a standard
Permagraph measurement method and extrapolated to the point of the maximum
excitation field.
Technical ferrites are usually manufactured to have a main axis for the preferred
direction of the magnetic flux. In this case the magnetization process not only
depends on the magnetizing field but also on the axis in which the field is active.
Those effects of anisotropy and isotropy on the simulation results will be shown
later in this paper.
182 C. Wang et al.

Fig. 3 The measured virgin curve of the analyzed magnet

3.2 Approximation Method to Describe the Hysteresis Effect

With the defined magnetization curve, the hysteresis effect of the magnet material
needs to be introduced in analysis, in order to evaluate the remanence of the magnet
after magnetization. However due to the limitations of the FEA-tool used, the
commonly used hysteresis models like Preisach model and Jiles Atherton model
are not supported in the simulation. Instead, the software supports method, which is
called “classic approach” or “linear approach” to describe the hysteresis [6].
The linear approach is based on the approximated linearity in the descending
branch of the hysteresis loop of magnet. When the magnetization field is removed,
the magnetic polarization J descends with constant slope, which is identical to the
slope in the saturated region [7], as indicated by the dashed line in the Fig. 3. The
intersection of the descending line and vertical curve is the remanence point, which
is around 415mT in the analyzed magnet. The slope of the descending part is equal
to the saturated permeability, which is μ0 μm for the curve of the magnetic flux
density B, and μ0 (μm − 1) for the curve of magnetic polarization J . For the ferrite
magnet used in this motor, the relative permeability in saturated region is μm ≈
1.05.
In addition, because of the linearity of the descending branch, only the operating
point with maximum excitation field is necessary to be simulated. Hence, the
magnetization analysis in the FEA-tool can be significantly simplified to a single
magnetostatic simulation with the maximum excitation field.
The Magnetization Analysis of Motor Magnet and Its Influence on Cogging Torque 183

4 Modelling of the Magnetization Device and Motor

Both the magnetization and cogging torque analyses are executed in the 3D
environment in FEA-tool. The ring magnet is magnetized in the magnetization
analysis first and then its remanent field will be transmitted into the motor model
to calculate the cogging torque.
A full 3D model of the magnetization unit is built and the magnet inserted as
shown in Fig. 4, where a 90◦ slice of the whole model is depicted.
The excitation current in the windings is generated by the discharge of a
connected capacitor. However, in the analysis only the peak value of the current
impulse is needed when the linear approach is used. Moreover, the possible effects
of eddy currents due to the transient current impulse are not considered in this
analysis.
Figure 5 shows a 2D magnetic field distribution in the middle cross section of the
model, with the maximum excitation current applied. It is obvious that the magnet is
not uniformly magnetized due to the field distribution imposed by the magnetizing
unit. There are different areas in the magnet (particularly areas close to the pole
transition zones) that reach different parts of the virgin curve and therefore will
have different remanent inductions after the magnetizing field is removed.
After the magnetization analysis the magnetized magnet is available for cogging
torque analysis of the motor via an internal datalink in the FEA software. The motor
model is built up as in Fig. 1. The magnetization of the magnet in the motor is
identical to the magnetization achieved in the magnetization calculation.
Since the cogging torque is evaluated at no load conditions the coils of the motor
are omitted in the analysis to reduce the number of finite elements and therefore
computing time. Moreover, due to the symmetry of the motor model, only a section
of 30◦ is necessary to be analyzed.

Fig. 4 Model structure of the magnetization device

184 C. Wang et al.

Fig. 5 The B field distribution in middle section with maximum excitation current

5 Result of Analysis

The motor cogging torque was calculated for three different magnet models:
1. A calculated magnetization profile based on isotropic magnet material.
2. A calculated magnetization profile based on anisotropic magnet material.
3. An ideal magnetization profile.
The 3D FEA solver of the tool applied uses tetrahedral mesh elements. A fine
mesh was applied to the magnet in the magnetization as well as the motor analysis.
In each model a mesh of approximately 60k tetrahedra for the magnet was achieved
and a maximum energy error of 0.5% for the whole system was set as solver
criterion.
In the isotropic magnet model, the pre-defined magnetization curve is valid for
all magnetization directions, but only valid in the radial direction for the anisotropic
model. The response of the materials to a magnetizing field in the simulation can
be seen in Fig. 6. Whereas the resulting magnetization of the isotropic material is a
vector that is parallel to the applied magnetizing field, the resulting magnetization
of the anisotropic material is a vector representing the component of the applied
magnetizing field in the preferred direction of the material.
The Magnetization Analysis of Motor Magnet and Its Influence on Cogging Torque 185

Fig. 6 Material response to magnetizing field in simulation

Compared to the calculated magnetization fields in the last section, the ideal
field is defined as a purely radial and homogeneous magnetic field with constant
magnitude. The magnetic fields of these three cases are shown in Fig. 7.
Results of the cogging torque analysis for the three cases mentioned above are
shown in Fig. 8
From the comparison, it can be seen that the difference among the curves for all
different magnet settings is slight. The curve of the anisotropic magnet is closer to
the curve under ideal condition, the reason is the similarity between both magnetic
fields. Both, the anisotropic field and ideal field only have a radial component of the
field vector, the only difference between both fields is the magnitude in the transition
zone.
For a detailed analysis, the curves can be transformed into frequency domain by
using FFT analysis. The result is depicted in Fig. 9.
From this figure, it can be seen that the differences appear mainly in the 24th
and 48th harmonics, which are up to 25% difference in the amplitude of order 48.
Compared to the torque curves in time domain, the difference in frequency domain
is much more prominent. The 24th harmonic is the main order of the torque curve.
It is caused by the interaction between the first harmonic of the magnetic field in the
air gap and the stator teeth.
Consequently, using the magnetic field from the magnetization analysis can
improve the accuracy of the motor optimization, because the cogging torque
fundamental and harmonics are mainly responsible for the coast down noise.
186 C. Wang et al.

Fig. 7 3D and 2D figures of the remanence field of all three cases. (a) 3D Field of the isotropic
magnet. (b) 2D Field of the isotropic magnet (Z=0, 0 ≤ ϕ ≤ 90◦ ). (c) 3D Field of the anisotropic
magnet. (d) 2D Field of the anisotropic magnet (Z=0, 0 ≤ ϕ ≤ 90◦ ). (e) 3D Field of the ideal
magnetization. (f) 2D Field of the ideal magnetization (Z=0, 0 ≤ ϕ ≤ 90◦ )
The Magnetization Analysis of Motor Magnet and Its Influence on Cogging Torque 187

Fig. 8 Comparison of the cogging torque curves

Fig. 9 Comparison the cogging torque in frequency domain

6 Summary

The calculation of the magnetization of the permanent magnet poles of an outer rotor
BLDC motor results in a more realistic field distribution in the motor simulation
and therefore allows a more accurate prediction of the cogging torque of the motor.
This supports the overall design and optimization process of these machines. For
proprietary reasons in this paper, the method is described using a ring magnet
motor design. With real motor samples, an improved agreement of prediction (3D
188 C. Wang et al.

field simulation) and measurement was found. It has to be pointed out that the
ferrites used are usually manufactured be anisotropic i.e. to have a main axis
and a lateral axis that respond differently to a magnetizing field. However due to
material imperfections and variances in the manufacturing process the material is
not perfectly anisotropic but has isotropic regions as well.
Therefore, to further improve the accuracy of the cogging torque analysis, an
improved material definition that accounts for anisotropic as well as isotropic
material properties needs to be developed.

References

1. I. Coenen, M. van der Giet, K. Hameyer, Manufacturing tolerances: estimation and prediction of
cogging torque influenced by magnetization faults. IEEE Trans. Mag. 48(5), 1932–1936 (2012)
2. M. Flankl, A. Tüysüz, J.W. Kolar, Cogging torque shape optimization of an integrated generator
for electromechanical energy harvesting. IEEE Trans. Ind. Electron. 64(12), 9806–9814 (2017)
3. C. Breton, J. Bartolome, J.A. Benito, G. Tassinario, I. Flotats, C.W. Lu, B.J. Chalmers, Influence
of machine symmetry on reduction of cogging torque in permanent-magnet brushless motors.
IEEE Trans. Magn. 36(5), 3819–3823 (2000)
4. Z.Q. Zhu, D. Howe, C.C. Chan, Improved analytical model for predicting the magnetic field
distribution in brushless permanent-magnet machines. IEEE Trans. Mag. 38(1), 229–238 (2002)
5. J.R. Hendershot, T.J.E. Miller, Design of Brushless Permanent-Magnet Machines (Motor Design
Books, Venice, 2010). ISBN 0-98406-870-8, 9-780-98406-870-8
6. ANSYS Maxwell support: Using the Hysteresis Model-Based Magnetization Approach
7. ANSYS Maxwell support: Compute Remanent Br from B-H curve
 Part IV
Mathematical and Computational Methods
A Combination of Model Order
Reduction and Multirate Techniques
for Coupled Dynamical Systems

M. W. F. M. Bannenberg, A. Ciccazzo, and M. Günther

Abstract Coupled dynamical systems are often encountered in the field of circuit
simulation. To drastically reduce the simulation cost of these systems a coupling
of model order reduction and multirate techniques is applied. The subject of this
method is a nonlinear coupled thermal-electrical system. By applying a combination
of the slowest first multirate technique with the nonlinear proper orthogonal
decomposition model order reduction the system is solved. Results yield a decrease
in simulation time whilst maintaining accuracy.

1 Introduction

Building an integrated circuit, for instance a microchip, is a complex process.

For the construction of such integrated circuits, first the silicon components are
made and then connected through a conducting metal. Before these circuits can
be constructed they need to be designed and tested by a computer. To this end
the integrated circuits are described in mathematical form by differential-algebraic
equations (DAEs). However, in the simulation of these circuits there are many
more phenomena to consider besides only the connections and interactions between
the components. Including these other phenomena occurring inside a microchip,
like thermal or electromagnetic coupling, greatly improves the accuracy of the
simulation. These considerations lead to a system of Partial Differential-Algebraic

M. W. F. M. Bannenberg ()
Bergische Universität Wuppertal, Wuppertal, Germany
STMicroelectronics, Catania, Italy
e-mail: [email protected]
A. Ciccazzo
STMicroelectronics, Catania, Italy
M. Günther
Bergische Universität Wuppertal, Wuppertal, Germany
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 191
M. van Beurden et al. (eds.), Scientific Computing in Electrical Engineering,
Mathematics in Industry 36, https://doi.org/10.1007/978-3-030-84238-3_19
192 M. W. F. M. Bannenberg et al.

Equations (PDAEs). Where DAEs and partial differential equations, describing the
spatially distributed elements and effects, are coupled via source terms or boundary
conditions. Both physical and structural characteristics of these PDAEs can be
exploited to increase the simulation efficiency. For instance by applying techniques
such as Multirate (MR) time integration and Model Order Reduction (MOR), as
will be presented in the following sections. Circuit simulation has been a driving
force for the application of MOR and MR techniques, see for instance [3, 10].
Much less attention has been given to the combination of these two techniques,
[13], and only with respect to linear model order reduction. In this paper a twofold
approach is presented in which the PDAEs are integrated using MR time integration
and parts of the system are reduced. This is done to increase the computational
efficiency, whilst maintaining accuracy. In Sect. 2, the mathematical methodology
is formulated for the circuit simulation and the multirate and MOR techniques
are described. Section 3 presents the experimental setup and the numerical results
obtained from the implementations of the previous two sections. Conclusions are
drawn and an outlook is given in Sect. 4.

2 Methodology

In this section the different mathematical concepts and techniques that are needed
for the simulation of electronic circuits are presented. Although most equations are
purposely stated in their most general form, some of them will be restricted by
assumptions with the specifics combination of MR and MOR in mind.

2.1 Mathematical Modelling

The Modified Nodal Analysis (MNA) approach for modelling electronic circuits
yield time-dependent systems of DAEs,

d
A C q + AR r(ATR e) + ALiL + AV iV + AI i(t) = 0,  (1)
dt
d  
φ − ATL e = 0, (2)
dt
v(t) − A TV e = 0 (3)

q − qC (A TC ) = 0, (4)

φ − φL (iL ) = 0. (5)
Combined MR-MOR for Coupled Systems 193

Where e, iL , V are the node voltages and branch currents through inductors and
voltage sources, and the charges and fluxes q, φ. The functions r, qC and φL are
predetermined. Independent current sources iI and voltage sources vV may appear.
The incidence matrices AC , AL , AR , AV , AI follow from the topology of the circuit.
This system can be written in the general semi-explicit DAE form, [7, 12].

f : Rn × Rm × I → Rn , g : Rn × Rm × I → Rm . (6)

y˙ = f(
y , z, t), y(0) = y0 , (7)
0 = g(
y , z, t), z(0) = z0 . (8)

With y : I → Rn and z : I → Rm denoting the differential and algebraic solutions

on time-interval [t0 , t1 ], respectively. Furthermore y0 and z0 need to be consistent
initial conditions. Secondly, other phenomena can be included via PDEs, which are
denoted in general form given by

L : D × I × V → Rm , L (
x , t, u) = 0. (9)

where L is a differential operator, D ⊂ Rd , with d ∈ {1, 2, 3} the spatial domain

and V a function space to which u : D × I → Rm belongs. These two systems
consisting of DAEs and PDEs can be coupled together in PDAEs. After applying
a suitable space discretization to the PDAEs the following initial value problem of
semi-explicit DAEs is obtained

y˙ = f(
y , z, u, t), y(0) = y0 , (10)
0 = g(
y , z, u, t), z(0) = z0 , (11)
u˙ = h(
 y , z, u, t), u(t0 ) = u 0 . (12)

2.2 Multirate

Since the coupled system (10)–(12) is constructed by the combination of two

different processes it can be assumed that they act within different time scales.
To exploit this characteristic, the total system is partitioned into fast and slow
subsystems, with xF = y, xS = u and zF = z,

x˙F = fF (
xF , zF , xS ), xF (0) = xF,0 , (13)
x˙S = fS (
xF , zF , xS ), xS (0) = xS,0 , (14)
0 = gF (
xF , zF , xS ), zF (0) = zF,0 . (15)

With differential variables xF ∈ RnF , xS ∈ RnS and algebraic variables zF ∈ RnZ ,
subscripts {F, S} indicating fast or slow dynamics, for t ∈ [t0 , t1 ] with consistent
194 M. W. F. M. Bannenberg et al.

initial conditions. The system is guaranteed to be of index-1 by assuming that the

Jacobian

gFz (
xF , zF , xS ) is invertible (16)

in a neighbourhood of the solution of the system (13-15). The algebraic constraints

are partitioned into the fast subsystem. This type of coupling lets us consider
electrical circuits with a differential index up to 1, coupled with slower ODE
systems. The total index-1 system can be integrated with the stiffly accurate implicit
Euler method. To exploit the assumed different time scales, a multirate integration
method is proposed. This approach is analogous to [14] but with the algebraic
constraint in the fast subsystem, taking the subsequent MOR into account. The
integration of the coupled system (13-15) for one macro-step tn → tn+1 = tn + H
is defined as

xF,n+(l+1)/m , zF,n+(l+1)/m , x¯S,n+(l+1)/m ),

xF,n+(l+1)/m = xF,n+l/m + hfF (
(17)
xS,n+1 = xS,n + H fS (x¯F,n+1 , z¯ F,n+1 , xS,n+1 ), (18)
xF,n+(l+1)/m , zF,n+(l+1)/m , x¯S,n+(l+1)/m ).
0 = gF ( (19)

With l = 0, . . . , m − 1 for the micro grid and the coupling variables denoted by x¯F ,
z¯ F , x¯s . The coupling strategy is chosen to be the Coupled-Slowest-First approach as
this is shown to have a consistency of order 1 for the problem posed in [14]. First
the whole system is solved for the macro-step.
∗
xF,n+1 = xF,n + H fF ( ∗
xF,n+1 ∗
, zF,n+1 , xS,n+1 ), (20)

xS,n+1 = xS,n + H fS ( ∗

xF,n+1 ∗
, zF,n+1 , xS,n+1 ), (21)
∗ ∗
0 = gF (
xF,n+1 , zF,n+1 , xS,n+1 ). (22)

Where the step size H is chosen according to the slow dynamics. From this it follows
∗
that the fast solutions, xF,n+1 ∗
and zF,n+1 , are not accurate and discarded. Following
the micro-step integration the fast solutions are computed for l = 0, . . . , m − 1,
using linearly interpolated values for the slow variables.

2.3 Model Order Reduction

Applying a spatial discretization to the PDE can result in large nonlinear ODE
systems. To reduce the computational effort needed in each time step to solve
this system MOR techniques are used. Due to the nonlinearity of the ODE most
conventional MOR techniques can be discarded as they are only applicable to linear
Combined MR-MOR for Coupled Systems 195

systems. Hence the chosen method for this system is a reduction by a Galerkin
projection, with a basis constructed by Proper Orthogonal Decomposition (POD),
[4]. This is then extended by the application of the Discrete Empirical Interpolation
Method (DEIM), [5], using a QR selection procedure (Q-DEIM), [6]. By using
a Galerkin projection a reduced model is constructed, [6]. Let Vr denote an r-
dimensional subspace spanned by the columns of V ∈ RnS ×r . The full state of
the slow subsystem xS is then approximated by xS ≈ V xS,r using model reduction
basis V . The reduced model of (13)–(15) is then defined by

x˙F = fF (
xF , zF , V xS,r ), xF (0) = xF,0 , (23)
x˙S,r = fS,r (
xF , zF , xS,r ), xS,r (0) = xS,r,0, (24)
0 = gF (
xF , zF , V xS,r ), zF (0) = zF,0 . (25)

With fS,r (
xF , zF , xS,r ) = V T fS (
xF , zF , V xS,r ). The reduced basis V is
constructed through POD. First a numerical simulation of the full system is
performed. From the numerical results of this simulation snapshots x1 , xi , . . . , xNS
are obtained, with xi = x(ti ) ∈ RnS for i = 1, . . . , NS . Then the POD snapshot
matrix is

X = [x1 , . . . , xNS ] ∈ RnS ×NS . (26)

From this the thin Singular Value Decomposition (SVD) is computed

X = ZΣY T , (27)

where Z ∈ RnS ×k , Y ∈ RNS ×k are orthogonal and Σ = diag(σ1 , . . . , σk ) ∈ Rk×k

with k = min(nS , NS ). Now a reduction basis V is constructed by taking the leading
r singular vectors of Z corresponding to the r largest singular values. However there
is a problem with the Galerkin projection that causes computational inefficiencies.
The reduced term fS,r (xF , zF , xS,r ) has a computational complexity that depends
on the non-reduced full order size nS . To reduce the number of evaluations Q-DEIM
is applied. Consider the nonlinear function fS : T → RnS with T ⊂ RnS , the
coupled terms are dropped from the notation as these are not reduced, and matrix
U ∈ RnS ×m of rank m. Then the DEIM approximation of fS is defined by, [5,
Definition 3.1],

fˆS (τ ) = U (SU )−1 ST fS (τ ). (28)

where S is a selection matrix of size nS × m by selecting columns of identity matrix

I of size nS × nS . Then the reduced nonlinear function fS,r is approximated with
the QDEIM approach by

fS,r (
xS,r ) ≈ V T U (SU )−1 ST fS (V xS,r ). (29)
196 M. W. F. M. Bannenberg et al.

Using the interpolation of general nonlinear functions, outlined in Sect. 3.5 of [5],
a general nonlinear function can be represented as

[F ( y ) = Fi (
y )]i = Fi ( y (ji )),
yj i , yj i , . . . , yjni ) = Fi ( (30)
1 2 i

where Fi : Yi → R, Yi ⊂ Rni , with integer vector ji = [j1i , j2i , . . . , jni i ]

denoting the indices of the components required to evaluate Fi . The numerical
implementation of this allows to compute (29) without the full evaluation of fS .
Depending on the underlying nonlinear function it might even be possible to
compute (29) without lifting xS,r to the full dimension nS . This however depends
on the dependencies of the individual functions.

2.4 Combining MR and MOR

To maximise the effectiveness of the MR and MOR combination the following steps
are taken:
• Perform a benchmark simulation using a very large number of time steps to
obtain a very accurate snapshot matrix X.
• The reduced bases V and U are then constructed by taking the appropriate
columns of Z obtained through POD, and selection matrix S is constructed by
the Q-DEIM approach.
• Using the reduced bases, the reduced order system is integrated through time
using the Coupled-Slowest-First MR approach.
The computational approach of this is done by first using the scheme of (17)–(19)
with fS replaced by fS,r , as in (23)–(25), and then incorporating the Coupled-
Slowest-First approach. The coupling for the fast intermediate time-step is done by
using linear interpolated values. As these values don’t change during the Newton
iteration of solving the faster subsystem, computation time can be saved. By
computing the coupling values once for the first and last value and interpolating
between these values, expensive function evaluation of the lifted state vector can be
avoided.

3 Results

In this section the previously described MR-MOR integration scheme is imple-

mented and applied to the thermal-electric test problem of [2]. For the simulation
the algorithm has been implemented in C++ using various packages, GSL 2.6 [8],
LAPACK [1] and Eigen [9]. Visualisation of the results is done by using MATLAB
2019b under the Academic Student License.
Combined MR-MOR for Coupled Systems 197

3.1 Experimental Setup

To test the accuracy and convergence of the MR-MOR integration scheme the
system is simulated in three settings:
• The full system with singlerate time integration.
• The full system with multirate time integration.
• The system with a reduced slow part and multirate time integration.
This is done for with intermediate micro-steps m = 5. Furthermore the POD-
QDEIM reduction factors r and g are chosen to be equal to the number of largest
singular values with σi > 1e − 15. The step sizes are obtained by integrating
the system with Nt = [8 16 32 64 128 256 512 1024]. For the simulation a
thermodynamic discretisation is chosen to have N = 101. The circuit is simulated
over a time interval from t0 = 0 to tN = 0.01125 seconds. The input signal v(t)
πt
is set to sin( 2.5e−3 )mV . The rest of the circuit and thermal settings are set to the
values as described in [2]. The reference solution is obtained from an SR integration
with N = 32, 000.
In Fig. 1 we see the difference between the reference solution and the simulated
solution in the final time-step. This is done for the output node u3 of the thermal-
electrical circuit. It clearly shows that the MR scheme outperforms the SR, as
for the same order of error the MR approach has a slower computation time.

-11
SR, MOR, MR-MOR comparison
10
SR
MR
MR-MOR

10 -12
Error in output

-13
10

-14
10
10 -2 10 -1 10 0 10 1
Computation time

Fig. 1 Accuracy of SR, MOR and MR-MOR

198 M. W. F. M. Bannenberg et al.

Incorporating the POD-QDEIM MOR method results in an even further reduction

of the computational integration effort. However, there needs to be done a snapshot
simulation before this can be used. Thus the speedup has a one time extra cost.

4 Conclusion

From the numerical results it shows that the multirate implicit Euler scheme
combined with POD/Q-DEIM model order reduction results in an accurate solution
with a reduced computation time. The approximation errors seem to converge along
with the MR errors. An expected positive result is that for similar computation
times the application of MR improves the accuracy of the solution. Furthermore,
the additional application of POD/Q-DEIM reduction has a trivial impact on the
approximation error whilst even further reducing the computation time. Although
these results are positive a side note should be made. The reduction in computation
time of the POD/Q-DEIM reduction shows to be decreasing for smaller time steps
with much larger systems. This is likely due to the coupling structure of the test
problem but further investigation is needed. Other next steps will focus on numerical
analysis for a proof of convergence for the MR-MOR scheme and the extension to
general integration schemes. Besides the further investigation of DAE-ODE coupled
systems, first steps have been made towards a MR-MOR scheme for a DAE-DAE
coupled system.

Acknowledgments The authors are indebted to the funding given by the European Union’s
Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie Grant
Agreement No. 765374.

References

1. E. Anderson et al., LAPACK Users’ Guide, vol. 9 (SIAM, New York, 1999)
2. A. Bartel, M. Günther, From SOI to abstract electric-thermal-1D multiscale modeling for first
order thermal effects. Math. Comput. Modell. Dynam. Syst. 9(1), 25–44 (2003)
3. P. Benner, M. Hinze, E.J.W. Ter Maten (eds.) Model Reduction for Circuit Simulation
(Springer, Berlin, 2011)
4. G. Berkooz, P. Holmes, J.L. Lumley, The proper orthogonal decomposition in the analysis of
turbulent flows. Annu. Rev. Fluid Mech. 25(1), 539–575 (1993)
5. S. Chaturantabut, D.C. Sorensen, Nonlinear model reduction via discrete empirical interpola-
tion. SIAM J. Sci. Comput. 32(5), 2737–2764 (2010)
6. Z. Drmac, S. Gugercin, A new selection operator for the discrete empirical interpolation
method—improved a priori error bound and extensions. SIAM J. Sci. Comput. 38(2), A631–
A648 (2016)
7. D. Estévez Schwarz, C. Tischendorf, Structural analysis of electric circuits and consequences
for MNA. Int. J. Circ. Theory Appl. 28(2), 131–162 (2000)
8. M. Galassi, GNU scientific library (2002). http://www.gnu.org/software/gsl/
Combined MR-MOR for Coupled Systems 199

9. G. Guennebaud, B. Jacob, Eigen: a c++ linear algebra library (2014). http://eigen.tuxfamily.

org. Accessed 22 November 2014
10. M. Günther (ed.) Coupled Multiscale Simulation and Optimization in Nanoelectronics
(Springer, Berlin, Heidelberg, 2015)
11. M. Günther, U. Feldmann, CAD-based electric-circuit modeling in industry I: mathematical
structure and index of network equations. Surv. Math. Ind. 8(2), 97–130 (1999)
12. M. Günther, U. Feldmann, J. ter Maten, Modelling and discretization of circuit problems, in
Handbook of Numerical Analysis, vol. 13 (Elsevier, Amsterdam, 2005), pp. 523–659
13. C. Hachtel et al., Multirate DAE/ODE-simulation and model order reduction for coupled field-
circuit systems, in Scientific Computing in Electrical Engineering (Springer, Cham, 2018),
pp. 91–100
14. C. Hachtel et al., Multirate implicit Euler schemes for a class of differential–algebraic
equations of index-1. J. Comput. Appl. Math. 2019, 112499 (2019)
Waveform Relaxation for Low Frequency
Coupled Field/Circuit
Differential-Algebraic Models of Index 2

Idoia Cortes Garcia, Jonas Pade, Sebastian Schöps, and Caren Tischendorf

Abstract Motivated by the task to design quench protection systems for super-
conducting magnets in particle accelerators we address a coupled field/circuit
simulation based on a magneto-quasistatic field modeling. We investigate how a
waveform relaxation of Gauß-Seidel type performs for a coupled simulation when
circuit solving packages are used that describe the circuit by the modified nodal
analysis. We present sufficient convergence criteria for the coupled simulation of
FEM discretised field models and circuit models formed by a differential-algebraic
equation (DAE) system of index 2. In particular, we demonstrate by a simple
benchmark system the drastic influence of the circuit topology on the convergence
behavior of the coupled simulation.

1 Introduction

Lumped circuit models, such as modified nodal analysis (MNA), are well-
established in electrical engineering. However, they neglect the spatial dimension
and therefore distributed phenomena like the skin effect. For certain devices, this
may lead to inaccuracies of unacceptable magnitude in the simulation, e.g. for
electric machines [14] or the quench protection system of superconducting magnets
in particle accelerators [1]. These cases call for field/circuit coupling [2, 16].
To solve such coupled systems, it is often advisable to use waveform relaxation
(WR) [7], since this iterative method allows for dedicated step sizes and suitable
solvers for the different subsystems, and even for the use of proprietary blackbox
solvers. The coupled field/circuit model considered here is a DAE in the time
domain after space discretisation of the field system. It is well-known that WR

I. C. Garcia · S. Schöps
Technical University of Darmstadt, CEM Group, Darmstadt, Germany
e-mail: [email protected]; [email protected]
J. Pade () · C. Tischendorf
Department of Mathematics, Humboldt University of Berlin, Berlin, Germany
e-mail: [email protected]; [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 201
M. van Beurden et al. (eds.), Scientific Computing in Electrical Engineering,
Mathematics in Industry 36, https://doi.org/10.1007/978-3-030-84238-3_20
202 I. C. Garcia et al.

can suffer from instabilities for DAEs unless an additional contraction criterion is
satisfied [7, 12]. This work presents coupled field/circuit models, which are DAEs
of index 2 [5], for the case where WR is convergent and the case where it diverges.
Furthermore, generalizing a convergence criterion of [12], a topological and easy-
to-check criterion is provided. Finally, we present numerical simulations verifying
the topological convergence criterion.

2 Field/Circuit Model

To describe the electromagnetic (EM) field part, we consider a magnetoquasistatic

approximation of Maxwell’s equations in a reduced magnetic vector potential
formulation [4]. This leads to the curl-curl eddy current partial differential equation
(PDE). The circuit side is formulated with the MNA [6]. For the numerical
simulation of the coupled system, the method of lines is used with a finite element
(FE) discretisation. Altogether, this leads to a time-dependent coupled system of
DAE initial value problems (IVPs), described by

M ȧ + K(a)a − Xim = 0, X ȧ = vc , (1)

E(x)ẋ + f (t, x) = P im , P  x − vc = 0. (2)

The first Eq. (1) represents the space-discrete field model based on the matrices
 
(M)ij = σ ωi · ωj dV, (K(a))ij = ν(a)∇ × ωi · ∇ × ωj dV , (3)
Ω Ω

which follow from the Ritz-Galerkin approach using a finite set of Nédélec basis
functions ωi [10] defined on the domain Ω; σ denotes the space-dependent
electric conductivity and ν(a) the magnetic reluctivity that can additionally depend
nonlinearly on the unknown magnetic vector potential a. The current through the
field device is described by im . The excitation matrix is computed from a winding
density function χj modelling the j -th stranded conductor [15] as

(X)ij = χj · ωi dV . (4)
Ω

Definition 1 A function f : Rn → Rn is strongly monotone and a square matrix

M(x) is uniformly positive definite, if

∃μf : (x2 − x1 ) (f (x2 ) − f (x1 )) ≥μf x2 − x1 2 , ∀x1 , x2 ∈ Rn ,

∃μM : y  M(x)y ≥μM y2 , ∀x ∈ Rn , y ∈ Rm .

The space-discretization is supposed to meet the following properties.

WR for Field/Circuit DAEs of Index 2 203

Assumption 2 It holds (a) M is symmetric, (b) the matrix pencil λM + K is

symmetric and positive definite for λ > 0, (c) X has full column rank and (d) the
function a
→ K(a)a is strongly monotone.
The assumptions are in agreement with previous formulation in the literature, e.g.
[3, 15]. The first Assumption 2a follows naturally if a Ritz-Galerkin formulation (3)
is chosen. The second Assumption 2b will be guaranteed by appropriate boundary
and gauging conditions. Thirdly, the full column rank Assumption 2c follows from
the fact that the columns are discretisations of different coils that are located in
spatially disjoint subdomains. Finally, the monotonicity Assumption 2d follows
from the strong monotonicity of the underlying nonlinear material law, i.e. the BH-
curve [13]. In general, the field model is a multiport element such that the circuit
coupling is established via multiple currents and voltages, i.e., vector-valued im and
vc . However, for simplicity of notation we assume a two-terminal device in the
following.
The circuit Eq. (2) can be expanded into
⎛ ⎞ ⎛ ⎞ ⎛ ⎞
LC (e) 0 0 gR (e) + AL iL + AV iV + qi (t) Am
⎜ ⎟ ⎜  ⎟ ⎜ ⎟
E(x) = ⎝ 0 −L(iL ) 0⎠ , f (t, x) = ⎝ AL e ⎠ , P = ⎝ 0 ⎠
0 0 0 A
V e − qv (t) 0
(5)

using the definitions LC (e) := AC C(A  

C e)AC , gR (e) := AR g(AR e) and x =
(e, iL , iV ) where A are the usual incidence matrices and L(·), C(·) are state-
dependent square matrices describing inductances and capacitances. The position
of the field device in the circuit is described by Am , and the voltage over the device
by vc . The function g(·) describes the voltage–current relation of resistive elements
and qi , qv are the input current and voltage. Finally, x collects all node potentials
e, currents through branches with voltage sources iV and inductors iL . The circuit
system shall fulfill the following properties:
Assumption 3 It holds (a) g, C and L are Lipschitz continuous, g is strongly
monotone and C, L are uniformly positive definite,
 (b) qi and qv are continuously
differentiable, (c) AV has full column rank and AC AV AR AL has full row rank.
Assumption 3a reflects the global passivity of the respective elements [8]. Con-
sidering well-known relations between incidence matrices and circuit topology,
Assumption 3c excludes the electrically forbidden configurations of loops of
voltage sources and cutsets of current sources [5].
204 I. C. Garcia et al.

3 Waveform Relaxation and Convergence

We consider the Gauß-Seidel WR method. Applied to the coupled system (1)–(2)

for given consistent initial values, this yields the scheme

M ȧ k + K(a k )a k − Xim
k
= 0, X ȧ k = vck−1 , (6)
E(x k )ẋ k + f (t, x k ) = P im
k
, P  x k − vck = 0. (7)

The coupling variables are the current through and the voltage over the field device
im and vc , where im k is computed in (6) and is then given to (7) as input, and vice
k
versa for vc . The superscript k denotes the iteration index. A common choice for the
initial guess vc0 is constant extrapolation of the initial value.
We shall proceed as follows:
1. Lemmata 4 and 6 provide a DAE-decoupling of the EM field DAE (1) and the
MNA DAE (2), respectively.
2. Definition 5 introduces the concept of parallel CVR paths. Assuming their
existence and exploiting the previous decoupling Lemmata, Lemma 7 yields a
DAE-decoupling of the coupled WR iteration (6)–(7). Notably, it reveals the
structure of its inherent ODE, given by φ in Eq. (11).
3. The convergence Theorem 8 is a simple consequence of the previous Lemmata;
it shows that the existence of parallel CVR paths guarantees convergence of the
WR scheme (6)–(7).
For visual reasons, we shall write column vectors as (a, b, c).
Lemma 4 Let Assumption 2 hold. Then, for a given source term vc , there exists a
coordinate transformation (w, u) = T −1 a and a system of the form

u̇ + A1 u = A2 vc , w = Bu, im = G1 u + G2 vc (8)

such that (a, im ) solves Eq. (1) if and only if (u, w, im ) solves Eq. (8).
Proof For better readability and shortness we present the proof only for the slightly
more restrictive case where X M = 0, which is usually satisfied.
We equivalently transform the field DAE with new coordinates T α = a:

T  MT α̇ + T  K(T α)T α − T  Xim = 0,

(9)
X T α̇ = vc .

The transformation matrix T := (Tker X T⊥ ) is constructed such that the columns

of Tker and T⊥ form a basis of ker M ∩ ker X and (ker M)⊥ , respectively. It
is nonsingular indeed, since its construction and Assumption 2 combined with
XT M = 0 guarantee that imX ⊥ imTker and imT⊥ ⊥ im(Tker X).
WR for Field/Circuit DAEs of Index 2 205

With α = (w, u) and u = (u1 , u2 ), the transformed DAE (9) has the detailed
form
 
Tker K(T α)Tker w + Tker K(T α)(X T⊥ )u = 0,

X K(T α)T α − X Xim = 0,

T⊥ MT⊥ u̇2 + T⊥ K(T α)T α = 0,

X X u̇1 = vc .

The underlined matrices are nonsingular due to Assumption 2, and Eq. (8) is
obtained by inversion and insertion.
Definition 5 A CVR path in a circuit is a path which consists of only capacitances,
voltages sources and resistances. An element has a parallel CVR path, if its incident
nodes are connnected by a CVR path.
Lemma 6 Let Assumption 3 hold. Then, for a given source term im , there exists a
coordinate transformation (y, z1 , z2 ) = T −1 x and a system of the form

ẏ = f0 (t, y, z, z2 , u), z1 = g1 (t, y, z2 , ż2 , u), z2 = g2 (t) + QP im , (10a)

vc = P  T (y, z1 , z2 ) (10b)

with f0 , g1 , g2 uniformly globally Lipschitz continuous ∀t and g2 ∈ C 1 such that

1. (x, vc ) solves Eq. (2) if and only if (y, z1 , z2 , vc ) solves Eq. (10),
2. QP = 0 if each EM field element has a parallel CVR-path.
A detailed proof can be found in [11], where Q is shown to have the form (Q1 ∗ ∗)
with imQ1 = ker(AC AV AR ) . Hence, if each field element has a parallel CVR-
path, each column of Am can be written as a sum of columns of (AC AV AR ) and it
follows Q1 Am = 0, thus QP = 0.
Lemma 7 Let Assumptions 2 and 3 hold. If each EM field element has a parallel
CVR path, then there exists a coordinate transformation (r, s) = T −1 (a, x) and a
system of the form

ṡ k = φ(t, s k , s k−1 ), r k = ϕ(t, s k ) (11)

with φ uniformly globally Lipschitz continuous ∀t and φ, ϕ continuous such that

(a k , im
k , x k , v k ) solves Eqs. (6)–(7) if and only if (s k , r k ) solves Eq. (11).
c
206 I. C. Garcia et al.

Proof We apply Lemmata 4, 6 to the iterated subsystems (6), (7). This yields an
equivalent system

u̇k = −A1 uk + A2 vck−1 , wk = Buk , k

im = G1 uk + G2 vck−1 ,
(12)
ẏ k = f0 (t, y k , zk , z2k , uk ), z1k = g1 (t, y k , z2k , ż2k , uk ), z2k = g2 (t), (13)
vck = P  T (y k , z1k , z2k ).
(14)

Since each field element has a parallel CVR path, z2k = g(t) does not depend on uk
anymore.
We insert vck−1 = P  x k−1 = P  T (y k−1 , z1k−1 , z2k−1 ) and z1k−1 and z2k−1 therein
to obtain, with g̃1 (t, y k−1 , uk−1 ) = g1 (t, y k−1 , g2 (t), ġ2 (t), uk−1 ),

u̇k = φ2 (t, uk , y k , uk−1 , y k−1 ) := −A1 uk + A2 P  T (y k−1 , g̃1 (t, y k−1 , uk−1 ), g2 (t)).

Insertion of z1k , z2k , ż2k into f0 yields

ẏ k = φ1 (t, uk , y k ) := f0 (t, y k , g1 (t, y k , g2 (t), ġ2 (t), uk ), g2 (t), uk ).

Hence, defining s k := (uk , y k ) and φ := (φ1 , φ2 ), the sequence (uk , y k ) is given

implicitly by an ODE recursion of the form ṡ k = φ(t, s k , s k−1 ).
The algebraic constraint of Eq. (11) is obtained with r k = (wk , i k , z1k , z2k , vck ),
s = (uk , y k ) and
k

ϕ(t, s) = (Bu, Gu, g1 (t, y, g2 (t), ġ2 (t), u), g2 (t)).

Clearly, (s k , r k ) solves Eq. (11) if and only if α̃ k := (uk , wk , im

k , y k , zk , zk , v k )
1 2 c
k k
solves Eqs. (12)–(14), and α̃ solves (12)–(14) if and only if (a , im k , x k , v k ) solves
c
Eqs. (6)–(7).
We deduce the main result of this work:
Theorem 8 If each EM field element of the coupled system (1)–(2) has a parallel
CVR path, then the WR scheme (6)–(7) is uniformly convergent to the exact solution
of (1)–(2).
Proof The ODE part of Eq. (11) is a WR scheme for ODEs with Lipschitz
continuous vector field φ. It is well-known that such schemes are unconditionally
convergent on bounded time intervals [7]. The convergence of s k clearly implies
the convergence of (s k , r k ) defined by (11). Due to the equivalence provided by
Lemma 7, it follows that the original scheme (6)–(7) is convergent.
WR for Field/Circuit DAEs of Index 2 207

Remark 9 The convergence result holds for arbitrary continuous initial guesses x 0
and for bounded intervals of arbitrary size, see e.g. [7, 11].
Remark 10 The MNA decoupling given in Lemma 6 shows that g1 depends on
z2 and the derivative ż2 . Hence, the system is most sensitive to perturbations of z2 .
The input of the EM field subsystem in the WR scheme is in fact a perturbation.
Therefore, the condition QP = 0 from Lemma 6 is crucial to derive Theorem 8.
If at least one EM field element has no parallel CVR path, then QP = 0. Then,
analogously to Lemma 7 and its proof, we find ṡ k = φ(t, s k , s k−1 , ṡ k−1 ), which is
guaranteed to converge only if φ is contractive in ṡ k−1 , see [7, 11].

4 Numerical Examples

To illustrate the convergence behaviour of the WR scheme according to the derived

criteria, we consider the toy example circuits in Fig. 2a and b. Both are described
with MNA (2) and the (arbitrary) parameters R = 1Ω, L = 5H, C = 1F, is (t) =
sin(2t) + 5 sin(20t) and vs (t) = sin(t) + sin(20t) are set. The eddy current Eq. (1)
is solved on the single phase isolation transformer shown in Fig. 1. For simplicity,
a zero current is imposed on the secondary coil (dark orange) and only the primary
coil is coupled to the circuit.
The WR algorithm is applied on the simulation time window I = [0 0.8] s
and the internal time integration is performed with the implicit Euler scheme
with time step size δt = 10−2 s. The theoretical result is illustrated by the
successful simulation, see Fig. 3a, of the model shown in Fig. 2a which satisfies the
convergence criterion of Theorem 8. However, numerical simulations of the model
shown in Fig. 2b show that WR can diverge indeed if the criterion is not satisfied
(Fig. 3).

Fig. 1 Single phase isolation

transformer Primary coil
(‘MyTransformer’), see [9]
208 I. C. Garcia et al.

EM EM

G L G L
C C
qv qi qv qi

n3 n3
(a) (b)

Fig. 2 Field/circuit coupling with model from Fig. 1 (CVR path is dashed). (a) Convergent case.
(b) Divergent case

mon mon
k=1 k=1
200 k=2 200 k=2
Potential e / V
Potential e / V

100 100

0 0

0 0.2 0.4 0.6 0.8 0 0.2 0.4 0.6 0.8

time t / s time t / s
(a) (b)

Fig. 3 Monolithic (“mon”) and WR solution for k = 1, 2 iterations. (a) Convergent case. (b)
Divergent case

5 Conclusions

In this work, we have presented a space-discretised coupled field/circuit model,

which is a DAE of index 2, and a simulation of this model by means of WR.
Furthermore, we have provided an easy-to-check topological convergence criterion
for a class of coupled DAE/DAE systems of index 2.

Acknowledgments This work is supported by the ‘Excellence Initiative’ of the German Federal
and State Governments, the Graduate School of CE at TU Darmstadt and DFG grant SCHO1562/1-
2. Further, we acknowledge financial support under BMWi grant 0324019E and by DFG under
Germany’s Excellence Strategy – The Berlin Mathematics Research Center MATH+ (EXC-2046/1,
ID 390685689).
WR for Field/Circuit DAEs of Index 2 209

References

1. L. Bortot et al., STEAM: a hierarchical co-simulation framework for superconducting

accelerator magnet circuits. IEEE Trans. Appl. Super. 28(3) (2018). 4900706
2. I. Cortes Garcia et al., Optimized field/circuit coupling for the simulation of quenches in
superconducting magnets. IEEE J. Multiscale Multiphys. Comput. Tech. 2(1), 97–104 (2017)
3. I. Cortes Garcia, H. De Gersem, S. Schöps, A structural analysis of field/circuit coupled
problems based on a generalised circuit element. Numer. Algorithm. 83(1), 373–394 (2020)
4. C.R.I. Emson, C.W. Trowbridge, Transient 3d eddy currents using modified magnetic vector
potentials and magnetic scalar potentials. IEEE Trans. Magn. 24(1), 86–89 (1988)
5. D. Estévez Schwarz, C. Tischendorf, Structural analysis of electric circuits and consequences
for MNA. Int. J. Circ. Theor. Appl. 28(2), 131–162 (2000)
6. C.-W. Ho, A.E. Ruehli, P.A. Brennan, The modified nodal approach to network analysis. IEEE
Trans. Circ. Syst. 22(6), 504–509 (1975)
7. E. Lelarasmee, A.E. Ruehli, A.L. Sangiovanni-Vincentelli, The waveform relaxation method
for time-domain analysis of large scale integrated circuits. IEEE Trans. Comput. Aided Des.
Integr. Circ. Syst. 1(3), 131–145 (1982)
8. M. Matthes, Numerical Analysis of Nonlinear Partial Differential-Algebraic Equations: A
Coupled and an Abstract Systems Approach. Dissertation, Universität zu Köln, 2012
9. D. Meeker, Finite Element Method Magnetics, version 4.2 (25feb2018 build) edition, 2018
10. P. Monk, Finite Element Methods for Maxwell’s Equations (Oxford University Press, Oxford,
2003)
11. J. Pade, Analysis and waveform relaxation for a differential-algebraic electrical circuit model.
Dissertation, Humboldt University of Berlin, 2021
12. J. Pade, C. Tischendorf, Waveform relaxation: a convergence criterion for differential-algebraic
equations. Numer. Algorithm. 81, 1327–1342 (2019)
13. C. Pechstein, Multigrid-Newton-methods for nonlinear-magnetostatic problems. Master’s
thesis, Universität Linz, 2004
14. S.J. Salon, Finite Element Analysis of Electrical Machines (Kluwer, Boston, 1995)
15. S. Schöps, Multiscale Modeling and Multirate Time-Integration of Field/Circuit Coupled
Problems (VDI Verlag. Fortschritt-Berichte VDI, Reihe, 2011)
16. S Schöps, H. De Gersem, A. Bartel, A cosimulation framework for multirate time-integration
of field/circuit coupled problems. IEEE Trans. Magn. 46(8), 3233–3236 (2010)
Splitting Methods for Linear Circuit
DAEs of Index 1 in port-Hamiltonian
Form

Malak Diab and Caren Tischendorf

Abstract Operator splitting is a powerful method for numerical investigation of

complex models. This method was successfully used for ordinary and partial differ-
ential equations (ODEs and PDEs). In constrained dynamical problems as electric
circuits or energy transport networks, differential-algebraic equations (DAEs) arise.
The constraints prevent a simple transfer of operator splitting from ODEs to DAEs.
Here, we present an approach for splitting linear circuit DAEs of index 1 based
on a port-Hamiltonian modeling that we derive from loop and cutset equations by
a topological decoupling. Finally, we present convergence results for the proposed
DAE operator splitting.

1 Introduction

The idea of operator splitting methods is based on the splitting of a complex problem
into a sequence of simpler sub-problems. Usually, one exploits some structural
properties of the separated operators belonging to the sub-problems, for example,
the linear behavior, the symmetric behavior or the stiff behavior that allows the
application of efficient integration methods to the sub-problems, see for instance
[8, 10–12]. For dynamical problems like ODEs or parabolic PDEs, additive operator
splitting are well established and appropriate. However, for constrained problems an
additive operator splitting method would usually fail. This becomes obvious when
comparing the simple problems

u = Au = A1 u + A2 u and Ax = A1 x + A2 x = b.

M. Diab () · C. Tischendorf

Humboldt Universität zu Berlin, Institute of Mathematics, Berlin, Germany
e-mail: [email protected]; [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 211
M. van Beurden et al. (eds.), Scientific Computing in Electrical Engineering,
Mathematics in Industry 36, https://doi.org/10.1007/978-3-030-84238-3_21
212 M. Diab and C. Tischendorf

Solving u = A1 u and afterwards using its solution as an initial condition to solve

u = A2 u yields an approximate solution of u = Au, while solving Ax = b in the
same manner, does not make sense. Here, a multiplicative splitting Ax = A1 A2 x
would be appropriate to solve Ax = b by A1 y = b and afterwards A2 x = y. It
shows us that there is no simple extension of operator splitting for ODEs to DAEs.
One has to adapt the operator splitting for DAEs to the different nature of inherent
DAE parts.
The next section describes the branch oriented circuit modeling. It provides a
natural decoupling of the circuit DAEs that can be exploited for a suitable operator
splitting. Section 3 describes our operator splitting approach for linear circuit DAEs
of index 1 [15]. It includes a convergence analysis and a discussion of some
structural properties of the subsystems (for instance Hamiltonian structure of the
first subsystem). Finally we demonstrate numerical results for a benchmark circuit
in Sect. 4.

2 Circuit Modeling

In contrast to standard circuit modeling using the modified nodal analysis [9] we
consider the branch oriented loop-cutset modeling [3, 4]. It allows us to split the
operators in a natural way exploiting physical properties.
For a given circuit graph G with n node and b branches, select any tree and
remove all its links. Then replace each link once at a time, it will form a loop that is
called as fundamental loop. We select an orientation of the loop to coincide with that
of the link completing it. On the other hand, a fundamental cutset with reference to
a tree is a cutset formed with one tree branch and remaining links. The orientation
of a cutset is the same of that of the tree branch.
Definition 1 The fundamental loop matrix B ∈ Rb−(n−1)×b is defined by its entries
⎧
⎪
⎪
⎨1, if the branch j has the same orientation of fundamental loop i
bij = −1, if the branch j has the opposite orientation of fundamental loop i
⎪
⎪
⎩0, else.

Lemma 1 (Loop Equations, KVL [3]) Let v be the vector of branch voltages in
an electric network, then we have

Bv = 0 (1)

In general, matrix B is arranged such that the first columns correspond to entries of
links and the columns correspond to entries of tree branches, therefore
   
B = Bl Bt = I Bt
Splitting Methods for Linear Circuit DAEs of Index 1 in port-Hamiltonian Form 213

Definition 2 The fundamental cutset matrix Q ∈ R(n−1)×b is defined by its entries

⎧
⎪
⎪
⎨1, if the branch j has the same orientation of cutset i
qij = −1, if the branch j has the opposite orientation of cutset i
⎪
⎪
⎩0, else.

Lemma 2 (Cut-set Equations, KCL [3]) Let i be the vector of branch currents in
an electric network, then we have

Qi = 0 (2)
   
and similar to the columns re-arrangement of B, we get Q = Ql Qt = Ql I .
Theorem 1 (Orthogonality Relation [3]) For a given connected graph G , the
orthogonality relation between the fundamental loop matrix B and the fundamental
cutset matrix Q is given by BQ = 0.
The circuit equations consist of the loop equations (1) and cutset equations (2)
reflecting the Kirchhoff’s laws together with elements constitutive equations

iC = CvC , vL = LiL , iG = GvG , vR = RiR , iI = is (t), vV = vs (t).

(3)

For simplicity, we consider only RLC circuits since our focus is to demonstrate the
new splitting approach. We assume that all resistances, conductances, capacitances
and inductances show a globally passive behavior, i.e. their corresponding matrices
R, G, C and L are positive definite. In addition, the independent functions vs and
is for voltage and current sources are assumed to be continuously differentiable.
Notice that, we used in our approach the conductive description for all resistances
that belong to the tree and the resistive description for all resistances that does not
belong to the tree, see below.
An index-1 circuit DAE models a circuit network that does neither have an LI-
cutset nor a CV-loop, see [5]. Then we can construct a tree as follows [14]:
1. All capacitive elements and voltage sources belong to the tree.
2. All inductive elements and current sources do not belong to the tree.
3. Split resistors in such a way that all G-resistances belong to the tree and all R-
resistances do not belong to the tree.
214 M. Diab and C. Tischendorf

Then, the loop and cutset equations have the form

⎛ ⎞ ⎛ ⎞ ⎛ ⎞ ⎛ ⎞
vL vC iL iC
⎝vR ⎠ + Bt ⎝vG ⎠ = 0, Ql ⎝iR ⎠ + ⎝iG ⎠ = 0.
vI vV iI iV

Inserting the element constitutive equations we get the DAE system

⎛ ⎞ ⎛ ⎞ ⎛ ⎞ ⎛  ⎞
LiL vC iL CvC
⎝RiR ⎠ + Bt ⎝ vG ⎠ = 0, ⎝
Ql iR ⎠ + GvG ⎠ = 0.
⎝
vI vs (t) is (t) iV

Notice that Ql = −Bt due to Lemmas 1, 2 and Theorem 1. Introducing

⎛ ⎞ ⎛ ⎞
BLC BLG BLV QCL QCR QCI
Bt =: ⎝BRC BRG BRV ⎠ , Ql =: ⎝QGL QGR QGI ⎠
BI C BI G BI V QV L QV R QV I

and reordering the equations, we obtain a system of the port-Hamiltonian form

Dx  (t) + J x(t) + My(t) = rx (t) (4a)

−M T x(t) + Sy(t) = ry (t) (4b)
z(t) + Kx x(t) + Ky y(t) = rz (t) (4c)
i  i   vI 
with x = vC , y = vG , z = iV ,
L R

L    
0 BLC 0 BLG R BRG
D= 0
0 C , J = QCL 0 , M= QCR 0 , S= QGR G

and
    
0 BI C 0 BI G BLV vs BRV vs BI V vs
Kx = QV L 0 , Ky = QV R 0 , rx = − QCI is , ry = − QGI is , rz = − QV I is .

Notice that J is skew-symmetric since BLC = −Q

CL . Furthermore,
& ' & '
R 0 0 BRG
S = S1 + S2 := +
0 G QGR 0

with the positive definite diagonal matrix S1 and the skew-symmetric matrix S2
since BRG = −Q GR . Consequently, S is not symmetric (unless BRG = 0) but
positive definite and hence non-singular. Furthermore, we see that system (4) is a
port-Hamiltonian DAE in the sense of the definitions given in [6] and [13]. For
[13], one can choose x̃ = (x, y), z̃(x̃) = x̃, ỹ := −z and ũ = (is , vs ) where the
Splitting Methods for Linear Circuit DAEs of Index 1 in port-Hamiltonian Form 215

tilde notation refers to the variables in [13]. For [6], one can choose the space V =
{(x, y, z) : z + Kx x + Ky y = rz } with x̄ = (x, y, z), z̄(x̄) = (x, y), ȳ := B̄  z̄(x̄)
and ū = (is , vs ), where the bar notation refers to the variables in [6]. Since (4c)
can be interpreted as output equation for z, we consider only the reduced DAE
system (4a)–(4b) in the following.

3 Operator Splitting for Index-1 Circuit DAEs

Regarding the fact that additive splitting makes no sense for solving the con-
straints (4b), we propose a splitting approach based on the inherent ODE. Therefore,
we rewrite the DAE system (4a)–(4b) equivalently as

Dx  + J x + MS −1 M  x = rx (t) − MS −1 ry (t) (5a)

 
y = S −1 ry (t) + M  x . (5b)

We split (5a), using Lie-Trotter splitting, into the subsystems Dx  + J x = 0 and

Dx  + MS −1 M T x = rx (t) − MS −1 ry (t). (6)

Next, we reformulate (6) with (5b) back as DAE and obtain the following splitting
approach (SADAE) for circuit index-1 DAEs.
1. Initialize x2 (t0 ) := x0 and n = 0.
2. Solve on [tn , tn+1 ] the first subsystem

Dx1 + J x1 = 0, x1 (tn ) = x2 (tn ) (splitDAE 1)

3. Solve on [tn , tn+1 ] the second subsystem

Dx2 (t) + My(t) = rx (t), x2 (tn ) = x1 (tn+1 ) (splitDAE 2a)

−M T x2 (t) + Sy(t) = ry (t). (splitDAE 2b)

4. Set n = n + 1 and go to 2. unless tn is the final time point.

3.1 Subsystem Properties

The first subsystem (splitDAE 1) is in fact a Hamiltonian ODE system with the
Hamiltonian
1  1 1
H (x) = x Dx = vC CvC + iL LiL =: H (vC , iL ) (7)
2 2 2
216 M. Diab and C. Tischendorf

describing the total energy stored in the capacitors and inductors. We have

d
H (x) = x  Dx  = −x  J x = 0
dt
since J is skew-symmetric. Obviously, H is a quadratic form. Consequently, we can
apply symplectic numerical methods to (splitDAE 1). They have the advantage to
preserve the total energy H stored in the capacitors and inductors [7].
The second subsystem (splitDAE 2a)–(splitDAE 2b) leads to non-symmetric but
positive definite linear systems after time discretization that allows the exploitation
of suitable iterative methods [2].

3.2 Convergence Analysis

In order to verify the convergence of DAE operator splitting method, one has to rely
on the convergence of the ODE operator splitting method. For this reason, we define
the non-homogeneous Cauchy problem

u (t) = A1 u(t) + A2 u(t) + r(t), u(t0 ) = u0 (8)

where the initial condition u0 and the source function r are bounded. Let Δt denotes
the time step such that the following stability condition is satisfied

||eΔt (A1 +A2 ) || ≤ 1, ||eΔt A1 || ≤ 1, and ||eΔt A2 || ≤ 1

After time discretization, apply the following operator splitting algorithm (OSA)

u1 (t) = A1 u1 (t), t ∈ [tn , tn+1 ] and u1 (tn ) = unsp
u2 (t) = A2 u2 (t) + r(t), t ∈ [tn , tn+1 ] and u2 (tn ) = u1 (tn+1 )

where u0sp = u0 , and the splitting solution at t = tn+1 is un+1

sp = u2 (tn+1 ).

Theorem 2 (See [1]) Under the boundedness and stability conditions formulated
above, the approximated splitting solution obtained from the operator splitting
algorithm (OSA) converges to the exact solution of the ODE (8).
If we denote by T (tn ) the solution operator of (8) at the n-th time step, and by Ts (tn )
the splitting solution operator, then we have: ||T (tn )u0 −Ts (tn )u0 || → 0 as Δt → 0.
Regarding the equivalence of the DAE system (splitDAE 2a)–(splitDAE 2b) to the
system

Dx2 + MS −1 M  x2 = rx (t) − MS −1 ry (t) (9a)

 
y = S −1 ry (t) + M  x2 . (9b)
Splitting Methods for Linear Circuit DAEs of Index 1 in port-Hamiltonian Form 217

we may directly conclude the following theorem from Theorem 2 (choosing A1 =

−D −1 J and A2 = −D −1 MS −1 M  and r = D −1 (rx − MS −1 ry )).
Theorem 3 Let the time stepsize Δt be sufficiently small, the initial currents and
voltages as well as the source functions of current and voltage sources be bounded.
Then, the approximated solution of the circuit DAE operator splitting approach
(SADAE) on page 215 converges to the exact solution of the DAE (4a)–(4b).

4 Numerical Simulation

We use a small RLC circuit example in order to demonstrate the operator splitting
approach for DAEs. It operates in a GHz regime as often used in chip design.
Using the tree in Fig. 1, we get for the circuit DAE system (4a)–(4c) the matrices
⎛ L1 0 0 0 0 0 ⎞ ⎛ ⎞ ⎛1
0 0 0 1 1 1 0 ⎞ ⎛ −vs ⎞
0 L2 0 0 0 0 0 0 0 0 −1 0 0 1 0
⎜ 0 0 L3 0 0 0 ⎟ ⎜ 0 −1 ⎟ ⎝0 −1 ⎠ ,
D=⎝ 0 0 0 C1 0 0 ⎠ , J = ⎝ −1
0
0
0
0
0
0
0
0 0 ⎠, M= 0 0 rx = ⎝ 0
0
⎠
0 0 0 0 C2 0 −1 1 0 0 0 0 0 0 0
0 0 0 0 0 C3 −1 0 1 0 0 0 0 0 0

and
& ' & '
G1 0   0
S= , Kx = −1 0 0 0 0 0 , Ky = 0, ry = , rz = 0.
0 G2 0

For comparison, we consider the following three variants of numerical simulation

of the circuit:
1. Solve (4a)–(4c) by implicit Euler method.
2. Solve (splitDAE 1) and (splitDAE 2a)–(splitDAE 2b) by implicit Euler method
3. Solve (splitDAE 1) by symplectic Euler and (splitDAE 2a)–(splitDAE 2b) by
implicit Euler method
In Fig. 2 we see the reference solution computed by time stepsize h = 1e − 13
and the error between the numerical solution for the three simulation variants with

C1
R1 L1

C2 L2

V
R2
C3 L3

Fig. 1 Benchmark RLC-circuit. The dashed branches form the tree considered for the model
equations
218 M. Diab and C. Tischendorf

Fig. 2 Reference solution for inductive currents for circuit in Fig. 1(left). Error for numerical
solution of the three simulation variants with time stepsize h = 1e − 11 (right)

time stepsize h = 1e − 11 and the reference solution. The results show that the
solution of the DAE splitting approach (variant 2) is almost the same as for the non-
splitted solution (variant 1). It means that the error caused by splitting is neglectable
in comparison with the numerical discretization error. The use of the DAE splitting
approach with the symplectic Euler method (variant 3) gives the best results and is
even faster than the other variants since the symplectic Euler method for the first
subsystem (5a) is an explicit method.

5 Conclusions and Outlook

In this paper, we extended the operator splitting method from ODEs to circuit
linear DAEs. Followed by the topological decoupling of circuit DAEs of index 1
in loop-cutset formulation, we were able to construct a suitable decomposition of
the matrices so that a natural port-Hamiltonian DAE structure is visible and can be
exploited for a convergent splitting approach that is explicit and energy preserving
in the dynamic part.

Acknowledgments This project has received funding from the European Union’s Horizon 2020
research and innovation program under grant agreement No 76504. Furthermore, we acknowledge
financial support by DFG under Germany’s Excellence Strategy – The Berlin Mathematics
Research Center MATH+ (EXC-2046/1, ID 390685689).

References

1. M. Bjórhus, Operator splitting for abstract cauchy problems. IMA J. Numer. Anal. 18, 419–443
(1988)
2. A.T. Chronopoulos, s-step iterative methods for (non)symmetric (in)definite linear systems.
SIAM J. Numer. Anal. 28(6), 1776–1789 (1991)
Splitting Methods for Linear Circuit DAEs of Index 1 in port-Hamiltonian Form 219

3. L.O. Chua, C.A. Desoer, E.S. Kuh, Linear and Nonlinear Circuits (McGraw-Hill, Singapore,
1987)
4. C.A. Desoer, E.S. Kuh, Basic Circuit Theory. International student edition (McGraw-Hill, New
York, 1984)
5. D. Estévez Schwarz, C. Tischendorf, Structural analysis of electric circuits and consequences
for MNA. Int. J. Circ. Theory Appl. 28(2), 131–162 (2000).
6. M. Günther, A. Bartel, B. Jacob, T. Reis, Dynamic iteration schemes and port-Hamiltonian
formulation in coupled circuit simulation, 2020. arXiv:2004.12951v1
7. E. Hairer, G. Wanner, C. Lubich, Geometric Numerical Integration, vol. 31. Springer Series
in Computational Mathematics (Springer, Berlin, Heidelberg, 2002)
8. E. Hansen, A. Ostermann, Dimension splitting for quasilinear parabolic equations. IMA J.
Numer. Anal. 30(3), 857–869 (2010)
9. C.-W. Ho, A. Ruehli, P. Brennan, The modified nodal approach to network analysis. IEEE
Trans. Circ. Syst. 22(6), 504–509 (1975)
10. M. Hochbruck, T. Jahnke, R. Schnaubelt, Convergence of an adi splitting for maxwell’s
equations. Numer. Math. 129(3), 535–561 (2015)
11. H. Holden, K. Karlsen, K. Lie, N. Risebro, Splitting Methods for Partial Differential Equations
with Rough Solutions (European Mathematical Society, Zürich, 2010)
12. W. Hundsdorfer, J.G. Verwer, A note on splitting errors for advection-reaction equations. Appl.
Numer. Math. 18(1), 191 – 199 (1995)
13. V. Mehrmann, R. Morandin, Structure-preserving discretization for port-Hamiltonian descrip-
tor systems (2019) arXiv:1903.10451v1
14. R. Riaza, Differential-Algebraic Systems: Analytical Aspects and Circuit Applications (World
Scientifc, Singapore, 2008)
15. C. Tischendorf, R. Lamour, R. März, Differential-Algebraic Equations. A Projector Based
Analysis (Springer, Hamburg, 2012)
Reduced Order Modelling for Wafer
Heating with the Method of Freezing

E. J. I. Hoeijmakers, H. Bansal, T. M. van Opstal, and P. A. Bobbert

Abstract Accurate and real-time temperature control for wafer heating is one of
the main challenges in semiconductor manufacturing processes. With reduced-order
modelling (ROM), the computational complexity of the mathematical model can be
decreased in order to solve the model quickly at a low computational cost, while
still maintaining the computational accuracy. However, the translating temperature
profile, due to moving sources, render the standard reduction approaches to be
ineffective. We propose to invoke the concept of the “Method of Freezing” and
use it in conjunction with the standard ROM approaches to obtain an effective low-
complexity model. We finally assess the effectiveness of the proposed approach on
the 2-dimensional heat equation with moving heat loads. Numerical results clearly
show the potential of the proposed approach over the standard one in terms of
computational accuracy and the dimension of the resulting reduced-order model.

1 Introduction

In photolithography, feature sizes are decreasing in effort for manufacturers to keep

up with Moore’s law. This has prompted the use of higher energy lasers, leading
to more wafer heating and, therefore, more thermal expansion. Accurate and real-
time prediction of the temperature distribution around the moving laser beam is
a necessity as this facilitates to correct the laser beam trajectory and to create

E. J. I. Hoeijmakers () · P. A. Bobbert

Department of Applied Physics, Technical University of Eindhoven, Eindhoven, Netherlands
e-mail: [email protected]
H. Bansal
Department of Mathematics and Computer Science, Technical University of Eindhoven,
Eindhoven, Netherlands
e-mail: [email protected]
T. M. van Opstal
Sioux Mathware, Eindhoven, Netherlands
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 221
M. van Beurden et al. (eds.), Scientific Computing in Electrical Engineering,
Mathematics in Industry 36, https://doi.org/10.1007/978-3-030-84238-3_22
222 E. J. I. Hoeijmakers et al.

the desired temperature at every place on the wafer [1]. However, this remains a
challenge since standard numerical methods take a lot of computational time, and
the increased resolution requirements due to the reduced feature sizes slow the
model down.
Reduced-order modelling (ROM) reduces the model complexity and aids in real-
time prediction of the quantity of interest. Translating temperature profiles, due to
moving sources, render the standard ROM approaches ineffective [2]. Hence, we
propose to invoke the “Method of Freezing” along with standard ROM approaches
in order to obtain an effective low-complexity model computable in real-time.
The concept of the “Method of Freezing” has been applied on parabolic and
hyperbolic problems in the past [3]. However, [4] is the only work which so far
exploits the “Method of Freezing” for non-linear reduced basis approximations.
This work considers a numerical experiment, which falls in the realm of hyperbolic
problems, namely the parameterized Burgers-type problem in 2D (without source
terms).
The main contribution of this work is to use the “Method of Freezing” in
conjunction with standard ROM approaches to facilitate accurate and real-time
prediction of the temperature. The “Method of Freezing” relies on an ansatz
that decomposes the original dynamics into shape and travelling dynamics. The
resulting shape dynamics is amenable for an efficient basis generation. We then
use these generated bases to apply Proper Orthogonal Decomposition (POD) on
the shape dynamics and, ultimately, obtain a reduced-order model. We finally
assess the performance of the combined approach of the “Method of Freezing” and
reduced basis approximations on a test-case of practical relevance, and discuss the
computational merits of the proposed ROM approach over the standard one.
The paper is organized as follows. In Sect. 2.1, we introduce the 2-dimensional
heat equation and discuss the numerical method for its discretization. We invoke the
idea of the “Method of Freezing”, reformulate the model problem and present the
corresponding discretized representation in Sect. 2.2. A Galerkin-type projection-
based ROM is performed on a semi-discrete model representation in Sect. 3. A
numerical case-study is presented in Sect. 4 to showcase the effectiveness of the
proposed approach. Finally, Sect. 5 ends with conclusions and future works.

2 Theory

In this section, we first introduce the model and the numerical method employed
for the spatio-temporal discretization. We then introduce the idea of the “Method of
Freezing” and present a model reformulation and its discrete representation.
Reduced Order modelling for Wafer Heating with the Method of Freezing 223

2.1 Model Introduction

To model the wafer heating, we use the well-known heat equation in two-
dimensions. As the height of the wafer is one order of magnitude less than the
length and the width of the wafer, the temperature gradient along the thickness
of the wafer is very small. This makes the 2-dimensional heat equation a good
approximation of the real situation. The 2-dimensional heat equation is governed
by:
& '
∂u ∂ 2u ∂ 2u
−α + 2 = Q(x, y, t), (x, y) ∈ Ω, t ∈ [0, tf ], (1)
∂t ∂x 2 ∂y
u(x, y, t = 0) = u0 , (2)
∂u ∂u
nx + ny = 0 on ∂Ω, (3)
∂x ∂y

where u represents the wafer temperature, u0 stands for a constant initial tempera-
ture, Ω stands for the spatial domain of interest, n = (nx , ny ) denotes the normal
to the boundary ∂Ω, tf indicates the final simulation time, and α is the thermal
diffusivity constant. The thermal diffusivity constant can be expressed with the
thermal conductivity k, the specific heat capacity Cp and the density ρ of the wafer
in the form α = ρCk p . Here, a moving heat load Q(x, y, t) is assumed to be of the
non-affine form:
 2 
x−cx t y−cy t 2
− 12 − 12
Q(x, y, t) = e σx σy
, (4)

where cx and cy are the speeds of the heat load in the x- and y-direction, respectively
and, the variance of the Gaussian distribution along the x- and the y-direction is
given by σx2 and σy2 , respectively.
After multiplying (1) by a smooth test-function w, integrating over the domain
and invoking Green’s theorem, a weak formulation of the 2-dimensional heat
equation can be constructed, resulting in:
 &  '  
∂u ∂u ∂w ∂u ∂w ∂u
wdA + α · dA + · dA − α wds = QwdA,
Ω ∂t Ω ∂x ∂x Ω ∂y ∂y ∂Ω ∂n Ω
(5)

where dA = dxdy and ds is a boundary surface element. Using (3), the fourth term
on the left-hand-side of (5) cancels out[5].
In order to solve (5) numerically, discretization in space and time is necessary.
We discretize the domain such that the structured mesh aligns with the orientation
of the features which need to be printed. We then employ a finite element method to
discretize in space. We approximate the solution with a summation over B-spline
224 E. J. I. Hoeijmakers et al.

N
basis-functions φi , u = i=1 ui (t)φi (x) [6]. Here, N is the number of finite
elements used in the domain discretization and ui is the weight of every basis
function. To discretize in time, the first-order backwards Euler method is applied
as is also used in Chap. 8 of [5]. Discretizing in both space and time results in the
following equation:

Muk+1 + ΔtαDuk+1 − Δt Q̃k+1 = Muk , (6)

where M is the mass matrix, D is the diffusion matrix, Q̃ is the source vector
representative of the moving heat loads and Δt indicates the time-step. Equation (6)
needs to be solved for every time instant k + 1.
The numerical solution will be at most first-order accurate if the first-order
backwards Euler method is applied in conjunction with the higher-order spatial
discretization. However, in this paper, we are not concerned about the order of
accuracy of the numerical solution, but intend to show the potential of the “Method
of Freezing”. To this end, the first-order temporal discretization is representative
enough for quantifying the numerical performance, while being simple to imple-
ment. The implementation of a higher-order temporal discretization is deferred to
future works.
We will now discuss a change of coordinates or so-called “Method of Freezing”
that we propose to use in conjunction with standard ROM techniques to obtain an
effective complexity reduction for problems with moving heat load(s).

2.2 Model Reformulation: Method of Freezing

The “Method of Freezing” maps all symmetry-related solutions to a single class

of solutions. This method separates the dynamics in the group direction from the
dynamics in the remaining directions of the phase space. The general idea of this
method is to perform a coordinate transformation of the form:

u(x, y, t) = v(x − cx t, y − cy t, t) = v(ξx , ξy , t), (7)

Incorporating (7) in (1) results in the following modified heat equation:

* +
∂v ∂v ∂v ∂ 2v ∂ 2v
− cx − cy −α + = Q(ξx , ξy ). (8)
∂t ∂ξx ∂ξy ∂ξx2 ∂ξy2

This modified heat equation is quite similar to the original equation given in (1),
except the additional second and third term on the left-hand side which represent an
Reduced Order modelling for Wafer Heating with the Method of Freezing 225

extra convection term. The weak formulation of (8) under zero Neumann boundary
conditions is given by:
  
∂v ∂v ∂v
wdξx dξy − cx wdξx dξy − cy wdξx dξy
Σ ∂t Σ ∂ξx Σ ∂ξy
&  ' 
∂v ∂w ∂v ∂w
−α dξx dξy + dξx dξy = Q(ξx , ξy )wdξx dξy ,
Σ ∂ξx ∂ξx Σ ∂ξy ∂ξy Σ
(9)
where Σ represents the transformed domain as per the coordinate transformation.
Discretizing (9) in space and time yields:

Mv k+1 + αΔtDv k+1 − ΔtCv k+1 − Δt Q̃k+1 = Mv k , (10)

where M and D are, respectively, the mass and diffusion matrix, and C is the
convection matrix.
Although we consider constant cx and cy , the “Method of Freezing” can handle
time-dependent speeds by adding an ingredient known as phase conditions; see [3].

3 Reduced Order Modelling

In this section, we build a reduced-order model both via the standard and the
proposed ROM approach. The standard and the proposed ROM approach, built upon
a Galerkin type projection-based ROM methodology [7], is discussed in Sects. 3.1
and 3.2, respectively.

3.1 Standard Reduced Order Modelling Approach

The numerical solution of the 2-dimensional heat equation can be written as a u-

snapshot matrix, where every column k represents the solution at the k-th time-
step. Upon performing singular value decomposition (SVD) on the snapshot matrix
composed of u, a projector P T : Uh → Ur is obtained and further used to build
a reduced-order model. Here, Uh is a h-dimensional high-fidelity space and Ur is
a r-dimensional reduced space spanned by the functions obtained from a truncated
singular value decomposition of the u snapshot matrix. The standard reduced-order
model is given by:

Mred uk+1 k+1

red + αΔtDred ured − P Δt Q̃
T k+1
= Mred ukred , (11)

where Dred = P T DP and Mred = P T MP are the reduced diffusion and mass
matrices, respectively.
226 E. J. I. Hoeijmakers et al.

3.2 Proposed Reduced Order Modelling Approach

The proposed novel ROM approach employs the “Method of Freezing” in conjunc-
tion with standard projection-based reduction techniques. We again employ SVD.
However, in this proposed framework, the SVD is performed on the v snapshot
matrix, instead of the u snapshot matrix. We now obtain a projector LT : Vh →
Vr where Vh is a h-dimensional high-fidelity space and Vr is a r-dimensional
reduced space spanned by the functions obtained from a truncated singular value
decomposition of the v snapshot matrix. Finally, the proposed (frozen) reduced-
order model is:
k+1 k+1 k+1
Mred,p vred + αΔtDred,p vred − ΔtCred,p vred − LT Δt Q̃k+1 = Mred,p vred
k
,
(12)

where Cred,p = LT CL represents the reduced matrix corresponding to the extra

convection term, and, Mred,p = LT ML and Dred,p = LT DL, respectively,
represent the reduced mass and diffusion matrices.

4 Numerical Results

In this section, we numerically test the proposed (Freezing-POD) approach and

show its effectiveness as a reduced-order modelling technique.
Let the domain of the wafer be given by Ωd = [−0.01, 0.02]m×[−0.02, 0.04]m.
The wafer is subdivided into 9 smaller rectangular sub-domains and each sub-
domain has the dimensions 1 by 2 cm. The heat load will move around one of
these sub-domains in practice. In order to not consider the boundary conditions
close to the boundary edges of the wafer, we consider that the laser only moves
over the middle sub-domain Ω, i.e., Ω = [0, 0.01]m × [0, 0.02]m. Motivated by
the application, we consider u0 in (2) to be equal to the room temperature, i.e.,
u0 = 298K. Furthermore, we spatially discretize a rectangular sub-domain by a
20 × 20 mesh, i.e., 400 finite-elements. Moreover, we consider a silicon wafer with
thermal diffusivity constant α = 8.8 · 10−5 m2 /s [8]. Additionally, we assume that
the laser has a surface of approximately 2 by 20 mm. As a result, the variance in
the x-direction, σx2 , is 0.002 m, and the variance in the y-direction, σy2 , is 0.02 m.
We take 50 steps in time for the scenario under consideration, i.e., t ∈ [0, 0.05]s
with a time step of 0.001s. A laser is considered to move along the x-direction with
a speed of 0.2 m/s for first 25 time steps, i.e., cx = 0.2 m/s and cy = 0 m/s for
t = [0, 0.025]s and along the y-direction with a speed of 0.2 m/s for next 25 time
steps, i.e., cx = 0 m/s and cy = 0.2 m/s for t = (0.025, 0.05]s.
We build the snapshot matrix composed of solution u obtained in (6) and another
snapshot matrix composed of shape dynamics v obtained in (10). We then perform
SVD on these snapshot matrices to obtain the corresponding singular values, whose
Reduced Order modelling for Wafer Heating with the Method of Freezing 227

Fig. 1 Singular value decay behavior for the proposed and the standard approach

decay behavior is known to give a good expectation about the possible reduction
in the dimensionality of the full-order model. In Fig. 1, the singular value decay
behavior for the proposed and the standard ROM approach is shown. It can be
observed that incrementing the number of POD modes by one yields a sharp
initial decrease in the singular values both for the proposed and the standard ROM
approach. However, post the sharp decay, we can see that the singular values
corresponding to the proposed approach decay faster than the one corresponding to
the standard approach. An initial sharp decrease is attributed to the fact that only a
single mode is representative enough to capture the mean temperature on the silicon
wafer. Other modes are required to accurately determine the change (with respect
to the mean) in the temperature due to the moving heat loads. The observed decay
behavior clearly indicates a possibility of an effective dimensionality reduction if
the “Method of Freezing” is used together with the standard ROM techniques.
Further computational benefits of the proposed approach over the standard one
can be clearly seen in Fig. 2, which shows the behavior of the reduced-order
modelling (ROM) error for increasing dimensions of the reduced-order model.
We assess the error of the standard and proposed approaches in the (absolute)
L2 -norm in space and time. The error via the standard approach corresponds
to the difference between the finite-element based numerical solution u and the
reconstructed solution obtained by lifting the standard reduced-order solution
ured , obtained in (11), to the high-dimensional problem space. And, the error via
228 E. J. I. Hoeijmakers et al.

Fig. 2 ROM error for the proposed and the standard approach versus varying dimensions of the
reduced-order model

the proposed approach corresponds to the difference between the finite-element

based numerical solution u and the reconstructed solution obtained by lifting and
shifting the reduced-order solution vred obtained in (12). It is clearly observable
that the (absolute) ROM error incurred upon using the proposed approach for
varying dimensions of the reduced-order model is lower than the error incurred
while using the standard approach. The proposed approach is expected to give a
lower ROM error as the shape dynamics is essentially localized around the initial
configuration. In principle, the ROM error is a function of the spatial and the
temporal discretization error. Given the fact that the shape dynamics is essentially
localized in the proposed approach, the amount of temporal discretization error is
significantly less than that obtained in the standard approach. We also claim that the
larger time-step size can be used to advance the reduced-order model built using the
“Method of Freezing” compared to the admissible time-step size in the standard
ROM framework. This claim is supported by the fact that the time-step size is
generally controlled by the CFL restrictions, which are dictated by the time-scale of
the problem. As a consequence of localized shape dynamics, the time-step size is not
too severely restricted in the proposed approach as in the standard approach, which
also eventually aids in temporal complexity reduction. As a result, the dimension
of the reduced-order model obtained by using the proposed approach will be much
Reduced Order modelling for Wafer Heating with the Method of Freezing 229

smaller than the counterpart obtained using the standard reduction approach in order
to have the same accuracy.

5 Conclusion and Future Outlook

We have proposed to employ the concept of the “Method of Freezing” in conjunction

with a Galerkin-type projection based methodology in order to overcome the limita-
tion of the standard projection-based reduced-order modelling (ROM) techniques
in dealing with moving heat loads. We have demonstrated the performance of
the proposed approach on a test-case of practical relevance that encompasses the
movement of the laser beam along both dimensions of the wafer.
This work focused on reproducing the results of the time-dependent heat equation
via standard and proposed ROM approaches. This reproduction step is essential
before attempting to develop a parametric reduced-order model as we cannot hope
to have an effective low-complexity reduced-order model if the numerical approach
does not fare well in the reproduction step. Furthermore, it should be emphasized
that the considered model is non-affine due to the nature of the moving heat load(s),
and that the projection alone is not sufficient to reduce the costs of constructing a
reduced-order model for such non-affine (and non-linear) problems. Moreover, there
might be other sources of non-affine and/or non-linear nature, such as radiative heat
fluxes, temperature-dependence of parameters, etc. These non-affine and non-linear
problems can be effectively dealt with the proposed ROM approach by using an
additional concept of hyper-reduction introduced in [9].
Future works will deal with a modification to the idea of the “Method of
Freezing” to eventually obtain a suitable decomposition ansatz that accounts for
the physical boundary conditions. In addition, the effectiveness of the proposed
approach will be investigated in terms of the computational speed-up. Moreover,
the “Method of Freezing” in conjunction with standard projection-based ROM
approaches and hyper-reduction will be used to develop a framework for parametric
ROM.

Acknowledgments G. van Zwieten, J. van Zwieten, C. Verhoosel, E. Fonn, T. van Opstal, & W.
Hoitinga. (2019, June 11). Nutils (Version 5.0). Zenodo. https://doi.org/10.5281/zenodo.3243447.

References

1. M. Rabus, A.T. Fiory, N.M. Ravindra, P. Frisella, A. Agarwal, T. Sorsch, J. Miner, E. Ferry, F.
Klemens, R. Cirelli et al., Rapid thermal processing of silicon wafers with emissivity patterns.
J. Electron. Mater. 35(5), 877–891 (2006)
2. M. Ohlberger, S. Rave, Reduced basis methods: Success, limitations and future challenges, in
Proceedings of the Conference Algoritmy (2016), pp. 1–12
230 E. J. I. Hoeijmakers et al.

3. W.J. Beyn, V. Thummler, Freezing solutions of equivariant evolution equations. SIAM J. Appl.
Dynam. Syst. 3(2), 85–116 (2004)
4. M. Ohlberger, S. Rave, Nonlinear reduced basis approximation of parameterized evolution
equations via the method of freezing. Comptes Rendus Mathematique 351(23–24), 901–906
(2013)
5. T.J.R. Hughes, The Finite Element Method: Linear Static and Dynamic Finite Element Analysis
(Courier Corporation, North Chelmsford, 2012)
6. L. Piegl, W. Tiller, Curve and surface constructions using rational B-splines. Comput.-Aided
Des. 19(9), 485–498 (1987)
7. P. Benner, W.H.A. Schilders, S. Grivet-Talocia, A. Quarteroni, G. Rozza, M. Silveira Luís,
Snapshot-Based Methods and Algorithms. Model Order Reduction, vol. 2 (De Gruyter, Berlin,
Boston, 2020)
8. R. Hull, Properties of crystalline silicon. No. 20. IET (1999)
9. M. Barrault, Y. Maday, N.C. Nguyen, A.T. Patera, An ‘empirical interpolation’ method:
application to efficient reduced-basis discretization of partial differential equations. Comptes
Rendus Math. 339(9), 667–672 (2004)
Multirate DAE-Simulation
and Its Application in System Simulation
Software for the Development of Electric
Vehicles

Michael Kolmbauer, Günter Offner, Ralf Uwe Pfau,

and Bernhard Pöchtrager

Abstract This work is devoted to the efficient simulation of large multi-physical

networks stemming from automated modeling processes in system simulation
software. The simulation of hybrid, battery and fuel cell electric vehicle applications
requires the coupling of electric, mechanic, fluid and thermal networks. Each
network is established by combining the connection structure of a graph with phys-
ical equations of elementary components and resulting in a differential algebraic
equation (DAE). In order to speed up the simulation a non-iterative multirate time
integration co-simulation method for the system of coupled DAEs is established.
The power of the multirate method is shown via two representative examples of a
battery powered electric vehicle with a cooling system for the battery pack and a
three phase inverter with a cooling system.

M. Kolmbauer ()
MathConsult GmbH, Linz, Austria
e-mail: [email protected]
G. Offner · R. U. Pfau
AVL List GmbH, Graz, Austria
e-mail: [email protected]; [email protected]
B. Pöchtrager
Radon Institute for Computational and Applied Mathematics (RICAM), Austrian Academy of
Sciences, Linz, Austria
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 231
M. van Beurden et al. (eds.), Scientific Computing in Electrical Engineering,
Mathematics in Industry 36, https://doi.org/10.1007/978-3-030-84238-3_23
232 M. Kolmbauer et al.

1 Background and Introduction

State-of-the-art modeling and simulation packages such as AVL CRUISE™M1 ,

Dymola2, or Amesim3 offer many concepts for the automatic generation of dynamic
system models. Modeling is done in a modularized way, based on a network
of subsystems which again consists of simple standardized subcomponents. For
instance, in case of HEVs (hybrid electric vehicles), BEVs (battery electric vehicles)
and FCEVs (fuel cell electric vehicles) these can be the vehicle chassis, the drive
line, the air path of the ICE (internal combustion engine) including combustion and
exhaust aftertreatment, the cooling and lubrication system of the ICE and battery
packs, the electrical propulsion system including the engine and a battery pack,
the air conditioning and passenger cabin models, waste heat recovery and finally
according control systems. Due to the complex interaction of the subsystems, the
challenges in the development of future power trains do not only lie in the design
of individual components but in the assessment of the power train as a whole.
On a system engineering level it is required to optimize individual components
globally and to balance the interaction of different subsystems. Due to the increasing
complexity of the models, the systems exhibit largely varying time scales and are
difficult in the numerical handle. A mainly automatized multirate approach is a
promising way to decrease the computational effort.
The structure of the work is the following: In Sect. 2 the individual physical
networks are introduced and the coupling conditions are stated in order to obtain a
fully coupled system of network DAEs. The multirate time integration technique for
the coupled system of network DAEs is described in Sect. 3 and the corresponding
numerical results are presented in Sects. 4 and 5. Finally we conclude in Sect. 6.

2 Problem Formulation

We consider a network that is composed of multi-physical elements. The network

elements describing the electric contribution are given by current sources, voltage
sources, nodes, ground, resistors, capacitors and inductors. The fluid network
consists of pipes, pumps, demands, junctions and reservoirs. The electro-thermal
coupling is established by lumped mass elements representing the pipe wall and the
masses from the battery and heat transfer connections. The individual components
are assembled to a network N , which is represented by a linear directed graph. The
graph structure is described by an incidence matrix A, which can be used for the
model descriptions, cf. [8]. In the following we state the DAEs for the three main
involved physical networks.

1 https://www.avl.com/de/cruise-m.
2 http://www.dynasim.com.
3 http://www.plm.automation.siemens.com.
Multirate DAE-Simulation of Electric Vehicles 233

Electric Network
We consider an electric network NE = {R, C, L, V , I, N, G, B} that is
composed of resistors R, capacitors C, inductors L, voltage sources V , current
sources I , nodes N, grounds G and batteries B. The DAE for the network in NE in
input-output form is given by: For given continuous inputs (uTR , uTC , uTB )T find the
potentials e = (eNT , e T )T , the currents j = (j T , j T , j T , j T , j T )T and the outputs
G R C L V B
y = (yR ) , such that
T T

AR jR + AC jC + AL jL + AV jV + AV jB + AI j¯I = 0
r(uR )jR − ATR e = 0
d(c(uC )ATC e)
jC − =0
dt
djL (1)
l − ATL e = 0
dt
ATV e = v̄V
ATB e = v̄B (jB , uB )
yR = jR ATR e

for given boundary conditions eG = 0 and given resistance r, capacitance c and

inductance l as well as prescribed currents j¯l and prescribed voltages v̄V and v̄B .
The coupling variables are expressed as temperature of the resistor uR , the capacitor
uC and the battery uB as well as the energy flux of the resistor yR .
Solid Network
We consider a solid network NS = {SW, LW, H T , H S, T B} that is composed
of solid walls SW , lumped walls LW , heat transfers H T , heat sources H S and
temperature boundaries T B. The DAE for the network NS in input-output form
is given by: For given continuous inputs (uTH sS , uTT bS )T , find the temperatures
T , T T )T , the heat fluxes (H T )T and the outputs (y T , y T , y T )T , such that
(TSw Lw H tS Sw Lw H tS

dTSw
mSw cp,Sw = ASw,H tS HH tS + ASw,H s HH s + ASw,H su uH sS
dt
0 = ALw,H tS HH tS + ALw,H s HH s + ALw,H su uH sS

HH tS = cH tS ATSw,H tS TSw + ATLw,H tS TLw + ATT b,H tS TT b + ATT bu ,H tS uT bS

ySw = |(ATSw,H su + AT bu ,H tS ATSw,H tS )|TSw

yLw = |(ATLw,H su + AT bu ,H tS ATLw,H tS )|TLw

yH tS = AT bu ,H tS HH tS
(2)
234 M. Kolmbauer et al.

for given boundary conditions HH s = H̄H s and TT b = T̄T b and given positive
definite coefficient matrices mSw , cp,Sw and cH tS . The coupling variables are
expressed as the energy fluxes uH sS and uT bS and the temperatures ySw , yLw and
yH tS .
Fluid Network
We consider a fluid network NF = {P I, P U, DE, V J, LJ, RE, H T , T B} that
is composed of pipes P I , pumps P U , demands DE, volume junctions V J , lumped
junctions LJ , reservoirs RE, heat transfers H T and temperature boundaries T B.
The DAE for the network NF in input-output form is given by: For given continuous
inputs (uTH sF , uTT bF )T , find the pressures (pLj
T , pT )T the mass flows (q T , q T )T ,
Vj Pi Pu
the temperatures (TVj , TLj ) , the heat fluxes (HHT tF , HPT u , HPT i )T and the outputs
T T T
T , y T , y T )T , such that
(yVj Lj H tF

dqP i 
= c1,P i ATJ c,P i pJ c + ATRe,P i pRe + c2,P i diag (|qP i |) qP i + c3,P i
dt
fP u (qP u ) = ATJ c,P u pJ c + ATRe,P u pRe

0 = AJ c,P i qP i + AJ c,P u qP u + AJ c,De qDe

dTVj
mVj cp,Vj = AVj,P i HP i + AVj,P u HP u
dt
+ AVj,De HDe + AVj,H tF HH tF + AVj,H su uH sF
0 = ALj,P i HP i + ALj,P u HP u
+ ALj,De HDe + ALj,H tF HH tF + ALj,H su uH sF
HP i = BJ c (qP i )TVj + BJ c (qP i )TLj + BJ c (qP i )TRe
HP u = BJ c (qP u )TVj + BJ c (qP u )TLj + BJ c (qP u )TRe

HH tF = cH tF ATVj,H tF TVj + ATLj,H tF TLj + ATT bu ,H tF uT bF

yVj = |(ATVj,H su + AT bu ,H tF ATVj,H tF )|TVj

yLj = |(ATLj,H su + AT bu ,H tF ATLj,H tF )|TLj

yH tF = (AT bu ,H tF + ATLj,H su ATLj,H tF + ATVj,H su ATVj,H tF )HH tF

(3)
for given boundary conditions qDe = q̄De , HDe = H̄De , pRe = p̄Re and TRe =
T̄Re and given coefficients c1,P i , c2,P i , c1,P i , mVj , cp,Vj and cH tF as well as given
functions fP u . The function BJ c checks for the sign of the mass flow qP i , cf. [4].
The coupling variables are expressed as the temperatures uH sF and uT bF and the
energy fluxes yVj , yLj and yH tF .
Multirate DAE-Simulation of Electric Vehicles 235

Multi-Physical Model
The multi-physical model is derived by combining (1), (2) and (3) with appropri-
ate coupling conditions. The coupling conditions describe the relation between the
inputs and outputs of the individual models. For the model used in Sects. 4 and 5,
the following coupling conditions are used, see e.g. [9].
⎛ ⎞ ⎛ ⎞⎛ ⎞
uR 0 CR,Sw 0 0 0 0 0 yR
⎜ u ⎟ ⎜ 0 ⎟⎜y ⎟
⎜ C ⎟ ⎜ CC,Sw 0 0 0 0 0 ⎟ ⎜ Sw ⎟
⎜ u ⎟ ⎜ 0 ⎟⎜y ⎟
⎜ B ⎟ ⎜ CB,Sw 0 0 0 0 0 ⎟ ⎜ Lw ⎟
⎜ ⎟ ⎜ ⎟⎜ ⎟
⎜ uH sS ⎟ = ⎜CH sS ,R 0 0 0 0 0 CH sS ,H tF ⎟ ⎜ yH tS ⎟
⎜ ⎟ ⎜ ⎟⎜ ⎟
⎜ uT b S ⎟ ⎜ 0 0 0 0 CT bS ,Vj 0 0 ⎟ ⎜ yVj ⎟
⎜ ⎟ ⎜ ⎟⎜ ⎟
⎝uH sF ⎠ ⎝ 0 0 0 CH sF ,Vj 0 0 0 ⎠ ⎝ yLj ⎠
uT b F 0 CT bF ,Sw 0 0 0 0 0 yH tF
(4)

The connectivity equation (4) represents the electro-thermal coupling of the electric
network and the cooling systems. Combining all subsystems and their connectivity
equations (4) yields a DAE:
Find

z := (uR , uC , uB , uH sS , uT bS , uH sF , uT bF , eN , eG , jR , jC , jL , jV , jB , TSw , TLw , HH tS ,

pLj , pVj qP i , qP u , TVj , TLj , HH tF , HP u , HP i , yR , ySw , yLw , yH tS , yVj , yLj , yH tF ),

* +
d(c(uC )ATC e) djL dTSw dqP i dTVj
ż := , , , , ,
dt dt dt dt dt

such that

F (ż, z, t) = 0. (5)

DAEs resulting from automated modeling software typically obtain a structure

with (differential) index greater 1, cf. [4–6] and hence are not suitable for a direct
simulation with standard solvers. In the setup of multiple physical networks it is not
sufficient, that the full DAE (5) can be reduced to a d-index (differential index) 1.
Additionally, each subsystem, to which a solver is applied, has to fulfill d-index 1
conditions as well, cf. [1, 3]. In our applications an automatic index reduction is
performed if the electric or the fluid system happens to be of d-index 2.
236 M. Kolmbauer et al.

3 Multirate Integration for Coupled Network DAEs

In our multirate approach the full DAE (5) is partitioned due to the physical
background to n ∈ N subsystems (typically n ' 2). Each subsystem is index
reduced according to the available literature, cf. [4–6]. Since in the global network
the individual subsystems are interacting with each other, i.e. inputs and outputs
are connected according the connectivity equation (4), it is necessary to put it into
an input-output form. For this purpose, each subsystem i = 1, . . . , n classifies its
inputs ui , state variables xi , algebraic variables ai and outputs yi . To conclude, this
approach yields a coupled system of n semi-explicit DAEs in input-output form of
(differential) index 1. For inputs ui given by Eq. (4), find xi , ẋi , ai and yi , such that

ẋi = fi (xi , ai , ui , t)
0 = ri (xi , ai , ui , t) (6)
yi = gi (xi , ai , ui , t)

for i = 1, . . . , n. A careful choice of the connectivity matrix given in (4) guarantees

that the coupled system obtains (differential) index 1 as well, cf. [3, 9]. E.g. one
possible choice is the usage of differential states, which are not involved in any
index reduction, as coupling variables.
For each subsystem (6) an arbitrary Runge-Kutta method with micro-step sizes
hi is used, cf. Figure 1. The choice of the actual integration technique depends on the
properties of the underlying system and can be explicit, implicit, fixed or adaptive.
The whole system (5) is integrated via a non-iterative co-simulation technique with
macro-step size H = max(hi ). All systems are updated at the end of each macro-
step. This principle relates to synchronous communication and we refer to these
points in time as synchronization times, cf. Fig. 1. The evaluation of each macro-
step of the subsystems is done in a sequential Gauss-Seidel-approach. The values ui
are handled with appropriate interpolation or extrapolation techniques, depending
on the slow or active characteristic of the interacting subsystems. Due to n ' 2
slow-first or fast-first strategies (cf. [2]) have been extended to strategies, that can
be used for an arbitrary number n of components.

Synchronization time

macro-step

micro-step

...
tk = tki 0 tki 1 tki 2 tki n = tk+1

Fig. 1 Macro-step of the i-th system from synchronization time tk to tk+1 .

Multirate DAE-Simulation of Electric Vehicles 237

4 Simulation of a BEV with Cooling System

We consider a BEV that demonstrates the modeling of an electrical system coupled

to the required cooling system, cf. Fig. 2. The model consists of an electrical
propulsion and two cooling circuits. An oil circuit is used for cooling of the electric
machine and a coolant circuit is used for cooling of the battery pack, inverter and low
voltage DC-DC converter. The involved subsystem of the coupled electro-thermal
model can be reduced to DAEs of (differential) index 1.
The multirate approach presented in Sect. 3 is put into context with the reference
solution of a single solver approach (both sequential/single CPU). In this example
eight thermal circuits, three mechanic circuits, an electric circuit, 14 gas circuits
and two fluid circuits are present which represent in total 461 equations. The
solvers for both, the single solver approach and the multirate approach are all

Fig. 2 Schematic representation of a BEV with cooling system in AVL CRUISE™M. The
corresponding results are displayed in Fig. 3 and Table 1
238 M. Kolmbauer et al.

Fig. 3 Comparison of elapsed time of a multirate case against a single solver case

Table 1 Comparison of singlerate and multirate approach corresponding CPU-time and average
real time factor (RTF)
Case CPU-time Avg RTF
Singlerate 144.98 0.805447
Multirate 21.03 0.116853

adaptive explicit solvers [7]. Hence the step size of the single solver is limited to
the minimum step size of all subdomains, while the multirate approach is limited
to the synchronization time or to the characteristic of its own domain. Here the
synchronization times are after each macro-step of 20 ms.
The simulation time of a singlerate case (in red) is compared with those of a
multirate case (in blue) using AVL CRUISE™ M, cf. Fig. 3. A significant speed up
in the calculation time can be achieved, while the accuracy of the solution is still
sufficiently high due to the adaptivity of the individual solvers (Table 1).

5 Simulation of a Three Phase Inverter with Cooling System

We consider a detailed physical model of an inverter with switches/transistors, an

RC (resistor-capacitor) filter as well as a 3 phase ohmic load. The inverter is used to
convert a DC (direct current) voltage through timed switching of the six transistors
into a PWM (pulse width modulation) signal. The RC filter then averages the PWM
and thus creates a 3 phase AC (alternating current) voltage.
Multirate DAE-Simulation of Electric Vehicles 239

Fig. 4 Comparison of elapsed time of a multirate case against a single solver case

Table 2 Comparison of singlerate and multirate approach corresponding CPU-time and average
real time factor
Case CPU-time Avg RTF
Singlerate 303.38 606.76734
Multirate 37.37 74.73045

In total this example consists of 178 equations which are spread over 20 solvers.
A fluid circuit, seven gas circuits and eleven thermal circuits are responsible for
modeling the cooling. In the multirate scheme each circuit is solved individually
with one scheme. For all of them an explicit fixed step method with a step size of
1ms is used. On the other hand the electric network is solved by its own scheme as
well. Again an explicit fixed step method is used, whereby the chosen step size is
now 1μs. The information exchange takes place after each macro-step of 1ms. This
model is of special interest, since the electric network and the fluid network run on
completely different time scales (of order O(1000)). Again significant speed up in
the calculation time can be achieved (Fig. 4 and Table 2).
240 M. Kolmbauer et al.

6 Conclusion

As shown, the multirate approach offers a possibility to reduce computation time

considerably. In order to ensure a stable simulation, automatic index reduction of the
physical networks, appropriate solver settings for each subsystem and an adequate
coupling procedure, play a decisive role. For the correct choice a significant speed
up can be achieved, while conserving the accuracy criteria.

Acknowledgments Part of this work has been supported by the government of Upper Austria
within the programme Innovatives Oberösterreich.

References

1. A. Bartel, M. Günther, PDAEs in refined electrical network modeling. SIAM Rev. 60, 56–91
(2018)
2. A. Bartel, M. Günther, Multirate Schemes - An Answer of Numerical Analysis to a Demand
from Applications. IMACM Preprint, No. 2019–12, University of Wuppertal (2019)
3. A. Bartel, M. Günther, Inter/extrapolation-based multirate schemes – a dynamic-iteration
perspective (2020). Available at https://arxiv.org/abs/2001.02310
4. A.-K. Baum, M. Kolmbauer, G. Offner, Topological solvability and DAE-index conditions for
mass flow controlled pumps in liquid flow networks. Electr. Trans. Num. Anal. 46, 395–423
(2017)
5. D. Estévez Schwarz, C. Tischendorf, Structural analysis of electric circuits and consequences
for MNA. Int. J. Circ. Theor. Appl. 28, 131–162 (2000)
6. S. Grundel, L. Jansen, N. Hornung, T. Clees, C. Tischendorf, P. Benner, Model order reduction
of differential algebraic equations arising from the simulation of gas transport networks, in
Progress in Differential-Algebraic Equations (Springer, Berlin, 2014), pp. 183–205
7. A. Hindmarsh, P. Brown, K. Grant, S. Lee, R. Serban, D. Shumaker, C. Woodward, SUNDIALS:
suite of nonlinear and differential/algebraic equation solvers. ACM Trans. Math. Softw. 31, 363–
396 (2005)
8. L. Jansen, C. Tischendorf, A unified (P)DAE modeling approach for flow networks, in Progress
in Differential-Algebraic Equations (Springer, New York, 2014), pp. 127–151
9. M. Kolmbauer, G. Offner, B. Pöchtrager, Topological index analysis and its application to multi-
physical systems in system simulation software (2020). Available at https://www.ricam.oeaw.ac.
at/files/reports/20/rep20-22.pdf
A Hysteresis Loss Model for Tellinen’s
Scalar Hysteresis Model

Jan Kühn, Andreas Bartel, and Piotr Putek

Abstract A particularly well-suited hysteresis loss model for Tellinen’s scalar

hysteresis model is defined. Important basic properties are discussed. The model
is based on the enclosed area of simple hysteresis loops as a measure of the energy
loss. It can be applied for any simple excitation in a nearly steady state condition.
The losses can be computed on-the-fly, during the field computations.

1 Introduction

Hysteresis is an important phenomenon for the simulation of magnetic fields in

ferromagnetic materials. There are several hysteresis models already available in
the literature, see e.g. [4]. One example is Tellinen’s scalar hysteresis model, which
was introduced in [8]. It is physically motivated but quite simple to evaluate. A
recent investigation by Steentjes et al. [6] showed that it is still quite competitive to
other, even more complex hysteresis models.
Moreover, there exists a thermal extension of Tellinen’s model, which models the
temperature depends of hysteresis in a Tellinen-like fashion [5]. Now, the work at
hand aims at constructing a dedicated loss model, which inherits the computational
benefits of Tellinen’s model. For this new development, we currently assume that
the magnetic fields are nearly in steady state.
Tellinen [8] provides a model to describe bh-curves. The thermal extension [5]
deals with the influence of temperature. This paper presents a loss model. Then, in
fact, the application of this model is based on a coupled problem, which consists
of the curl-curl equation (magnetic field) and a heat conduction (temperature). This
coupled system can be solved using distributed simulation techniques.

J. Kühn () · A. Bartel

Bergische Universität Wuppertal, Wuppertal, Germany
e-mail: [email protected]; [email protected]
P. Putek
Universität Rostock, Rostock, Germany
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 241
M. van Beurden et al. (eds.), Scientific Computing in Electrical Engineering,
Mathematics in Industry 36, https://doi.org/10.1007/978-3-030-84238-3_24
242 J. Kühn et al.

The outline of this work reads: First, we introduce Tellinen’s model [8]. Then,
we define the loss model for a steady state and apply it as an approximation for
any nearly steady state. Based on properties of the model, we justify the use of
steady state approximation also with numerical results. At the end, conclusions and
an outlook are given.

2 Tellinen’s Scalar Hysteresis Model

Original Model First, we report the isothermal hysteresis model introduced by

Tellinen [8]. To this end, we assume that the temperature has a fixed value and the
fields are described by the scalar magnetic field strength h and the scalar magnetic
flux density b. Then, any ferromagnetic material has a specific limiting saturation
curve. That is, for a transition from very small values of h ( 0 (i.e., full saturation)
to very large values h ' 0, we have that the relation b = b(h) is given by the
+ +
specific function Bsat = Bsat (h) ∈ C 1 . The reverse direction can be defined by
− + ±
Bsat (h) := −Bsat (−h), see Fig. 1. To have a physically compliant model, Bsat is
requested to form a loop:
+ −
 − +

Bsat (h) < Bsat (h) , lim Bsat (h) − Bsat (h) = 0 (1)
|h|→∞

and the derivative is bounded by the vacuum permeability μ0 from below:

d + +
B (h) ≥ μ0 > 0 , lim d Bsat (h) = μ0 . (2)
dh sat |h|→∞ dh

Any current state of the material (h0 , b0 ) has to belong to the loop region I
, ) + -
I = (h, b) ∈ R2 ) Bsat −
(h) ≤ b ≤ Bsat (h) . (3)

±
Fig. 1 Left: Example of Bsat and path starting from demagnetized state. Right: Schematic of
Tellinen’s model. Defining the values on the boundary and interpolate in between
A Hysteresis Loss Model for Tellinen’s Scalar Hysteresis Model 243

Now, Tellinen’s model defines the differential reluctivity μdiff = dh db

at (h0 , b0 ) via
vertical interpolation of the slopes on the saturation curves (see Fig. 1b) (notice that
−
on Bsat the slope in the forward direction (i.e., increasing h) is given by μ0 ) by
 +
μ+
dBsat (h)
= λ dh + (1 − λ)μ0 if h is increasing,
μdiff = diff
− (4)
− dBsat (h)
μdiff = λμ0 + (1 − λ) dh if h is decreasing,

where the interpolation parameter λ is given by

−
Bsat (h) − b
λ = λ(h, b) = − + ∈ [0, 1] . (5)
Bsat (h) − Bsat (h)
We note the following: (i) If h is regarded as the independent variable, the
corresponding b can be determined as the solution of the ordinary differential
equation (ODE) dh db
= μdiff with μdiff given by (4). (ii) An analytical solution
(h, b) of this ODE with initial value (h0 , b0 ) ∈ I stays inside the loop area I .
(iii) The intrinsic induction is defined by bi (h) = b(h) − hμ0 . The definition can be
transformed to the usage of bi .
Partition of States in Tellinen’s Model We introduce a partition of possible states
I . To this end, we observe that for each h there exists a unique value beq with
+ −
Bsat (h) < b eq < Bsat (h), such that holds (for λ, which correspond to that specific
eq
b ):
+
dBsat (h) dB − (h)
μ+
diff = λ + (1 − λ)μ0 = λμ0 + (1 − λ) sat = μ−
diff . (6)
dh dh
Solving this equation for λ, we obtain
− −
dBsat (h) dBi,sat (h)
− μ0
λ (h) = λ =
eq
+
dh
− = +
dh
, (7)
dBsat (h) dBsat (h) dBi,sat (h) dB − (h)
dh + dh − 2μ0 + i,sat
dh dh

which gives in turn the value of beq as

d −
− dh Bi,sat (h)  − +

beq (h) = Bsat (h) − d − d +
Bsat (h) − Bsat (h) . (8)
dh Bi,sat (h) + dh Bi,sat (h)

Now, the set of all possible states I can be split into disjoint sets I = , I < and I >
with
) )
I  = {(h, b) ∈ I ) μ+ − )
diff (h, b)  μdiff (h, b)} = {(h, b) ∈ I λ  λ }
eq
(9)
)
= {(h, b) ∈ I ) beq (h)  b} for all  ∈ {=, <, >}.

See Fig. 2 for an example.

244 J. Kühn et al.

±
Fig. 2 Two examples with different saturation curves Bsat and the resulting equilibrium curve beq
≶
as well as the sets I . We remark that the right example is a very academic version

3 Adapted Hysteresis Loss Model

There exists several hysteresis loss models for different kinds of hysteresis descrip-
tions [1–3, 7]. A prediction of the losses is already presented in the original
Tellinen model [8]. But the used method for the hysteresis losses is based upon
a posteriori evaluation of the simulated fields b, h. Our model differs in this
respect and provides a method for calculating losses at runtime. An overview and
classification of different hysteresis loss models is e.g. given in [4]. In principle, the
loss model presented below will work for other hysteresis models, too. However,
it is particularly well suited for Tellinen’s model [8] with the respective thermal
extension [5] due to its structure and properties. First, we define the loss model for
the steady state and then we extend it to almost steady state situations.
Idea We follow the approach of distributed simulation (Co-Simulation). A sim-
ulation of the heat equation describes the behavior of the temperature. The
presented loss model provides corresponding source terms. Often in applications,
e.g. electric machines, the rate of changes in the magnetic fields are several orders
of magnitude faster than changes in temperature. In this setting, the assumption of a
constant-temperature while handling magnetic fields is often exploited in distributed
simulation techniques.
For a simple, closed loop in the bh-plane, the enclosed area represents an energy
density (J/m3 ) and the material specific volumetric heat capacity cV (J/(m3 K))
provides the conversion into a temperature change ΔT (K). In a steady state, the
material periodically passes through the same phases over and over again and, for
this reason, runs on a closed bh curve. For memory reasons and the fact, that a priori
the stable loop is unknown, we do not want to save the complete history of the curve,
but calculate it from within the simulation on-the-fly, i.e., at runtime. To this end,
we reverse the hysteresis model computation to predict the return path of the curve
from a turning point at the same time as we compute the forward, see e.g. Fig. 4
right, where the curves p+ (forward) and p− (backward) will be computed at the
same time.
A Hysteresis Loss Model for Tellinen’s Scalar Hysteresis Model 245

b b

h h

Fig. 3 Left: An example with minor hysteresis loops. This case is excluded by simple excitation.
Right: An example with an intermediate intersection

(h1 , b1 ) (h1 , b1 )
b b
p− h p− h
p+ p+

(h0 , b0 ) (h0 , b0 )
Fig. 4 Left: paths p+ and p− form a simple loop without intersection. Right: an incomplete cycle
and its incremental section of the area. Notice, p− is traversed in reverse direction

Prerequisites First, we consider a simple excitation such that the magnetic field
strength is monotonously increased from h0 to h1 and then monotonously decreased
back to h0 . This ensures that there is no minor hysteresis loop (see left of
Fig. 3). Still, this is not sufficient to ensure that the bh-loop has no (intermediate)
intersections (see Fig. 3, right). Below, sufficient conditions are presented.
Now, let (h0 , b0 ) and (h1 , b1 ) denote the turning points of a simple loop (cf.
Fig. 4). Then, the loop can be split into two paths p+ , p− : [h0 , h1 ] → R with

p+ (h) < p− (h) for all h ∈ (h0 , h1 ) , p+ (hk ) = p− (hk ) = bk k ∈ {0, 1}.
(10)

On-the-Fly Algorithm To initiate the loss model, we assume that the current state
(h0 , b0 ) is a turning point of a simple hysteresis loop, where h is (wlog) increased.
Thus, the simulation will follow the curve p+ (using μ+ diff ), see Fig. 4. Now, a
second computation is simultaneously performed based on μ− diff to follow the reverse
direction, which results in a prediction of the return path p− . Both simulations might
use e.g. an ODE solver. For discrete steps, the resulting trapezoids (see Fig. 5) can
be summed up to approximate the loop area. The incorporated halving of the area
takes into account that there is a forward and backward phase. This continues until
the second reversal point is reached. The procedure is then restarted from this point.
246 J. Kühn et al.

b
(h1 , b1 )
p− h
p+

(h0 , b0 )
Fig. 5 For discrete points, the model results in trapezoids

b b
h h

Fig. 6 Nearly steady state: actual turning point after (left) and before (right) the intersection of
the curves p+ and p−

Non-steady State Loss Computation For simple closed bh-loops (steady state)
the proposed method is accurate up to numerical precision of the employed
solver. Now, if the turning point (h1 , b1 ) cannot be determined accurately by the
intersection of p+ and the predicted curve p− (as depicted in Fig. 6), we are not in
steady state. We can prove (via some fixed-point argument) that this model exhibits
convergent behavior for simple periodic inputs. Due to this, non-steady states are
converging to the steady state. Numerical examples are presented in Sect. 4. If the
difference between the actual and predicted reversal point is small enough (criteria
set by user), we consider our model as a valid approximation and say it is nearly
steady state.
Analytical Results Next, we develop criteria that guarantee the existence of at least
one further intersection point (h1 , b1 ) based on a turning point (h0 , b0 ) and the
corresponding paths p+ and p− . In a second step we then investigate when exactly
only the intersection points (h0 , b0 ) and (h1 , b1 ) exist. It is then shown that both
the Tellinen’s model [8] and the thermal extension [5] converge towards the steady
state.
A Hysteresis Loss Model for Tellinen’s Scalar Hysteresis Model 247

±
Fig. 7 Example with the same Bsat as in Fig. 2, two paths p+ and p− and beq . Left: Only two
intersections of p± . Right: More than two intersections of p±

As seen in Fig. 4 the intersection (h0 , b0 ) fulfills μ+ −

diff < μdiff , while the
+ −
intersection (h1 , b1 ) holds μdiff > μdiff . In general, we have
Lemma 1 Assuming that the paths p+ and p− have intersections and no contact
points, i.e., the common points are not on beq . Any sequence (hi , bi ) with i =
0, . . . , k and h0 < . . . < hk of all intersections of the paths p+ and p− must
alternating fulfill μ+ − + −
diff < μdiff and μdiff > μdiff , i.e., are alternating element of I
<
>
and I (see (9)), starting with the former and ending with the latter.
−
Lemma 2 (Existence of Intersection) For any (h0 , b0 ) ∈ I < with Bsat (h0 ) = b0 ,
there exists at least one other point (h1 , b1 ) ∈ I with h0 < h1 , such that the curves
>

p+ and p− intersect at this point.

The proof is mainly based on the consideration of the limit h → ∞ for the paths
p± (h) see Fig. 7 (left). As seen in Fig. 7 (right), there could be more than two
intersections of p± . These points are alternately above and below of beq .
Lemma 3 (Uniqueness of Intersection) Given (h0 , b0 ) as in Lemma 2 and
dbeq (h)
> μdiff(h) = μ±
eq eq
diff (h, b (h)) for all h (11)
dh
we have exactly one further intersection, i.e., (h1 , b1 ).
Proof We can restrict the paths p± to only go through beq from above to below. This
effectively limits the number of possible intersections of p± to two and therefore
makes (h1 , b1 ) unique. 
+
Lemma 4 If we assume Bsat ∈ C 2 , then the uniqueness restriction (11) can be
formulated as follows:
⎛ + ⎞
−
dBi,sat (h) & + − ' dBi,sat (h)
dBsat (h) dBsat (h)  −  d
+
dh
+ +
+ Bsat (h) − Bsat (h) ⎝ −dh ⎠ > 0 .
dBi,sat (h) dh dh dh dBi,sat (h)
dh dh
(12)
248 J. Kühn et al.

+
Lemma 5 Let Bsat fulfil (12) and the temperature T be constant. We start from
operation point (h0 , b0 ) ∈ I and h varies periodically between h0 and h1 with
h0 < h1 . The sequence of b-values Bk (k ∈ N) at the turning point given by h0
(computed by Tellinen’s model) is convergent for k → ∞. The resulting stable loop
is unique and depends only on the choice of h0 and h1 , but not on b0 .
Proof (sketch) First, we define one iteration. To this end, let b+ (h) be the solution
+ + +
+ +
dh = μdiff (h, b ) with b (h0 ) = b0 and μdiff as in (4). At h = h1 ,
of the ODE db
we have
 h1
b1 = b+ (h1 ) = b0 + μ+ +
diff (h, b (h))dh . (13)
h0

−
The reverse direction is the same. Let b− (h) be the solution of db dh =
μ−diff (h, b − (h)) with b − (h ) = b . Evaluated at h , this results in b := b − (h )
1 1 0 2 0
(analog to (13)).
+ − + −
Let ϕ : [Bsat (h0 ), Bsat (h0 )] → [Bsat (h0 ), Bsat (h0 )] denote the resulting b-value
at h0 after one iteration (starting from b̄), i.e.,
 h1  h0
ϕ(b̄) = b̄ + μ+ +
diff (h, b (h))dh + μ− −
diff (h, b (h))dh (14)
h0 h1

with b+ defined w.r.t. (h0 , b̄) (13) and b− to (h1 , b + (h1 )). Now, we construct a
sequence Bk via Bk+1 = ϕ(Bk ) and B0 = b0 . If ϕ(b̄) = b̄ holds, the resulting loop
would be closed, i.e., it is stable.
μ+ −
diff (h, b) and μdiff (h, b) are strictly monotone in the second component. This
+
causes two ODE solutions of db db = μdiff with different initial values to become
closer to each other. An analogous statement can be made for μ− diff .
Given (h0 , b0 ), (h0 , b1 ) ∈ I , we can prove, that q ∈ [0, 1) exists, such that

|ϕ(b1 ) − ϕ(b0 )| ≤ q|b1 − b0 | (15)

holds. So ϕ is a contraction and Banach’s fixed-point theorem applies. This ensure

convergence and a unique fixed-point, i.e., a stable loop.
A numerical example is given in Fig. 8. We conclude that under these assump-
tions, the model tends towards the steady state. If we now assume that the
temperature changes only slowly in comparison to the magnetic fields (thermal
extension of Tellinen’s model [5]), then the steady state is disturbed by this, but it is
still approximately maintained. Similar convergence considerations for the thermal
extension [5] also show the convergence towards a stable state. This is a kind of
confirmation of our model’s assumption and that after a certain initial phase we are
permanently in an almost steady state and our loss model is therefore applicable.
A Hysteresis Loss Model for Tellinen’s Scalar Hysteresis Model 249

Fig. 8 Example of convergence. A random starting point (h0 , b0 ) ∈ I and h1 > h0 is chosen.
Periodically and monotonously alternating between h0 and h1 converges to a stable loop

4 Numerical Results
±
As an academic example, a material is defined by Bsat depicted Fig. 8, left. We
choose h0 = −1.8e5 A/m, h1 = 2.5e5 A/m and b0 = −1.5 T. Starting at
(h0 , b0 ) ∈ I , the ordinary differential equation dh
db
= μ+diff (h, b) is solved numer-
ically on the interval [h0 , h1 ] by a Runge-Kutta method. Then, dh db
= μ−diff (h, b)
is solved backward from h1 to h0 , where the initial value is the final value of
the previous computation. This procedure is repeated n times. This results in the
sequence of b-values at h0 : B0 , . . . , Bn . As seen in Fig. 8, right, the absolute
difference |Bn − Bn−1 | converges. Even a low number of loops n results in a nearly
steady state and thus, would allow us to apply the loss model presented above.

5 Conclusion and Outlook

We have proposed a loss model with on-the-fly computation for Tellinen’s hysteresis
model. In steady state, it results in a precise computation of the model respective
hysteresis loss. Thus it is a valid approximation for nearly steady state. Moreover,
the convergence towards a stable bh-loop is proven. We note that our model is not
suited for complex waveforms or rotating fields, cf. [4]. As a model feature, we
stress that this hysteresis loss model employs only rough material data and needs
small computational and memory cost. Moreover, this model can be combined with
other than Tellinen’s hysteresis model, if the value of b can be computed for changes
in h. But the properties of the Tellinen’s model make it an almost optimal candidate.
Our next step will be the integration of this loss model into a finite element
magnetoquasistatic field simulation and its analysis. Special attention will be paid
to whether this model can be applied per node.
250 J. Kühn et al.

References

1. A.P.S. Baghel, S. Kulkarni, Dynamic loss inclusion in the Jiles–Atherton (JA) hysteresis model
using the original JA approach and the field separation approach. IEEE Trans. Magn. 50, 369–
372 (2014)
2. L.R. Dupre, R. Van Keer, J.A.A. Melkebeek, An iron loss model for electrical machines using
the Preisach theory. IEEE Trans. Magn. 33(5), 4158–4160 (1997)
3. G. Friedman, I.D. Mayergoyz, Hysteretic energy losses in media described by vector Preisach
model. IEEE Trans. Magn. 34(4), 1270–1272 (1998)
4. A. Krings, J. Soulard, Overview and comparison of iron loss models for electrical machines. J.
Electr. Eng. 10, 162–169 (2010)
5. J. Kühn, A. Bartel, P. Putek, A thermal extension of Tellinen’s scalar hysteresis model, in
Proceedings of the SCEE 2018, ed. by G. Nicosia, V. Romano. Springer (2020), pp. 55–63
6. S. Steentjes, K. Hameyer, D. Dolinar, M. Petrun, Iron-loss and magnetic hysteresis under
arbitrary waveforms in no electrical steel: A comparative study of hysteresis models. IEEE
Trans. Ind. Electron. 64(3), 2511–2521 (2017)
7. C.P. Steinmetz, On the law of hysteresis. Proc. IEEE 72(2), 197–221 (1984)
8. J. Tellinen, A simple scalar model for magnetic hysteresis. IEEE Trans. Magn. 34(4), 2200–2206
(1998)
Hybrid Modeling: Towards the Next
Level of Scientific Computing
in Engineering

Stefan Kurz

Abstract The integration of machine learning (Keplerian paradigm) and more

general artificial intelligence technologies with physical modeling based on first
principles (Newtonian paradigm) will impact scientific computing in engineering in
fundamental ways. Such hybrid models combine first principle-based models with
data-based models into a joint architecture. This paper will give some background,
explain trends and showcase recent achievements from an applied mathematics
and industrial perspective. Examples include characterization of superconducting
accelerator magnets by blending data with physics, data-driven magnetostatic field
simulation without an explicit model of the constitutive law, and Bayesian free-
shape optimization of a trace pair with bend on a printed circuit board.

1 Introduction: What Is Hybrid Modeling?

If we take a look at Gartner’s 2018 Hype Cycle of Emerging Technologies [1],

Deep Learning has been at the top of inflated expectations and should now be
moving towards the plateau of productivity. Indeed, machine learning and more
general artificial intelligence technologies recently have spurred a lot of interest in
the applied mathematics and industrial communities, see for instance [2–4], and [5]
for an introduction.
The idea of combining physics with data has a long history. Following [3, p. 57],
we call modeling based on first principles the Newtonian paradigm. Newton’s
laws of motion provided (within their range of validity) “for the first time a
unified quantitative explanation for a wide range of observations” [6]. Conversely,
Johannes Kepler started from astronomer Tycho Brahe’s and own measurement
data, and worked towards a mathematical description to fit the measured data.

S. Kurz ()
Bosch Center for Artificial Intelligence, Renningen, Germany
Centre for Computational Engineering, Technical University of Darmstadt, Darmstadt, Germany
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 251
M. van Beurden et al. (eds.), Scientific Computing in Electrical Engineering,
Mathematics in Industry 36, https://doi.org/10.1007/978-3-030-84238-3_25
252 S. Kurz

Following again [3], we call this approach the Keplerian paradigm. Both paradigms
complement each other. For a simple enough model system, Kepler’s laws can be
derived from Newton’s theory. Conversely, starting from a two-body model system,
actual trajectories of celestial bodies can be modeled by Newton’s laws plus data-
driven terms that correct for perturbations due to effects that are not present in the
model.
In modern terms, we call this complementary approach hybrid modeling.
Definition Hybrid models combine first principle-based models with data-based
models into a joint architecture, supporting enhanced model qualities, such as
robustness and explainability.

First principles express formalized domain knowledge. For the purpose of this
paper the domain knowledge results from physics. But there are other possibilities,
such as statistics (e.g., probabilistic graphical models [7, Ch. 8]) or discourse
(e.g., ontologies [8]). Data may be obtained from any source, in particular from
observation or simulation. We find also the somewhat narrower terms scientific
machine learning [9], physics-based machine learning [10] and predictive data
science [4], respectively.
Consider a high-dimensional manifold that contains some big data. It might
be that a submanifold can be identified, which is dictated to us by the laws of
physics, e.g. regarding admissible system dynamics. Learning algorithms can then
be used to project the data into this submanifold. In other words, the structure of
submanifold embeds physical constraints. A classical example is Kálmán filtering
[11], and an example is presented in Sect. 2. Kálmán filtering was actually an
enabling technology for the moon landing in 1969, where the goal was landing
within ≈500 m after ≈400,000 km of travel. More preference is given to physics
or data, depending on the level of uncertainty. Ensemble Kálmán filtering is used
in weather forecasting centres worldwide [12]. They have to deal with about 106
incoming data points per hour, and mathematical models with about 109 states.
Ensemble Kálmán filtering can be recognized as Gaussian hidden Markov model
[13]. This use case is similar to digital twinning, since data from the field is acquired
and used to update the models. Citing [4, p. 39]: “Learning from data through the
lens of models is a way to exploit structure in an otherwise intractable problem.”
Looking closer into engineering, we notice that a large class of physics models
can be decomposed into conservation laws and constitutive laws [14, Ch. 1.3], [15].
The conservation laws are of topological nature and can therefore be discretized
easily, leaving little room for data-driven techniques. The situation is different for
the constitutive relations, which are of metric nature, and encode phenomenological
material properties. Except for simple media (local, linear) there are many potential
complications (non-local, hysteretic, non-linear, multi-scale, multi-physics, etc.).
Here, data-driven models can be useful, provided that the models fulfil certain
admissibility criteria, which can often be expressed in terms of invariance with
respect to symmetry groups (orthogonal group, Lorentz group, etc.). This is
showcased in Sect. 3.
Hybrid Modeling 253

To sum up, hybrid modeling has the potential to improve the Pareto tradeoff
between simulation accuracy and simulation cost significantly, and therefore bring
scientifc computing in engineering to the next level. In the remainder of the paper
we will showcase this by some recent achievements.

2 Blending Data with Physics [16]

This section deals with the characterization of superconducting accelerator magnets

of CERN’s Large Hadron Collider. The magnets are characterized by measurements
and the field in their aperture is modeled by the Boundary Element Method (BEM).
Different methods for local magnetic field measurements are available. There are
mapper systems, based on 3D magnetic Hall probes, and coil-based systems. The
latter feature plane translating or saddle-shaped rotating coils. The coil-based sys-
tems benefit from straightforward calibration and linear transfer function, while the
mapper systems offer a small active area (≈0.01 mm2 ), a high positioning resolution
(µm scale), and admit measuring of all three field components simultaneously.
Figure 1 shows field quality maps of a dipole field in a rectangular magnet cross
section. The colours ranging from blue to red indicate the deviation from an ideal
dipole field, in logarithmic scale. The top image shows a field quality map that can
be obtained by interpolation from measurement data, while the bottom image was
obtained by reconstruction from a BEM model. In the sequel, the field is represented
by a double layer potential. The dipole layer is located on the boundary of the
rectangle. This reconstruction method enjoys a smoothing property. Moreover, the
reconstructed field is locally an exact magnetostatic solution. The double layer

40mm

360mm

Fig. 1 Field quality maps of a dipole field in a rectangular magnet cross section. Top: interpolation
of measurement data. Bottom: reconstruction from a BEM model. The colours ranging from blue
to red indicate the deviation from an ideal dipole field, in logarithmic scale
254 S. Kurz

Table 1 Bayesian update and Kálmán update. Quantities ν and d are to be understood as random
variables with probability distributions p(·). The number of degrees of freedom of the model is
denoted by N, and the dimension of measurement data by M
ν State vector of BEM model
ν ∼ N (ν, Q) ν ∈ RN mean values
Q ∈ RN×N covariance matrix, process noise
d measurement data vector
d|ν ∼ N (Mν, R) M ∈ RM×N discrete measurement operator
R ∈ RM×M covariance matrix, measurement noise

density is discretized by piecewise linear continuous boundary elements, giving rise

to a state vector ν, with N degrees of freedom.
A hybrid model is established, by estimating the state vector from measurement
data. This can be accomplished by Kálmán filtering [11]. The relevant quantities are
defined in Table 1. We start from a prior of the state p(ν), and measurement data
given the state, p(d|ν). Then, by Bayesian update, we infer the posterior of the state
given the measurement data,

p(ν|d) ∝ p(d|ν)p(ν) . (1)

Under normal distribution assumption this can be computed easily explicitly. Then,
the Bayesian update turns into a Kálmán update, which can be readily expressed in
terms of linear algebra operations,

ν
→ ν + K(d − Mν) , (2a)
Q
→ (I − KM)Q , (2b)

where
 −1
K := QM  MQM  + R ∈ RN×N (3)

is the Kálmán gain matrix. Matrix M is the measurement matrix. It maps the degrees
of freedom of the BEM model to the measured quantities. These are flux density
vectors in case of Hall probe measurements, and magnetic fluxes in case of coil-
based systems. Technically, this amounts to evaluating the integral operator of the
double layer potential, in terms of the discrete model. In actual applications this
approach is extended to a box-shaped domain in three dimensions, cf. Fig. 2. The
Kálmán update results in a three-step procedure.
1. We select some prior from previous measurements or simulations. In the simplest
case, we start from zero, with some estimate for the covariance matrix, i.e. ν ∼
N (0, Q). This is a so-called smoothing prior. In fact, the reconstruction of
the dipole layer from the measured field boils down to an inverse problem,
Hybrid Modeling 255

Fig. 2 Hybrid model combining measured data with the BEM. Top left: A translating induction
coil consists of multiple single-wire loops in x direction. For fixed y, it measures magnetic flux
increments ΔΦ between successive trigger points along the z axis. Bottom left: The covariance
matrix can be estimated from an ensemble of runs. The figure shows frequency distributions for
three exemplary positions. Right: Contour plots of posterior mean and variance field magnitudes

and selecting a smoothing prior of the form Q = σ 2 I is related to Tikhonov

regularization [17, Sect. 2.1.2].
2. The measurement is carried out by a translating induction coil, which consists
of multiple single-wire loops. The covariance matrix R is estimated from an
ensemble of runs.
3. Finally, the Kálmán update (2) yields the new state vector and its covariance
matrix. Since the state is modeled as random variable, we can immediately
propagate uncertainties to the quantities of interest, such as magnetic flux density,
magnetic vector potential or transfer maps. Moreover, data from different sensors
can be combined through this procedure.

3 Data-Driven Field Simulation [18]

This section is about data-driven field simulation for magnetostatic problems.

Here, data-driven simulation is meant in the context of simulations directly on
the material data. In that manner, data-driven computing bypasses “the empirical
material modeling step of conventional computing altogether” [19, p. 81]. Consider
Maxwell’s equations for magnetostatics,

curl H = j , div B = 0 in Ω , (4)

where H denotes the magnetic field strength, j the imposed source current density,
B the magnetic flux density, and Ω the considered domain. Moreover, we assume
256 S. Kurz

suitable boundary conditions on Γ = ∂Ω, for instance n · B = 0, where n is the

normal vector to the boundary. :  1
The phase space of the system is denoted by Z := z(x) := B(x), H (x) ,
x ∈ Ω. The set of all states that fulfill (4) and the boundary conditions is
denoted by M ⊂ Z , the set of Maxwell-conforming fields1 . To uniquely solve
equations (4), a relation between B and H is necessary. In a conventional approach,
B and H are connected through an empirically determined constitutive law. This
law is mostly constructed as a fit with splines or regression techniques through
raw measurement data. In contrast to the conventional approach, the data-driven
field solver
: acts directly on the data. The 1 measurement data is collected in a set
D ∗ := z∗i := B ∗i , H ∗i , i = 1, . . . , N , where N is the number: of measurement
points. This gives
1 rise to a set of discrete material states D := z ∈ Z | z(x) ∈
D ∗ ∀x ∈ Ω . “The material response is not known exactly and, instead, it is
imperfectly characterized” [19, p. 95] by the set D.
The solution is given by the states M ∩ D that fulfill Maxwell’s equations,
while being compatible with the material states. However, for a finite number of
data points, this set is very likely empty, M ∩ D = ∅. Therefore, we define the
solution S by the relaxed condition
: 1
S := argmin d(z, D) , z ∈ M , (5)

where the distance function

1
B − B ∗ 2 + 1
H − H ∗ 2
d(z, z∗ ) := ν̃ μ̃
(6)
2 2
is defined in terms of auxiliary norms  · ν̃,μ̃ . They do not represent material
properties but are rather chosen to improve the convergence of the numerical
scheme. The solution of (5) is organized as a fixed point iteration, cf. Fig. 3 right.
For any given state z ∈ Z a modified FE solver is used to compute the state
z* ∈ M such that d(z, z* ) = min. The solver is based on a variational principle
discussed in [20]. The idea is to solve Ampère’s and Faraday’s laws exactly and
shift the discretization error entirely into the constitutive relation.2 Given z* , a
discrete optimization selects the closest measurement data points, cf. Fig. 3 left.
These so-called active measurement data are associated with a state z× ∈ D such
that d(z× , z* ) = min. State z = z× is the starting point for the next iteration. Under
convexity assumptions this algorithm converges to the solution of (5). Furthermore,
it has been shown in [22] for linear elasticity that an increasing set of measurement
data recovers the conventional solution.

1We do not delve into regularity considerations or functional analytical frameworks here.
2This formulation was proposed on the Compumag Conference 1983 in Genoa. The related
variational principle was called “Ligurian”, in honor of the Genoa region, and in similarity to
“Lagrangian”. [21, p. 49].
Hybrid Modeling 257

Fig. 3 Iterative data-driven solver. Left: Measured B(H )-characteristic. Active measurement data
(red crosses) are closest to given field points (blue circles). Right: The outer fixed point iteration
combines solutions of a variational principle by a modified FE solver (blue circles) with discrete
optimizations that select states associated with active measurement data (red crosses)

In [18], a quadrupole magnet was analyzed in 2D by the proposed method.

As a baseline, a standard magnetostatic finite element solution was considered, by
discretizing 1/8 of the geometry with 6k piecewise linear elements. The non-linear
system was solved by 20 Newton steps. For the novel method, data was created
by an equidistant sampling of the given non-linear B(H )-characteristic, without
adding noise. The relative error of H decreased almost inversely with N. However,
depending on N = 102 . . . 104 the proposed method required 10 . . . 500 outer
iterations for convergence. As an advantage, the nonlinearity could be naturally
included, and exactly known (air) and data-driven (iron) material information could
be combined.

4 Bayesian Free-Shape Optimization [23]

Bayesian optimization (BO) is an optimization method to optimize a given function

which is expensive to evaluate [24]. It is built upon a hybrid architecture that
blends intricate physical models with a Bayesian machine learning technique, such
as Gaussian Process (GP) regression. The resulting surrogate models are cheap to
evaluate, including derivatives, and keep track of their interpolation uncertainty. The
core idea of BO is to successively refine those surrogates in regions of design space
that are close to optimal, which are however not known beforehand. Regions with
high surrogate uncertainty might be optimal even though the mean interpolation
says otherwise. Thus, surrogate refinement requires balancing exploration against
exploitation during sampling, the so-called bandit problem3. There are different
strategies for achieving a good balance. One strategy considers the best value

3 In this model problem, an array of slot machines is considered. The gambler must balance the
goal to find the slot machine with the highest gain (exploitation) with the goal to achieve good
results on every play (exploration).
258 S. Kurz

1 EEC 2
D 1 2 D
C 1’ 2’ C

Fig. 4 Schematic for system simulation: extracted equivalent electrical circuit (EEC) of the trace
pair, surrounded by mode converters

ports 1&1’
Gradient
fixed outer vectors at (ii) Mixed
trace
free surface: mode
inner trace w/ (i) Reflecton conversion
free surface
ports 2&2’

Fig. 5 Trace pair with bend. Left: The outer trace is fixed, the inner trace has a free interior surface.
Right: Shape gradient vectors at the free surface for the two objectives (i) and (ii)

achieved so far and computes the Expected Improvement (EI). The next sample is
taken at the point with the largest EI; this yields yet another optimization problem.
The BO algorithm stops if the EI drops below some threshold. The BO approach can
be generalized in various ways, such as BO with noise, BO in several dimensions,
and BO for several objectives.
As an industrial example we consider BO of a differential trace pair on a printed
circuit board. Differential signalling benefits from high immunity against electro-
magnetic interference and low crosstalk. However, bend discontinuities in transmis-
sion lines introduce (i) reflection and (ii) differential-to-common-mode conversion.
An optimal design hence requires multi-objective optimization of the geometry.
A parametric case was studied in [25], while we aim at free-shape optimization.
Figure 4 shows a schematic for system simulation. The trace pair with ports 1,1’
and 2,2’, respectively, is described by an equivalent electrical circuit (EEC). Mode
converters admit a separation of differential mode (D) and common mode (C) signal
components. The optimization objectives can be stated in terms of S-parameters: 4
! !
(i) reflection |SDD11| = min; (ii) mixed mode conversion |SCD21 | = min.
The geometric setting is depicted in Fig. 5 left. The outer trace is fixed, while the
inner trace has a free interior surface. The geometry is described by a finite element

4 For simplicity evaluated at a fixed frequency of f = 500 MHz.

Hybrid Modeling 259

mesh, and the free surface can be re-shaped by mesh morphing. This corresponds
to a high-dimensional design space with ≈200 dimensions. This should be put
in contrast to the six-dimensional design space that was considered in [25]. The
optimization problem is: Find the Pareto front for the shape of the free surface that
minimizes the objectives.
The ingredients for solving the optimization problem are: finite element electro-
magnetic field solver, EEC extraction, and adjoint sensitivity analysis. Figure 5 right
shows the shape gradient vectors at the free surface for the two objectives (i) and
(ii). The two gradient vector fields point in opposite directions, so the objectives are
conflicting. However, the gradient fields are not exactly negatives of each other, so
there is still subtle room for improvement.
The BO is extended to the multi-objective case as follows. The Pareto front is
approached via a sequence of auxiliary optimization problems, each with respect to
a certain 2D affine subspace of the high-dimensional design space. This particular
affine subspace is spanned by the adjoint-based gradients; it is the subspace of
maximum objective variance. For each optimization problem of the sequence, BO
learns and optimizes GP surrogate models for the objective functions, restricted to
this subspace. Once the intermediate Pareto front is converged in this subspace,
new subspaces may be chosen on the intermediate Pareto front. Figure 6 shows the
result of this algorithm, after only ≈100 design evaluations. Note that even subtle
improvement potentials will be exploited by the hybrid free-shape optimizer.
(ii) Mixed mode conversion

Low
Starng point:
reflecon
parametrically
Selected opmized
optmized design
design Low
conversion

(i) Reflecon

Fig. 6 Converged Pareto front for the trace pair with bend (green dots). The algorithm was started
with a parametrically optimized design (black dot). An optimized design from the Pareto front was
selected as an example (red dot)
260 S. Kurz

5 Summary and Outlook

We highlighted three examples of hybrid modeling, cf. Table 2. As an outlook, we

will discuss hybrid models from other applied mathematics and industrial domains.
They will soon find their way into scientific computing in electrical engineering,
too.
Physics-Informed Neural Networks [26] This method utilizes a fully-connected
neural network (NN) to map a space-time domain to the unknown solution of an
initial- and boundary-value problem. The NN is inserted into the governing partial
differential equation (PDE) or variational principle and symbolically differentiated.
This yields another NN, with modified activation functions but identical parameters,
a so-called physics-informed NN. No labelled data is required for training. Rather,
a combined loss function is minimized. One component is associated with the
initial and boundary conditions, the other either with the residual norm or the
variational functional of the PDE. The latter component enforces the structure
of the physics equation. The solution is mesh-free and analytical. More general
network architectures, such as convolutional encoder-decoder NN’s are discussed
in [27]. Physics-informed NN’s benefit from their prior knowledge (also known as
inductive bias) that helps them overcoming the challenges of generalization and
data-efficiency. In fact, such hybrid models require only relatively small training
data, typically a few hundred up to a few thousand points [26, p. 688].
Embedding Physics Simulation into Deep Learning [28] This work is motivated
by control engineering, in particular by the development of intelligent reinforcement
learning agents. “The end result is that we can embed an entire physical simulation
environment as a layer in a deep network, enabling agents to both learn the
parameters of the environments to match observed behavior and improve control
performance via traditional gradient based learning.” [28, p. 2]. The main ingredient
is an adjoint-based solver, which allows efficient backpropagation of gradients and
avoids their tedious computation by finite differences. Then, deep convolutional
neural networks can be integrated seamlessly with physics-based models in machine
learning platforms such as PyTorch and TensorFlow.

Table 2 Three examples of hybrid modeling

Section Physics Data Approach
2 2D magnetostatics: Fields measured by Synthesis of physics
BEM discretization Hall and data by Kálmán
probe or moving coil update
3 2D magnetostatics: Measured material Projection of data
Ampère’s & Gauss’s data   into admissible
law points B ∗i , H ∗i physics manifold
4 Electromagnetic Data sampling from Bayesian
Darwin physics model Optimization:
model GP machine learning
Hybrid Modeling 261

Fig. 7 Field inversion and machine learning. Step (1): Identify the model errors ε i (x, t) as defined
in (7) for given time series Y i , by solving PDE-constrained optimization problems. Step (2): Train
a NN that describes a mapping from state variables (more precisely: features thereof) to model
errors, by using the results from the previous step. Step (3): Include the correction operator ε̂ in the
state equation

Field Inversion and Machine Learning (FIML) [29] This method stems from
computational fluid dynamics (CFD). For turbulent flows one may either solve
Navier-Stokes equations by direct numerical simulation (DNS) or large eddy
simulation (LES). This approach is accurate but numerically expensive, since it
involves a range of space and time scales. On the other hand, one may use the
Reynolds-averaged Navier-Stokes (RANS) method, where turbulence effects are
accounted for by phenomenological models rather than first principles. This method
is much more efficient but less accurate. With the help of FIML, both approaches
can be combined.
Going beyond CFD, on an abstract level, let some system dynamics be governed
by a low-fidelity state equation of the form
 
∂t + D u(x, t) = ε(x, t) , (7)

where ∂t is the time derivative, D is the differential operator in space, u(x, t) is the
state variable, and ε(x, t) is the (unknown) model error. We consider a discretized
setting. Assume that an observable y(u) is defined by some functional of the state
variable, and several observed time series Y i , i = 1, . . . , N are available, either
measured or from high-fidelity simulation. The idea of FIML is to learn a correction
operator ε̂ to account for the model error, cf. Fig. 7. Note that the NN does not
directly operate on the state variable, but rather on some low-dimensional feature
set f (u). Some achievements, limitations and further developments of this method
applied to airfoil modeling can be found in [30].
Epilog We have discussed hybrid modeling mainly from a physics-based perspec-
tive, where significant advantages could be achieved by joining with data-driven
models. Conversely, hybrid modeling is also beneficial from the standpoint of
industrial AI. In contrast to consumer AI, industrial AI focuses on smart products
and their creation. Such AI should be robust, that is sufficiently tolerant against
perturbations, and explainable, that is, the AI function can be made comprehensible
262 S. Kurz

to humans. Industrial AI hence calls for inclusion of domain knowledge by hybrid

modeling.

Acknowledgments Support in preparing the examples as well as inspiring discussions with the
following colleagues are acknowledged: Armin Galetzka (TU Darmstadt); Andreas Klaedtke,
Xiaobai Li, Manuel Schmidt (Bosch Corporate Research); Melih Kandemir, Zico Kolter (Bosch
Center for Artificial Intelligence); Melvin Liebsch (CERN).

References

1. M. Walker, Hype cycle for emerging technologies, 2018. Tech. Rep. G00340159, Gartner
Research (2018). https://www.gartner.com/en/documents/3885468/hype-cycle-for-emerging-
technologies-2018
2. P.V. Coveney, E.R. Dougherty, R.R. Highfield, Big data need big theory too. Philos. Trans.
R. Soc. A: Math. Phys. Eng. Sci. 374(2080), 20160,153 (2016). https://royalsocietypublishing.
org/doi/abs/10.1098/rsta.2016.0153
3. W. E, Machine learning: Mathematical theory and scientific applications, in ICIAM – Inter-
national Congress on Industrial and Applied Mathematics (2019). https://web.math.princeton.
edu/~weinan/ICIAM.pdf
4. K. Willcox, Predictive data science for physical systems – from model reduction to scientific
machine learning (2019), in ICIAM – International Congress on Industrial and Applied Mathe-
matics (2019). https://kiwi.oden.utexas.edu/papers/Willcox-Predictive-Data-Science-ICIAM-
2019.pdf
5. C.F. Higham, D.J. Higham, Deep learning: an introduction for applied mathematicians. SIAM
Rev. 61(4), 860–891 (2019)
6. Wikipedia contributors: Newton’s laws of motion — Wikipedia, the free encyclope-
dia (2019). https://en.wikipedia.org/w/index.php?title=Newton27s_laws_of_motion&oldid=
915580692 [Online. Accessed 26 September 2019]
7. C.M. Bishop, Pattern Recognition and Machine Learning (Springer, New York, 2006)
8. N. Guarino, D. Oberle, S. Staab, What is an ontology?, in Handbook on Ontologies (Springer,
New York, 2009), pp. 1–17
9. S. Lee, N. Baker, Basic research needs for scientific machine learning: core technologies for
artificial intelligence. Tech. rep., USDOE Office of Science (SC)(United States) (2018). https://
www.osti.gov/servlets/purl/1484362
10. R. Swischuk, L. Mainini, B. Peherstorfer, K. Willcox, Projection-based model reduction:
Formulations for physics-based machine learning. Comput. Fluids 179, 704–717 (2019)
11. R.E. Kálmán, A new approach to linear filtering and prediction problems. Trans. ASME–J.
Bas. Eng. 82, 35–45 (1960)
12. A. Stuart, The legacy of Rudolph Kálmán – blending data and mathematical models, in Boeing
Distinguished Colloquia, Univ. Washington (2019). https://www.sfb1294.de/fileadmin/user_
upload/Kalman_Lectures/1st_Kalman_Lecture_2018_Andrew_Stuart.pdf
13. W. Pieczynski, F. Desbouvries, Kálmán filtering using pairwise Gaussian models. in 2003 IEEE
International Conference on Acoustics, Speech, and Signal Processing, 2003. Proceedings.
(ICASSP’03), vol. 6, pp. VI–57–VI–60 (IEEE, New York, 2003)
14. E. Tonti, The Mathematical Structure of Classical and Relativistic Physics (Springer, New
York, 2013)
15. E. Tonti, Discrete physics – algebraic formulation of physical fields (2014). http://www.
discretephysics.org/en/. [Online; Accessed 26 September 2019]
16. M. Liebsch, S. Russenschuck, S. Kurz, Boundary-element methods for field reconstruction in
accelerator magnets. IEEE Trans. Magn. 56(3), 1–4 (2020)
Hybrid Modeling 263

17. J.M. Bardsley, Computational Uncertainty Quantification for Inverse Problems, vol. 19
(SIAM, New York, 2018)
18. H. De Gersem, A. Galetzka, I.G. Ion, D. Loukrezis, U. Römer, Magnetic field simulation with
data-driven material modeling (2020). Preprint. arXiv: 2002.03715
19. T. Kirchdoerfer, M. Ortiz, Data-driven computational mechanics. Comput. Methods Appl.
Mech. Eng. 304, 81–101 (2016)
20. J. Rikabi, C. Bryant, E. Freeman, An error-based approach to complementary formulations of
static field solutions. Int. J. Numer. Methods Eng. 26(9), 1963–1987 (1988)
21. B. Trowbridge, Compumag conference – the first 25 years (2001). https://www.compumag.
org/wp/wp-content/uploads/2018/07/TwentyFiveYearsOfCompumag.pdf. [Online; Accessed
16 April 2020]
22. S. Conti, S. Müller, M. Ortiz, : Data-driven problems in elasticity. Arch. Ration. Mech. Anal.
229(1), 79–123 (2018)
23. S. Schuhmacher, A. Klaedtke, C. Keller, W. Ackermann, H. De Gersem, Adjoint technique for
sensitivity analysis of coupling factors according to geometric variations. IEEE Trans. Magn.
54(3), 1–4 (2018)
24. P.I. Frazier, Bayesian optimization, in Recent Advances in Optimization and Modeling of
Contemporary Problems (INFORMS, Catonsville, 2018), pp. 255–278
25. C. Gazda, I. Couckuyt, H. Rogier, D.V. Ginste, T. Dhaene, Constrained multiobjective
optimization of a common-mode suppression filter. IEEE Trans. Electromagn. Compatibil.
54(3), 704–707 (2012)
26. M. Raissi, P. Perdikaris, G.E. Karniadakis, Physics-informed neural networks: a deep learning
framework for solving forward and inverse problems involving nonlinear partial differential
equations. J. Comput. Phys. 378, 686–707 (2019)
27. Y. Zhu, N. Zabaras, P.S. Koutsourelakis, P. Perdikaris, Physics-constrained deep learning for
high-dimensional surrogate modeling and uncertainty quantification without labeled data. J.
Comput. Phys. 394, 56–81 (2019)
28. F. de Avila Belbute-Peres, K. Smith, K. Allen, J. Tenenbaum, J.Z. Kolter, End-to-end
differentiable physics for learning and control, in Advances in Neural Information Processing
Systems (2018), pp. 7178–7189
29. E.J. Parish, K. Duraisamy, A paradigm for data-driven predictive modeling using field inversion
and machine learning. J. Comput. Phys. 305, 758–774 (2016)
30. J.R. Holland, J.D. Baeder, K. Duraisamy, Towards integrated field inversion and machine
learning with embedded neural networks for RANS modeling, in AIAA Scitech 2019 Forum
(2019), p. 1884
Machine Learning for Initial Value
Problems of Parameter-Dependent
Dynamical Systems

Roland Pulch and Maha Youssef

Abstract We consider initial value problems of nonlinear dynamical systems,

which include physical parameters. A quantity of interest depending on the solution
is observed. A discretisation yields the trajectories of the quantity of interest in
many time points. We examine the mapping from the set of parameters to the
discrete values of the trajectories. An evaluation of this mapping requires to solve
an initial value problem. Alternatively, we determine an approximation, where the
evaluation requires low computation work, using a concept of machine learning.
We employ feedforward neural networks, which are fitted to data from samples of
the trajectories. Results of numerical computations are presented for a test example
modelling an electric circuit.

1 Introduction

We examine initial value problems of nonlinear dynamical systems consisting of

ordinary differential equations (ODEs) or differential-algebraic equations (DAEs).
The systems include physical parameters, which vary in a predetermined bounded
domain. A quantity of interest (QoI) is defined depending on the solution of
the dynamical system. Hence there is a mapping from the parameters onto the
trajectories of the QoI in the time domain. Our aim consists in the determination of
an approximation of this mapping, which can be evaluated with a low computational
effort. These approximations can be applied as surrogate models in uncertainty
quantification, see [7], for example.
Methods using polynomials and their orthogonal bases yield surrogate models
of parametric problems, see [9, 10]. Alternatively, we employ an approach of
machine learning using artificial neural networks (NNs), see [2, 3], to construct
an approximation. In [11, 12], proper orthogonal decomposition (POD) is applied

R. Pulch () · M. Youssef

Universität Greifswald, Institute of Mathematics and Computer Science, Greifswald, Germany
e-mail: [email protected]; [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 265
M. van Beurden et al. (eds.), Scientific Computing in Electrical Engineering,
Mathematics in Industry 36, https://doi.org/10.1007/978-3-030-84238-3_26
266 R. Pulch and M. Youssef

to data of solutions of parametric partial differential equations to obtain a reduced

basis. Consequently, a mapping between low-dimensional spaces is approximated
by NNs. In contrast, we discretise in time and use the data from the trajectories of
the QoI in many time points. A mapping from a low-dimensional parameter space
to a high-dimensional space is approximated by an NN now.
We apply feedforward NNs for this approximation. The determination of an NN
requires a training procedure using data. We obtain the data of the trajectories
by solving initial value problems of the dynamical systems for samples of the
parameters. The fitting of an NN represents a (nonlinear) optimisation problem. In
an NN, the number of neurons in a hidden layer is typically larger than the number
of neurons in the input layer or the output layer. Since the number of outputs is large
in our case, we also have high numbers of neurons in the hidden layers.
Finally, we demonstrate numerical results for a test example, which is a DAE
model of an electric circuit. The trajectories associated to some parameter samples
are illustrated. We show statistics of the approximation errors.

2 Parameter-Dependent Dynamical Systems

Let parameters p ∈ Π ⊂ Rq be given. We consider a nonlinear dynamical system

of the form

 x˙ (t, p) = f(t, x(t, p), p).

M(p) (1)

The mass matrix M  : Π → Rn×n and the right-hand side f : [t0 , tf ]×Rn ×Π → Rn
include the parameters. Thus the solution x : [t0 , tf ] × Π → Rn depends both on
time and the parameters. If the mass matrix is non-singular, then (1) represents a
system of ODEs. If the mass matrix is singular, then a system of DAEs is given. We
examine initial value problems (IVPs)

x(t0 , p) = x0 (p). (2)

In the case of DAEs, the initial values have to be consistent, see [5]. Consistent initial
values often depend on the parameters. We define a QoI y : [t0 , tf ] × Π → Rn by a
function g : Rn → R via

y(t, p) = g(
x (t, p)). (3)

Each selection of the parameters yields a trajectory of the QoI in the time domain.
We obtain the mapping

p
→ {(t, y(t, p)) : t ∈ [t0 , tf ]} (4)
Machine Learning for Parameter-Dependent Dynamical Systems 267

for any p ∈ Π. Our aim is to construct an approximation of the mapping (4), which
can be evaluated cheap, in particular, without solving IVPs (1), (2) any more.
The following approach can also be used for boundary value problems (BVPs)
of dynamical systems, because we only include the trajectories of the QoI in the
method. The trajectories are computed from either IVPs or BVPs.

3 Time Discretisation

We discretise the trajectories of the QoI (3) in the time domain [t0 , tf ]. Let

t0 < t1 < t2 < · · · < tm−1 < tm ≤ tf . (5)

Equidistant time points can be used. We consider the mapping Θ : Π → Rm

⎛ ⎞
y(t1 , p)
⎜ .. ⎟
Θ : Π → Rm , Θ(p) = ⎝ . ⎠. (6)
y(tm , p)

Each evaluation of (6) requires to solve an IVP (1), (2) followed by the extraction of
the QoI (3). The IVPs of the dynamical systems are solved by numerical methods,
see [4, 5], like Runge-Kutta schemes and linear multistep methods. The methods
yield approximations of the solution in discrete time points, which are typically
determined by a local error control. Thus these time points are not identical to our
choice (5). Nevertheless, we obtain the solution in the points (5) by an interpolation
or a dense output in time.
Stiff systems of ODEs and all DAEs require implicit methods in the time
integration. Therein, a nonlinear system of algebraic equations has to be solved
in each time step. Thus the computational effort becomes large. Our goal is to
determine an approximation of the mapping (6), whose evaluation is cheap.

4 Machine Learning

We arrange an artificial NN, see [3], to approximate the mapping (6). An NN

consists of an input layer, an output layer and additional hidden layers. Figure 1
illustrates the schematic of an NN. Let Nj be the number of neurons in the j th layer
for j = 0, 1, . . . , J . Therein, j = 0 and j = J represent the input layer and the
output layer, respectively. Thus there are J − 1 hidden layers. It holds that N0 = q
and NJ = m and thus NJ ' N0 in our problem.
268 R. Pulch and M. Youssef

Fig. 1 Artificial NN with input layer (red), hidden layers (green), and output layer (blue)

The mathematical model Ψ : RN0 → RNJ of an NN consists in a chain of

operators

Ψ = TJ ◦ ρ ◦ TJ −1 ◦ ρ ◦ TJ −2 ◦ · · · ◦ ρ ◦ T2 ◦ ρ ◦ T1 . (7)

Each operator Tj : RNj−1 → RNj is an affine-linear function

Tj (z) = A j z + bj

including a matrix Aj ∈ RNj ×Nj−1 , a vector bj ∈ RNj , and the input z ∈ RNj−1 . In
the context of machine learning, the entries of A j and bj are denominated as weights
and biases, respectively. The operator ρ is a nonlinear transfer function ρ : R → R
(also called activation function). Typical choices are, for example, the hyperbolic
tangent sigmoid function

ρ(x) = 2
1+e−2x
−1 (8)

and the hard-limit function

.
0 for x < 0,
ρ(x) = (9)
1 for x ≥ 0.

In (7), the function ρ is applied to vectors in each component separately.

In the fitting of an NN, the ideal is to minimise the distances Θ(p) − Ψ (p) for
any p ∈ Π. The degrees of freedom are the weights and biases, i.e., (Aj , bj )Jj=1 .
Since a nonlinear optimisation problem appears, iterative methods are required to
obtain numerical solutions.
In practise, the fitting involves three sample sets for training, validation, and test:

Strain = {p1 , . . . , pk } ⊂ Π
Svalid = {
q1 , . . . , qk  } ⊂ Π (10)
Stest = {r1 , . . . , rk  } ⊂ Π.
Machine Learning for Parameter-Dependent Dynamical Systems 269

For example, random samples can be chosen, where a uniform probability distri-
bution is assumed in the parameter domain Π. The minimisation is based on the
differences Θ(pi )−Ψ (pi ) for parameter tuples pi from the training set. The (vector-
valued) differences are measured using the mean squared error or the mean absolute
error. The error measure decreases monotone for the parameters in the training set
due to the minimisation. The validation set is included to prevent an overfitting. If
the error measure of the validation set increases, then the training is stopped and the
best previous case is put out. The test set is not involved in the minimisation at all.
Hence this set allows for an estimate of the quality of the trained NN.

5 Numerical Results for Test Example

All numerical computations were performed within the software MATLAB [8]
using the Deep Learning Toolbox.
We investigate an electric circuit introduced in [1], which is illustrated by Fig. 2.
This circuit performs a voltage doubling for specific choices of parameters and input
voltage. A mathematical modelling yields a nonlinear system of DAEs (1) with
n = 3 equations for the three unknown node voltages presented in [1]:

C1 ẋ1 = − Rx12 + F (−(x1 + x3 ))

C2 ẋ2 = − R11 (x2 + x3 + uin ) (11)
0 = − R11 (x2 + x3 + uin ) + F (−(x1 + x3 )) − F (x3 ).

3 2 1

R1 C2

uin C1

0
R2

Fig. 2 Diagram of electric circuit

270 R. Pulch and M. Youssef

The current–voltage relation of the diodes reads as F (u) = γ (exp(δu) − 1). We

use the constant parameters γ = 4.067 · 10−8 and δ = 5.634 · 10−2 given in [6].
The differential index of the DAE system (11) is one. We choose the second node
voltage as QoI (3).
We consider variations in four physical parameters: the two capacitances and the
two resistances. The ranges Cj ∈ [2·10−9, 3·10−9 ] for j = 1, 2, R1 ∈ [106 , 2·106],
R2 ∈ [108, 2 · 108 ] form the parameter domain Π ⊂ R4 .
As input voltage, we supply the harmonic oscillation

uin (t) = A sin 2π
T t

with amplitude A = 500 and period T = 0.1. The total time interval of our
simulations is [t0 , tf ] = [0, 0.5]. The initial values (2) are set to zero, which
represents a consistent case in this example. The backward differentiation formulas
(BDF), see [4], yield the numerical solutions of the IVPs. High accuracy requests
are imposed in the local error control with relative tolerance εrel = 10−4 and
absolute tolerance εabs = 10−6 . The error control generates approximations on
a non-uniform grid in time. We extract the trajectories of the QoI in m = 200
equidistant time points t = Δt for = 1, . . . , m with Δt = tf −t m
0
by
interpolation. The order of accuracy coincides for both the uniform grid and non-
uniform grid. The associated error of the time integration is negligible in comparison
to the approximation error of the NNs below. Figure 3 gives an impression of the
variability within the trajectories of the QoI for our parameter domain.
We select the number of samples as k = k  = k  = 500 in the sets (10).
Often the validation set and the test set are chosen smaller than the training set
due to a restricted amount of data. In contrast, we are able to use larger sets,
since a high number of trajectories can be produced by numerical simulations. In
particular, a large test set provides reliable statistics in the error analysis. A pseudo
random number generator yields the parameter samples in the multidimensional
cuboid Π. Our NNs include two hidden layers with 400 neurons in each layer.

Fig. 3 Twenty trajectories of

the QoI for different
parameter samples
400

200

-200

0 0.1 0.2 0.3 0.4 0.5

Machine Learning for Parameter-Dependent Dynamical Systems 271

106 108
Train Train
Validation
105 107 Validation
Test Test
Best Best
106
104
105
103
104
2
10
103

101 102
0 100 200 300 400 500 600 700 0 50 100 150 200 250

(i) (ii)

Fig. 4 Mean squared errors during the fitting of the two NNs in the iterative minimisation (The
green line of the validation set is mostly located behind the red line of the test set.)

Table 1 Number of iteration steps and mean squared error (MSE) of test set for different training
methods in NNs with hard-limit transfer function (i) and purely linear transfer function (ii)
(i) (ii)
Steps MSE Steps MSE
Conjugate gradient method 728 1067.5 254 277.21
One-step secant method 1696 1062.2 328 277.24
Gradient descent method 10000 1323.1 1413 277.74

Using more hidden layers or more neurons did not improve the results significantly.
We investigate two NNs, which differ only in the choice of the transfer function:
(i) hard-limit transfer function (9),
(ii) purely linear transfer function.
In the training, a conjugate gradient backpropagation method iteratively solves the
minimisation problem. Figure 4 shows the performance of the training procedure.
In the case of the hard-limit transfer function, the training is stopped at the 728th
iteration step, because the error of the validation set increases slightly. In the case of
the linear transfer function, the training is terminated at the 254th iteration step due
to a too small step size. These two NNs are used in the following error analysis.
In addition, we tried two other backpropagation techniques in the training of
the NNs: a one-step secant method (quasi Newton method) and a gradient descent
method with momentum and adaptive learning rate. More information on all three
methods can be found in [2], for example. Table 1 demonstrates the number of
iteration steps (until a termination criterion applies) as well as the final mean squared
error of the test set for the three techniques. We observe that the conjugate gradient
method exhibits the best performance.
Figure 5 illustrates several trajectories of the test set. A comparison of the exact
trajectories from the time integration and the approximations from the two NNs is
shown. An interesting property is that NN (i) with the nonlinear transfer function
272 R. Pulch and M. Youssef

400 400

200 200

0 0

-200 -200

0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5

400 400

200 200

0 0

-200 -200

0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5

Fig. 5 Trajectories of the QoI for some samples from the test set

exhibits an oscillatory behaviour of the approximations in time, whereas NN (ii)

with the linear transfer function generates more smooth approximations in time.
The incorrect behaviour is not a transient effect, because all discrete time points (5)
are treated equally in the NNs, i.e., without the consideration of their ordering in
time. We also tried other nonlinear transfer functions (like (8)), which still caused
oscillations in time.
Finally, we quantify the relative errors of the approximations for each sample
trajectory using a discrete L 1 -norm in time. The error associated to the ith
parameter sample reads as

tf − t0  |ỹ(t , pi ) − y(t , pi )|
m
Ei = , (12)
m |y(t , pi )|
=1

where y is the original value from the time integration and ỹ denotes the approxima-
tion from an NN. The initial value is not included due to its value zero. The statistics
of the errors are depicted in Table 2. We discuss the resulting mean values. In NN (i),
a smaller mean error is achieved in the training set, whereas the other two sets show
larger errors in comparison to NN (ii). Moreover, the mean errors are balanced for
all three sets in NN (ii). This behaviour is in agreement to the performance of the
training demonstrated by Fig. 4.
Machine Learning for Parameter-Dependent Dynamical Systems 273

Table 2 Mean value and standard deviation of relative errors in discrete L 1 -norm, see (12), for
the three parameter sets within the two trained NNs
Mean St.dev.
NN (i) NN (ii) NN (i) NN (ii)
Training set 0.044 0.086 Training set 0.135 0.228
Validation set 0.130 0.086 Validation set 0.210 0.158
Test set 0.120 0.079 Test set 0.155 0.146

6 Conclusions

We arranged a mapping from a set of parameters to discrete values of a QoI obtained

from IVPs of dynamical systems. We approximated this mapping by artificial NNs.
A test example was investigated, where two NNs were trained. In both NNs, the
quality of the approximations is moderate with respect to the mean values of the
errors. However, the NN including a linear transfer function yielded more smooth
discretised trajectories in time, whereas the NNs with nonlinear transfer functions
produced incorrect oscillations in time.

References

1. B. Barz, E. Suschke, Numerische Behandlung eines Algebro-Differentialgleichungssystems.

RZ-Mitteilungen, vol. 7 (Humboldt-Universität, Berlin, 1994)
2. K.-L. Du, M.N.S. Swamy, Neural Networks and Statistical Learning (Springer, London, 2014)
3. I. Goodfellow, Y. Bengio, A. Courville, Deep Learning (MIT Press, London, 2017)
4. E. Hairer, S.P. Nørsett, G. Wanner, Solving Ordinary Differential Equations I: Nonstiff
Problems, 2nd edn. (Springer, Berlin, 1993)
5. E. Hairer, G. Wanner, Solving Ordinary Differential Equations II: Stiff and Differential-
Algebraic Equations, 2nd edn. (Springer, Berlin, 1996)
6. W. Kampowsky, P. Rentrop, W. Schmidt, Classification and numerical solution of electric
circuits. Surv. Math. Ind. 2, 23–65 (1992)
7. O.P. Le Maître, O.M. Knio, Spectral Methods for Uncertainty Quantification (Springer,
Dordrecht, 2010)
8. MATLAB, version 9.7.0.1190202 (R2019b) (The Mathworks Inc., Natick, MA, 2019)
9. R. Pulch, Polynomial chaos for the computation of failure probabilities in periodic problems,
in Scientific Computing in Electrical Engineering SCEE 2008 ed. by J. Roos, L. Costa.
Mathematics in Industry, vol. 14 (Springer, Berlin, 2010), pp. 191–198
10. T.J. Sullivan, Introduction to Uncertainty Quantification (Springer, Cham, 2015)
11. Q. Wang, J.S. Hesthaven, D. Ray, Non-intrusive reduced order modeling of unsteady flows
using artificial neural networks with application to a combustion problem. J. Comput. Phys.
384, 289–307 (2019)
12. J. Yu, J.S. Hesthaven, Flowfield reconstruction method using artificial neural network. AIAAJ
57, 1–17 (2019)
Modeling of Thermoelectric Generator
via Parametric Model Order Reduction
Based on Modified Matrix Interpolation

Ananya Roy, Gunasheela Sadashivaiah, Chengdong Yuan, M. Nabi,

and Tamara Bechtold

Abstract In this paper, finite element modeling of a simplified human tissue

model consisting of muscle, fat and skin layers is carried out. A thermoelectric
generator is placed in the fat layer and functioned as a power supply for electrically
active implants. As the finite element method produces a large number of ordinary
differential equations, model order reduction becomes the only way out of the
computational complexity. In this work, the height of the thermocouples, which has
a large influence on the generated power, is considered as a geometrical parameter.
A modified matrix interpolation based parametric model order reduction method is
used to construct the reduced order model valid for an arbitrary parameter value.

1 Introduction

During the last years with the development of technology, energy harvesting systems
have become a popular area of research. They ensure longevity, eco-friendly
operation, low maintenance and have a wide range of applications from aircraft,
biosensors [1] to telemetry systems [2] etc. Here presented, thermoelectric generator
(TEG) acts as an alternative source of energy to provide stable power to electrically
active implants [3]. Choosing the proper geometry for the TEG is very important
aspect [4]. To come up with an adequate design, the impact of geometrical param-

A. Roy () · M. Nabi

Indian Institute of Technology, Delhi, New Delhi, India
e-mail: [email protected]; [email protected]
G. Sadashivaiah
University of Rostock, Rostock, Germany
e-mail: [email protected]
C. Yuan · T. Bechtold
University of Rostock, Rostock, Germany
Jade University of Applied Sciences, Wilhelmshaven, Germany
e-mail: [email protected]; [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 275
M. van Beurden et al. (eds.), Scientific Computing in Electrical Engineering,
Mathematics in Industry 36, https://doi.org/10.1007/978-3-030-84238-3_27
276 A. Roy et al.

eters needs to be analyzed. It has been seen that the thermoelectric performance
of the TEG is dependant on the height of the thermocouples [5]. With change in
height of the thermocouples, temperature difference between the top and the bottom
level changes and so the generated power. We wish to investigate this influence via
mathematical method of model order reduction (MOR).
The TEG is modelled as distributed parameter system via partial differential
equations. Finite element analysis converts these partial differential equations into
large-scale ordinary differential equation systems, which solution is computation-
ally costly. MOR derives the low-dimensional approximation of the higher order
original system [6, 7]. Furthermore, during design optimization, the system needs
to be simulated repeatedly for different values of geometrical parameters. If these
parameters can be preserved within the reduced models, then the full-scale system
must not be repeatedly synthesized and reduced at each parameter value. This idea
gives rise to parametric model order reduction (pMOR). In this paper, parametric
modeling of a TEG is carried out with modified matrix interpolation based pMOR
method. The parameter considered is the height of thermocouples.

2 Model Description

The human body is a thermal energy source. When the surrounding temperature
varies, the body temperature varies between 23 ◦ C (at the skin surface) and 37
◦ C (in the body core). Implantable TEG utilizes the temperature difference in the

body and generates electrical power. It is made of an array of thermocouples,

and each thermocouple consists of a p-type and n-type Bismuth Telluride. In this
work, the TEG model contains 8×8 thermocouple legs with cross-sectional area 1
mm2 , which are electrically connected in series through copper interconnects. The
thermocouples are thermally connected in parallel between two 17.5×17.5 mm2
ceramic plates. A schematic of TEG is shown in Fig. 1 together with the Peltier
height (the parameter of interest).

Heat absorption
Heat
flow

N P Peltier height
Copper
Ceramic plates interconnect

Thermoelectric
Heat emission components

(a) (b)

Fig. 1 (a) Schematic of a TEG; (b) schematic of a thermocouple

Modeling of Thermoelectric Generator 277

Based on Seebeck effect, electrons and holes in the thermocouples start moving
when a temperature gradient occurs. As a result, thermal energy is being converted
into electrical energy, which can be utilized to power electrical implants. The voltage
(V) generated by the TEG is given by:

V = n · ΔT · (α1 − α2 ) (1)

where, ΔT is the temperature difference between the top level and bottom level of
the TEG, n is the number of thermocouples, α1,2 are the Seebeck coefficients of the
thermocouple legs.
Here a simplified human tissue model is considered to study the behavior of the
TEG inside human tissue. The human tissue model consists of muscle, fat, and skin
layers as shown in Fig. 2. The TEG is surrounded with a 40×40 mm2 housing made
of Teflon and placed within the fat layer, as maximum temperature difference occurs
there [4].
The material properties of various parts in TEG and different human tissue are
shown in Tables 1 and 2.

Fig. 2 Schematic of a TEG

embedded in the fat layer of
the tissue model Housing Skin

Fat
8mm
35mm

2mm Muscle

Table 1 Material properties of various parts of TEG

Density Specific heat Thermal conductivity Seebeck coefficient
(kg/m3 ) (J/kg/K) (W/m/K) (µV/K)
Housing 2250 1000 0.25 –
Ceramic plates 3720 880 25 –
p-type thermocouple leg 7700 90 1.6 200
n-type thermocouple leg 7700 90 1.6 −200
278 A. Roy et al.

Table 2 Material properties of various tissue types

Density Specific heat Thermal conductivity Perfusion rate Metabolic heat
(kg/m3 ) (J/kg/K) (W/m/K) (1/s) (W/m3 )
Muscle 1090.4 3421.2 0.4949 3.37e−4 498.52
Fat 911 2348.3 0.2115 3.01e−4 279.8
Skin 1109 3390.5 0.3722 9.05e−4 841.57

The heat conduction in the human tissue is described by the Pennes’ bioheat
model [8] expressed as:

∂T
∇(κ∇T ) + ρb cb ω(Ta − T ) +Qm = ρc (2)
! "# $ ∂t
Qb (T )

where T is the resulting temperature field and κ, ρ and c are the thermal conduc-
tivity, density and specific heat of the tissue, respectively. The heat generation rates
provided by metabolism and perfusion are described by Qm and Qb (T ). The density
and specific heat of blood are expressed as ρb = 1049.75 kg/m3 and cb = 3617
J/kg/K. ω is the blood perfusion rate in different tissue layers. Blood temperature
Ta = 37 ◦ C is set as temperature boundary condition at bottom surface of the tissue
model. Note that the temperature dependent perfusion effect Qb (T ) can be applied
as the ‘convection-type’ effect as introduced in [9]. The value of the metabolic heat
generation rates Qm in different tissue layers are introduced in Table 2.
The heat is dissipated by convection at the skin surface as the external heat loss
effect:

qconv = h · (T − Tamb ) (3)

where qconv is the heat flux normal to the boudary skin surface. The heat transfer
coefficient is expressed by h and the ambient temperature by Tamb . The steady
state solution of the system (2), (3) is taken as the initial condition for the transient
thermal simulation.
The finite element discretization of Eq. (2) with convection boundary condi-
tion (3) leads to a following large-scale system of ordinary differential equations:

EṪ (t) = AT (t) + Bu(t) (4)

y(t) = CT (t)

where, E, A ∈ Rn×n are the heat capacity and heat conductivity matrices,
respectively, B ∈ Rn×m is the input distribution matrix and C ∈ Rp×n is the user-
defined output matrix. In this work, the order of the model, n ≈ 4 × 104 changes
with the Peltier height, is very large, and T (t) ∈ Rn is the state vector of unknown
nodal temperatures.
Modeling of Thermoelectric Generator 279

3 Parametric Model Order Reduction with Matrix

Interpolation

The performance of the TEG is influenced by the height of the thermocouple.

In Fig. 3, it can be seen that if the height changes, the temperature difference
between the top level and the bottom level of the thermocouple changes, as well
as the generated voltage. Hence the length of the thermocouple is considered as a
geometric parameter here.
Ideally, the pMOR should be able to cope with an arbitrary number of parameters
and allow for situations in which the matrix dependence on parameters can not be
expressed analytically. One of such matrix-interpolation-based pMOR schemes has
been introduced in [10] and later refined in [11]. This refined scheme is applied in
the modeling of TEG.
The system is sampled at k different values of parameter of interest. For each
value of parameter a single large-scale finite element model is generated as:

Ei Ṫi (t) = Ai Ti (t) + Bi u(t) (5)

yi (t) = Ci Ti (t)

k locally reduced order models (ROMs) are generated by projecting each large-scale
system onto a lower order subspace. We have used one sided Arnoldi algorithm [12],
which generates a transformation matrix Vi ∈ Rni ×r , where ni is the order of each
large-scale system and r ( ni is the order of the corresponding ROM.

Fig. 3 Effect of peltier height of TEG

280 A. Roy et al.

On the basis of the method introduced in [10] and [11], the globally reduced
models after applying modified matrix interpolation method, can be written as:
E∗r,i A∗r,i B∗
# $! " # $! " # $! "
r,i
−1 ∗ −1 ∗
Mi Er,i Ti Ṫr (t) = Mi Ar,i Ti Tr (t) + Mi Br,i u(t) (6)
y(t) = Cr,i T−1 T ∗ (t)
! "# i $ r
C∗r,i

where Mi ∈ Rr×r and Ti ∈ Rr×r are transformation matrices which should be

chosen appropriately.
Once the globally reduced models (E∗r,i , A∗r,i , B∗r,i , C∗r,i ) at all the discrete points
i = 1, 2, . . . , k are obtained, reduce model at any parameter value p can be obtained
by using a weighted interpolation of the matrices of these local models as:

Er Ṫr∗ (t) = Ar Tr∗ (t) + Br u(t) (7)

y(t) = Cr Tr∗ (t)
k ∗ k ∗ k ∗
where, Er = i=1 wi Er,i , Ar = i=1 wi Ar,i , Br = i=1 wi Br,i , Cr =
k ∗
 k
i=1 wi Cr,i with i=1 wi = 1.

4 Simulation Results

In this work, a simplified human tissue model consisting of muscle, fat, and
skin layers is considered. The TEG is placed within the fat layer. A geometrical
parameter, the height of the thermocouple is varied from 3.65 mm to 3.95 mm
and discretized at 3.65, 3.75, 3.85, 3.95 mm. Large-scale finite element models are
generated at these discrete points by using ANSYS Mechanical [13]. Subsequently,
the corresponding ROMs of order 31, are generated by using “Model Reduction
inside ANSYS” [14]. Utilizing these ROMs through modified matrix interpolation
based pMOR algorithm, a global reduced order model is generated.
To verify the proposed method, an intermediate point is chosen at p = 3.8 mm.
A global reduced model is interpolated at this point and compared to the full-scale
model of order n = 44,942, with 3.8 mm Peltier height. To study the influence
of the convection boundary condition as mentioned in (3), an initial state of the
TEG is obtained with heat transfer coefficient h = 8.8 W/m2 /K through steady
state simulation. Afterwards, a transient simulation with heat transfer coefficient
h = 11.18 W/m2 /K is carried out for 7000 s. The ambient temperature is set as
constant Ta = 25 ◦ C. The transient thermal response at the top and the bottom
of the thermocouples, at intermediate point p = 3.8 mm, is shown in Fig. 4. For
ease of measurement, we have calculated the average temperature of the top and
bottom surfaces of the thermocouples. A comparative analysis of relative errors of
average temperatures are shown in Fig. 5. It can be seen that the relative error at the
Modeling of Thermoelectric Generator 281

Fig. 4 Transient thermal response at selected top and bottom nodes of thermocouple at interme-
diate point

Fig. 5 Relative error at different parameter values

intermediate point is around 0.1624%, which is accurate enough for the problem at
hand. Furthermore, the reduced models obtained at every discrete points, produce
results with still higher accuracy. This is expected, because the reduced models are
obtained at these very discrete points, while the model at the intermediate point is
calculated through interpolation.
282 A. Roy et al.

5 Conclusions and Outlook

In this paper, the potential design optimization strategy, based on modified matrix
interpolation pMOR, for a human TEG has been investigated. As the thermoelectric
performance of the device is highly affected by its Peltier height, we have chosen
this height as a parameter of interest. Modified matrix interpolation based pMOR
is applied to reduce computational complexity. A reduced model valid for an
arbitrary Peltier height is generated through this method. Numerical simulations
of the original large-scale model and its interpolated surrogate prove the efficacy of
the proposed method.
In the future work, we will incorporate the cross-sectional area of thermocouple
as another parameter and perform multiple-parameter model order reduction to get
an optimal design of the TEG.

Acknowledgments Financial support of the CRC 1270 ELAINE (Electrically Active Implants) is
acknowledged.

References

1. M. Koplow, A. Chen, D. Steingart, P.K. Wright, J.W. Evans, Thick film thermoelectric
energy harvesting systems for biomedical applications, in Proceedings of the 5th International
Summer School Symposium on Medical Devices and Biosensors (2008), pp. 322–325
2. S. Dalola, V. Ferrari, M. Guizzetti, D. Marioli, E. Sardini, M. Serpelloni, A. Taroni,
Autonomous sensor system with power harvesting for telemetric temperature measurements
of pipes. IEEE Trans. Instrum. Measure. 58(5), 1471–1478 (2009)
3. Y.W. Chong, W. Ismail, K. Ko, C.Y. Lee, Energy harvesting for wearable devices: a review.
IEEE Sensors J. 19(20), 9047–9062 (2019)
4. O. Jadhav, C.D. Yuan, D. Hohlfeld, T. Bechtold, Design of a thermoelectric generator for
electrical active implants, in MikroSystemTechnik Congress (2017), pp. 1–4
5. B. Jang, S. Han, J.Y. Kim, Optimal design for micro-thermoelectric generators using finite
element analysis. Microelectron. Eng. 88(5), 775–778 (2011)
6. W.H.A. Schilders, H.A. Van der Vorst, J. Rommes, Model Order Reduction: Theory, Research
Aspects and Applications, vol. 13 (Springer, Berlin, 2008)
7. B. Lohmann, B. Salimbahrami, Introduction to Krylov subspace methods in model order
reduction, in Methods and Applications in Automation (2000), pp. 1–13
8. C.K. Charny, Mathematical models of bioheat transfer, in Advances in Heat Transfer, vol. 22
(Elsevier, Amsterdam, 1992), pp. 19–155
9. C.D. Yuan, S. Kreß, G. Sadashivaiah, E.B. Rudnyi, D. Hohlfeld, T. Bechtold, Towards efficient
design optimization of a miniaturized thermoelectric generator for electrically active implants
via model order reduction and submodeling technique. Int. J. Numer. Methods Biomed. Eng.
36(4) (2020). https://doi.org/10.1002/cnm.3311
10. H. Panzer, J. Mohring, R. Eid, B. Lohmann, Parametric Model Order Reduction by
Matrix Interpolation. Automatisierung-stechnik Methoden und Anwendungen der Steuerungs-
, Regelungs- und Informationstechnik 58(8), 475–484. ISSN (Print) 0178–2312, August 2010.
https://doi.org/10.1524/auto.2010.0863
11. A. Roy, M. Nabi, Efficient simulation of electro-thermal micro-gripper using PMOR, in Indian
Control Conference (ICC), Kanpur, 2018
Modeling of Thermoelectric Generator 283

12. R.W. Freund, Krylov-subspace methods for reduced-order modeling in circuit simulation. J.
Comput. Appl. Math. 123(1–2), 395–421 (2000)
13. Ansys, Academic Research Mechanical, 2019 R3. ANSYS, Inc.
14. E.B. Rudnyi, Mor for ANSYS, in System-Level Modelling of MEMS, ed. by T. Bechtold, G.
Schrag, L. Feng. Wiley-VCH Book Series on Advanced Micro and Nanosystems (Wiley-VCH,
Weinheim, 2013), pp. 425–438
Nonlinear Model Order Reduction
of a Thermal Human Torso Model

Gunasheela Sadashivaiah, Chengdong Yuan, and Tamara Bechtold

Abstract Considering the aging populations in Europe, electrically active implants

play a major role in the modern medical field. Main drawback is their limited
battery period. Here presented thermoelectric generator, which is based on Seebeck
effect, utilizes temperature gradient inside the human body to generate electric
voltage. It is aimed to prolong the life-time of implants. Its design optimization
relies on an accurate thermal human torso model. In this work, we account
for physiologically correct thermal transfer effects such as, internal heat transfer
described by Pennes’ bioheat equation and external heat transfer due to convection,
radiation and sweating. Furthermore, realistic and temperature-dependent material
properties are applied to the human tissues and to the components of thermoelectric
generator. The goal of this work is to find an efficient low-rank approximation of
this complex, nonlinear model by proper orthogonal decomposition and dynamic
mode decomposition techniques.

1 Introduction

In the last couple of years, aging of the population is the main concern especially in
European countries [1]. Concerning this, various developments in the medical sector
were provoked. The development of electrically active implants is a special boon
in regeneration therapies and deep brain stimulation to treat movement disorders.
Among the various factors affecting the performance of implants, their limited

G. Sadashivaiah ()
Institute for Electronic Appliances and Circuits, University of Rostock, Rostock, Germany
e-mail: [email protected]
C. Yuan · T. Bechtold
Institute for Electronic Appliances and Circuits, University of Rostock, Rostock, Germany
Department of Engineering, Jade University of Applied Sciences, Wilhelmshaven,
Wilhelmshaven, Germany
e-mail: [email protected]; [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 285
M. van Beurden et al. (eds.), Scientific Computing in Electrical Engineering,
Mathematics in Industry 36, https://doi.org/10.1007/978-3-030-84238-3_28
286 G. Sadashivaiah et al.

battery period is a major drawback. Several energy harvesting technologies have

been proposed to prolong the operational time of the implants [2, 3]. One promising
device solution is a thermoelectric generator (TEG). It is an energy harvesting device
that transforms thermal into electrical energy by Seebeck effect when embedded in
the human body [4].
The design optimization of such TEG device and its driving circuitry depends
on an accurate, efficient thermal tissue model. The authors in [5] integrated a
TEG model within a simplified linear 3D cubic tissue model. Krylov-subspace
based model order reduction (MOR) method [6] was successfully employed to
obtain the low-rank approximation of the model. In [7], the authors considered a
realistic human torso model with temperature-dependent heat transfer effects. To
apply the conventional MOR method, linearization techniques were adopted during
model reduction. In this work, we consider more realistic material properties, heat
transfer effects and boundary conditions in the torso model, which are temperature-
dependent. To enable efficient design optimization, we investigate the feasibility of
nonlinear MOR methods such as proper orthogonal decomposition (POD) [8, 9] and
dynamic mode decomposition (DMD)[10, 11].

2 Case Study

In this section, we present the model of TEG incorporated in the fat tissue in the
chest region of the human torso model. The aim of numerical simulations is to find
the temperature difference across the TEG.
Figure 1a represents the setup of an electrically active implant and Fig. 1b
describes the model of TEG that was constructed in ANSYSWorkbench [12]
based on the available commercial TEGs. The geometry consists of top and bottom
ceramic plates made of aluminum oxide, with the cross-sectional area of 24.6×24.6
mm2 and height of 0.565 mm. The junction between two plates encloses an array
of 16 × 16 p-type and n-type thermocouple legs. The legs are made of temperature-
dependent bismuth telluride, each with height of 2.27 mm and cross-section of

Fig. 1 (a) Schematic of a TEG integrated inside the human tissue; (b) TEG model with 16×16
thermocouple legs and housing
Nonlinear Model Order Reduction of a Thermal Human Torso Model 287

Fig. 2 (a) TEG embedded in the fat layer of chest region; (b) human torso model with internal
organs

0.8 ×0.8 mm2 . Additionally, TEG is surrounded by a disk-like housing made of

Teflon and with radius of 19 mm. The Seebeck voltage produced by the TEG is
usually of few microvolts and is given by: Vt eg = n2 · ΔT (α1 − α2 ) where, n
is the number of thermocouple legs, ΔT is the temperature difference across the
thermocouple legs, and α1 , α2 are the Seebeck coefficients of p and n doped bismuth
telluride.
To achieve the realistic temperature distribution effect, the 3D human torso model
was modeled based on the segmented magnetic resonance imaging data [13] as
shown in Fig. 2a. The model consists of solid internal organs, blood vessels, skeleton
and tissue layers as shown in Fig. 2b. The realistic material properties for tissues are
applied, as given in [14]. The internal heat transfer in living tissues is given by
Pennes’ bioheat model [15]:

∂T
ρc = ∇κ∇T + ρb cb ω(Ta − T (r , t)) +Qm , (1)
∂t ! "# $
Qb (T )

where, ρ, c, and κ are the density, specific heat, and thermal conductivity of tissues
respectively. ρb , cb , and ω denote the density, specific heat, and perfusion rate of
blood. To maintain the core temperature at 37 ◦ C, heat transfer occurs internally due
to metabolic heat generation Qm , temperature dependent blood perfusion Qb , and
thermal conduction. Ta is the arterial blood temperature and T (r , t) is the unknown
temperature of the human tissues.
To maintain the natural balance in the body, excess heat is transferred to the
ambient environment through the skin surface, which is given by:

qskin = hc (TΓskin − Tamb ) + σ (TΓ4skin − Tamb

4
) + he (Pskin − Pa ), (2)
! "# $ ! "# $ ! "# $
qconv qrad qeva
288 G. Sadashivaiah et al.

where, qconv , qrad and qeva are the convection, radiation, and evaporation effects
applied as heat flux inputs to the skin surface in ANSYS. Tamb represents the
ambient temperature. The variable hc represents the heat transfer coefficient in
W/m2 /K, σ = 5.6705 × 10−8 W/m2 /K4 represents the Stefan-Boltzmann constant
and  = 0.95 represents the emissivity. Heat loss due to evaporation occurs mainly
in the form of sweating through the skin surface. In the evaporation term Pskin and
Pa represents the saturated vapour pressure at skin temperature and partial vapour
pressure respectively. In accordance with the Lewis relation [16], the evaporation
coefficient he can be represented in terms of the heat transfer coefficient as:

he K
= 16.5 . (3)
hc kPa

The relative humidity is φ = PPsaa . Antoine’s equation for saturated vapour

pressure at skin surface Pskin and saturated vapour pressure Psa are defined as:

4030.18
Pskin = 0.1exp(18.956 − ) in kPa, (4)
Tskin + 235

4030.18
Psa = 0.1exp(18.956 − ) in kPa. (5)
Tamb + 235

Therefore, the final equation for the evaporation heat loss is represented by:

4030.18 4030.18
qeva = 1.65hc w{exp(18.956 − ) − φ · exp(18.956 − )},
Tskin + 235 Tamb + 235
(6)

where, w represents the skin wettedness and its value range between 0.06 − 1.
The spatial discretization of the model (1) with boundary conditions (2) at
the skin surface leads to the following large-scale system of nonlinear ordinary
differential equations (ODEs):
⎧
⎪
 ⎨ E · Ṫ (t) = A(T ) · T (t) + !B · u(T (t)),
"# $
F (T (t )) (7)
⎪
⎩ y(t) = C · T (t),
N

where, A(T ) ∈ RN×N is the temperature-dependent global heat conductivity matrix

and E ∈ RN×N is the constant heat capacity matrix. B ∈ RN×m , C ∈ Rp×N are
the input and output matrices respectively, with m as the number of inputs and p as
the number of outputs. y(t) ∈ Rp defines the output vector and u(T (t)) ∈ Rm is the
temperature-dependent load vector. The system is nonlinear due to radiation effect,
which corresponds to fourth power of temperature, evaporation effect in the form
of sweating and temperature-dependent material properties of the thermocouple
Nonlinear Model Order Reduction of a Thermal Human Torso Model 289

legs. The length of unknown state vector T (t) ∈ RN is N = 1, 340, 734, which
defines the dimension of the full system (7). We would like to emphasize that the
nonlinear input effects can be linearized, and conventional Krylov-subspace based
MOR can be employed. But in our case, the heat conductivity of thermocouple legs
is considered temperature-dependent and hence, the nonlinear MOR methods have
to be applied.

3 Model Order Reduction

In this work, we employ two different MOR approaches to compute the reduced
order model (ROM) of system (7), the proper orthogonal decomposition and the
dynamic mode decomposition.

3.1 Proper Orthogonal Decomposition

One of the most common methods for reducing the dimensionality of the nonlinear
systems is POD, which is also known as Karhunen-Loève (KL) decomposition or
the principal component analysis. The method employs singular value decompo-
sition (SVD) to construct the optimal projection subspace, also called the reduced
basis, which captures most of the dynamics of the given data-set [8, 9]. In this work,
we compute the projection subspace φpod by employing the POD technique, which
is used in conjunction with the Galerkin projection [17] to obtain the ROM:

 Er · T˙r (t) = Ar (Tr ) · Tr (t) + φpod
T F (φ
pod Tr (t)),
(8)
r
y(t) = Cr · Tr (t),

where, Er = φpod
T Eφ
pod , Ar (Tr ) = φpod A(T )φpod , Cr = Cφpod are the reduced
T

matrices and accuracy between the full system and ROM is given by  T (t) −
φpod Tr (t) .

3.2 Dynamic Mode Decomposition

Dynamic mode decomposition is a data-driven method, i. e. it uses the measured

data for constructing the ROM. According to Schmid [10], DMD is a special case
of Koopman theory [18] and thus we begin with the definition of Koopman operator
for the nonlinear system (7), written in symbolic form:

dT (t)
= Nl {T (t)}, (9)
dt
290 G. Sadashivaiah et al.

where, T (t) ∈ M , an N-dimensional manifold and Nl is a nonlinear operator.

The Koopman operator K , acts on set of observable functions g : M → C so
that: K g(T (t)) = g(Nl {T (t)}). Thus, K is a linear operator that maps nonlinear
system in the state space to linear system in the observable space. The method yields
the matrix An , which is the approximation of the Koopman operator:

d T̃ (t)
= An T̃ (t), (10)
dt

where, T̃ (t) ∈ RN is the approximate solution and An ∈ RN×N is the matrix,

which defines the best-fit linear dynamics using only the measured data from the
full system. The main objective of the method is to find the linear approximation
operator An , so that the true and approximated solution remain close in a least square
sense, i.e T (t) − T̃ (t).
In general, the computed matrix An is highly ill-conditioned. The speciality
is that, the method predicts the future state of the system by exploiting the low-
rank structures, associated from the eigendecomposition of An . Therefore, the
approximated solution via the associated low-rank structures is given by:


r
T̃ (t) = bi ψidmd exp(ωi t), (11)
i=1

where, ψidmd is the DMD basis of rank r, ωi is the eigenvalues of the matrix An and
bi is the initial condition.

4 Numerical Simulation Results

In this work, all the computations are performed on a PC with an Intel Xeon
E5-2680, 2.5 GHz, 128 GB RAM, 4 active cores processor. Initially, steady-state
thermal simulation of the model is conducted with Tamb = 25 ◦ C and hc = 3.1
W/m2 /K. The result of the steady-state simulation is considered as the initial values
to conduct the transient thermal simulation for the new value of hc = 5.48 W/m2 /K.
For both approaches, snapshot matrix is built out of 20 equidistant snapshots with
step size Δt = 350 s for t ∈ [0, 7000]. The singular values obtained by performing
SVD of the snapshot matrix are shown in the Fig. 3a.
Figure 3b and c represents the maximum relative error (%) between the full order
model of dimension N = 1, 045, 923 and ROMs of dimension r=3 (POD) and r=4
(DMD). The maximum relative error of the POD approach amounts to 0.051%, and
of DMD approach amounts to 0.071%. The runtime for transient thermal simulation
of the full system amounts to 5460 s. In both approaches, the time required to
construct snapshot matrix and to perform SVD in offline stage is 5460 + 137.5 s,
Nonlinear Model Order Reduction of a Thermal Human Torso Model 291

a 10
0
b
0.10
Singular values Error in POD Approx. TEG-Top Node
Error in DMD Approx.
–1
10
0.08

Relative Error (%)

–2
10
0.06
log(σk)

–3
10
0.04
–4
10

10
–5 0.02

–6
10 0.00
0 5 10 15 20 0 1000 2000 3000 4000 5000 6000 7000
Snapshots Time (s)

c 0.10 Error in POD Approx. TEG-Bottom Node

Error in DMD Approx.

0.08
Relative Error (%)

0.06

0.04

0.02

0.00
0 1000 2000 3000 4000 5000 6000 7000
Time (s)

Fig. 3 (a) Singular values σk of the full human torso model, (b) and (c) represents the maximum
relative error at selected nodes of full and reduced order model

but in online stage POD requires only 63.1 s and DMD only 22.7 s for computation
of ROMs.

5 Conclusion and Outlook

In this work, the low-rank approximation of the large-scale nonlinear thermal human
torso model was generated via POD and DMD methods. The main advantage of
the DMD method over POD is, that it does not require any information about
the governing equations of the system to be solved. However, in terms of error
convergence rate, POD-based MOR method is superior compared to DMD method
due to the fact that DMD modes are not orthogonal. In future, both approaches will
292 G. Sadashivaiah et al.

be tested on the parameterized human torso models, with skin wettedness w and
ambient temperature Tamb as parameters.

References

1. C. Casey, J. Gullo, 2018 Aging readiness and competitiveness report. AARP Int. J. 12, 14–15
(2019). https://doi.org/10.26419/int.00036.003
2. S. Priya, D.J. Inman, Energy Harvesting Technologies (Springer, New York, 2009)
3. M.A. Hannan, M. Saad, S.A. Samad, A. Hussain, Energy harvesting for the implantable
biomedical devices: issues and challenges. Biomed. Eng. Online 13, 79 (2014)
4. P. Miao, P.D. Mitcheson, A.S. Holmes, E.M. Yeatman, T.C. Green, B.H. Stark, Mems inertial
power generations for biomedical applications. Microsyst. Technol. 12(10–11), 1079–1083
(2006)
5. O. Jadhav, C.D. Yuan, D. Hohlfeld, T. Bechtold, Design of a thermoelectric generator for
electrical active implants, in MikroSystemTechnik Congress (2017), pp. 1–4
6. R.W. Freund, Krylov-subspace methods for reduced-order modeling in circuit simulation. J.
Comput. Appl. Math. 123, 395–421 (2000)
7. C. Yuan, S. Kreß, G. Sadashivaiah, E.B. Rudnyi, D. Hohlfeld, T. Bechtold, Towards efficient
design optimization of a miniaturized thermoelectric generator for electrically active implants
via model order reduction and submodeling technique. Int. J. Numer. Methods Biomed. Eng.
36, e3311 (2020)
8. P. René, Model Reduction via Proper Orthogonal Decomposition (Springer, Berlin, Heidel-
berg, 2008), pp. 95–109
9. A. Quarteroni, A. Manzoni, F. Negri, Reduced Basis Methods for Partial Differential Equations
(Springer, Cham, 2016), pp. 115–140
10. P. Schmid, Dynamic mode decomposition of numerical and experimental data. J. Fluid Mech.
656, 5–28 (2010)
11. A. Alla, J.N. Kutz, Nonlinear model order reduction via dynamic mode decomposition. SIAM
J. Sci. Comput. 39, 1–20 (2016)
12. Ansys®Academic Research Mechanical, Release 2020 R1, Workbench
13. S. Makarov, G. Noetscher, J. Yanamadala, VHP-Female Datasets. NEVA Electromagnetics,
LLC:VHP-Female 2.2 edn. (2015)
14. P.A. Hasgall, G.F. Di, C. Baumgartner, IT’IS database for thermal and electromagnetic
parameters of biological tissues, IT’IS Foundation, Switzerland, (2018)
15. H.H. Pennes, Analysis of tissue and arterial blood temperatures in the resting human forearm.
J. Appl. Physiol. 1(2), 93–122 (1948)
16. W.K. Lewis, The evaporation of a liquid into a gas. 44, 445–446 (1922)
17. P. Benner, S. Gugercin, K. Willcox, A survey of projection-based model reduction methods for
parametric dynamical systems. SIAM Rev. 57, 483–531 (2015)
18. B.O. Koopman, Hamiltonian systems and transformation in Hilbert space. Proc. Natl. Acad.
Sci. 17, 315–318 (1931)
Multi-Level Iterations for Microgrid
Control with Automatic Level Choice

Robert Scholz, Armin Nurkanović, Amer Mešanović, Jürgen Gutekunst,

Andreas Potschka, Hans Georg Bock, and Ekaterina Kostina

Abstract Microgrids are considered a key technology for the energy transition,
but the rising penetration of renewable energy sources is pushing current control
approaches to their limits. Nonlinear model predictive control (NMPC) is a promis-
ing approach to address this issue, although achieving real-time feasibility with
standard schemes is challenging. Therefore, we propose to use the Multi-Level
Iteration NMPC scheme with a novel automatic level choice. This allows us to
always use the most accurate linearizations available while being real-time feasible,
even during strongly transient phases where a fixed level choice may be too slow.
We use a realistic-sized microgrid to illustrate the capabilities of this method.

1 Introduction

Microgrids (MG) are small, local electrical networks comprising heterogeneous

components, such as generators, storage systems and loads. They are managed
autonomously and are operated either as an islanded network or within a connected
larger network. This allows the MG to be considered as a single controllable entity
in the utility grid. MGs are a promising approach to handle the rising number of
renewable energy sources (RES). Their ability to handle unforseen disturbances on
a local level is likely to make them a key technology to enable more distributed and
heterogeneous networks[1].
Tight operational bounds on frequency and voltage make MG control a chal-
lenging task, especially during demanding load scenarios. State-of-the-art methods
rely on a hierarchical control structure, using proportional-integral controllers and

R. Scholz () · J. Gutekunst · A. Potschka · H. G. Bock · E. Kostina

Interdisciplinary Center for Scientific Computing (IWR), Heidelberg University, Heidelberg,
Germany
e-mail: [email protected]
A. Nurkanović · A. Mešanović
Siemens AG, Munich, Germany
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 293
M. van Beurden et al. (eds.), Scientific Computing in Electrical Engineering,
Mathematics in Industry 36, https://doi.org/10.1007/978-3-030-84238-3_29
294 R. Scholz et al.

various filters on different levels. However, experience shows that this control
paradigm reaches its limits under high penetration of RES [1].
Nonlinear model predictive control (NMPC) is a general model-based methodol-
ogy to control dynamical processes. Measurements of the process are embedded in
an optimal control problem, which is solved repeatedly with respect to an objective
function and operational limits. Since NMPC offers a flexible control framework, it
appears to be well suited for the control of MGs.
One of the main challenges for NMPC in the field of power engineering is
the real-time requirement. Transient electrical dynamics involve high-frequency
oscillations, which are costly to simulate, and a high sampling rate is necessary to
react to disturbances in time. Therefore, in the literature, NMPC is mainly used on a
higher control level, relying on traditional integral or droop controllers to handle the
electrical dynamics [2]. To deal with the fast electrical behavior of MGs directly
with NMPC, tailored schemes are necessary, like the Advanced Step Real-Time
Iteration [3].
In this paper, we propose to use the Multi-Level Iteration (MLI) scheme for
MG control. This approach is based on the well-established Real-Time Iteration
[4], which eliminates the need to solve the underlying optimal control problems
until convergence. Additional computation time is saved by updating the problem
linearization only partially in every iteration using cheap update formulas to increase
the feedback rates and render NMPC applicable for MG control.

2 Nonlinear Model Predictive Control

NMPC is a general framework to control dynamic processes modeled by

differential- algebraic equations (DAE). At each time tk within a given sequence
of sampling points t0 < t1 < . . . a feedback signal uξk is computed based on the
current system state ξk . In the traditional NMPC setting, this is done by solving an
optimal control problem (OCP) with fixed time horizon of length T of the form:

min Φ(x(·), z(·), u(·)) (1a)

x(·),z(·),u(·)

s.t. ẋ(t) = f (x(t), z(t), u(t)), 0 = h(x(t), z(t), u(t)), (1b)

x(tk ) = ξk , t ∈ [tk , tk + T ] , (1c)
x lo ≤ x(t) ≤ x up , zlo ≤ z(t) ≤ zup , ulo ≤ u(t) ≤ uup . (1d)

The differential and algebraic states x(t) and z(t) are subject to the DAE system (1b)
with initial value set to the current system state ξk (1c). The control inputs of the
process are represented by u(t) and the objective is defined by the functional Φ.
The NMPC feedback signal applied in the interval [tk , tk+1 ) is the first part of the
solution uξk (t) = u∗k (t). We discretize the problem with the direct multiple shooting
Multi-Level Iterations for Microgrid Control with Automatic Level Choice 295

method introduced by Bock [5] and obtain a finite dimensional, structured nonlinear
program (NLP) of the compact form

min l(w) s.t. b(w) + Eξk = 0, wlo ≤ w ≤ wup . (2)

w

Here l is the discretized objective function (1a), the function b together with the
constant matrix E represent the discretized DAE system (1b) with the initial value
embedding constraint (1c) and wlo and wup are the lower and upper bounds on states
and controls. A sequential quadratic programming (SQP) method is used to solve
this NLP, which generates a sequence of primal-dual iterates (wj , λj , μj )j ∈N based
on the quadratic program (QP)

1 b(wj ) + Eξk + BΔw = 0,
min Δw AΔw + a  Δw s.t. (3)
Δw 2 wlo ≤ Δw + wj ≤ wup .

The matrix A is the Hessian (or an approximation thereof) with respect to w of the
Lagrangian L (wj , λj , μj ) of the NLP (2). The linear objective term is defined by
the objective gradient a = ∇w l(wj ) and the constraints are linearizations based on
b(wj ) and its Jacobian B = ∇b(wj ) . The solution (ΔwQP , λQP , μQP ) of QP (3)
is used to update the primal-dual variables:

wj +1 = wj + ΔwQP , λj +1 = λQP , μj +1 = μQP . (4)

Under mild assumptions, local quadratic convergence of the SQP method is

guaranteed. As a consequence, once the iterates are sufficiently close to the true
solution, only one iteration per sampling time is sufficient to obtain excellent
solution approximations for the NLPs at subsequent sampling times [4]. The Real-
Time Iteration, introduced by Diehl [4, 5], allows a further speedup as it exploits
that the initial value ξk only enters linearly in the QP (3). This means that most of
the QP data can be prepared based on the current iterate (wj , λj , μj ), before ξk is
known. As soon as ξk becomes available, only the QP solution step is necessary to
generate the feedback signal.

3 Multi-Level Iterations

Depending on the application, the Real-Time Iteration still requires a high compu-
tational effort in every iteration. To set up the QP (3), the constraints, the objective
gradient, the constraint Jacobian and the Hessian (corresponding to b, a, B, A in (3))
have to be computed. MLI can reduce this computational effort drastically and thus
speed up the feedback process by only updating parts of the QP.
The MLI scheme is based on the fact that Newton-type methods (such as the SQP
method described in the previous section) do not require the exact computations of
296 R. Scholz et al.

Table 1 Computations and update formulas for the QP data for the different Levels
Necessary computations Update formula for QP data
Level b(wj ) a(wj ) B(wj ) A(wj , λj ) b a B A
D ✓ ✓ ✓ ✓ b(wj ) a(wj ) B(wj ) A(wj , λj )
C ✓ ✓ (✓)a ✗ b(wj ) a(wj ) + (B̄C − B(wj ) )λj B̄C ĀC
B ✓ ✗ ✗ ✗ b(wj ) āB + ĀB (wj − w̄B ) B̄B ĀB
A ✗ ✗ ✗ ✗ b̄A āA B̄A ĀA
Only the vector-matrix product λ B needs to be computed in an adjoint fashion
a

derivatives in A and B to remain locally convergent. This can be exploited to avoid

the expensive evaluation of the Hessian and the Jacobian in every iteration.
Instead, the different components of the QP (3) are updated in four hierarchical
levels with descending computational complexity. Each level stores a reference
point (w̄, λ̄, μ̄) and the corresponding QP data b̄, ā, B̄ and Ā. Every level is working
on its own set of iterates, which are independent of the other levels. Table 1 explains
which data is computed in each iteration and how the QP data is updated. The
convergence is usually analyzed for a fixed system state ξk . Level D corresponds
to a full SQP step and therefore inherits its local quadratic convergence. Level C
avoids the full Jacobian evaluation, but is still converging to an optimal point of
the original NLP (2). Level B abandons the sensitivity generation completely and
convergences to a feasible point. Level A refers to linear MPC, since all QP data is
fixed. It provides feedback with the lowest computational effort, but for nonlinear
models convergence can not be guaranteed. A detailed description of the levels can
be found in [6, 7] and their convergence properties are analyzed in [5].

3.1 Automatic Level Choice

In practice, the presented levels are operated simultaneously. The lower levels are
used to give fast feedback and the higher, computationally more expensive levels
provide accurate linearizations of the NLP (2). Usually this is done by a sequence
of levels, which is fixed in advance. To ensure real-time feasibility, the required
computation time must be estimated and the sequence chosen accordingly [6].
This approach turns out to be inflexible, because the evaluation of a level needs to
be finished before the next evaluation is scheduled. In order to be real-time feasible,
the worst-case computation time needs to be treated. If an adaptive integration
method is used, the integration time may vary strongly between the steady state
and transients. This leads to an unnecessary conservative scheduling of higher
levels, even when the computation time is low and would allow a faster rate. To
overcome this issue, we propose to choose the levels automatically online instead.
In this method every level is operated in parallel. In the beginning of the simulation,
the evaluation of all levels is triggered. As soon as an evaluation is finished, the
Multi-Level Iterations for Microgrid Control with Automatic Level Choice 297

corresponding level is marked as ready. When the simulation of the controlled

process reaches the next sampling point, the linearization of the highest level, which
is marked as ready, is applied to the QP (3). The lower levels are reinitialized with
the updated QP data b̄, ā, B̄ and Ā, and a new evaluation starts. Since level A
includes no reevaluation of the QP data, its computation time is very low and it
can always be used as a fallback, if no other level is ready. With this approach, we
can always use the most accurate available linearizations whilst at the same time
having the guarantee of remaining real-time feasible.

4 Dynamic Microgrid Model

To deal with dynamics on different time scales, MG control is typically organised

in three hierarchical levels. Primary control relies on local internal controllers
of the components and has the goal to stabilize the system. Secondary control
is responsible for eliminating any steady-state error introduced by the primary
control. Long term planning and the incorporation of weather forecast and load
prediction is done by the tertiary control. Our proposed controller is working on
the secondary control level. It receives control signals from the tertiary control and
sets the reference values of primary controllers of the individual components. The
test MG model comprises not only the physical components, but also the primary
control level.

4.1 Scenario and Model Description

We use the model of a test MG which is comprehensively presented in [3]. Here,

we give a short overview. The structure of the MG system is depicted in Fig. 1. It
comprises two identical diesel generators (DG), a battery (BA), a photovoltaic plant
(PV), and a passive PQ-load. The DGs consist of a synchronous generator actuated
by a diesel engine with a governor for frequency stabilization (IEEE DEGOV1) and
are equipped with an Automatic Voltage Regulator (IEEE AC5A). The setpoints for
frequency ωref and voltage Vref serve as control variables of the MLI-controller. The
battery is modeled as a constant DC voltage source connected to an inverter with an
ref and voltage V ref .
internal droop. It is controlled by the setpoints for frequency ωBA BA
The base power of the MG is Sgrid = 100 kVA, the nominal power of the generator
is SDG = 325 kVA, and the nominal power of the battery is SBA = 150 kVA. The
complete MG is given as a DAE system of index 1 with 37 differential and 42
algebraic states and 6 control inputs.
To demonstrate the capabilities of the proposed controller, we apply a challeng-
ing load scenario. At the beginning, the system is in a steady state and the reference
ref ref
values for the battery are set to PBA = QBA = 0p.u.. The generators share the load
298 R. Scholz et al.

P,Q

DG DG BA PV
Fig. 1 Topology of the test MG

of Pload = 5p.u. and Qload = 1p.u. equally. After 1 s a sudden unscheduled load
step of 40% in active and reactive power takes place, which exceeds the capacity of
the generators. To ensure that the operational limits are satisfied, the battery needs
to leave the provided reference values and serve the missing load. The simulation
has an overall length of 8 s.
In MG control, we need to consider several objectives with different priorities.
This is modeled by a continuous least squares objective functional
 ti +T
Φ(x, z, u) = r(x(t), z(t), u(t))2 dt (5)
ti

of OCP (1) with a weighted norm and a residual function r(x, z, u). The most
important goal is to steer the frequency ω(t) and voltage at the load Vload (t) to
the nominal value 1p.u. after a disturbance. During transients, we want to utilize the
battery to stabilize frequency and voltage. In steady state, the performance of the
ref ref
battery should follow setpoints PBA , QBA from a higher control level, in order to
charge or discharge the battery. The generators are supposed to share the remaining
load equally. These goals are achieved by tracking terms

r1 (x, z, u) = ω − 1, r2 (x, z, u) = Vload − 1,

ref ref
r3 (x, z, u) = PBA − PBA , r4 (x, z, u) = QBA − QBA ,
r5 (x, z, u) = P1 − P2 , r6 (x, z, u) = Q1 − Q2 .
Multi-Level Iterations for Microgrid Control with Automatic Level Choice 299

5 Numerical Results

We discretize OCP (1) with two multiple shooting intervals and the length of the
prediction horizon is fixed to T = 1 s. The length of the first shooting interval
corresponds to the sampling time of 100 ms and the second to 900 ms. The numerical
simulations are carried out with the NMPC framework MLI [6]. For integration and
sensitivity generation, the SolvIND integrator suite is used and the QPs are solved
by qpOASES [8].
The continuous least squares objective function (5) enables us to use a Gauß-
Newton approximation of the Hessian in QP (3). Besides its favorable numerical
properties, its main advantage is, that it relies only on first-order derivatives.
Therefore, we do not have to compute second-order derivatives, which is the most
costly task when evaluating QP (3).
We compare our proposed MLI-controller with a typical state-of-the-art control
setup for small microgrids: The generators are equipped with an integral controller
for steady-state error elimination of the frequency with a settling time of approxi-
mately 20 s and a sampling time of 100 ms. The voltage setpoint Vref is kept constant
during the full simulation time.
In Fig. 2, the performance of the proposed MLI-controller is shown in com-
parison to the traditional control approach. The MLI-controller steers back the

Frequency Voltage at Load

1
1
ω [p.u.]

V [p.u.]

0.98 I-Controller
MLI-Controller
0.96
0.99
0.94
0 2 4 6 8 0 2 4 6 8
Apparent Power Generator #1 Apparent Power Battery

3.5
operational limit 1
S [p.u.]
S [p.u.]

3 0.5

0
2.5
0 2 4 6 8 0 2 4 6 8
time [s] time [s]

Fig. 2 Control performance of the MLI-controller in comparison to a traditional control approach.

In the top row, the frequency and the voltage at the load is depicted. In the bottom row, the apparent
power of the generators and the battery is shown. The MLI-controller is able to steer frequency
and voltage back to the nominal value faster and with a lower initial drop, while respecting the
operational limits
300 R. Scholz et al.

computation time [s]

100 Level D
Level B
10− 1

10− 2
0 1 2 3 4 5 6 7 8
sampling time [s]

Fig. 3 Computation times and scheduling of Level B and D. The elapsed computation time
is depicted by the height of the bars while the width shows in which sampling intervals the
computations were performed

frequency faster with a lower initial drop after the unforeseen disturbance. The
voltage gets stabilized faster and the steady state offset is eliminated. In the
beginning, the battery follows the setpoints and does not contribute in load sharing.
After the load drop, the fast reacting dynamics of the battery are used to stabilize
the system. Since the overall load exceeds the operational bounds of the generators,
the battery temporarily deviates from its reference value and instead serves the
necessary additional load. In contrast to this, the integral-controller is not able
to obey the operational limits of the generators. If there are no safety measures
installed, the generators are overloaded, which may cause physical damage. In
Fig. 3, the computation time and the scheduling of the different MLI levels are
shown. In the beginning, the system is in a steady state and the computation time
is low. After the load jump at t = 1 s, the system is in a transient phase and the
computation time rises sharply, which leads to less level D evaluations. Afterwards,
the system gets steered back to a steady state and the computation time decreases.
As the evaluation time for level B is always below the sampling time, no level A
occurs. Level C is not used, because the Gauß-Newton approximation of the Hessian
implies that the difference in computation time between level C and D is low.

6 Conclusion

In this paper a novel MG controller based on Multi-Level Iterations is presented.

We used a realistic-sized example MG, modeled with a DAE, to perform numerical
experiments and compared the performance with a traditional control approach.
In the example shown, the MLI-controller did not only outperform the traditional
controller, but also was able to respect operational bounds.

Acknowledgments This research was funded by the German Federal Ministry of Education and
Research (BMBF) in the research project MOReNet. (Grant No 05M18VHA)
Multi-Level Iterations for Microgrid Control with Automatic Level Choice 301

References

1. M. Ilić, R. Jaddivada, X. Miao, Modeling and analysis methods for assessing stability of
microgrids, in 20th IFAC World Congress, vol. 50, IFAC-PapersOnLine (2017), pp. 5448–5455
2. C. Bordons, F. Garcia-Torres, M.A. Ridao, Model predictive control of microgrids, in Advances
in Industrial Control (Springer International Publishing, Cham, 2019)
3. A. Nurkanović, A. Mešanović, A. Zanelli, G. Frison, J. Frey, S. Albrecht, M. Diehl, Real-
time nonlinear model predictive control for microgrid operation, in 2020 American Control
Conference (ACC) (2020), pp. 4989–4995
4. M. Diehl, Real-Time Optimization for Large Scale Nonlinear Processes. Dissertation, Heidel-
berg University, 2001
5. H.G. Bock, M. Diehl, E. Kostina, J. Schlöder, Constrained optimal feedback control of systems
governed by large differential algebraic equations, in Real-Time PDE-Constrained Optimization
(SIAM, Philadelphia, 2007), pp. 3–22
6. L. Wirsching, Multi-level iteration schemes with adaptive level choice for nonlinear model
predictive control. Dissertation, Heidelberg University, 2018
7. A. Nurkanović, S. Albrecht, M. Diehl, Multi level iterations for economic nonlinear model
predictive control, in Recent Advances in Model Predictive Control: Theory, Algorithms, and
Applications, ed. by T. Faulwasser, M.A. Müller, K. Worthmann (Springer, Cham, 2021),
pp. 65–105
8. H.J. Ferreau, C. Kirches, A. Potschka, H.G. Bock, M. Diehl, qpOASES: A parametric active-set
algorithm for quadratic programming. Math. Program. Comput. 6, 327–363 (2014)
Multi-Level Inversion Based on Mesh
Decoupling

Benny Shachor, Hadi Hajibeygi, and Domenico Lahaye

Abstract Accurate material characteristics in many real-field engineering appli-

cations can only be determined via inverse modeling. This is due to the fact
that often times measuring field quantities in the resolution required to model
and predict engineering processes are impossible. Despite several advancements,
still it remains a challenge as to how to characterise material properties through
an inverse modeling approach. Special challenge is driven when the scale of the
field is large, while the parameters are expected to be defined at high resolution.
Such challenge demands for scalable inverse modeling techniques, clearly beyond
the scope of classical single-level approaches. In this paper, we propose a new
multi-level approach based on mesh decoupling of the state and design variables.
This approach allows for treating the design variables on various scales without
comprising the accuracy of the state and adjoint equation solve. The performance
of the new method is investigated for estimation of the heterogeneous parameters of
one-dimensional and two-dimensional elliptic equations. Results illustrate that the
mesh decoupling technique provides a promising framework for solving large-scale
heterogeneous systems.

1 Introduction

Inverse modeling is an important step in defining system characteristics when

available measurement data are insufficient to fully describe the parameters. Special
challenge is imposed on real-life engineering applications, as the scale of the
problem is very large while the necessary parameters are expected to be defined
on high resolutions. This challenge makes classical inverse modeling approaches

B. Shachor () · H. Hajibeygi

Department of Geoscience and Engineering, TU Delft, Delft, Netherlands
e-mail: [email protected]
D. Lahaye
Delft Institute of Applied Mathematics, TU Delft, Delft, Netherlands
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 303
M. van Beurden et al. (eds.), Scientific Computing in Electrical Engineering,
Mathematics in Industry 36, https://doi.org/10.1007/978-3-030-84238-3_30
304 B. Shachor et al.

computationally too expensive to be applicable. Recently, a multiscale inverse

strategy has been introduced [3], where the state variable was represented in a
coarser resolution, while parameters stayed in the fine scale.
The current research develops a new approach based on the multi-level decou-
pling technique. Through sequences of parameter restrictions, we first optimise the
mismatch between the observation and measurements at the coarsest space. These
coarse-scale parameters provide good approximations for the initial parameters
at a higher resolution. We perform this procedure until the fine-scale converged
parameters are obtained. Note that at any stage, depending on the value of the
objective function, one can stop iterations. As such, convergence on the fine-scale
resolution is not necessary. This is quite relevant for field applications where only a
good approximation of the parameters is acceptable, if they are found efficiently.

2 Problem Description

In this section we subsequently introduce the partial differential equation con-

strained least squares problems we intend to solve, a classical discrete adjoint
solution method and a Newton Trust-Region algorithm [2]. The discrete adjoint
method is explained in more detail in [10]. The method is compared with a
continuous adjoint method in [6] and the references cited therein. The literature
on least squares optimization methods with partial differential equation constraints
is very vast. Recent monographs include [1, 5, 9].
Non-linear Least-Squares Problem We intend to solve non-linear least squares
problems in which the state variable u(x) is constrained by a partial differential
equation [5, 10]. As state equation we consider the diffusion equation with spatially-
varying diffusion coefficient k(x) and source term f (x) on a computational domain
Ω with homogeneous Dirichlet boundary conditions. For notational convenience
we assume Ω to be the unit interval or the init square. This assume is not restrictive
for the subsequent arguments we wish to make. Our goal is to recover the spatial
variation of k(x) by solving the state equation for multiple sources fj (x) and
evaluating the computed solution at various receiver locations. Let Ns and Nr
denoted the number of sources and receivers, respectively. The set of differential
equations that we solve can then for 1 ≤ j ≤ Ns be written as

∇ · (k(x)∇uj ) = fj on Ω. (1)

The source function f (x) is defined as a set of point sources at the source
locations. The state equation for u(x) is solved as many times as the number of
sources. At each solve, one source is set to have amplitude one and all other source to
amplitude zero. After each solve, the computed solution is evaluated at the receiver
locations. After Ns state equation solves, Ns vectors u j of size Nr are available.
Each of these vectors depends on k(x).
Multi-Level Inversion Based on Mesh Decoupling 305

A set of measured values at receiver locations is assumed to be available for each

source. This set is a set of Ns vectors dj of size Nr . We will work with synthetic
data generated by solving the state equation assuming the diffusion equation to be
equal to the exact value.
The cost function that we intend to minimize consists of two terms. The first term
is the sum over j of the Euclidean norm of the discrepancies uj − dj . The second
term is a regularization term that penalizes large variations in k(x). It is proportional
to the Euclidean norm of the gradient of k(x). The second term is multiplied with
a weight α before being added to the least squares part of the cost functional.
The objective is to find the argument k(x) that minimizes the cost functional and
thus to recover the heterogeneity of the medium that best describe the observations.
The set of Ns state equations act as partial differential equation constraints in this
minimization problem.
Discrete Adjoint Method We adopt a classical discrete adjoint solution approach.
This means that we first discretize the cost functional and the differential equation
constraints on Ω. We subsequently differentiate the discrete Lagrangian to obtain
the non-linear system that defines the first order critical points. This non-linear
system is solved by a Newton method.
We adopt a standard Galerkin finite element method using linear shape functions
on segments in one dimension and triangles in two dimensions. We represent the
diffusion coefficient by a constant per element. The resulting vector k is the discrete
set of design variables. In each of the j linear systems for the discrete state u j ,
the coefficient matrix Aj depends on the discrete diffusion coefficient k.  We add to
the discretized cost functional the sum over j of the discrete Lagrange parameters
vj times the residual vector in uj . The discrete Lagrangian obtained depends on
 First order critical points of the Lagrangian are
the 2 Ns + 1 vectors u j , vj and k.
obtained by setting the derivatives with respect to these vectors equal to zero.
Trust-Region Newton Method The non-linear system that defines the first-order
critical points is solved in two steps. First the Ns decoupled linear state and adjoint
equations are solved. In solving for the adjoint variables vj , the linear algebra of
solving for uj is reused. In this paper the LU-factorization of the coefficient matrix
 is recycled. Given that the state equation (1) depends on j through the source
Aj (k)
term fj (x) only, only one LU-factorization for all j ’s is required. The large scale
applications of the multi-level inversion that we here propose does however require
adopting iterative solution methods as proposed in [3, 4]. In the second step the non-
linear system for k is solved by trust-region globalized Newton method starting from
an initial guess. The gradient vector is computed from the available state and adjoint
vectors at negligible computational cost. The Hessian matrix is computed column-
wise by finite difference approximation of the gradient vector. This requires Ns
additional state and adjoint solves for each component of the vector k.  The Hessian
computes therefore dominates the computational cost of the procedure. In future
work, this can be alleviated by a Gauss-Newton or a Broyden-Fletcher-Goldfarb-
Shanno (BFGS) rank-one iterative approximation of the Hessian.
306 B. Shachor et al.

3 Nested Iteration Based on Mesh Decoupling

In this section we describe a multi-level optimization procedure that exploits a

decoupling of the mesh for state and design variables.
Mesh Decoupling Procedure The computational cost of the procedure outlined in
the previous section is dominated by the number of design variables given by the
number of segments in 1D or triangles in 2D. We wish to accelerate the solution of
the optimization problem by reducing the mesh size without lowering the accuracy
of the computed solution. We achieve this goal by representing the design variables
k on a more global scale while at the same time preserving the fine scale resolution
for the state uj and adjoint variables vj . We therefore introduce two decoupled
meshes on the domain of computation Ω.
The first mesh coincides with the classical notion of a mesh to solve for uj and
 The second mesh serves to represent k.
vj given values for k.  In this work we keep
the first mesh fixed at a sufficiently fine level and is denoted by Ωf ine . The second
mesh is defined independently from the first by a hierarchy denoted by denoted by
Ωk . The coarsest mesh Ω1 in the design variables k has four elements along x in the
one-dimensional problem and two elements along x and y in the two-dimensional
problem. The finest mesh in k is comparable in size to the mesh in uj and vj . On
each mesh Ωk , the diffusion coefficient is represented by element-wise constant
values. The coefficient values are transferred from the mesh Ωk to the next finer
mesh Ωk+1 and from the mesh Ωk to the mesh Ωf ine using injection.
On the coarse meshes Ωk the gradient and the Hessian have lesser components
and are therefore easier to compute. The problem on the coarser mesh has fewer
variables and is expected to converge in less iterations.
Components of the Multi-Level Optimization Algorithm The hierarchy of
meshes Ωk allows to define a multi-level solution procedure by as nested iteration.
This procedure is started by solving the optimization problem on the mesh Ω1 using
the constant value of 0.5. Having solved the optimization problem on the mesh
Ωk−1 , the design variables are injected to the finer mesh Ωk and used as initial
guess for the Newton method. The procedure is repeated until reaching the mesh
Ωf ine . The trust-region method acts as a smoother on the meshes Ωk and Ωf ine .
Choice of the Regularization Parameter on Multiple Levels The least-squares
problem that we solve is regularized by adding a term that penalizes large variations
 This spatial variation is measured by the Euclidean
in the spatial distribution of k.
norm of the gradient of the design variables k.  This norm is equal to the energy

norm of k and scales with the mesh width of Ωk [10]. The factor that weighs
the regularization term in the cost functional requires to be adjusted accordingly
when traversing the meshes Ωk in the multi-level algorithm. We employ a numerical
scheme such that the energy norm of Ωk scales with the square of the mesh width of
Ωk . We therefore divide the weighting factor by four in moving from Ωk to Ωk+1 in
the multi-level algorithm. On finer meshes Ωk , the norm of variations in k is larger
and should be given less weight.
Multi-Level Inversion Based on Mesh Decoupling 307

The Algorithm The Multi-Level algorithm is presented below.

Algorithm 4: The multi-level algorithm

Input: Initial guess, α
Output: Nf ine design variables
1 k1 := initial guess
2 for l = 1 : Nlevel do
3 while Trust-Region Newton does not converge do
4 Inject k from Ωk → Ωf ine
5 Compute the Cost, Gradient and Hessian functions on Ωk
6 end
7 Inject k from Ωk → Ωk+1
8 α = α4
9 end

4 Numerical Results

In this section we compare the performance of the single-level and multi-level

algorithms applied to the one-dimensional and two-dimensional test problems.
The one-dimensional and two-dimensional test problem were taken from [10]
and [5], respectively. Details of both test problems are given in Table 1. The assumed
exact spatial dependence of the design variables is given by

k(x) = −0.2e−72(x−0.45) ,
2
(2)

in the one-dimensional problem and by

k(x, y) = 15, on ((x + 0.5)2 + (y + 0.5)2 ) ≤ (0.32)

k(x, y) = 300, on ((x − 0.5)2 + (y − 0.5)2 ) ≤ (0.32) (3)
k(x, y) = 10, otherwise

Table 1 One-dimensional and two-dimensional problem definition

One-dimensional Two-dimensional
Computational domain 0≤x ≤1 0 ≤ x, y ≤ 1
# elements in Ωf ine 64 1368
# sources (Ns ) 2 5
Sources Locations [ 13 ],[ 23 ] [-0.5,-0.5],[-0.5,0.5],
[0,0],[0.5,-0.5],[0.5,0.5]
Source Expression Dirac Delta at the node Dirac Delta at the node
# Receiver (Nr ) 5 5
Observations Locations [0.1, 0.3 , 0.5, 0.7, 0.9] Location of the sources
308 B. Shachor et al.

in the two-dimensional problem. In the one-dimensional problem, the mesh Ω1 is

refined four times such that the mesh Ω5 coincides Ωf ine . In the two-dimensional
problem, the mesh Ω1 is refined four times such that each element in Ωf ine overlaps
with only one element in Ω5 . In the final stage of the multi-level algorithm applied
to the two-dimensional problem, the design variable distribution is defined on the
unstructured triangular mesh Ωf ine .
The converge of the single-level and multi-level optimization algorithm are
presented in Figs. 1 and 2 for the one-dimensional and two-dimensional problem,
respectively. Both figures show the decrease in the cost-functional vs. a measure of
the computational cost. The computational cost of a single PDE solve is given by the
number of elements in the mesh Ωf ine and is equal in both the single and multi-level
algorithm. We therefore adopt the number of PDE solves as a fair metric to compare
the single-level and multi-level algorithm. The numerical differentiation to obtain
the Hessian is responsible for the bulk of the number of PDE solves. The color in
the graphs corresponds to the various levels. In the last stage of the convergence,
both the single and multi-level algorithm operate on the same mesh in k as indicated
by the color coding.
Both Figs. 1 and 2 show that the single-level algorithm converges initially slow
and subsequently very fast. The trust-region method brings the Newton algorithm
in the region of super-linear convergence around the solution. The multi-level

Computational efficiency SL and ML

100
SL
Level 1
Level 2
Level 3
10-2 Level 4
Level fine
Cost Function

10-4

10-6

10-8

10-10
100 101 102 103
Number of PDE solves

Fig. 1 Comparison of the single-level (SL) and multi-level (ML) algorithm by considering the
convergence history in the cost functional as function of the number of PDE solves in the
one-dimensional problem. The single-level method converges slowly until reaching the basin of
attraction. Subsequently is converges faster. The multi-level method converges super-linearly on
each level
Multi-Level Inversion Based on Mesh Decoupling 309

Computational efficiency SL and ML

101
SL
Level 1
Level 2
100 Level 3
Level 4
Level 5
Level fine
10-1
Cost Function

10-2

10-3

10-4

10-5
100 102 104
Number of PDE solves

Fig. 2 Comparison of the single-level (SL) and multi-level (ML) algorithm by considering the
convergence history in the cost functional as function of the number of PDE solves in the two-
dimensional problem. Similar observation as in the one-dimensional problem in Fig. 1

algorithm is seen not to suffer from slow convergence on the finest level. The method
is on the contrary seen to significantly reduce the cost functional in a limited number
 The solution from the coarser level provides
of PDE solves in the coarsest levels in k.
a sufficiently good initial approximation for the Newton algorithm to converge
superlinearly on a given level. The multi-level algorithm requires considerably less
PDE solves to reach solutions with moderate accuracy. However, to reach the final
solution, both the single and multi-level algorithm require the same number of PDE
solves. This is due to the fact that on the finest level both the single and multi-level
algorithm converge very fast. Further research is required to circumvent this issue.
Figure 3 shows the converge of the single-level and multi-level algorithm in
terms of the evolution of the design variables k.  The figure shows that the multi-
level algorithm avoids premature small scale variations in k.  The figure also shows
that the single and multi-level algorithm converge to the same solution. The same
message is conveyed in Fig. 4 for the two-dimensional problem.
310 B. Shachor et al.

Iteration 0 Iteration 1 Iteration 3 Iteration 12

1.5 0.626 0.9 1.05

0.624 0.88 1
1
0.622
0.86 0.95
0.62
K

K
0.5
0.84 0.9
0.618
0
0.616 0.82 0.85

0.614
-0.5 0.8 0.8
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
X X X X

(a) (b) (c) (d)

Level 1 iteration 0 Level 1 iteration 1 Level 4 iteration 5 Level 5 iteration 5

1.5 1.05 1.05

0.6245
1 1
1
0.624

0.6235 0.95 0.95

K

K
0.5
0.623 0.9 0.9
0.6225
0
0.85 0.85
0.622

-0.5 0.6215 0.8 0.8

0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
X X X X

(e) (f) (g) (h)

Fig. 3 Comparison of the single-level (SL) and multi-level (ML) algorithm by considering the
convergence history in the design variables k starting from the same initial guess in the one-
dimensional problem. The second column of pictures shows that the ML algorithm avoids
premature small scale variations in k. (a) SL, iteration 0. (b) SL, iteration 1. (c) SL, 384 PDE
solves. (d) SL, final result. (e) ML, level 1, Iteration 0. (f) ML, level 1, iteration 1. (g) ML, 392
PDE solves. (h) ML, final result

(a) Exact solution (b) Single-Level (c) Multi-Level

Fig. 4 Assumed exact distribution of the design variables k (left) and the solution found by the
single-level (middle) and multi-level algorithm (right)

5 Conclusions

We presented a multi-level algorithm to solve non-linear least square problems

with a Poisson equation for the state variable as constraint. The algorithm exploits
the discretization of the state and adjoint variable on a fixed fine mesh and the
representation of the design variables on a mesh hierarchy with varying spatial scale.
This decoupling allows to coarsen the space of design variables while at the same
time preserving accurate state and adjoint solves. Numerical results show that the
multi-level algorithm avoids small scale variations of the design variables in early
stages of the algorithms. The multi-level algorithm therefore converges significantly
faster than its single-level counterpart on coarse scales in the design variables. The
Multi-Level Inversion Based on Mesh Decoupling 311

multi-level algorithm is thus valuable to adopt in often occurring scenarios in which

a coarse scale representation of the design variables yields valuable information.
The development and results are publicly available at [7] and [8].

References

1. D. Chavent, Nonlinear Least Squares for Inverse Problems (Springer, New York,, 2010)
2. A.R. Conn, N.I. Gould, P.L. Toint, Trust Region Methods (SIAM, Philadelphia, 2000)
3. R. de Moraes, H. Hajibeygi, J.D. Jansen, A multiscale method for data assimilation. Comput.
Geosci. 24, 425–442 (2020)
4. D. Echeverría, P.W. Hemker, Manifold mapping: a two-level optimization technique. Comput.
Vis. Sci. 11(4–6), 193–206 (2008)
5. E. Haber, Computational Methods in Geophysical Electromagnetics (SIAM, Philadelphia,
2014)
6. D. Lahaye, W. Mulckhuyse, Adjoint sensitivity in PDE constrained least squares problems as a
multiphysics problem. COMPEL: Int J Comput. Math. Electri. Electron. Eng. 31(3), 895–903
(2012)
7. B. Shachor, Multi-Level Inversion Based On Mesh Decoupling. Master’s thesis, TU Delft,
2019
8. B. Shachor, Multi-Level Inversion Based on Mesh Decoupling for Poisson inverse problems
with Dirichlet BC. https://github.com/Bennyshachor/Multi-Level. Cited 12 Jan 2020
9. A. Tarantola, Inverse Problem Theory and Methods for Model Parameter Estimation (SIAM,
Philadelphia, 2005)
10. C.R. Vogel, Computational Methods for Inverse Problems (SIAM, Philadelphia, 2002)