This action might not be possible to undo. Are you sure you want to continue?
MOBK058FM MOBK058Sarris.cls December 15, 2006 7:56
Adaptive Mesh Reﬁnement
for TimeDomain Numerical
Electromagnetics
i
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK058FM MOBK058Sarris.cls December 15, 2006 7:56
Copyright ©2007 by Morgan & Claypool
All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted in
any form or by any means—electronic, mechanical, photocopy, recording, or any other except for brief quotations
in printed reviews, without the prior permission of the publisher.
Adaptive Mesh Reﬁnement for TimeDomain Numerical Electromagnetics
Costas D. Sarris
www.morganclaypool.com
ISBN: 1598290789 paperback
ISBN: 9781598290783 paperback
ISBN: 1598290797 ebook
ISBN: 9781598290790 ebook
DOI: 10.2200/S00052ED1V01Y200609CEM011
A Publication in the Morgan & Claypool Publishers series
SYNTHESIS LECTURES ON COMPUTATIONAL ELECTROMAGNETICS #11
Series Editor: Constantine A. Balanis, Arizona State University
ISSN: 19321252 (Print)
ISSN: 19321716 (Electronic)
First Edition
10 9 8 7 6 5 4 3 2 1
ii
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK058FM MOBK058Sarris.cls December 15, 2006 7:56
Adaptive Mesh Reﬁnement
for TimeDomain Numerical
Electromagnetics
Costas D. Sarris
University of Toronto
Ontario, Canada
SYNTHESIS LECTURES ON COMPUTATIONAL ELECTROMAGNETICS #11
M
&C
Morgan
&
Claypool Publishers
iii
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK058FM MOBK058Sarris.cls December 15, 2006 7:56
iv
ABSTRACT
This monograph is a comprehensive presentation of stateoftheart methodologies that can
dramatically enhance the efﬁciency of the ﬁnitedifference timedomain (FDTD) technique,
the most popular electromagnetic ﬁeld solver of the timedomain form of Maxwell’s equations.
These methodologies are aimed at optimally tailoring the computational resources needed for
the wideband simulation of microwave and optical structures to their geometry, as well as
the nature of the ﬁeld solutions they support. That is achieved by the development of robust
“adaptive meshing” approaches, which amount to varying the total number of unknown ﬁeld
quantities in the course of the simulation to adapt to temporally or spatially localized ﬁeld
features. While mesh adaptation is an extremely desirable FDTD feature, known to reduce
simulation times by orders of magnitude, it is not always robust. The speciﬁc techniques pre
sented in this book are characterized by stability and robustness. Therefore, they are excellent
computer analysis and design (CAD) tools.
The book starts by introducing the FDTD technique, along with challenges related to
its application to the analysis of reallife microwave and optical structures. It then proceeds
to developing an adaptive mesh reﬁnement method based on the use of multiresolution anal
ysis and, more speciﬁcally, the Haar wavelet basis. Furthermore, a new method to embed a
moving adaptive mesh in FDTD, the dynamically adaptive mesh reﬁnement (AMR) FDTD
technique, is introduced and explained in detail. To highlight the properties of the theoretical
tools developed in the text, a number of applications are presented, including:
• Microwave integrated circuits (microstrip ﬁlters, couplers, spiral inductors, cavities).
• Optical power splitters, Yjunctions, and couplers
• Optical ring resonators
• Nonlinear optical waveguides.
Building on ﬁrst principles of timedomain electromagnetic simulations, this book presents
advanced concepts and cuttingedge modeling techniques in an intuitive way for programmers,
engineers, and graduate students.
It is designed to provide a solid reference for highly efﬁcient timedomain solvers, em
ployed in a wide range of exciting applications in microwave/millimeterwave and optical
engineering.
KEYWORDS
Computational electromagentics, ﬁnite difference time domain (FDTD), multiresolution time
domain (MRTD), adaptive mesh reﬁnement, microwave simulation, optical waveguides.
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK058FM MOBK058Sarris.cls December 15, 2006 7:56
v
Contents
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
2. ANumerical Interface Between FDTDand Haar MRTD: Formulation
and Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Multiresolution Analysis: A brief Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2.1 General . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2.2 Wavelet Bases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.3 Derivation of TimeDomain Schemes by the Method of Moments . . . . . . . . . . 15
2.4 TwoDimensional Hybrid ArbitraryOrder Haar MRTD/FDTD
Scheme: Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.4.1 MRTD, FDTD Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.4.2 Connection Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.5 Numerical Results: Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.6 Numerical Results: Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.6.1 Metal FinLoaded Cavity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.6.2 Mesh Truncation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3. Efﬁcient Implementation of Adaptive Mesh Reﬁnement in the Haar
Waveletbased MRTDTechnique. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.2 WaveletBased FrontTracking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.3 Adaptive Haar Wavelet Simulation of Pulse Compression in an
Optical Fiber Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.3.1 Formulation of Haar MRTD Scheme. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.3.2 Boundary Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.3.3 Validation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.3.4 Implementation and Performance of Adaptive MRTD . . . . . . . . . . . . . . 57
3.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK058FM MOBK058Sarris.cls December 15, 2006 7:56
vi CONTENTS
4. The Dynamically Adaptive Mesh Reﬁnement (AMR)FDTDTechnique:
Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.2 AMRFDTD: Overview of the Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.3 Mesh Tree and Field Update Procedure in AMRFDTD . . . . . . . . . . . . . . . . . . . 70
4.3.1 Categorization of Boundaries of Child Meshes . . . . . . . . . . . . . . . . . . . . . 70
4.3.2 Update of the Mesh Tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.3.3 Field Updates on CPBType Boundaries . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.3.4 Field Updates on SBType Boundaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.3.5 Field Updates at Junctions of SBType Boundaries . . . . . . . . . . . . . . . . . . 75
4.3.6 Field Updates on PBType Boundaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.3.7 Transition from Old to New Mesh Tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.4 Adaptive Mesh Reﬁnement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.4.1 Detection of Cells That Need Reﬁnement . . . . . . . . . . . . . . . . . . . . . . . . . . 78
4.4.2 Modeling of Wave Propagation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.4.3 Clustering. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
4.4.4 Guidelines for Choosing AMRFDTD Parameters . . . . . . . . . . . . . . . . . 80
4.5 AMRFDTD and MRTD: Similarities and Differences . . . . . . . . . . . . . . . . . . . . 82
5. Dynamically Adaptive Mesh Reﬁnement in FDTD: Microwave Circuit
Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.2 Microstrip LowPass Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.3 Microstrip Branch Coupler . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
5.4 Microstrip Spiral Inductor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
5.5 Discussion: Stability and Accuracy of AMRFDTD results. . . . . . . . . . . . . . . . . . 96
5.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
6. Dynamically Adaptive Mesh Reﬁnement in FDTD: Optical Applications
and Error Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
6.1 Multilevel AMRFDTD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
6.2 Dielectric Waveguide with a Corrugated Permittivity Proﬁle . . . . . . . . . . . . . . . 106
6.3 Dielectric Waveguide Power Splitter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
6.4 Dielectric Waveguide YJunction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
6.5 Dielectric Ring Resonator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
6.6 Numerical Error Estimation and Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
6.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK058FM MOBK058Sarris.cls December 15, 2006 7:56
vii
Acknowledgments
First and foremost, I would like to express my gratitude to Prof. Constantine Balanis, for asking
me to contribute to this series, and to Joel Claypool for his support to this project.
The research included in the part of this monograph that refers to the MRTD technique
was conducted at the University of Michigan, Ann Arbor, under the inspiring guidance of
Prof. Linda Katehi, to whom the author is ever indebted for her mentorship and example of
achievement through excellence. Financially, this work was supported by the Army Research
Ofﬁce (and Dr. James F. Harvey should be personally acknowledged for his continued interest)
and the U.S. Army CECOM (with Dr. Barry Perlman deserving a special acknowledgment as
well).
The work on the dynamically adaptive mesh reﬁnement ﬁnitedifference timedomain
(AMRFDTD) technique is the result of a productive collaboration of the author with
Dr. Yaxun Liu at the Edward S. Rogers Sr. Department of Electrical and Computer Engineer
ing, University of Toronto. It was primarily supported by the Natural Sciences and Engineering
Research Council of Canada (NSERC) through a Discovery Grant to the author. Dr. Liu was
partially supported by the Ontario Centers of Excellence through a Collaborative Research
Grant and NSERC through a Collaborative Research and Development (CRD) Grant. All
this generous support is gratefully acknowledged.
Costas D. Sarris, Toronto, Ontario, September 2006
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK058FM MOBK058Sarris.cls December 15, 2006 7:56
viii
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK058FM MOBK058Sarris.cls December 15, 2006 7:56
ix
List of Tables
2.1 Position of and Field Values at the MRTD Equivalent Grid Points of Fig. 2.10. . . . . . 19
3.1 Validation Data for Nonadaptive MRTD Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
5.1 Computation Time and SParameter Error Metric E
S
for the Microstrip
LowPass Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
5.2 Computation Time and SParameter Error Metric E
S
for the Microstrip
Branch Coupler . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
5.3 Computation Time and SParameter Error Metric E
S
for the Microstrip
Spiral Inductor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
6.1 Computation Time and TimeDomain Error Metric E
t
for the Optical
Waveguide of Fig. 6.1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
6.2 Computation Time and TimeDomain Error Metric E
t
for the Dielectric
Waveguide of Fig. 6.4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
6.3 Computation Time and TimeDomain Error Metric E
t
for the Dielectric
Waveguide Power Splitter for Fields Recorded at Port 2. . . . . . . . . . . . . . . . . . . . . . . . . . 110
6.4 Computation Time and TimeDomain Error Metric E
t
for the Dielectric
Waveguide YJunction for Fields Recorded at Port 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
6.5 Computation Time and TimeDomain Error Metric E
t
for the Dielectric Ring
Resonator for Fields Recorded at Port 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
6.6 Resonant Frequencies (in THz) of the Dielectric Ring Resonator. . . . . . . . . . . . . . . . . 121
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK058FM MOBK058Sarris.cls December 15, 2006 7:56
x
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK058FM MOBK058Sarris.cls December 15, 2006 7:56
xi
List of Figures
2.1 Coarse rectangular mesh for a layered coplanar waveguide structure, with cells
containing variable dielectric permittivity and electric conductivity proﬁles . . . . . . . . . . . 8
2.2 Projections of a function f on three successive approximation spaces V
0
, V
1
, V
2
. . . . . 10
2.3 Haar scaling and mother wavelet function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.4 Demonstration of the multiresolution principle as implemented by the Haar basis . . . 13
2.5 Haar scaling and mother wavelet functions in the Fourier domain. . . . . . . . . . . . . . . . . . 13
2.6 Battle–Lemarie scaling and mother wavelet function. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.7 Battle–Lemarie scaling and mother wavelet functions in the Fourier domain . . . . . . . . 14
2.8 Equivalent grid points (e.g.p.) within a scaling cell as introduced by a ﬁrstorder
wavelet expansion in x, zdirections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.9 Explanation of deﬁnition (2.22) for the Haar basis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.10 Mesh reﬁnement in the Haar basis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.11 Electric/magnetic ﬁeld equivalent grid points for zeroorder Haar MRTD in
one dimension, under the common convention of half a cell offset between
electric/magnetic scaling cells . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.12 Electric/magnetic ﬁeld equivalent grid points for zeroorder Haar MRTD in
one dimension, with electric/magnetic scaling cell offset chosen in consistence
with (2.33) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.13 Onedimensional interface between FDTD and a ﬁrstorder Haar MRTD
scheme. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.14 Reﬂectionless propagation through a onedimensional FDTD/fourthorder
MRTD interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.15 FDTD update from MRTD data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.16 MRTD update from FDTD data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.17 Empty rectangular cavity geometry and interface of a 3 by 3order MRTD/FDTD
mesh conﬁguration (dimensions are given in FDTD cells, MRTD mesh is 2 × 2) . . . 31
2.18 Empty rectangular cavity geometry and interface of a 3 by 4order MRTD/FDTD
mesh conﬁguration (dimensions are given in FDTD cells, MRTD mesh is 1 × 1) . . . 31
2.19 Time and frequency domain patterns of electric ﬁeld (E
y
) sampled within the cavity
of Fig. 2.17, as obtained by FDTD and the hybrid scheme . . . . . . . . . . . . . . . . . . . . . . . . 32
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK058FM MOBK058Sarris.cls December 15, 2006 7:56
xii MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
2.20 Time and frequency domain patterns of electric ﬁeld (E
y
) sampled within the
cavity of Fig. 2.18, as obtained by FDTD and the hybrid scheme . . . . . . . . . . . . . . . . . . 33
2.21 Empty square cavity geometry and interface of a 4 by 4order MRTD/FDTD
mesh conﬁguration (dimensions are given in FDTD cells, MRTD mesh is 1 × 1) . . . 33
2.22 Electric ﬁeld pattern for the TE
22
mode of the empty cavity of
Fig. 2.21 (4 by 4 MRTD) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.23 Electric ﬁeld pattern for the TE
22
mode of the empty cavity of Fig. 2.21
(4 by 4 MRTD) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.24 Metal ﬁnloaded cavity geometry and interface of a 2 by 2order MRTD/FDTD
mesh conﬁguration (dimensions are given in FDTD cells, MRTD mesh is 4 × 4) . . . 35
2.25 Electric ﬁeld distribution in the metal ﬁnloaded cavity (dominant TEmode) . . . . . . 35
2.26 Electric ﬁeld sampled within the ﬁnloaded cavity as a function of time for
time steps 18,000–20,000 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.27 Concept of MRTD mesh termination via an FDTD/MRTD interface . . . . . . . . . . . . . 37
2.28 Slab waveguide geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.29 Waveforms of incident and reﬂected electric ﬁeld (scaling and wavelet terms of
order 0–2), sampled in front of the dielectric slab of Fig. 2.28. . . . . . . . . . . . . . . . . . . . . . 38
2.30 Numerical and theoretical S
11
for the slab geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.31 Case study for the interfacebased termination of a twodimensional MRTD
domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.32 FDTDUPML interfaced with a 0 by 0order MRTD (MRTD mesh is 32 × 32) . . 40
2.33 FDTDUPML interfaced with a 2 by 2order MRTD (MRTD mesh is 8 × 8) . . . . 40
2.34 FDTDUPML interfaced with a 4 by 4order MRTD (MRTD mesh is 2 × 2) . . . . 41
2.35 Comparison of FDTD and MRTD/FDTD interface results for the three case
studies of Figs. 2.32–2.34 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.36 Electric ﬁeld at point B of Fig. 2.33, when a 4, 8, 16 FDTD UPML is used
to terminate the MRTD mesh. An FDTD solution (with a 32cell UPML)
is appended for comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
2.37 Spatial ﬁeld distribution in the interfaceterminated domain of Fig. 2.31
at time steps 20, 50. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
2.38 Spatial ﬁeld distribution in the interfaceterminated domain of Fig. 2.31
at time steps 80, 600. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.39 Concept of ABC termination of an MRTD mesh; the FDTD region terminated
into the ABC should at least extend over one scaling MRTD cell . . . . . . . . . . . . . . . . . . 43
2.40 Incident and reﬂected ﬁeld waveforms, for Mur’s ﬁrstorder ABC (implemented via
an FDTD/MRTD interface), in the case of onedimensional wave propagation
(Fig. 2.39), for MRTD orders 0–3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK058FM MOBK058Sarris.cls December 15, 2006 7:56
LISTOF FIGURES xiii
2.41 Reﬂection coefﬁcient for Mur’s ﬁrstorder ABC (implemented via an
FDTD/MRTD interface), in the case of onedimensional wave propagation
(Fig. 2.39) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.1 Unthresholded to total number (%) of wavelet coefﬁcients of orders 0, 1, for a
fourthorder Haar MRTD simulation of the problem of Fig. 2.28 . . . . . . . . . . . . . . . . . 49
3.2 Unthresholded to total number (%) of wavelet coefﬁcients of orders 2, 3, for a
fourthorder Haar MRTD simulation of the problem of Fig. 2.28 . . . . . . . . . . . . . . . . . 49
3.3 Comparison of S
11
, derived by adaptive MRTD with = 10
−5
, 10
−4
for the
dielectric slab incidence problem of Fig. 2.28 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.4 Evolution of wavefronts in the dielectric slab simulation (Fig. 2.28) and regions
of unthresholded wavelet coefﬁcients (boundaries are deﬁned by blue dots,
a threshold = 10
−4
is used) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.5 Evolution of wavefronts in the dielectric slab simulation (continued) . . . . . . . . . . . . . . . 52
3.6 Haar MRTD equivalent grid points at the beginning and the end of the ﬁber . . . . . . . 55
3.7 MRTD and FDTD results for the forward ﬁeld intensity at the beginning,
the middle, and the end of the optical ﬁber ﬁlter (1000 grid points) . . . . . . . . . . . . . . . . 57
3.8 Spatiotemporal propagation and compression of the nonlinear optical pulse
(forward ﬁeld intensity is shown) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.9 Spatiotemporal evolution of forward ﬁeld wavelet coefﬁcients (square of magnitude
of wavelet coefﬁcients is shown) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.10 Graphical presentation of the waveletbased, fronttracking algorithm. . . . . . . . . . . . . . 60
3.11 Incident and transmitted forward ﬁeld intensity for a threshold of 10
−7
. . . . . . . . . . . . . 61
3.12 Transmitted forward ﬁeld intensity for thresholds of 10
−4
, 10
−1
. . . . . . . . . . . . . . . . . . . 61
3.13 CPU time economy for the adaptive MRTD code, with respect to FDTD,
for absolute thresholds from 10
−7
to 0.1 applied to Haar MRTD . . . . . . . . . . . . . . . . . . 62
3.14 Relative error (%) in the peak intensity of the transmitted forward ﬁeld for
adaptive Haar MRTD. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.1 (a) Geometry and (b) data structure of a threelevel mesh tree. . . . . . . . . . . . . . . . . . . . . . 68
4.2 Sampling points of electric and magnetic ﬁeld components in a Yee cell. Solid
arrows are for electric ﬁeld, whereas hollow arrows are for magnetic ﬁeld.
The shaded areas represent reﬁned Yee cells. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.3 Types of boundaries of a child mesh. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.4 Sampling points for electric ﬁeld on CPBtype boundary. The solid arrows are for
the child mesh. The hollow arrows are for the parent mesh. . . . . . . . . . . . . . . . . . . . . . . . 73
4.5 Interface between two meshes of the same level. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
4.6 Typical SBtype boundaries. A, B, C, D, and E are meshes of the same level. . . . . . . . 75
4.7 Transition from old mesh tree to new mesh tree. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK058FM MOBK058Sarris.cls December 15, 2006 7:56
xiv MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
4.8 Splitting and shrinking of boxes (child meshes). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
5.1 Microstrip lowpass ﬁlter geometry. P1 and P2 denote ports 1 and 2, respectively. . . . 84
5.2 AMR coverage and number of child meshes at different time steps for the
lowpass ﬁlter simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
5.3 Vertical electric ﬁeld magnitude at z = 0.4 mm and t = 0. One child mesh;
AMR coverage is 4.3% . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
5.4 Vertical electric ﬁeld magnitude at z = 0.4 mm and t = 100t. One child mesh;
AMR coverage is 13.4% . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
5.5 Vertical electric ﬁeld magnitude at z = 0.4 mm and t = 200t. One child mesh;
AMR coverage is 52.5% . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
5.6 Vertical electric ﬁeld magnitude at z = 0.4 mm and t = 500t. Three child
meshes; AMR coverage is 44.2% . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
5.7 Vertical electric ﬁeld magnitude at z = 0.4 mm and t = 800t. Three child meshes;
AMR coverage is 28.0% . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
5.8 Vertical electric ﬁeld magnitude at z = 0.4 mm and the center of the microstrip
line, 3 mm from the right edge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
5.9 Comparison of the scatteringparameters of the lowpassﬁlter of Fig. 5.1,
obtained by FDTD and AMRFDTD. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
5.10 Effect of threshold for gradient θ
g
. The rest of the AMR parameters are:
θ
e
= 0.1, θ
c
= 0.7, N
AMR
= 10, σ
AMR
= 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
5.11 Effect of threshold for energy θ
e
. The rest of the AMR parameters are:
θ
g
= 0.01, θ
c
= 0.7, N
AMR
= 10, σ
AMR
= 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
5.12 Effect of threshold for coverage efﬁciency θ
c
. The rest of the AMR
parameters are: θ
g
= 0.01, θ
e
= 0.01, N
AMR
= 10, σ
AMR
= 2 . . . . . . . . . . . . . . . . . . . . . 91
5.13 Effect of interval of AMR N
AMR
. The rest of the AMR parameters are:
θ
g
= 0.01, θ
e
= 0.01, θ
c
= 0.7, σ
AMR
= 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
5.14 Effect of spreading factor of AMR σ
AMR
. The rest of the AMR parameters are:
θ
g
= 0.01, θ
e
= 0.01, θ
c
= 0.7, N
AMR
= 10 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
5.15 Microstrip branch coupler geometry. P1, P2, P3, and P4 denote port 1, 2, 3, and 4 . . 93
5.16 S
11
for the microstrip branch coupler geometry of Fig. 5.15, as determined
by FDTD and AMRFDTD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
5.17 S
21
for the microstrip branch coupler geometry of Fig. 5.15, as determined
by FDTD and AMRFDTD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
5.18 S
31
for the microstrip branch coupler geometry of Fig. 5.15, as determined
by FDTD and AMRFDTD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK058FM MOBK058Sarris.cls December 15, 2006 7:56
LISTOF FIGURES xv
5.19 S
41
for the microstrip branch coupler geometry of Fig. 5.15, as determined
by FDTD and AMRFDTD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
5.20 Vertical electric ﬁeld magnitude at z = 0.6 mm and the center of the microstrip
line of port 3, 6 mm from the edge. The excitation is imposed at 3 mm from
the edge of port 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
5.21 Spiral inductor geometry. P1 and P2 denote port 1 and port 2, respectively . . . . . . . . . 98
5.22 Vertical ﬁeld evolution and associated mesh reﬁnement in the microstrip spiral
inductor, simulated by a twolevel dynamic AMRFDTD. . . . . . . . . . . . . . . . . . . . . . . . . 99
5.23 Sparameters for the spiral inductor geometry of Fig. 5.21, as determined by
FDTD and AMRFDTD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
5.24 Vertical electric ﬁeld magnitude at z = 0.6 mm and the center of the microstrip
line, 6 mm from the right edge. The excitation is imposed at 3 mm from the left
edge of the microstrip line . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
6.1 Optical waveguide geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
6.2 Vertical electric ﬁeld magnitude distribution across the optical waveguide of
Fig. 6.1, at t = 100t. One level2 child mesh and one level3 child mesh
are shown . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
6.3 Temporal waveform of the vertical electric ﬁeld magnitude at the center of the
optical waveguide of Fig. 6.1, 1.5 μm from the right edge . . . . . . . . . . . . . . . . . . . . . . . . 105
6.4 Geometry of the dielectric waveguide with a corrugated permittivity proﬁle.
Dimensions are given in units of μm. P1 and P2 denote ports 1 and 2,
respectively . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
6.5 Electric ﬁeld at the center of the dielectric waveguide of Fig. 6.4, 1.2 μm from
the right edge. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
6.6 Electric ﬁeld at the center of the dielectric waveguide of Fig. 6.4, 1.2 μm from the
right edge, obtained with the dynamic AMRFDTD, for 40,000 time steps,
indicating the absence of latetime instability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
6.7 Geometry of the dielectric waveguide power splitter. Dimensions are given in units
of μm. P1, P2, P3, and P4 denote ports 1, 2, 3, and 4, respectively . . . . . . . . . . . . . . . . 109
6.8 Electric ﬁeld at 1 μm from the edge of port 2 of the power splitter of Fig. 6.7 . . . . . . 110
6.9 Electric ﬁeld at 1 μm from the edge of port 3 of the power splitter of Fig. 6.7 . . . . . . 111
6.10 Electric ﬁeld obtained with the dynamic AMRFDTD at 1 μm from the edge
of port 2 of the power splitter of Fig. 6.7, obtained with the dynamic
AMRFDTD, for 40,000 time steps, indicating the absence of
latetime instability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK058FM MOBK058Sarris.cls December 15, 2006 7:56
xvi MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
6.11 Electric ﬁeld magnitude distribution across the power splitter of Fig. 6.7,
at t = 100t. There are three meshes in total: one level1, one level2,
and one level3 mesh . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
6.12 Electric ﬁeld magnitude distribution across the power splitter of Fig. 6.7,
at t = 300t. There are three meshes in total: one level1, one level2,
and one level3 mesh . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
6.13 Electric ﬁeld magnitude distribution across the power splitter of Fig. 6.7,
at t = 400t. There are eight meshes in total: one level1, three level2,
and four level3 meshes (tracking transmitted and reﬂected wavefronts) . . . . . . . . . . . 113
6.14 Geometry of the dielectric waveguide Yjunction. The dielectric constant of the
waveguide is 9 and the dielectric constant of the surrounding medium is 1.
Dimensions are given in units of μm. P1, P2, and P3 denote ports
1, 2, and 3, respectively . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
6.15 Electric ﬁeld at the center of the dielectric waveguide Yjunction of Fig. 6.14,
2 μm from the edge of port 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
6.16 Electric ﬁeld at the center of the dielectric waveguide Yjunction of Fig. 6.14,
2 μm from the edge of port 2, obtained with the dynamic AMRFDTD, for
25,000 time steps, indicating the absence of latetime instability . . . . . . . . . . . . . . . . . . 115
6.17 Electric ﬁeld magnitude distribution across the Yjunction of Fig. 6.14,
at t = 100t. There are three meshes in total: one level1, one level2,
and one level3 mesh . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
6.18 Electric ﬁeld magnitude distribution across the Yjunction of Fig. 6.14,
at t = 400t. There are three meshes in total: one level1, one level2,
and one level3 mesh (enclosed in the black solid line) . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
6.19 Electric ﬁeld magnitude distribution across the Yjunction of Fig. 6.14,
at t = 600t. There are three meshes in total: one level1, one level2,
and one level3 mesh (enclosed in the black solid line). Note that level2 and
level3 meshes track the pulse into the junction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
6.20 Electric ﬁeld magnitude distribution across the Yjunction of Fig. 6.14,
at t = 800t. There are four meshes in total: one level1, two level2,
and one level3 mesh (enclosed in the black solid line) . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
6.21 Geometry of the dielectric ring resonator coupled to two dielectric waveguides.
Dimensions are in units of μm. P1, P2, P3, and P4 denote ports 1, 2, 3, and 4,
respectively . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
6.22 Electric ﬁeld at 2 μm from port 2 of the ring resonator of Fig. 6.21 . . . . . . . . . . . . . . . 118
6.23 Electric ﬁeld at 2 μm from port 2 of the ring resonator of Fig. 6.21 . . . . . . . . . . . . . . . 119
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK058FM MOBK058Sarris.cls December 15, 2006 7:56
LISTOF FIGURES xvii
6.24 Electric ﬁeld at 2 μm from port 2 of the ring resonator of Fig. 6.21, obtained with
the dynamic AMRFDTD, for 100,000 time steps, indicating the absence of
latetime instability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
6.25 S
21
obtained with the fourlevel AMRFDTD. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
6.26 Error versus simulation time for variable (a) N
AMR
, (b) σ
AMR
, (c) θ
c
. . . . . . . . . . . . . . . 122
6.27 Dependence of −l og
10
E on θ
e
and θ
g
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
6.28 Curveﬁtted dependence of −l og
10
˜
E on θ
e
and θ
g
based on (6.6) . . . . . . . . . . . . . . . . . 124
6.29 Curveﬁtted dependence of error bound −l og
10
E
b
on θ
e
and θ
g
based on (6.6) . . . . . 125
6.30 Geometry of a directional coupler (not drawn to scale). Dimensions are given in
μm. The surrounding medium is air. The dielectric constant of the waveguide
and thecenter piece is 3.24 and 4.41, respectively . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
6.31 Electric ﬁeld at port 3 and 4 of the directional coupler . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK058FM MOBK058Sarris.cls December 15, 2006 7:56
xviii
P1: IML/FFX P2: IML
MOBK05801 MOBK058Sarris.cls December 15, 2006 7:29
1
C H A P T E R 1
Introduction
Recent advances in solidstate device technology have enabled the development of planar mono
lithic microwave integrated circuits (MMICs) in multilayer, densely integrated architectures.
These provide lowcost, lightweight and reproducible alternatives to conventional, lowloss
but bulky components. In addition, circuit miniaturization allows for the implementation of
systemonachip concepts, that serve the need of the wireless industry for light and small
transceiver structures well.
The next steps beyond the current stateoftheart include the monolithic integration
of silicon/silicon germanium (SiGe) circuits, microelectromechanical switch (MEMS) de
vices, micromachined components (e.g., ﬁlters, resonators, multiplexers), and digital CMOS
signal processing units, for high power/lowloss, tactical and commercial communication
systems.
As a result, approximate circuit design approaches, which are based on representing
multielement structures as an assembly of cascaded, individually functioning blocks, do not
constitute a reliable simulation tool for cuttingedge applications, due to their failure to account
for the inevitable electromagnetic coupling of closely spacedcomponents. Yet, the latter produces
parasitic effects such as crosstalk and signal distortion, that generate signal integrity issues with
an adverse impact on the overall performance of wireless communication systems. The profound
need to model and eventually mitigate such effects fuels the research on fullwave techniques
of computational electromagnetics (CEM). Methods that are referred to as such, proceed to
the formulation of a solution to a given boundary value electromagnetic problem, with no
assumptions on the shape of the wavefronts involved. On the other hand, highfrequency
asymptotics assume the reduction of wavefronts to rays, facilitating the application of optics
principles to problems of scattering from electrically large, unconventionally shaped objects [1].
In essence, numerical schemes are developed by formulating an operator problem and
subsequently pursuing its solution by computational (as opposed to analytical) means. The
question of the operator itself being either an integral or a differential one, classiﬁes accordingly
the related scheme. Electromagnetic problems are cast in the form of an integral equation by
use of a Green function, which assumes in general the form of a dyad G( ¯ r , ¯ r
). Despite the fact
P1: IML/FFX P2: IML
MOBK05801 MOBK058Sarris.cls December 15, 2006 7:29
2 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
that special forms of integral equations can be treated by analytical techniques, such as, among
others, the WienerHopf method [2], the dominant integral equation solver is arguably the
numerical method of moments [3].
A Green function incorporates an intuitively rich description of the physics of a given
problem. In addition, it inherently accounts for Sommerfeld’s radiation condition [4], rendering
the use of absorbing boundaries unnecessary. However, the use of Green dyadics becomes ex
tremely complex, if tractable at all, in realistic cases, naturally giving way to differential methods,
that trade versatility with computational effort. In fact, techniques such as the ﬁnite element
method (FEM, [5]), the ﬁnite difference timedomain (FDTD, [6]), or the transmission line
matrix (TLM, [7]) method that solve partial differential equations in a mesh, translate into
largescale algebraic systems. Nevertheless, recent advances in highperformance and parallel
computing have made these methods a feasible alternative, thus prompting an everincreasing
research interest in those.
This monograph focuses on the solution of Maxwell’s equations in the timedomain.
To this end, a broadband impulse excitation, such as a Gaussian pulse, is injected as an
initial condition to the computational domain. Then, a marchingintime procedure is fol
lowed to calculate the evolution of its wavefront, until the steady state is reached. On the
contrary, frequencydomain techniques extract ﬁeld information at a given frequency point.
Clearly, the latter are preferred to the former, when narrowband simulations are pursued. In
addition, frequency sweep techniques, such as the asymptotic waveformevaluation (AWE, [8]),
have enhanced the capability of frequencydomain techniques to produce broadband results in
certain cases, suffering though from general issues of accuracy and efﬁciency.
Timedomain differential methods are becoming increasingly popular among the elec
tromagnetic community as a consequence of their versatility and their ability to provide sim
ulation results that are intuitively meaningful to circuit designers and microwave engineers.
In particular, the ﬁnitedifference timedomain (FDTD) technique offers a mathematically
straightforward analysis method, suitable for arbitrary electromagnetic geometries. However,
since Yee’s scheme [6] is secondorder accurate only, and sensitive to numerical dispersion, a
dense discretization of at least ten, but usually twentyﬁve points per wavelength is necessary for
the extraction of a convergent solution. Therefore, the FDTD treatment of either electrically
large geometries or ﬁne detail structures typically results in a computationally intensive, memory
and execution time consuming calculation. This fact naturally poses the signiﬁcant question of
deriving new, reasonably accurate techniques, with computational requirements that can be met
by nowadays computer performance.
As an alternative to the conventional FDTD, several higher order numerical techniques
have been developed [9], aimed at the discretization of electromagnetic structures at rates that
P1: IML/FFX P2: IML
MOBK05801 MOBK058Sarris.cls December 15, 2006 7:29
INTRODUCTION 3
may even approach the Nyquist limit. Furthermore, the incorporation of subgridding algorithms
in the FDTD scheme [10] has indicated the potential of disconnecting the mean cell size
within a domain, from the size of the smallest geometric detail that is contained in the latter.
Yet, standard subgridding techniques involve spatiotemporal interpolations or extrapolations
at the boundaries of different resolution parts of the numerical grid, that render the rigorous
enforcement of the divergence free nature of the magnetic ﬁeld and the continuity conditions
a rather subtle issue.
If allowing for a relatively coarse mesh and taking into account localized geometric details
are two challenges that novel numerical schemes are expected to meet, a third, but equally
important one, is adaptivity. The concept of adaptivity implies the possibility of dynamic mesh
reﬁnement at regions of the computational domain that are electromagnetically active.
As the timedomain characterization of planar structures typically registers the history of
a wideband (narrow) pulse propagation, adaptively imposing dense gridding conditions only in
and around the pulse and the products of its retroreﬂections can further extend the efﬁciency
of a technique, beyond the limits that a static subgridding scheme can provide for.
Recently, waveletbased timedomain methods, henceforth referred to as multiresolu
tion timedomain (MRTD), employing Battle–Lemarie, Daubechies, biorthogonal, and Haar
wavelets have been presented, for example, in [11–16] and references therein. According to
MRTD, the electromagnetic ﬁeld components are spatially expanded in a basis composed of
scaling and wavelet functions, that effectively deﬁne a discrete mesh in space. In addition,
pulse functions are used for the temporal ﬁeld expansion. Upon substitution of these expan
sions into Maxwell’s equations, Galerkin method is applied to derive ﬁnite difference equations
with respect to the ﬁeld scaling and wavelet spatiotemporal coefﬁcients. These ﬁnite differ
ence equations are solved by “marching in time,” in the FDTD sense. It is noted that spatial
variations of ﬁeld components are reﬂected on the magnitude of their respective wavelet coef
ﬁcients, which are either stored in computer memory or removed according to certain criteria
(thresholding). In fact, several types of MRTD schemes with distinct numerical properties can
be derived, depending on the wavelet basis that is employed for the discretization of Maxwell’s
equations.
In [11], Battle–Lemarie scaling functions enhanced by one wavelet level (zeroorder
wavelet functions) were employed for the formulation of the socalled WMRTD scheme.
Signiﬁcant memory economy was shown to be associated to this method, because of its highly
linear dispersion properties, that allowed for the modeling of electromagnetic structures at
discretization rates that approached the Nyquist limit (λ/2). It is worth noting though, that
these computational gains were strictly associated to the Battle–Lemarie scaling function rather
than the Battle–Lemarie wavelet function basis. Furthermore, space adaptivity via an a posteriori
P1: IML/FFX P2: IML
MOBK05801 MOBK058Sarris.cls December 15, 2006 7:29
4 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
thresholding of wavelet coefﬁcients was pursued in [17]. Additionally, the range of WMRTD
applications that have been demonstrated so far, includes printed transmission line analysis
[18], nonlinear circuit modeling [19], complex airdielectric boundary problems, and antenna
geometries [20].
Because of their relative simplicity, Haar MRTDschemes present an attractive alternative
to their entire domain counterparts and have therefore become the subject of several studies [14,
21]. Since Haar MRTDconstitutes the “wavelet extension” of the FDTDtechnique, a great deal
of FDTD algorithms can be modiﬁed in order to resolve Haar MRTD modeling issues. Also,
ﬁnite difference equations in Haar MRTDinvolve only “nearest neighbor” interactions, which is
an attractive feature of an algorithm especially when it is combined with parallelization/domain
decomposition techniques. For these reasons, the popularity of Haar MRTDhas recently grown
in the computational electromagnetics community.
It is worth noting that wavelet methods meet all aforementioned challenges for time
domain numerical schemes. First, by employing a highorder basis (such as cubic splines),
a coarse mesh can be established, even when electrically small geometric features are to be
included. Then, by adding wavelets and using adaptivity algorithms, similar to the ones that have
been demonstrated in signal processing applications [22], a straightforward implementation of
adaptive gridding is obtained. Moreover, an inherent contradiction related to MRTD schemes
is that although their efﬁciency is expected to increase with the order of the multiresolution
expansion at hand for homogeneous domains, the complexity of conductor, dielectric, and
boundary modeling that they present, also increasing with wavelet order, seriously compromises
their potential applicability to stateoftheart devices. This complexity is associated with the
incorporation of highorder wavelets into the ﬁnite difference expression of the constitutive
relations for dielectrically inhomogeneous or conducting media.
This monograph is aimed at presenting two efﬁcient solutions to the problem of estab
lishing mesh adaptive timedomain numerical electromagnetic techniques.
First, a numerical interface between FDTD and the Haar waveletbased MRTD is pre
sented in detail, in Chapter 2. It is shown that the interface can greatly facilitate the application
of both localized hard boundary conditions and the implementation perfectly matched layer
(PML) absorbers for mesh truncation. It is worth noting that no interpolations or extrapola
tions (spatial or temporal) are necessary for the establishment of this interface. In addition, no
stability issues are associated with its application.
Then, Chapter 3 deals with the most fundamental, applicationoriented question, related
towavelet methods: Howtoefﬁciently implement the widely advertisedmeshadaptivity, without
defeating the purpose of obtaining operation savings. To this end, wavefronttracking ideas
are introduced and applied to onedimensional linear and nonlinear optical wave propagation
problems.
P1: IML/FFX P2: IML
MOBK05801 MOBK058Sarris.cls December 15, 2006 7:29
INTRODUCTION 5
Following up to the technique of Chapter 3, Chapter 4 presents a computationally ﬂexible
and efﬁcient method that can achieve mesh adaptivity, while fully preserving the versatility
of FDTD. This technique, called the dynamically adaptive mesh reﬁnement (AMR)FDTD,
essentially translates the AMRtechnique of computational ﬂuid dynamics to threedimensional
electromagnetic problems. Its exceptional performance is demonstrated through microwave and
optical applications in Chapters 5, 6.
P1: IML/FFX P2: IML
MOBK05801 MOBK058Sarris.cls December 15, 2006 7:29
6
P1: IML/FFX P2: IML
MOBK05802 MOBK058Sarris.cls December 15, 2006 7:32
7
C H A P T E R 2
ANumerical Interface Between
FDTDand Haar MRTD:
Formulation and Applications
A numerically stable interface between the Haar waveletbased MRTD technique and FDTD
is presented in this section. Such a hybridization of MRTD facilitates the application of the
method to open structures and inhomogeneous circuit geometries, where the use of highorder
wavelets signiﬁcantly complicates the formulation of a pure MRTD scheme. Furthermore, it
allows for the straightforward enforcement of localized and perfectly matched layer (PML)
type absorbing boundary conditions. The fact that the implementation of the proposed interface
involves no spatial or temporal interpolations indicates its potential to efﬁciently connect FDTD
and Haar MRTD.
2.1 INTRODUCTION
The timedomain characterization of microwave structures, often encountered in wireless front
end applications such as ﬁlter, resonator or feed components, usually includes the modeling of
ﬁne detail complex boundaries and regions of dynamically varying ﬁeld distributions. For this
purpose, dense gridding conditions are necessary for the extraction of an accurate solution to the
problemat hand. Nevertheless, the implementation of a uniformly dense mesh for an electrically
large structure, translates to computationally burdensome simulations, as a result of both the
size of the domain (in cells) and the stabilityrelated restriction on the time step of a ﬁnite
difference timedomain (FDTD)type of scheme, imposed by the small dimensions of the unit
cell. Hence, the incorporation of local mesh reﬁnement techniques into conventional time
domain solvers is motivated, as a means of alleviating the overall cost of modeling structures
with localized geometric details.
Typically, subgridding techniques are implemented via spatiotemporal interpolations or
extrapolations. Furthermore, a higherorder static subgridded FDTD algorithm, that needs no
interpolatory operations has been recently proposed in [23]. However, waveletbased numerical
P1: IML/FFX P2: IML
MOBK05802 MOBK058Sarris.cls December 15, 2006 7:32
8 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
FIGURE2.1: Coarse rectangular mesh for a layered coplanar waveguide structure, with cells containing
variable dielectric permittivity and electric conductivity proﬁles
algorithms provide a natural framework for the implementation of dynamically adaptive sub
gridding, which holds the promise of signiﬁcantly accelerating conventional CAD tools, even
those with local subgridding features.
For homogeneous domains, the computational efﬁciency of waveletbased schemes, such
as MRTD, is expected to increase with the order of the multiresolution expansion [24]. Yet,
the complexity of conductor, dielectric, and boundary modeling that they present, also increas
ing with wavelet order, seriously compromises their potential applicability to stateoftheart
devices. In particular, whenever a single cell includes variable material properties (such as di
electric and/or conducting layers), direct MRTD update equations are replaced by matrix ex
pressions, resulting from the discretization of constitutive relations [11, 14]. Such cases are
encountered in typical microwave devices, such as microstrip lines and coplanar waveguides
(Fig. 2.1). In addition, the most effective mesh truncation technique nowadays, the perfectly
matched layer (PML) absorber, is itself an inhomogeneous, uniaxially anisotropic (both elec
trically and magnetically) material.
The aforementioned contradiction is addressed here by means of a hybrid approach that
connects the FDTD technique with the Haar waveletbased MRTD, via a numerical interface.
The purpose of this approach is to establish an efﬁcient algorithm, allowing for the combination
of the versatility of FDTD with the adaptivity of MRTD, employing the ﬁrst in geometrically
complex parts of the domain and the second in homogeneous regions.
In the past, interfaces between different numerical methods have been developed in order
to address speciﬁc problems of interest. In [25], the frequencydomain method of moments
(MoM) was coupled to FDTD, for the analysis of ground penetrating radar (GPR) problems,
where GPR antenna operation was modeled by MoM, while the inhomogeneous ground (in
cluding dielectric stratiﬁcation) was incorporated in an FDTD mesh. In [26], a Daubechies
scaling functionbased Wavelet–Galerkin scheme was coupled to FDTD for the modeling of
P1: IML/FFX P2: IML
MOBK05802 MOBK058Sarris.cls December 15, 2006 7:32
ANUMERICAL INTERFACEBETWEENFDTDANDHAARMRTD 9
wedgeloaded twodimensional waveguide structures, apparently leading to signiﬁcant compu
tational savings compared to its pure FDTD counterpart. Still, the general applicability and
efﬁciency of the scheme was questionable, given the largely different dispersion characteris
tics of the two methods. Moreover, a hybridization of the transmission line matrix (TLM)
method and MRTD was pursued in [27], where the issue of dealing with disparate dispersion
methods was explicitly addressed via a conditionally stable space–time interpolation scheme.
Finally, [28] demonstrated a technique for including FDTDmodeled lumped elements in a
threedimensional domain simulated by the Haar waveletbased MRTD formulation that was
introduced in [14] and was restricted to one wavelet level.
This chapter proposes an interface between a Haar wavelet MRTD scheme of arbitrary
number of wavelet levels and FDTD [29]. Under certain conditions, the two schemes are
characterizedby the same dispersionproperties, a fact that is utilizedinorder tocouple themwith
no interpolations or extrapolations and with absolutely no spurious reﬂections at their interface.
The structure of this chapter is as follows: First, a brief overview of multiresolution
analysis (that the MRTD technique is based upon) is provided, along with a presentation of
how the technique can be systematically formulated. Then, the formulation of a Haar MRTD
scheme (supporting anarbitrarily highspatial resolution) is presented. Noting that the dispersion
equation for the latter coincides with the dispersion equation of an FDTD scheme of the same
spatial resolution, a simple connection algorithm between the two is proposed. Emphasis is
placed on the implementation problems that arise in the application of update equations at
the boundaries of the two domains. Validation studies include the analysis of twodimensional
dielectric cavity structures, under various FDTD/MRTD mesh conﬁgurations. Finally, the
method is applied for mesh truncation in open domain problems and the modeling of structures
with metal inserts such as a ﬁnloaded cavity.
2.2 MULTIRESOLUTIONANALYSIS: ABRIEF OVERVIEW
The theoretical foundation of waveletbased numerical techniques is the socalled multiresolu
tion analysis theory, that is often encountered under the headline of wavelets. A brief overview
of this subject is provided here for the purpose of completeness. A detailed study of the topic is
contained in [30].
2.2.1 General
Wavelet bases are a mathematical tool for hierarchical decomposition of functions in an orthog
onal expansion, according to the general scheme:
f (ξ) =
¸
k
c
k
φ(ξ −k) +
¸
j
¸
k
d
j,k
ψ(2
j
ξ −k) (2.1)
P1: IML/FFX P2: IML
MOBK05802 MOBK058Sarris.cls December 15, 2006 7:32
10 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
FIGURE2.2: Projections of a function f on three successive approximation spaces V
0
, V
1
, V
2
where f is assumed to belong to L
2
(R).
1
In (2.1), the ﬁrst sum represents the projection of
f (ξ) onto a subspace V
0
, that corresponds to an approximation of f at a “coarse” level of
resolution. The basis of V
0
is generated by orthogonal translations of φ(ξ) which is called
the scaling function. The resolution of V
0
is successively reﬁned by the second sum, which
consists of projections of f onto the subspaces W
j
, each one being spanned by a wavelet basis
¸
ψ(2
j
ξ −k)
¸
. The function ψ(ξ) is called the mother wavelet, since all other wavelet basis
functions ψ
j,k
are simply produced by dilations (for adjustment of resolution) and translations
of ψ. It is noted that a basic property of the subspaces V
0
, W
0
is
V
0
W
0
= V
1
(2.2)
where V
1
corresponds to a level of resolution twice that of V
0
. This means that scaling and zero
wavelet functions form a basis of orthogonal functions that spans V
1
. The notion of a projection
of f onto approximation spaces at successive levels of resolution is explained in Fig. 2.2.
Recursively,
V
0
W
0
W
1
· · ·
W
k−1
= V
k
(2.3)
1
Hilbert space of square integrable functions.
P1: IML/FFX P2: IML
MOBK05802 MOBK058Sarris.cls December 15, 2006 7:32
ANUMERICAL INTERFACEBETWEENFDTDANDHAARMRTD 11
In other words, adding one wavelet level, is equivalent to improving the resolution of an ap
proximation, like the one in Eq. (2.1), by a factor of two.
2
Hence, any desired resolution of
approximation can be achieved by adding an appropriate number of wavelet levels.
The advantage of this approach lies in the dependence of the magnitude of wavelet
coefﬁcients on the local regularity of f . In particular, wavelet coefﬁcients assume large values
near discontinuities or singularities. Indeed, an explicit relationship between the magnitude
and the “Lipschitz uniformity” of a square integrable function [31] is given by the following
theorem [32]: Let 0 < α < n be a real number that is not an integer. The function f (x) is uniformly
Lipschitz of order α over the interval [a, b] if and only if for any n ∈ Z and j ∈ Z such that
2
−j
n ∈ (a, b),  < f, ψ
j,n
>  = O
2
−(α+1/2) j
.
3
Therefore, once a numerical analysis of a problem is performed by expanding the un
known function in a wavelet basis, highorder wavelet coefﬁcients, which are located away from
discontinuities or singularities, typically assume insigniﬁcant values, that can be thresholded to
zero. Thresholding of a wavelet coefﬁcient implies its elimination from subsequent operations,
or—equivalently—it’s being treated as if it were equal to zero. By this procedure, an adaptive,
temporally moving mesh is implemented.
2.2.2 Wavelet Bases
The simplest (and oldest) orthogonal wavelet system is the Haar basis [33]. Its study is useful
from a theoretical point of view because it offers an intuitive understanding of many multires
olution properties. Furthermore, due to its simplicity, this basis is widely employed in a series
of applications, hence its study is also practically interesting.
The Haar scaling function is deﬁned as a pulse of unit length:
φ(ξ) =
¸
1, 0 ≤ ξ < 1
0, otherwise
(2.4)
and the scaling function basis is produced by orthogonal translations of the latter:
φ
k
(ξ) = φ(ξ −k) (2.5)
The Haar mother wavelet function is deﬁned as:
ψ(ξ) =
⎧
⎪
⎪
⎨
⎪
⎪
⎩
1, 0 ≤ ξ < 1/2
− 1, 1/2 ≤ ξ < 1
0, otherwise.
(2.6)
2
Dyadic multiresolution analyses are considered here. In general, modifying the dilation factor for the generation of
the wavelet basis from the mother wavelet can yield other types of MRA.
3
The notation < f, g >implies inner product. If f (h) and g(h) are two functions of h, it is said that f (h) = O(g(h)),
as h →0, if there is some constant C, such that:  f (h) < Cg(h), for h sufﬁciently small.
P1: IML/FFX P2: IML
MOBK05802 MOBK058Sarris.cls December 15, 2006 7:32
12 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
The wavelet basis is then produced by both translations and dilations, as shown in the deﬁnition:
ψ
j,k
(ξ) = 2
j/2
ψ(2
j
ξ −k). (2.7)
It is noted that the same deﬁnition holds for the derivation of any dyadic wavelet basis, from its
generating mother wavelet. In all these deﬁnitions ξ is a normalized, dimensionless variable. For
example, if φ or ψ functions represent a variation along the xdirection of a computational mesh
with a cell size x, then simply: ξ = x/x. The Haar scaling and mother wavelet functions
are depicted in Fig. 2.3, and the mechanism, in which multiresolution analysis is brought about
by employing this basis (in the sense of Eq. (2.3)), is qualitatively explained in Fig. 2.4.
The Haar basis has an excellent localization in space (or time) domain and poor localiza
tion in the corresponding Fourier domain, as shown in Fig. 2.5. Therefore, the Haar basis does
not share the typical wavelet property of combining good localization in both domains (within
the limits of the uncertainty principle [22]).
A basis that has also been used in applications, is the cubic spline Battle–Lemarie basis
[34]. The Battle–Lemarie scaling and mother wavelet (Fig. 2.6) are entire domain functions
and therefore schemes that are developed in this basis have to be truncated with respect to
space. However, Battle–Lemarie scaling and wavelets have an excellent localization both in
space and Fourier domains (Fig. 2.7), a feature that permits an a priori estimate of the necessary
levels of resolution for correct ﬁeld modeling.
−0.2 0 0.2 0.4 0.6 0.8 1
−0.5
0
0.5
1
scaling function
−0.2 0 0.2 0.4 0.6 0.8 1
−1
−0.5
0
0.5
1
Normalized space variable (e.g. x/Δ x)
mother wavelet
FIGURE2.3: Haar scaling and mother wavelet function
P1: IML/FFX P2: IML
MOBK05802 MOBK058Sarris.cls December 15, 2006 7:32
ANUMERICAL INTERFACEBETWEENFDTDANDHAARMRTD 13
V
0
W
0
W
1
V
1
V
0
W
0
0
V W
0
W
1
V
2
FIGURE2.4: Demonstration of the multiresolution principle as implemented by the Haar basis
−40 −30 −20 −10 0 10 20 30 40
0
0.2
0.4
0.6
0.8
1
Normalized wavenumber variable (e.g. k
x
Δ x)
F{φ}(k)
F{ψ}(k)
FIGURE2.5: Haar scaling and mother wavelet functions in the Fourier domain
P1: IML/FFX P2: IML
MOBK05802 MOBK058Sarris.cls December 15, 2006 7:32
14 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
Normalized space variable (e.g. x/Δ x)
FIGURE2.6: Battle–Lemarie scaling and mother wavelet function
Figures 2.5 and 2.7 also provide an intuitive explanation of the successive approximation
procedure that is connected to multiresolution analysis. Evidently, the scaling function
has the Fourier domain pattern of a low—pass ﬁlter, while the mother wavelet is by the
same token a band—pass ﬁlter. Hence, scaling functions alone provide a description of “low
frequency” characteristics of a signal, while the addition of wavelets enhances the capability of
−15 −5 −5 −15
0
0.4
0.8
1.2
Normalized wavenumber variable (e.g. k
x
Δ x)
F{φ}(k)
F{ψ}(k)
FIGURE2.7: Battle–Lemarie scaling and mother wavelet functions in the Fourier domain
P1: IML/FFX P2: IML
MOBK05802 MOBK058Sarris.cls December 15, 2006 7:32
ANUMERICAL INTERFACEBETWEENFDTDANDHAARMRTD 15
E E E E
k M3 k M2 k M1 k M
P E C
= 0
φ
ψ
(m1) cell mth cell
m
m
m1
φ
ψ
m1
FIGURE2.8: Equivalent grid points (e.g.p.) within a scaling cell as introduced by a ﬁrstorder wavelet
expansion in x, zdirections
a scheme to describe the highfrequency content of a signal. This argument also explains the
adaptivity property of waveletbased numerical schemes: wavelet coefﬁcients correspond to
highfrequency signal variations, in the absence of which, the value of these coefﬁcients decays
to insigniﬁcant levels.
2.3 DERIVATIONOF TIMEDOMAINSCHEMES BY THE
METHODOF MOMENTS
In [35], a systematic way to formulate timedomain numerical schemes for Maxwell’s equations
is provided. This technique, which is also employed in the present monograph for the derivation
of MRTD schemes, is explained here, by considering ﬁrst, the following onedimensional wave
equation example:
∂ E(z, t)
∂z
=
1
c
∂ E(z, t)
∂t
(2.8)
For the numerical solution of this equation, a discrete space–time mesh is introduced and the
values of the electric ﬁeld E(mz, kt) =
k
E
m
are sought by marching in time. The FDTD
scheme for this equation is derived by approximating the partial derivatives of each side of the
P1: IML/FFX P2: IML
MOBK05802 MOBK058Sarris.cls December 15, 2006 7:32
16 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
equation at a space–time mesh point (m z, k t), by centered differencing [36]:
∂ E(mz, kt)
∂z
=
E((m +1)z, kt) − E((m −1)z, kt)
2z
+ O(z
2
)
≈
k
E
m+1
−
k
E
m−1
2z
(2.9)
∂ E(mz, kt)
∂t
=
E(mz, (k +1)t) − E(mz, (k −1)t)
2t
+ O(t
2
)
≈
k+1
E
m
−
k−1
E
m
2t
(2.10)
The previous expressions stem from a reformulation of a Taylor series expansion for E(z, t)
around m z and k t, with respect to space and time. Then, simple algebraic manipulations
lead to the following secondorder accurate approximation of (2.8):
k+1
E
m
=
k−1
E
m
+
c t
z
(
k
E
m+1
−
k
E
m−1
) (2.11)
=
k−1
E
m
+ s (
k
E
m+1
−
k
E
m−1
) (2.12)
where s = c t/z is the CFL (Courant–Friedrichs–Levy) number [36]. For this scheme to
be stable, the CFL number has to be less than one. Alternatively, for the discretization of (2.8)
via the method of moments, the electric ﬁeld E(z, t) is expanded in pulse basis functions (or
equivalently: Haar scaling functions)
¸
h
n
(z)
¸
,
¸
h
n
(t)
¸
in space and time:
E(z, t) =
¸
k, m
k
E
m
h
m
(z) h
k
(t) (2.13)
with
h
n
(x) =
¸
1, nx < x < (n +1)x
0, otherwise
(2.14)
Substituting into (2.8), yields:
¸
k, m
k
E
m
dh
m
(z)
dz
h
k
(t) =
1
c
¸
k, m
k
E
m
h
m
(z)
dh
k
(t)
dt
(2.15)
Then, (2.15) is sampled in space and time using the complex conjugate of the basis
functions themselves as testing functions (Galerkin method). To carry out the testing procedure,
the following integrals are employed:
+∞
−∞
h
n
(x)
dh
n
(x)
dx
dx =
1
2
(δ
n
,n+1
−δ
n
,n−1
) (2.16)
+∞
−∞
h
n
(x) h
n
(x) dx = δ
n
,n
x (2.17)
P1: IML/FFX P2: IML
MOBK05802 MOBK058Sarris.cls December 15, 2006 7:32
ANUMERICAL INTERFACEBETWEENFDTDANDHAARMRTD 17
where δ
n
,n
is Kronecker’s delta:
δ
n
,n
=
¸
1, if n
= n
0, if n
= n
Hence, testing the left handside of (2.15), yields:
+∞
−∞
+∞
−∞
dz dt h
m
(z) h
k
(t)
∂ E(z, t)
∂z
=
¸
k
,m
k
E
m
(δ
m
,m+1
−δ
m
,m−1
) δ
k
,k
t
2
= (
k
E
m+1
−
k
E
m−1
)
t
2
. (2.18)
Similarly, testing the right handside of (2.15):
+∞
−∞
+∞
−∞
dz dt h
m
(z) h
k
(t)
1
c
∂ E(z, t)
∂t
=
¸
k
,m
k
E
m
δ
m
,m
(δ
k
,k+1
−δ
k
,k−1
)
z
2c
= (
k+1
E
m
−
k−1
E
m
)
z
2c
. (2.19)
Thus, it is easily concluded that Eq. (2.12) is again derived and therefore the two methods for
approximating (2.8) in a discrete space, are shown to be equivalent.
In a similar manner, expanding the electric and magnetic ﬁelds in scaling and wavelet
functions and applying the method of moments for the discretization of Maxwell’s equations
(as indicated in the previous section), the MRTD scheme is derived. In this case, the ﬁeld
expansion in (2.13) is written as:
E(z, t) =
¸
k, m
k
E
φ
m
φ
m
(z) h
k
(t) +
R
max
¸
r =0
2
r
−1
¸
p=0
k
E
ψ
r, p
m
ψ
r
m, p
(z) h
k
(t), (2.20)
where R
max
is the maximum waveletorder that has been introduced. Also,
φ
m
(z) = φ
z
z
−m
(2.21)
is a scaling basis produced by a scaling function φ and:
ψ
r
m, p
(z) = 2
r/2
ψ
2
r
z
z
−m
− p
(2.22)
with p = 0, 1, . . . , 2
r
−1. Hence, a slightly modiﬁed deﬁnition with respect to (2.7) is used
for the wavelet basis here, in order to keep a correspondence between the cell index m of the
scaling and the wavelet functions. This deﬁnition is schematically explained for the case of the
Haar basis in Fig. 2.9.
The introduction of wavelets gives rise to a reﬁnement of the scaling functionbased
approximation of E(z, t) by a factor 2
R
max
+1
. To explain the mechanics of this mesh reﬁnement,
P1: IML/FFX P2: IML
MOBK05802 MOBK058Sarris.cls December 15, 2006 7:32
18 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
ψ ψ
ψ
m, 0
1 1
m, 1
m, 0
0
(z)
mΔz z (Δ m+1)
φ
m
FIGURE2.9: Explanation of deﬁnition (2.22) for the Haar basis
the simple example of the Haar basis is employed. In Fig. 2.10, two scaling cells, deﬁned by
pulse functions are shown. Assuming the use of a ﬁrstorder scheme (two wavelet levels), the
following wavelet terms are included within the nth cell: ψ
0
n,0
, ψ
1
n,0
, ψ
1
n,1
. Linear combinations
of their coefﬁcients with the scaling coefﬁcient, produce four equivalent grid points (e.g.p.)
within the cell. The electric ﬁeld values at the equivalent grid points and the positions of the
(x)
x x x
n n+1 n+2
× × ×
x
n  MRTD cell (n + 1)  MRTD cell
: MRTD equivalent grid points
: Scaling cell limits
FIGURE2.10: Mesh reﬁnement in the Haar basis
P1: IML/FFX P2: IML
MOBK05802 MOBK058Sarris.cls December 15, 2006 7:32
ANUMERICAL INTERFACEBETWEENFDTDANDHAARMRTD 19
TABLE2.1: Position of and Field Values at the MRTDEquivalent Grid Points
of Fig. 2.10
POSITION FIELDNODAL VALUE
x
n
+0.125 x E
φ
n
+ E
ψ
0,0
n
+
√
2E
ψ
1,0
n
x
n
+0.375 x E
φ
n
+ E
ψ
0,0
n
−
√
2E
ψ
1,0
n
x
n
+0.625 x E
φ
n
− E
ψ
0,0
n
+
√
2E
ψ
1,1
n
x
n
+0.875 x E
φ
n
− E
ψ
0,0
n
−
√
2E
ψ
1,1
n
latter within the nth cell are provided in Table 2.1. A mathematically elegant and compact way
to express the relationship between ﬁeld nodal values and MRA expansion coefﬁcients is the
following matrix formulation:
⎡
⎢
⎢
⎢
⎢
⎢
⎣
ˆ
E
1
n
ˆ
E
2
n
ˆ
E
3
n
ˆ
E
4
n
⎤
⎥
⎥
⎥
⎥
⎥
⎦
=
⎡
⎢
⎢
⎢
⎢
⎢
⎣
+1 +1 +
√
2 0
+1 +1 −
√
2 0
+1 −1 0 +
√
2
+1 −1 0 −
√
2
⎤
⎥
⎥
⎥
⎥
⎥
⎦
·
⎡
⎢
⎢
⎢
⎢
⎢
⎣
E
φ
n
E
ψ
0,0
n
E
ψ
1,0
n
E
ψ
1,1
n
⎤
⎥
⎥
⎥
⎥
⎥
⎦
(2.23)
The matrix on the right handside of (2.23) is the onedimensional wavelet transform matrix
for a twolevel wavelet scheme. This formulation will be later on employed in the explicit
enforcement of localized boundary conditions and the development of a numerical interface
between the FDTD method and MRTD.
The modiﬁed ﬁnite difference equations for the MRTD approximation of (2.8) are now
dependent on the form that the following integrals assume in a given wavelet basis:
I
φ, φ
n,n
=
+∞
−∞
φ
n
(z)
dφ
n
(z)
dz
dz
I
φ, ψ
r
p
n,n
=
+∞
−∞
φ
n
(z)
dψ
r
n
, p
(z)
dz
dz
I
ψ
r
p
, φ
n,n
=
+∞
−∞
ψ
r
n, p
(z)
dφ
n
(z)
dz
dz
I
ψ
r
p
, ψ
r
p
n,n
=
+∞
−∞
ψ
r
n, p
(z)
dψ
r
n
, p
(z)
dz
dz .
(2.24)
P1: IML/FFX P2: IML
MOBK05802 MOBK058Sarris.cls December 15, 2006 7:32
20 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
Letting l = m
−m, the following set of update equations for scaling and wavelet terms, cast
in a generic form, is deduced:
k+1
E
φ
m
=
k−1
E
φ
m
+
2c t
z
+∞
¸
l =−∞
¸
I
φ, φ
m,m+l
k
E
φ
m+l
+
R
max
¸
r =0
2
r
−1
¸
p=0
I
φ, ψ
r
p
m,m+l
k
E
ψ
r, p
m+l
¸
k+1
E
ψ
r, p
m
=
k−1
E
ψ
r, p
m
+
2c t
z
+∞
¸
l =−∞
¸
I
ψ
r, p
, φ
m,m+l
k
E
φ
m+l
+
R
max
¸
r
=0
2
r
−1
¸
p
=0
I
ψ
r
p
,ψ
r
p
m,m+l
k
E
ψ
r
, p
m+l
¸
(2.25)
The inﬁnite sums with respect to the position index in (2.25), are either reduced to ﬁnite ones,
if ﬁnite domain basis functions are involved, or simply approximated as ﬁnite ones, if entire
domain basis functions are involved. A term widely used in the wavelet literature, as a formal
substitute for the concept of entire or ﬁnite domain functions is that of support. The deﬁnition
of this term is provided here, for completeness: The support of a function f is the closure of the
set {x : f (x) = 0}, or {x : f (x) = 0} [31]. In the case of inﬁnitely supported basis functions
(e.g., the Battle–Lemarie basis employed in [11]), the variation of the index l of the inﬁnite
sums is restricted within a ﬁnite range [−N, N], where N is called the stencil of the method.
The computation of the integrals of (2.24) can be carried out either analytically or com
putationally, depending on the integrands themselves. It is often the case, that wavelet functions
are deﬁned in the Fourier domain, rather than directly. Therefore, Fourier calculus has to be
employed for the extraction of their value.
As a ﬁnal note, the clariﬁcation of the use of the term “order” of an MRTD scheme, as
opposed to its common use in numerical methods, is necessary. The numerical discretization of
(2.8), given in (2.12), is called secondorder accurate, or—simply— a secondorder scheme, because
it implies a numerical approximation of the spatial and temporal derivatives of (2.8), with an
error proportional to z
2
and t
2
, respectively. This fact becomes evident by pure inspection
of expressions (2.10). On the other hand, an MRTD scheme is of nth order, if wavelets up to
order n are introduced. It is noted that wavelets are solely connected to mesh reﬁnement and
do not affect the order of accuracy of the underlying scheme.
2.4 TWODIMENSIONAL HYBRIDARBITRARYORDERHAAR
MRTD/FDTDSCHEME: FORMULATION
2.4.1 MRTD, FDTDEquations
In this section, MRTD is applied to the following system of twodimensional TE
z
Maxwell’s
equations:
∂ E
y
∂t
(ρ, t) =
1
∂ H
x
∂z
(ρ, t) −
∂ H
z
∂x
(ρ, t)
(2.26)
P1: IML/FFX P2: IML
MOBK05802 MOBK058Sarris.cls December 15, 2006 7:32
ANUMERICAL INTERFACEBETWEENFDTDANDHAARMRTD 21
∂ H
x
∂t
(ρ, t) =
1
μ
∂ E
y
∂z
(ρ, t) (2.27)
∂ H
z
∂t
(ρ, t) = −
1
μ
∂ E
y
∂x
(ρ, t) (2.28)
with ρ = x ˆ x + zˆ z. Based on the method outlined in [11], update equations are derived by
the method of moments, assuming a spatial expansion of electromagnetic ﬁeld components
in scaling and wavelet functions of an arbitrary basis and up to arbitrary orders r
x,max
, r
z,max
in x, zdirections respectively. However, our analysis is restricted to dyadic wavelet trans
forms, as they are the most commonly used for the purpose of adaptively solving partial dif
ferential equations. In the subsequent development, the discretization of a twodimensional
domain (in which ﬁeld solutions are sought) in cells of x by z is pursued, by means of
a wavelet basis, deﬁned by the scaling function φ and the socalled mother wavelet ψ [30].
Then, φ
m
(ξ) = φ (ξ/ξ −m) denotes the mth scaling function in ξ = x, zdirections. Ac
cordingly, the wavelet functions of order r that recursively reﬁne the resolution of φ
m
are deﬁned
as: ψ
r
m, p
= 2
r/2
ψ (2
r
(x/x −m) − p), where p = 0, · · · 2
r
−1. A basis of pulse functions
h
n
(t) = h(t/t −n) (deﬁned as in [11]), is employed for ﬁeld expansion in time, where t
denotes the time step, limited by the choice of x and zthrough the stability condition. Given
these deﬁnitions, E
y
, H
x
, H
z
are expressed in the formof the following orthogonal expansions:
E
y
( ¯ ρ, t) =
¸
n
h
n
(t)
¸
i,m
¸
n
E
y, φ φ
i,m
φ
i
(x) φ
m
(z)
+
¸
r
z
, p
z
n
E
y, φ ψ
rz, pz
i,m
φ
i
(x) ψ
r
z
m, p
z
(z)
+
¸
r
x
, p
x
n
E
y, ψ
rx , px
φ
i,m
ψ
r
x
i, p
x
(x) φ
m
(z)
+
¸
r
x
, p
x
¸
r
z
, p
z
n
E
y, ψ
rx , px
ψ
rz, pz
i,m
ψ
r
x
i, p
x
(x) ψ
r
z
m, p
z
(z)
¸
(2.29)
H
x
( ¯ ρ, t) =
¸
n
h
n
(t)
¸
i,m
¸
n
H
x, φ φ
i,m
φ
i
(x) φ
m
(z)
+
¸
r
z
, p
z
n
H
x, φ ψ
rz, pz
i,m
φ
i
(x) ψ
r
z
m
, p
z
(z)
+
¸
r
x
, p
x
n
H
x, ψ
rx , px
φ
i,m
ψ
r
x
i, p
x
(x) φ
m
(z)
+
¸
r
x
, p
x
¸
r
z
, p
z
n
H
x, ψ
rx , px
ψ
rz, pz
i,m
ψ
r
x
i, p
x
(x) ψ
r
z
m
, p
z
(z)
¸
(2.30)
P1: IML/FFX P2: IML
MOBK05802 MOBK058Sarris.cls December 15, 2006 7:32
22 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
H
z
( ¯ ρ, t) =
¸
n
h
n
(t)
¸
i,m
¸
n
H
z, φ φ
i
,m
φ
i
(x) φ
m
(z)
+
¸
r
z
, p
z
n
H
z, φ ψ
rz, pz
i
,m
φ
i
(x) ψ
r
z
m, p
z
(z)
+
¸
r
x
, p
x
n
H
z, ψ
rx , px
φ
i
,m
ψ
r
x
i
, p
x
(x) φ
m
(z)
+
¸
r
x
, p
x
¸
r
z
, p
z
n
H
z, ψ
rx , px
ψ
rz, pz
i
,m
ψ
r
x
i
, p
x
(x) ψ
r
z
m, p
z
(z)
¸
(2.31)
where n
= n +1/2, i
= i +s
x
, m
= m +s
z
. Thus, while half a time step offset between
the update of electric and magnetic ﬁeld terms is kept (as in FDTD), the offset of electric and
magnetic scaling cells in x and zdirections is left as a parameter under investigation. Most
MRTD studies, with the notable exception of [37], adopt the choice of s
x
= s
z
= 1/2, based
on the FDTD practice. In this section, a systematic way of determining these offsets is set
forth, by introducing the notion of equivalent MRTD grid points.
Assume that a certain scaling function basis generates electric ﬁeld grid points in one
dimension (ξ = x, z) : i · ξ, i = 1, 2, . . . , N
ξ
. Then, the introduction of wavelets of orders
r = 0, 1, . . . , r
ξ,max
, reﬁnes the mesh in ξdirection by a factor of ρ
ξ
= 2
r
ξ,max
+1
, since each
wavelet level successively doubles the resolution of the underlying approximation. In mathe
matical terms, the E
y
expansion of (2.30) can be cast in the next equivalent form [30]:
E
y
( ¯ ρ, t) =
¸
n
h
n
(t)
¸
i,m
¸
n
E
y, φ
Rx
φ
Rz
i,m
φ
R
x
i
(x) φ
R
z
m
(z)
¸
, (2.32)
where R
x
= r
x,max
+1, R
z
= r
z,max
+1 and
¸
φ
R
n
(ξ)
¸
=
¸
2
R/2
φ(2
R
(ξ/ξ) −n)
¸
is the scaling
basis that produces by itself an approximation of the same resolution as (2.30). In fact, the
latter corresponds to the hierarchical multiresolution decomposition of the former and the
coefﬁcients in (2.30) can directly be deduced from the ones in (2.32) via the wavelet transform.
This waveletinduced mesh reﬁnement can also be perceived as a procedure of generating
equivalent grid points that give rise to a mesh of cell sizes x/ρ
x
, z/ρ
z
. Similar observations
hold for (2.31), (2.31).
This argument is demonstrated for a zeroorder Haar MRTD scheme, utilizing Haar
scaling and zeroorder wavelet functions (Fig. 2.3), in Fig. 2.11 and 2.12, which depict the
equivalent grid points that are generated in both cases, in one dimension. It is noted that in
general, linear combinations of scaling and wavelet functions yield electric ﬁeld values at points:
i +( p +0.5)/2
r
x,max
+1
x, with p = 0, 1 · · · 2
r
x,max
+1
−1. In our twodimensional example,
these ﬁgures represent zcuts of the mesh, including E
y
and H
z
grid points (which are necessary
P1: IML/FFX P2: IML
MOBK05802 MOBK058Sarris.cls December 15, 2006 7:32
ANUMERICAL INTERFACEBETWEENFDTDANDHAARMRTD 23
(x)
(x)
φ
ψ
H
2i1
E
2i1
E
H H H
E E
2i 2i+1 2i+2
2i 2i+1 2i+2
i − cell (i − 1) − cell (i + 1) − cell
: Electric field equivalent grid point
: Magnetic field equivalent grid point
FIGURE 2.11: Electric/magnetic ﬁeld equivalent grid points for zeroorder Haar MRTD in one di
mension, under the common convention of half a cell offset between electric/magnetic scaling cells
for the approximation of xpartial derivatives involved in E
y
updates). Up to a normalization
multiplicative constant, the ﬁeldvalues at equivalent gridpoints withineachcell, are computedas
the sumand the difference of scaling and zeroorder wavelet terms, respectively. Accordingly, H
z
equivalent grid points are located at
i +s
x
+( p +0.5)/2
r
x,max
+1
x. However, the purpose
(x)
(x)
E
2i1
E E E
2i 2i+1 2i+2
i−cell (i − 1)−cell (i + 1)−cell
: Electric field equivalent grid point
: Magnetic field equivalent grid point
H
2i1
H
2i
H
2i+1
H
2i+2
FIGURE 2.12: Electric/magnetic ﬁeld equivalent grid points for zeroorder Haar MRTD in one di
mension, with electric/magnetic scaling cell offset chosen in consistence with (2.33)
P1: IML/FFX P2: IML
MOBK05802 MOBK058Sarris.cls December 15, 2006 7:32
24 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
of using wavelets is to implement in this “new” mesh of equivalent grid points, the method
produced by scaling functions only, at a resolution that is ﬁner (in xdirection) by the wavelet
reﬁnement factor ρ
x
. Hence, if this method deﬁnes half a scaling cell offset between electric and
magnetic nodes, the wavelet augmented method has to deﬁne half an e qui val e nt cell offset
between the equivalent electric/magnetic grid points in this direction. The choice of s
x
and s
z
that is consistent with this requirement is given below:
s
x
= 0.5ρ
x
= 2
−r
x,max
−2
s
z
= 0.5ρ
z
= 2
−r
z,max
−2
(2.33)
Figure 2.11 shows the node arrangement in zeroorder Haar MRTD, under the common
convention of half a cell offset between electric/magnetic scaling cells. Apparently, equivalent
electric and magnetic equivalent nodes are now collocated. On the other hand, if the separation
of electric and magnetic scaling cells is chosen in consistence with (2.33), which in this case
yields an offset of one quarter of a cell (Fig. 2.12), the equivalent grid points are leapfrogged in
space and correctly correspond to the mesh of an FDTDscheme of cell size x/2. As discussed
in [38], this latter approach leads to MRTD schemes with consistent numerical dispersion
properties (as opposed to the former approach) and therefore, it is adopted in the following
derivations.
Upon substitution of the MRTD expansions of all ﬁeld components into Eq. (2.26)–
(2.28), Galerkin’s method is applied for the derivation of ﬁeld update equations. In the following,
the update equations for E
y
, h
x
= H
x
x and h
z
= H
z
z coefﬁcients are provided. The scaling
cell dimensions are x ×z. Also, n
= n +1/2, i
= i +s
x
, m
= m +s
z
, ζ = ψ
r
x
, p
x
and
η = ψ
r
z
, p
z
. Finally, the coefﬁcients D
0
, D
1
are given in Section 2.8.
E
y
update equations
n+1
E
y,φφ
i,m
=
n
E
y,φφ
i,m
+
t
xz
¸
n
h
x,φφ
i,m
−
n
h
x,φφ
i,m
−1
+
¸
r
2
r/2
n
h
x,φψ
r,2
r
−1
i,m
−1
−
n
h
x,φψ
r,2
r
−1
i,m
−
¸
n
h
z,φφ
i
,m
−
n
h
z,φφ
i
−1,m
+
¸
r
2
r/2
n
h
z,ψ
r, p
φ
i
−1,m
−
n
h
z,ψ
r, p
φ
i
,m
¸
(2.34)
n+1
E
y,φψ
r, p
i,m
=
n
E
y,φψ
r, p
i,m
+
t
xz
¸
2
r/2
δ
p,0
n
h
x,φφ
i,m
−
n
h
x,φφ
i,m
−1
+
¸
r
, p
¸
D
2
(r, p, r
, p
)
n
h
x,φψ
r
, p
i,m
+D
3
(r, p, r
, p
)
n
h
x,φψ
r
, p
i,m
−1
¸
−
¸
n
h
z,φψ
r, p
i
,m
−
n
h
z,φψ
r, p
i
−1,m
+
¸
q
2
q/2
n
h
z,ψ
q,2
q
−1
ψ
r, p
i
,m−1
−
n
h
z,ψ
q,2
q
−1
ψ
r, p
i
−1,m
¸
(2.35)
P1: IML/FFX P2: IML
MOBK05802 MOBK058Sarris.cls December 15, 2006 7:32
ANUMERICAL INTERFACEBETWEENFDTDANDHAARMRTD 25
n+1
E
y,ψ
r, p
φ
i,m
=
n
E
y,ψ
r, p
φ
i,m
+
t
xz
¸
n
h
x,ψ
r, p
φ
i,m
−
n
h
x,ψ
r, p
φ
i,m
−1
+
¸
q
2
q/2
n
h
x,ψ
r, p
ψ
q,2
q
−1
i,m
−1
−
n
h
x,ψ
r, p
ψ
q,2
q
−1
i,m
−
¸
2
r/2
δ
p,0
n
h
z,φφ
i
,m
−
n
h
z,φφ
i
−1,m
+
¸
r
, p
¸
D
2
(r, p, r
, p
)
n
h
z,ψ
r
, p
φ
i
,m
+D
3
(r, p, r
, p
)
n
h
z,ψ
r
, p
φ
i
−1,m
¸
(2.36)
n+1
E
y,ψ
r, p
ψ
q,w
i,m
=
n
E
y,ψ
r, p
ψ
q,w
i,m
+
t
xz
¸
2
q/2
δ
w,0
n
h
x,ψ
r, p
φ
i,m
n
h
x,ψ
r, p
φ
i,m
−1
(2.37)
+
¸
q
,w
¸
D
2
(q, w, q
, w
)
n
h
x,ψ
r, p
ψ
q
,w
i,m
+ D
3
(q, w, q
, w
)
n
h
x,ψ
r, p
ψ
q
,w
i,m
−1
¸
−
¸
2
r/2
δ
p,0
n
h
z,φψ
q,w
i
,m
−
n
h
z,φψ
q,w
i
−1,m
+
¸
r
, p
¸
D
2
(r, p, r
, p
)
n
h
z,ψ
r
, p
ψ
q,w
i
,m
+D
3
(r, p, r
, p
)
n
h
z,ψ
r
, p
ψ
q,w
i
,m−1
¸
¸
H
x
update equations
n
h
x,φφ
i,m
=
n
−1
h
x,φφ
i,m
+
tx
μz
¸
n
E
y,φφ
i,m+1
−
n
E
y,φφ
i,m
+
¸
0≤r ≤r
z,max
2
r/2
n
E
y,φψ
r,0
i,m+1
−
n
E
y,φψ
r,0
i,m
¸
(2.38)
n
h
x,ζ φ
i,m
=
n
−1
h
x,ζ φ
i,m
+
tx
μz
¸
n
E
y,ζ φ
i,m+1
−
n
E
y,ζ φ
i,m
+
¸
0≤r ≤r
z,max
2
r/2
n
E
y,ζ ψ
r,0
i,m+1
−
n
E
y,ζ ψ
r,0
i,m
¸
(2.39)
n
h
x,φη
i,m
=
n
−1
h
x,φη
i,m
+
tx
μz
¸
2
r
z
/2
δ
p
z
,2
rz
−1
n
E
y,φφ
i,m
−
n
E
y,φφ
i,m+1
+
¸
r
z
, p
z
D
0
(r
z
, p
z
,
r
z
, p
z
n
E
y,φψ
r
z
, p
z
i,m+1
+
¸
r
z
, p
z
D
1
(r
z
, p
z
, r
z
, p
z
)
n
E
y,φψ
r
z
, p
z
i,m
¸
(2.40)
n
h
x,ζ η
i,m
=
n
−1
h
x,ζ η
i,m
+
tx
μz
¸
2
r
z
/2
δ
p
z
,2
rz
−1
n
E
y,ζ φ
i,m
−
n
E
y,ζ φ
i,m+1
+
¸
r
z
, p
z
D
0
(r
z
, p
z
,
r
z
, p
z
n
E
y,ζ ψ
r
z
, p
z
i,m+1
+
¸
r
z
, p
z
D
1
(r
z
, p
z
, r
z
, p
z
)
n
E
y,ζ ψ
r
z
, p
z
i,m
¸
(2.41)
P1: IML/FFX P2: IML
MOBK05802 MOBK058Sarris.cls December 15, 2006 7:32
26 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
H
z
update equations
n
h
z,φφ
i
,m
=
n
−1
h
z,φφ
i
,m
−
tz
μx
¸
n
E
y,φφ
i +1,m
−
n
E
y,φφ
i,m
+
¸
0≤r ≤r
x,max
2
r/2
n
E
y,ψ
r,0
φ
i +1,m
−
n
E
y,ψ
r,0
φ
i,m
¸
(2.42)
n
h
z,φη
i
,m
=
n
−1
h
z,φη
i
,m
−
tz
μx
¸
n
E
y,φη
i +1,m
−
n
E
y,φη
i,m
+
¸
0≤r ≤r
x,max
2
r/2
n
E
y,ψ
r,0
η
i +1,m
−
n
E
y,ψ
r,0
η
i,m
¸
(2.43)
n
h
z,ζ φ
i
,m
=
n
−1
h
z,ζ φ
i
,m
−
tz
μx
¸
2
r
x
/2
δ
p
x
,2
rx
−1
n
E
y,φφ
i,m
−
n
E
y,φφ
i +1,m
+
¸
r
x
, p
x
D
0
(r
x
, p
x
,
r
x
, p
x
n
E
y,ψ
r
x
, p
x
φ
i +1,m
+
¸
r
x
, p
x
D
1
(r
x
, p
x
, r
x
, p
x
)
n
E
y,ψ
r
x
, p
x
φ
i,m
¸
(2.44)
n
h
z,ζ η
i
,m
=
n
−1
h
z,ζ η
i
,m
−
tz
μx
¸
2
r
x
/2
δ
p
x
,2
rx
−1
n
E
y,φη
i,m
−
n
E
y,φη
i +1,m
+
¸
r
x
, p
x
D
0
(r
x
, p
x
,
r
x
, p
x
n
E
y,ψ
r
x
, p
x
η
i +1,m
+
¸
r
x
, p
x
D
1
(r
x
, p
x
, r
x
, p
x
)
n
E
y,ψ
r
x
, p
x
η
i,m
¸
(2.45)
A numerical dispersion analysis of this system of equations yields the following expression [38]:
¸
1
u
p
t
sin
ωt
2
¸
2
=
¸
1
x
eff
sin
k
x
x
eff
2
¸
2
+
¸
1
z
eff
sin
k
z
z
eff
2
¸
2
,
(2.46)
For completeness, the wellknown FDTD equations for TE
z
waves are provided below.
An FDTD cell size of δx ×δz and time step δt is assumed. Then:
n+1
E
y
i,m
=
n
E
y
i,m
+
δt
δxδz
¸
n+
1
2
h
x
i,m+
1
2
−
n+
1
2
h
x
i,m−
1
2
−
n+
1
2
h
z
i +
1
2
,m
+
n+
1
2
h
z
i −
1
2
,m
¸
n+
1
2
h
x
i,m+
1
2
=
n−
1
2
h
x
i,m+
1
2
+
δtδx
μδz
¸
n
E
y
i,m+1
−
n
E
y
i,m
¸
n+
1
2
h
z
i +
1
2
,m
=
n−
1
2
h
z
i +
1
2
,m
−
δtδz
μδx
¸
n
E
y
i +1,m
−
n
E
y
i,m
¸
(2.47)
P1: IML/FFX P2: IML
MOBK05802 MOBK058Sarris.cls December 15, 2006 7:32
ANUMERICAL INTERFACEBETWEENFDTDANDHAARMRTD 27
A dispersion analysis for the system of FDTD equations (2.47), leads to the following expres
sion [39]:
¸
1
u
p
δt
sin
ωδt
2
¸
2
=
¸
1
δx
sin
k
x
δx
2
¸
2
+
¸
1
δz
sin
k
z
δz
2
¸
2
. (2.48)
By inspection of Eq. (2.46) and (2.48), it is readily concluded that under the condition:
x
eff
≡ δx, z
eff
≡ δz, (2.49)
the FDTD and Haar MRTD schemes present the same dispersion properties. Hence, a
reﬂectionless interface between the two methods can be established. Time marching is carried
out concurrently in the FDTD and MRTD regions, since the same time step is chosen for both
methods. The common dispersion equation of the two schemes also implies common stability
properties. As a result, a stable time step for the MRTD region is also stable for FDTD.
2.4.2 Connection Algorithm
A simple, onedimensional example of an interface between a ﬁrstorder Haar MRTD and
the corresponding FDTD scheme is shown in Fig. 2.13. For the update of the electric ﬁeld
MRTDcoefﬁcients E
φ
5
, E
ψ
5
, E
ψ
10
5
, E
ψ
11
5
, magnetic scaling and wavelet terms one cell to the left,
within the FDTD region, are necessary. For their calculation, the FDTD nodal values of the
magnetic ﬁeld H
1
, H
2
, H
3
, H
4
are used as an input to a recursive fast wavelet transform (FWT).
Thus, the wavelet decomposition of the magnetic ﬁeld at that cell is deduced and employed
for the electric ﬁeld updates of MRTD, at the FDTD/MRTD boundary. In the case of an
FDTD to MRTD transition, an inverse fast wavelet transform (IFWT) is applied, to derive
FDTD coefﬁcients (nodal ﬁeld values) from MRTD terms. It is noted that both FWT and
IFWT are characterized by an optimal complexity of O(N). In addition, their computational
implementation is relatively simple.
As a computational validation of these interface principles, the propagation of a 0–5 GHz
Gaussian pulse through an FDTD/fourthorder MRTD interface is shown in Fig. 2.14. The
cell size for the FDTD region (and effective cell size for MRTD) is 2.4 mm, while the MRTD
scaling cell size is 76.8 mm. The time step is 0.9 of the Courant limit. The interface for the
electric ﬁeld nodes is located at cell 1440. Smooth and stable transition from the MRTD to
FDTD region is observed, as expected.
In the following, a twodimensional FDTD/MRTD connection algorithm is presented,
for the case where an FDTD region encloses an MRTD one. An application of interest is the
termination of an MRTD mesh in an FDTD perfectly matched layer absorber, that allows for
MRTD domain truncation via existing absorber codes. Similar concepts can be employed for
interfaces of other types.
P1: IML/FFX P2: IML
MOBK05802 MOBK058Sarris.cls December 15, 2006 7:32
28 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
x x x x
MRTD FDTD
Efield
Hfield
x : FDTD E  grid points
: MRTD E  equivalent
grid points
grid points
: MRTD H  equivalent
: FDTD H  grid points
H H H H
E
E E
φ
ψ ψ
10 11
ψ
E
5
5 5
5
4 3 2 1
FIGURE2.13: Onedimensional interface between FDTD and a ﬁrstorder Haar MRTD scheme
To begin with, data transfer from MRTD to FDTD is addressed. For the update of all
FDTD grid points whose stencil extends into the MRTD region, it sufﬁces to retrieve the
nodal ﬁeld values of the electric ﬁeld one FDTD cell within the MRTD region (Fig. 2.15).
Thus, the complete determination of the tangential electromagnetic ﬁeld components along the
boundary of the FDTD domain is allowed for. Then, the independent solution of this region
becomes possible, since the MRTD calculated tangential ﬁelds are used as boundary conditions
for FDTD. In consequence, the problem is reduced to calculating nodal ﬁeld values along the
boundary of the MRTD region, which is exactly the function of IFWT.
Conversely, for the update of the MRTD boundary magnetic ﬁeld coefﬁcients, FDTD
electric ﬁeld nodes extending over one scaling cell within the FDTD region are wavelet trans
formed, via a twodimensional FWT routine. If this scaling cell extends beyond the domain,
zero ﬁeld values are used. This is possible and physically correct for both closed and open
domain problems, since a perfect conductorbacked absorber is typically used for the simula
tion of an open boundary. Figure 2.16 schematically explains this procedure, for a case where
r
x,max
= r
z,max
= 1. It is noted that no FDTD grid points are sought for at the (shaded) corner
P1: IML/FFX P2: IML
MOBK05802 MOBK058Sarris.cls December 15, 2006 7:32
ANUMERICAL INTERFACEBETWEENFDTDANDHAARMRTD 29
1250 1300 1350 1400 1450 1500 1550 1600
0
0.5
1
t
=
5
5
0
Δ
t
1250 1300 1350 1400 1450 1500 1550 1600
0
0.5
1
t
=
6
5
0
Δ
t
1250 1300 1350 1400 1450 1500 1550 1600
0
0.5
1
t
=
7
7
0
Δ
t
MRTD cells FDTD cells
FIGURE 2.14: Reﬂectionless propagation through a onedimensional FDTD/fourthorder MRTD
interface
FDTD/MRTD
boundary
: H FDTD field node
grid point
: E MRTD equivalent
: E FDTD field node
: H FDTD field node
MRTD region
FDTD region
(z)
(x)
2D Yee’s cell
y
x
z
y
FIGURE2.15: FDTD update from MRTD data
P1: IML/FFX P2: IML
MOBK05802 MOBK058Sarris.cls December 15, 2006 7:32
30 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
y
y
FDTD/MRTD
boundary
MRTD scaling
cell
: E
y
FDTD field node
(necessary for the
MRTD update)
: E
y
MRTD equivalent
grid point
(z)
(x)
MRTD region
FDTD region
Δ x
Δ z
FIGURE2.16: MRTD update from FDTD data
region, shown in Fig. 2.16. This is due to the fact that MRTD, just as FDTD, uses a cross
shaped stencil for the update of all grid points. Computing the electric ﬁeld MRTDcoefﬁcients
from the FDTD data via an FWT and updating the tangential magnetic ﬁeld component
coefﬁcients via the standard MRTDﬁnite difference equations, determines again the tangential
electromagnetic ﬁeld components along the boundary of the MRTD region. This, in turn, is
sufﬁcient for its independent, MRTDbased solution.
Evidently, all operations that implement this connection algorithm are performed at
the same time step, during the update of the electric ﬁeld coefﬁcients in both regions, in an
absolutely stable fashion, due to the matching of the dispersion properties of the two schemes.
Furthermore, the same principles lead to interfaces between wavelet schemes of an arbitrary
basis and the ones that are formulated by the corresponding scaling functions only, provided
that the effective resolutions in the two regions are kept the same. It is also noted that the
extension of the interface algorithm to three dimensions is accomplished by treating each face
of Yee’s cell according to the method that has been set forth in this work.
2.5 NUMERICAL RESULTS: VALIDATION
Several twodimensional airﬁlled square resonators have been analyzed in various
FDTD/MRTD mesh conﬁgurations for validation purposes. The choice of those conﬁgu
rations was made in order to demonstrate that the stability and accuracy of the algorithm was
preserved under any gridding conditions, these being either related to the order of the MRTD
P1: IML/FFX P2: IML
MOBK05802 MOBK058Sarris.cls December 15, 2006 7:32
ANUMERICAL INTERFACEBETWEENFDTDANDHAARMRTD 31
40
36
2
7
: FDTD/MRTD
interface
(z)
(x)
FIGURE 2.17: Empty rectangular cavity geometry and interface of a 3 by 3order MRTD/FDTD
mesh conﬁguration (dimensions are given in FDTD cells, MRTD mesh is 2 × 2)
scheme or the proximity of the MRTD region to the hard boundary of the domain. The di
mensions of the cavity structures in all ﬁgures are given in FDTD cells (or MRTD equivalent
grid points). Figures 2.17 and 2.18 depict two case studies for MRTD/FDTD mesh conﬁgu
rations. In the ﬁrst case (Fig. 2.17), an MRTD scheme of order 3 in both x and zdirections
forms a mesh of 2 by 2 scaling cells corresponding to 32 × 32 FDTD cells, asymmetrically
placed within an FDTD region, that is terminated into the metal boundaries of the cavity.
The whole domain corresponds to 36 ×40 FDTD cells of dimension 1 cm × 1 cm. A pure
MRTD scheme would model these electric walls by means of image theory, necessitating the
introduction of several images of highorder wavelet terms [11]. On the other hand, enclosing
the MRTD region in an FDTD one, facilitates the treatment of these hard boundaries, whose
FDTD modeling amounts to setting the tangential to PEC electric ﬁeld nodal values equal to
zero. Similarly, in the second case, an MRTD scheme of orders 3 and 4 in x and zdirections
respectively forms a 1 by 1 scaling mesh and is interfaced with FDTD of cell size 1 cm×1 cm.
32
5
16
2
2
2
interface
: FDTD/MRTD
(z)
(x)
FIGURE 2.18: Empty rectangular cavity geometry and interface of a 3 by 4order MRTD/FDTD
mesh conﬁguration (dimensions are given in FDTD cells, MRTD mesh is 1 × 1)
P1: IML/FFX P2: IML
MOBK05802 MOBK058Sarris.cls December 15, 2006 7:32
32 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
1.8 1.85 1.9 1.95 2
x 10
4
−2
0
2
Time step
E
l
e
c
t
r
i
c
f
i
e
l
d
E
y
Interface
FDTD
0 0.5 1 1.5 2 2.5 3 3.5
10
−2
10
−1
10
0
Frequency [GHz]
E
y
[
S
p
e
c
t
r
a
l
d
e
n
s
i
t
y
]
Interface
FDTD
FIGURE2.19: Time and frequency domain patterns of electric ﬁeld (E
y
) sampled within the cavity of
Fig. 2.17, as obtained by FDTD and the hybrid scheme
The total domain corresponds to 20 × 39 FDTD cells. Cavity resonances that are derived via
a pure FDTD scheme and the proposed interfacebased method agree well (in both time and
frequency domain), as shown in Figs. 2.19 and 2.20.
In all cases the time step was set equal to 0.9 of the Courant stability limit.
A third validation case is shown in Fig. 2.21. In that case, the MRTD region is symmet
rically placed within the FDTD mesh and only two FDTD cells away from the boundary. The
MRTD scheme is of order 4 in both directions, hence giving rise to a single scaling cell mesh
in the MRTD region. The effective cell size is again 1 cm × 1 cm and the time step equal to
0.9 of the Courant limit. The patterns of the TE
21
 and TE
22
modes, derived via the interface
algorithm, are shown in Figs. 2.22 and 2.23. The smooth patterns conﬁrm the correctness of
the scheme and the absence of any type of spurious effects that would corrupt its performance.
This observation is also in agreement with the Haar MRTD dispersion analysis of [38].
2.6 NUMERICAL RESULTS: APPLICATIONS
2.6.1 Metal FinLoaded Cavity
The method of this chapter is applied for the simulation of a metal ﬁnloaded cavity, similar to
the one presented in [27]. This structure is chosen for the reason that the presence of the metal
ﬁn within the domain, restricts the order of the MRTD scheme that can be employed for its
analysis. In particular, whenever a scaling cell greater than the ﬁn dimensions is chosen, utmost
care is necessary for the compensation of the unphysical coupling of the regions belowand above
the ﬁn, caused by the scaling function deﬁning the ﬁn cell (or wavelets extending beyond the ﬁn
P1: IML/FFX P2: IML
MOBK05802 MOBK058Sarris.cls December 15, 2006 7:32
ANUMERICAL INTERFACEBETWEENFDTDANDHAARMRTD 33
1.8 1.85 1.9 1.95 2
x 10
4
−2
0
2
Time step
E
l
e
c
t
r
i
c
f
i
e
l
d
E
y Interface
FDTD
0.5 1 1.5 2 2.5
10
−2
10
−1
10
0
Frequency [GHz]
E
y
[
S
p
e
c
t
r
a
l
d
e
n
s
i
t
y
]
Interface
FDTD
FIGURE2.20: Time and frequency domain patterns of electric ﬁeld (E
y
) sampled within the cavity of
Fig. 2.18, as obtained by FDTD and the hybrid scheme
limits). However, the strategies that are followed in this case (e.g., domain split [40]), result in
a local increase of operations and consumption of computational resources. Furthermore, they
bring about practical implementation problems, especially when one is interested in developing
generic waveletbased CAD tools.
Haar MRTD
FDTD
36
32
(z)
(x)
FIGURE 2.21: Empty square cavity geometry and interface of a 4 by 4order MRTD/FDTD mesh
conﬁguration (dimensions are given in FDTD cells, MRTD mesh is 1 × 1)
P1: IML/FFX P2: IML
MOBK05802 MOBK058Sarris.cls December 15, 2006 7:32
34 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
0
10
20
30
40
0
20
40
−1
−0.5
0
0.5
1
zaxis
xaxis
E
l
e
c
t
r
i
c
f
i
e
l
d
E
y
FIGURE2.22: Electric ﬁeld pattern for the TE
22
mode of the empty cavity of Fig. 2.21 (4 by 4 MRTD)
For the interfacebased solution of the problem, the gridding conditions are given in
Fig. 2.24. In the MRTD region, a second by secondorder scheme is employed (4 by 4 scaling
cells). The time step is set at 0.8 of the Courant limit and FDTD cells of 1 cm × 1 cm are
used. As an excitation, a Gaussian pulse, with its 3 dB frequency chosen to be equal to f
c
/2,
f
c
being the cutoff frequency of the TE
11
mode of the cavity, is applied in the FDTD region
(at the plane z = 5 cm). Under these conditions, an absolutely stable performance of the code
0
10
20
30
40
0
10
20
30
40
−1
−0.5
0
0.5
1
zaxis
xaxis
E
l
e
c
t
r
i
c
f
i
e
l
d
E
y
FIGURE2.23: Electric ﬁeld pattern for the TE
22
mode of the empty cavity of Fig. 2.21 (4 by 4 MRTD)
P1: IML/FFX P2: IML
MOBK05802 MOBK058Sarris.cls December 15, 2006 7:32
ANUMERICAL INTERFACEBETWEENFDTDANDHAARMRTD 35
40
36
2
7
(z)
(x)
interface
: FDTD/MRTD
FIGURE2.24: Metal ﬁnloaded cavity geometry and interface of a 2 by 2order MRTD/FDTD mesh
conﬁguration (dimensions are given in FDTD cells, MRTD mesh is 4 × 4)
was obtained. The deduced electric ﬁeld spatial distribution is shown in Fig. 2.25. Moreover,
in order to demonstrate the stability of the solution, the electric ﬁeld, sampled at the point
(x = 1.55 cm, z = 2.35 cm), is plotted as a function of time for an arbitrary interval of 18,000–
20,000 time steps, in Fig. 2.26. It is noted that no latetime instabilities were observed over as
many as 100,000 time steps.
0
10
20
30
40
0
10
20
30
40
0
0.25
0.5
0.75
1
zaxis
xaxis
E
l
e
c
t
r
i
c
f
i
e
l
d
E
y
FIGURE2.25: Electric ﬁeld distribution in the metal ﬁnloaded cavity (dominant TEmode)
P1: IML/FFX P2: IML
MOBK05802 MOBK058Sarris.cls December 15, 2006 7:32
36 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
1.8 1.85 1.9 1.95 2
x 10
4
−1.5
−1
−0.5
0
0.5
1
1.5
Time step
E
l
e
c
t
r
i
c
f
i
e
l
d
E
y
FIGURE 2.26: Electric ﬁeld sampled within the ﬁnloaded cavity as a function of time for time steps
18,000–20,000
2.6.2 Mesh Truncation
2.6.2.1 PML Absorber
A typical problem related to higherorder and multiresolution formulations of ﬁnite difference
schemes is the inherently complex modeling of boundary and material conditions. For high
order Haar wavelet schemes, a formulation for a matched layer absorber that was proposed
in [41] highlighted the difﬁculty of such efforts, that essentially stems from the sampling of
varying electric and magnetic conductivities by a multilevel basis. On the other hand, in [37],
the approach of omitting wavelet operations in the absorber region was adopted. However, the
obvious compromise in accuracy that is related to this approach has negligible effect only in
radiation problems, under the condition that the absorber has been placed far enough from the
source. In this case, ﬁeld wavelet coefﬁcients at the absorber are small enough and the numerical
error produced by their omission is controllable. On the contrary, this method is not applicable
to open guiding structures because of the signiﬁcant ﬁeld values that typically impinge upon
the absorber region of their computational domain.
Furthermore, it may seem necessary that the absorber region of an MRTD mesh be
discretized by at least the degrees of freedom of a single MRTD cell. However, for highorder
schemes, this would lead to absorber regions that are much longer than the ones typically used in
FDTDapplications. For example, terminating an MRTDwith ﬁve wavelet levels (fourthorder
scheme) in just a single cell of it, is equivalent to using an FDTD absorber of 64 grid points per
direction. Nevertheless, excellent performance of 8–16 cell, optimized FDTDuniaxial perfectly
P1: IML/FFX P2: IML
MOBK05802 MOBK058Sarris.cls December 15, 2006 7:32
ANUMERICAL INTERFACEBETWEENFDTDANDHAARMRTD 37
matched layer (UPML) absorbers has been recently demonstrated [42], prompting the quest
for MRTD absorbers that may extend over a fraction of an MRTD scaling cell.
In this work, the concept of the FDTD/MRTD interface is employed for the implemen
tation of a UPML termination of an MRTD domain. In particular, an FDTDUPML region
encloses an MRTD one and the interface algorithm is applied for the reﬂectionless connection
of the two. In case the FDTDregion corresponds to a fraction of a scaling cell, the application of
MRTD update equations needs FDTD grid points beyond the conductor that backs the PML.
The latter are simply zeroed out and fed back as such to the FWT routine. This concept is
explained for a onedimensional case of a secondorder MRTD scheme (with eight equivalent
grid points per cell), terminated into a fourcell FDTD PML (half a scaling cell), in Fig. 2.27.
Using this method, the waveguide structure of Fig. 2.28, also presented in [17], is solved
by a fourthorder MRTD scheme, truncated with a 6 and 8 grid point PML corresponding
to 0.1875 and 0.25 of a scaling cell, which in this case is 8 mm. A 0–30 GHz Gaussian pulse
excitation is used, and the reﬂection coefﬁcient from the slab is calculated. Hence, the scaling
cell is 1.28 ×λ
min
, while the ﬁve wavelet levels successively reﬁne the resolution of the scheme
to λ
min
/25. In Fig. 2.29, waveforms of incident and reﬂected electric ﬁelds, analyzed in their
scaling and wavelet constituents are shown. Their correctness is readily veriﬁed, by noticing that
they assume the form of spatial ﬁeld derivatives, as expected. Additionally, Fig. 2.30 depicts
Zero field grid points
P
E
C
2ndorder MRTD equivalent grid points
FDTD grid points
MRTD cell
MRTD region FDTDPML
region
FIGURE2.27: Concept of MRTD mesh termination via an FDTD/MRTD interface
P1: IML/FFX P2: IML
MOBK05802 MOBK058Sarris.cls December 15, 2006 7:32
38 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
air
Excitation plane
8 mm
air
(z)
(x)
interface
MRTD/FDTD PML
PML
C
P
E r
= 2.56 ε
FIGURE2.28: Slab waveguide geometry
comparative plots of the numerical results derived by the two termination types, along with the
theoretical S
11
form derived by transmission line theory. Evidently, all three sets of results are
in good correlation with each other.
Then, a twodimensional, 8 cellUPML with theoretical reﬂection coefﬁcient R =
exp(−16) and fourthorder polynomial conductivity variation is used to terminate MRTD
meshes that correspond to a 64×64 FDTD domain (Fig. 2.31). The three case studies are
600 650 700 750 800 850
0
0.5
1
I
n
c
i
d
e
n
t
scal
0−wav
1−wav
2−wav
2800 2850 2900 2950 3000 3050 3100 3150 3200 3250 3300
−0.2
0
0.2
0.4
Time step
R
e
f
l
e
c
t
e
d
scal
0−wav
1−wav
2−wav
FIGURE 2.29: Waveforms of incident and reﬂected electric ﬁeld (scaling and wavelet terms of order
0–2), sampled in front of the dielectric slab of Fig. 2.28
P1: IML/FFX P2: IML
MOBK05802 MOBK058Sarris.cls December 15, 2006 7:32
ANUMERICAL INTERFACEBETWEENFDTDANDHAARMRTD 39
0 10 20 30
−70
−50
−30
−10
Frequency [GHz]
S
1
1
[
d
B
s
]
8 point PML
6 point PML
TL theory
FIGURE2.30: Numerical and theoretical S
11
for the slab geometry
depicted in Figs. 2.32–2.34 and correspond to MRTD orders 0 by 0, 2 by 2, and 4 by 4,
respectively. An electric current excitation of the form:
J
y
(t) =
1
x z
4 (t/t
0
)
3
−(t/t
0
)
4
exp(−t/t
0
) (2.50)
with t
0
= 1/(4π f
0
) and f
0
=1 GHz, is applied in the middle of the domain. The FDTD cell
size is set equal to 2.5 mm × 2.5 mm and the time step is t = 4.5ps . Timedomain ﬁeld
waveforms are then sampled at the points indicated as A, B, Cin the three ﬁgures, corresponding
J
y
FDTD UPML
FDTD UPML
F
D
T
D
U
P
M
L
F
D
T
D
U
P
M
L
MRTD
(z)
(x)
64
8
8
8 8
64
FIGURE2.31: Case study for the interfacebased termination of a twodimensional MRTD domain
P1: IML/FFX P2: IML
MOBK05802 MOBK058Sarris.cls December 15, 2006 7:32
40 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
(x)
(z)
A
FIGURE2.32: FDTDUPML interfaced with a 0 by 0order MRTD (MRTD mesh is 32 × 32)
to FDTD cells (31, 31), (17, 17), (1, 1) of the 64 ×64 domain. The results, shown in Fig. 2.35,
demonstrate an excellent agreement between the pure FDTD scheme and the FDTDUPML
terminated MRTD. It is noted that the UPML regions in these three examples extend over
4, 1, and 0.25 MRTD cells respectively, demonstrating the ability of the interface to provide
MRTD absorbing boundary conditions with efﬁciency and complexity that are independent of
the order of the underlying wavelet expansion.
In addition, Fig. 2.36 depicts the electric ﬁeld waveform sampled at point B of the
MRTD domain of Fig. 2.33, when terminated at 4, 8, and 16 FDTDUPML cells (0.5, 1, and
2 MRTDcells respectively). The same waveformis determined via the FDTDmethod, applied
(x)
(z)
B
FIGURE2.33: FDTDUPML interfaced with a 2 by 2order MRTD (MRTD mesh is 8 × 8)
P1: IML/FFX P2: IML
MOBK05802 MOBK058Sarris.cls December 15, 2006 7:32
ANUMERICAL INTERFACEBETWEENFDTDANDHAARMRTD 41
(x)
(z)
C
FIGURE2.34: FDTDUPML interfaced with a 4 by 4order MRTD (MRTD mesh is 2 × 2)
at a 64×64 mesh, terminated at 32 UPMLcells. Evidently, all four curves agree well. Moreover,
the broad time window over which the results are given, shows the absence of any signiﬁcant
retroreﬂections from the absorbers in all three termination schemes. Finally, Figs. 2.37 and
2.38 depict the spatial evolution of the pulse, its propagation toward and its absorption fromthe
0 50 100 150 200 250 300 350
−1
0
1
x 10
4
MRTD+Interf.
FDTD
0 50 100 150 200 250 300 350
−4000
−2000
0
2000
4000
E
l
e
c
t
r
i
c
f
i
e
l
d
E
y
0 50 100 150 200 250 300 350
−4000
−2000
0
2000
Time step
FIGURE 2.35: Comparison of FDTD and MRTD/FDTD interface results for the three case studies
of Figs. 2.32–2.34
P1: IML/FFX P2: IML
MOBK05802 MOBK058Sarris.cls December 15, 2006 7:32
42 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
0 100 200 300 400 500 600 700 800
−6000
−4000
−2000
0
2000
4000
Time step
E
y
[
V
/
m
]
4 PML cells
8 PML cells
16 PML cells
FDTD (ref.)
FIGURE2.36: Electric ﬁeld at point B of Fig. 2.33, when a 4, 8, 16 FDTD UPML is used to terminate
the MRTD mesh. An FDTD solution (with a 32cell UPML) is appended for comparison
FDTDUPML boundaries, that are interfaced to the MRTD domain. In fact, the spatial ﬁeld
distribution at time step 600, essentially represents the level of errors due to artiﬁcial reﬂections
from the absorber. This level appears to be at −80 dB.
Inconclusion, one canconsider the proposedmethodas aneffective alternative for MRTD
mesh termination, since it employs existing UPML implementations, linking them to MRTD
codes via the simple connection algorithm that was earlier described.
FIGURE 2.37: Spatial ﬁeld distribution in the interfaceterminated domain of Fig. 2.31 at time steps
20, 50
P1: IML/FFX P2: IML
MOBK05802 MOBK058Sarris.cls December 15, 2006 7:32
ANUMERICAL INTERFACEBETWEENFDTDANDHAARMRTD 43
FIGURE 2.38: Spatial ﬁeld distribution in the interfaceterminated domain of Fig. 2.31 at time steps
80, 600
2.6.2.2 Absorbing Boundary Conditions
Evidently, even the application of simple absorbing boundary conditions (ABCs) assumes sig
niﬁcant complexity in the context of multiresolution techniques. The reason for that is that
ABCs impose mathematical conditions at planes that include only a fraction of the equivalent
grid points contained within a terminal cell. The concept of an interfacebased solution to this
question is depicted in Fig. 2.39. For the purpose of applying an absorbing boundary condition,
an FDTD region—no shorter than a single MRTD scaling cell—is introduced. Then, FDTD
region is terminated into a classically implemented Mur’s ABC. Thus, while the boundary
condition itself takes a single cell, now the FDTD region needs to extend over the degrees of
freedom of an MRTD scaling cell—not a fraction of it, as in the case of a PECbacked PML
absorber. Because of this tradeoff, this mesh truncation method is more suitable to loworder
MRTD schemes.
Excitation plane
MRTD/FDTD interface
ABC termination
MRTD region
FIGURE 2.39: Concept of ABC termination of an MRTD mesh; the FDTD region terminated into
the ABC should at least extend over one scaling MRTD cell
P1: IML/FFX P2: IML
MOBK05802 MOBK058Sarris.cls December 15, 2006 7:32
44 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
50 100 150 200 250
0
0.25
0.5
0.75
1
I
n
c
i
d
e
n
t
1950 2000 2050 2100 2150 2200
−1
0
1
2
x 10
−4
Time step
R
e
f
l
e
c
t
e
d
Order=0
Order=1
Order=2
Order=3
FIGURE 2.40: Incident and reﬂected ﬁeld waveforms, for Mur’s ﬁrstorder ABC (implemented via
an FDTD/MRTD interface), in the case of onedimensional wave propagation (Fig. 2.39), for MRTD
orders 0–3
Results froma onedimensional implementation of these concepts are shown in Figs. 2.40
and 2.41. A 0–5 GHz Gaussian pulse propagation, in a TEM waveguide mode, is considered.
Four Haar MRTD schemes of orders 0–3, with effective cell size of 0.0024 m and time step
equal to 0.9 of their Courant limit, are terminated into Mur’s ﬁrstorder ABC. The FDTD
region occupies two MRTD scaling cells for each scheme. Therefore, the thickness of the
FDTD region for each case study is: 0.0096, 0.0192, 0.0384, 0.0768 m, respectively. That is
also the reason for the phase difference of the reﬂected waveforms in Fig. 2.40. Finally, Fig. 2.41
presents the reﬂection coefﬁcient of each of the four terminations. Obviously, the obtained ABC
performance is similar in all MRTD cases and conﬁrms the capability of the method to provide
an efﬁcient means of implementing an ABCtruncation of wavelet grids.
2.7 CONCLUSIONS
A numerical interface between an arbitraryorder Haar MRTD and FDTD was developed
and applied in this chapter. The two unique features of the proposed technique are that ﬁrst,
P1: IML/FFX P2: IML
MOBK05802 MOBK058Sarris.cls December 15, 2006 7:32
ANUMERICAL INTERFACEBETWEENFDTDANDHAARMRTD 45
0 1 2 3
−140
−120
−100
−80
−60
Frequency [GHz]
S
1
1
[
d
B
s
]
MRTD order = 0
MRTD order = 1
MRTD order = 2
MRTD order = 3
FIGURE 2.41: Reﬂection coefﬁcient for Mur’s ﬁrstorder ABC (implemented via an FDTD/MRTD
interface), in the case of onedimensional wave propagation (Fig. 2.39)
it does not employ any interpolations or extrapolations and second, it is applicable for any
MRTD order. Hence, the developed algorithm constitutes a computationally efﬁcient tool for
jointly exploiting the advantages of FDTD and MRTD, which is critical for the acceleration of
timedomain schemes, when applied to largescale problems of current microwave technology
structures.
APPENDIX: COEFFICIENTS FORHAARMRTDUPDATE
EQUATIONS
Analytical expressions of integrals that are encountered in the application of the method of
moments for the derivationof Haar MRTDupdate equations are providedbelow. The maximum
waveletorder that determines the mutual shift of electric/magnetic ﬁeld nodes inaxial directions
is denoted as R, while s
R
= 1/2
R+2
is the shift itself.
+∞
−∞
dφ
k
dt
(t)φ
k+s
R
(t) dt = δ
k
,k+1
−δ
k
,k
(A.1)
+∞
−∞
dψ
r
k
, p
dt
(t)φ
k+s
R
(t) dt = 2
r/2
(δ
k
,k+1
−δ
k
,k
) δ
p,0
(A.2)
+∞
−∞
dφ
k
dt
(t)ψ
r
k+s
R
, p
(t) dt = 2
r/2
(δ
k
,k
−δ
k
,k+1
) δ
p,2
r
−1
P1: IML/FFX P2: IML
MOBK05802 MOBK058Sarris.cls December 15, 2006 7:32
46 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
+∞
−∞
dψ
r
k
, p
dt
(t)ψ
r
k+s
R
, p
(t) dt = D
0
(r, p, r
, p
)δ
k
,k+1
+D
1
(r, p, r
, p
)δ
k
,k
+∞
−∞
dφ
k
+s
R
dt
(t)φ
k
(t) dt = δ
k
,k
−δ
k
,k−1
+∞
−∞
dψ
r
k
+s
R
, p
dt
(t)φ
k
(t) dt = 2
r/2
(δ
k
,k−1
−δ
k
,k
) δ
p,2
r
−1
+∞
−∞
dφ
k
+s
R
dt
(t)ψ
r
k, p
(t) dt = 2
r/2
(δ
k
,k
−δ
k
,k−1
) δ
p,0
(A.3)
+∞
−∞
dψ
r
k
+s
R
, p
dt
(t)ψ
r
k, p
(t) dt = D
2
(r, p, r
, p
)δ
k
,k
+D
3
(r, p, r
, p
)δ
k
,k−1
with
D
0
(r, p, r
, p
) = 2
(r +r
)/2
¸
ψ(ξ
1
) −2ψ(ξ
2
) +ψ(ξ
3
)
¸
D
1
(r, p, r
, p
) = 2
(r +r
)/2
¸
ψ(ξ
1
) −2ψ(ξ
2
) +ψ(ξ
3
)
¸
D
2
(r, p, r
, p
) = −D
1
(r, 2
r
− p −1, r
, 2
r
− p
−1)
D
3
(r, p, r
, p
) = −D
0
(r, 2
r
− p −1, r
, 2
r
− p
−1)
(A.4)
and
ξ
1
= 2
r
1 + p
/2
r
−1/2
R+2
− p
ξ
2
= 2
r
1 +( p
+0.5)/2
r
−1/2
R+2
− p
ξ
3
= 2
r
1 +( p
+1)/2
r
−1/2
R+2
− p
ξ
1
= 2
r
p
/2
r
−1/2
R+2
− p
ξ
2
= 2
r
( p
+0.5)/2
r
−1/2
R+2
− p
ξ
3
= 2
r
( p
+1)/2
r
−1/2
R+2
− p
(A.5)
Finally, δ
κ,λ
is the wellknown Kronecker delta and ψ the Haar mother wavelet function. The
previous expressions are readily programmable and allowfor the development of arbitraryorder
Haar wavelet MRTD codes. Yet, code efﬁciency is greatly enhanced by a priori recognizing the
nonzero coefﬁcients D
0
, D
1
, D
2
, D
3
and omitting operations that involve multiplications by
zero in the main timestepping loop. This is done at the preprocessing stage of an MRTDcode.
P1: IML/FFX P2: IML
MOBK05803 MOBK058Sarris.cls December 15, 2006 7:45
47
C H A P T E R 3
Efﬁcient Implementation of Adaptive
Mesh Reﬁnement in the Haar
Waveletbased MRTDTechnique
The main attractive feature of waveletbased, timedomain techniques is the simple implemen
tation of adaptive meshing, through the application of a thresholding procedure to eliminate
wavelet coefﬁcients that attain relatively insigniﬁcant values, at a limited compromise of ac
curacy. However, little attention has been devoted so far to the investigation of computational
costs and accuracy tradeoffs in order to obtain thresholdingrelated operation savings. This
chapter presents an efﬁcient implementation of thresholding applied to a nonlinear problem
and reports signiﬁcant execution time savings compared to the conventional FDTD technique,
that the application of the proposed method has led to.
3.1 INTRODUCTION
Multiresolution analysis (MRA) concepts have already been employed in a wide range of appli
cations [11,14,15,37]. Amotivating force for this research activity is the fact that waveletbased
methods provide the most natural framework for the implementation of adaptive grids, dynam
ically following local variations and singularities of solutions to partial differential equations. In
particular, it can be proven that the decay of wavelet expansion coefﬁcients of a square integrable
function depends on the local smoothness of the latter [32]. Hence, signiﬁcant wavelet values are
expected at space–time regions, where high variations in the numerical solution evolve. In this
sense, sparing the arithmetic operations on wavelet coefﬁcients belowa certain threshold—small
enough according to accuracy requirements—amounts to imposing coarse gridding conditions
at those regions, while allowing for a denser mesh at parts of the domain where the solution
varies less smoothly.
Several approaches to waveletbased mesh reﬁnement have been presented in the litera
ture. In [37, 43], Haar wavelets were used to selectively reﬁne the resolution of an underlying
P1: IML/FFX P2: IML
MOBK05803 MOBK058Sarris.cls December 15, 2006 7:45
48 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
FDTDscheme at parts of the domain where this seemed to be a priori necessary, in a static sense.
However, the resulting method was strictly equivalent to a subgridded FDTD and presented
obvious accuracy disadvantages, since it was not enhanced by the interpolatory operations, typi
cally employed in subgridded FDTD[44]. In general, the necessity to incorporate wavelets into
a timedomain simulation technique is always related to dynamic rather than static subgridding,
since the latter—having been extensively studied in the literature—can be nowadays efﬁciently
implemented by several existing engines.
On the contrary, adaptive, waveletbased meshing was introduced in [32] and applied to
electromagnetic structures in [12–15], always in conjunction with highorder basis functions,
these being either Daubechies or Bsplines. In both cases, the complexity of the proposed
scheme made clear that “. . . the construction of waveletbased discretizations with a robust
ness and ﬂexibility comparable to FDTD is still a challenging task” [15]. The current work
meets this challenge by building up a twolevel scheme on the simplest wavelet basis, the Haar
basis and attempting a simple and explicit implementation of an algorithm for the thresh
olding of wavelet coefﬁcients, based on ideas originally related to shockwave problems of
computational ﬂuid dynamics. Execution time measurements for the algorithm as applied to
a nonlinear optics problem, show that this procedure can actually lead to fasterthanFDTD
simulations.
3.2 WAVELETBASEDFRONTTRACKING
In this section, the case study of the incidence of a 0–30 GHz Gaussian pulse on a dielectric slab
(Fig. 2.28, Section 2.6.2.1) is revisited. Thresholding of Haar wavelet coefﬁcients is applied,
based on the condition:
n
E
ψ
r, p
j
≤ 2
−r/2
,
where is a predeﬁned threshold. Hence, the amplitude of the threshold is adapted to the order
r of each wavelet coefﬁcient, through the use of a normalization term 2
−r/2
. This choice is
due to the fact that the maximum absolute value of the basis function ψ
r
j, p
(x) is 2
r/2
. Then,
thresholding amounts to omitting wavelet terms whose maximum contribution to the ﬁeld
value is less than . For two different threshold values, = 10
−6
and 10
−4
, the number of
wavelet coefﬁcients that are above threshold is computed at each time step. The resultant plots,
for wavelets of orders 0–3 are shown in Figs. 3.1 and 3.2. Evidently, despite small differences
between them, all four curves have a similar pattern: Around time step 2000, unthresholded
wavelet coefﬁcients are doubled, as a result of the pulse incidence on the dielectric slab, that
generates an additional (reﬂected) wavefront. Then, the PMLabsorption of the two wavefronts,
symmetrically generated by the source in the middle of the domain, causes a stepwise decrease
P1: IML/FFX P2: IML
MOBK05803 MOBK058Sarris.cls December 15, 2006 7:45
EFFICIENTIMPLEMENTATIONOF ADAPTIVEMESHREFINEMENT 49
FIGURE3.1: Unthresholded to total number (%) of wavelet coefﬁcients of orders 0, 1, for a fourthorder
Haar MRTD simulation of the problem of Fig. 2.28
in the number of active wavelet coefﬁcients. Finally, the absorption of the reﬂected wavefront
signals the end of the simulation and the decay of the number of unthresholded coefﬁcients
to almost zero. Overall, thresholding of wavelet coefﬁcients yields a compression in memory
requirements by 64.6% (in 8192 time steps). Also, Fig. 3.3 depicts the reﬂection coefﬁcient
Time step Time step
S
e
c
o
n
d

o
r
d
e
r
w
a
v
e
l
e
t
s
T
h
i
r
d

o
r
d
e
r
w
a
v
e
l
e
t
s
Thres = 10
−6
Thres = 10
−4
Thres = 10
−6
Thres = 10
−4
FIGURE3.2: Unthresholded to total number (%) of wavelet coefﬁcients of orders 2, 3, for a fourthorder
Haar MRTD simulation of the problem of Fig. 2.28
P1: IML/FFX P2: IML
MOBK05803 MOBK058Sarris.cls December 15, 2006 7:45
50 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
MRTD
TLtheory
MRTD
TLtheory
0 5 10 15 20 25 30
Frequency [GHz]
0 5 10 15 20 25 30
Frequency [GHz]
0
−20
−30
−40
−50
−60
−70
−10
S
1
1
[
d
B
s
]
0
−20
−30
−40
−50
−60
−70
−80
−10
S
1
1
[
d
B
s
]
FIGURE 3.3: Comparison of S
11
, derived by adaptive MRTD with = 10
−5
, 10
−4
for the dielectric
slab incidence problem of Fig. 2.28
S
11
for the dielectric slab, as computed via adaptive MRTD. The excellent comparison of the
numerical results to transmission line theory, demonstrates that wavelet thresholding has a
limited effect on the accuracy of the algorithm.
Yet, this aspect of wavelet adaptivity will not be further investigated in this work, because
taking advantage of wavelet compression at runtime, presupposes the application of memory
operations (reshufﬂe of ﬁeld arrays) that have an adverse effect on execution time. It is noted
though, that there are applications, where ﬁeld values at all time steps for a certain domain need
to be stored in memory. A posteriori processing those and wavelet thresholding them, can lead
to signiﬁcant memory economy [45].
Moreover, Figs. 3.4 and 3.5 depict six snapshots of electric ﬁeld distribution in space,
for the same problem, solved with a onelevel Haar MRTD scheme (under same gridding
conditions as before). The source is located in the middle of the domain and produces two
symmetric, left and rightpropagating waves. At each time step, the regions of unthresholded
wavelet coefﬁcients (for = 10
−4
) are found and their boundaries are plotted (in blue dots).
Two observations are now in order: First, wavelets follow successfully the ﬁeld wavefronts,
throughout the domain. Second, in order to keep track of the evolution of wavelet coefﬁcients,
one has to essentially study the movement of the boundaries of the unthresholded wavelet
regions. This task is limited to tracking the direction of the velocity of those boundaries.
Indeed, one can interpret the Courant stability condition (c t ≤ x), as follows: Wavefront
position can change at each time step at most by one cell. In the following, these observations
are exploited for the efﬁcient application of thresholding in MRTD.
P1: IML/FFX P2: IML
MOBK05803 MOBK058Sarris.cls December 15, 2006 7:45
EFFICIENTIMPLEMENTATIONOF ADAPTIVEMESHREFINEMENT 51
0 500 1000 1500 2000 2500 3000 3500 4000
−0.5
0
0.5
1
1.5
MRTD equivalent grid point
E
l
e
c
t
r
i
c
f
i
e
l
d
[
V
/
m
]
200 400 600 800 1000 1200 1400 1600 1800 2000
0
0.5
1
1.5
2
Scaling cell
F
r
o
n
t
b
o
u
n
d
a
r
i
e
s
0 500 1000 1500 2000 2500 3000 3500 4000
−0.5
0
0.5
1
1.5
MRTD equivalent grid point
E
l
e
c
t
r
i
c
f
i
e
l
d
[
V
/
m
]
200 400 600 800 1000 1200 1400 1600 1800 2000
0
0.5
1
1.5
2
Scaling cell
F
r
o
n
t
b
o
u
n
d
a
r
i
e
s
0 500 1000 1500 2000 2500 3000 3500 4000
−0.5
0
0.5
1
1.5
MRTD equivalent grid point
E
l
e
c
t
r
i
c
f
i
e
l
d
[
V
/
m
]
200 400 600 800 1000 1200 1400 1600 1800 2000
0
0.5
1
1.5
2
Scaling cell
F
r
o
n
t
b
o
u
n
d
a
r
i
e
s
0 500 1000 1500 2000 2500 3000 3500 4000
−0.5
0
0.5
1
1.5
MRTD equivalent grid point
E
l
e
c
t
r
i
c
f
i
e
l
d
[
V
/
m
]
200 400 600 800 1000 1200 1400 1600 1800 2000
0
0.5
1
1.5
2
Scaling cell
F
r
o
n
t
b
o
u
n
d
a
r
i
e
s
FIGURE 3.4: Evolution of wavefronts in the dielectric slab simulation (Fig. 2.28) and regions of un
thresholded wavelet coefﬁcients (boundaries are deﬁned by blue dots, a threshold = 10
−4
is used)
3.3 ADAPTIVEHAARWAVELETSIMULATIONOF PULSE
COMPRESSIONINANOPTICAL FIBERFILTER
As a vehicle for the demonstration of the algorithms developed in this study, the propagation of
an optical pulse through a ﬁber ﬁlter and its gradual compression is simulated. This nonlinear
phenomenon is based on a selfphase modulation (SPM) induced negative dispersion that the
pulse experiences along the ﬁber and was originally reported and studied in [46]. It is noted that
P1: IML/FFX P2: IML
MOBK05803 MOBK058Sarris.cls December 15, 2006 7:45
52 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
0 500 1000 1500 2000 2500 3000 3500 4000
−0.5
0
0.5
1
1.5
MRTD equivalent grid point
E
l
e
c
t
r
i
c
f
i
e
l
d
[
V
/
m
]
200 400 600 800 1000 1200 1400 1600 1800 2000
0
0.5
1
1.5
2
Scaling cell
F
r
o
n
t
b
o
u
n
d
a
r
i
e
s
0 500 1000 1500 2000 2500 3000 3500 4000
−0.5
0
0.5
MRTD equivalent grid point
E
l
e
c
t
r
i
c
f
i
e
l
d
[
V
/
m
]
200 400 600 800 1000 1200 1400 1600 1800 2000
0
0.5
1
1.5
2
Scaling cell
F
r
o
n
t
b
o
u
n
d
a
r
i
e
s
FIGURE3.5: Evolution of wavefronts in the dielectric slab simulation (continued)
the refractive index in the ﬁber assumes the form: n(z) = n
0
+n
1
cos(2β
0
z) +n
2
E
2
, where
the cosinusoidal term is due to a periodic structure (grating) written within the core of the ﬁber.
Numerical modeling of this case was also pursued in [47], by means of the Battle–Lemarie cubic
splinebased SMRTD technique, resulting in execution time that was reported to be larger
than FDTD by a factor of 1.5. In this work, a one waveletlevel Haar MRTD scheme is used,
for the purpose of estimating the improvement in computational performance that exclusively
originates from the dynamic adaptivity of MRTD rather than the highorder of an underlying
scaling basis.
3.3.1 Formulation of Haar MRTDScheme
The electric ﬁeld in the ﬁber is decomposed in forward and backward propagating waves as:
E(z, t) = E
F
(z, t) e
j (βz−ωt)
+ E
B
(z, t) e
−j (βz+ωt)
. (3.1)
Substituting this expression into Maxwell’s equations and discarding terms with spatial variation
faster than e
j 2βz
(slowly varying envelope approximation), the following system of equations is
deduced:
∂ E
F
∂z
+
n
0
c
∂ E
F
∂t
= j κ E
B
e
−2 j βz
+ j γ
E
F

2
+2E
B

2
E
F
(3.2)
∂ E
B
∂z
−
n
0
c
∂ E
B
∂t
= −j κ E
F
e
2 j βz
− j γ
E
B

2
+2E
F

2
E
B
, (3.3)
P1: IML/FFX P2: IML
MOBK05803 MOBK058Sarris.cls December 15, 2006 7:45
EFFICIENTIMPLEMENTATIONOF ADAPTIVEMESHREFINEMENT 53
with γ = πn
2
/λ, κ = πn
1
/λ and β(ω) = n
0
(ω)ω/c −2πn
0
/λ
0
and λ
0
the freespace wave
length that satisﬁes the Bragg condition for the grating.
The FDTD update equations for the system of (3.2), (3.3) can be retrieved from [47]. In
this work, Haar MRTD equations are derived, by ﬁrst expanding both forward and backward
ﬁelds in terms of Haar scaling φ
m
(z) = φ(z/z −m) and wavelet ψ
m
(z) = ψ(z/z −m) func
tions in space and pulse functions h
k
(t) = h(t/t −k) in time. For these functions, deﬁnitions
(2.4), (2.6), (2.14) are followed. Field expansions read:
E
x
(z, t) =
∞
¸
m,k=−∞
k
E
x,φ
m
φ
m
(z) +
k
E
x,ψ
m
ψ
m
(z)
h
k
(t), (3.4)
where x = F, B. Upon substitution of (3.4) into (3.2), (3.3) and application of the method of
moments, the system of Haar MRTD update equations is formulated. In particular, update
equations for the mth cell values of E
F,φ
, E
F,ψ
, E
B,φ
, E
B,ψ
at the kth time step read:
k+1
E
F, φ
m
=
k−1
E
F, φ
m
−s
¸
k
E
F,φ
m+1
−
k
E
F,φ
m−1
+
k
E
F,ψ
m+1
−2
k
E
F,ψ
m
+
k
E
F,ψ
m−1
¸
+2 j κs ze
−2 j β(m+1/2)z
¸
Sa(βz)
k
E
B,φ
m
+ j Sa
(βz/2)
k
E
B,ψ
m
¸
+2 j γ s z
¸
k
A
m k
E
F,φ
m
+
k
B
m k
E
F,ψ
m
¸
,
(3.5)
k+1
E
F, ψ
m
=
k−1
E
F, ψ
m
+s
¸
k
E
F,ψ
m+1
−
k
E
F,ψ
m−1
+
k
E
F,φ
m+1
−2
k
E
F,φ
m
+
k
E
F,φ
m−1
¸
+2 j κs ze
−2 j β(m+1/2)z
¸
Sa(βz)
k
E
B,ψ
m
+ j Sa
(βz/2)
k
E
B,φ
m
¸
+2 j γ s z
¸
k
A
m k
E
F,ψ
m
+
k
B
m k
E
F,φ
m
¸
,
(3.6)
k+1
E
B, φ
m
=
k−1
E
B, φ
m
+s
¸
k
E
B,φ
m+1
−
k
E
B,φ
m−1
+
k
E
B,ψ
m+1
−2
k
E
B,ψ
m
+
k
E
B,ψ
m−1
¸
+2 j κs ze
2 j β(m+1/2)z
¸
Sa(βz)
k
E
F,φ
m
− j Sa
(βz/2)
k
E
F,ψ
m
¸
+2 j γ s z
¸
k
C
m k
E
B,φ
m
+
k
D
m k
E
B,ψ
m
¸
,
(3.7)
k+1
E
B, ψ
m
=
k−1
E
B, ψ
m
−s
¸
k
E
B,ψ
m+1
−
k
E
B,ψ
m−1
+
k
E
B,φ
m+1
−2
k
E
B,φ
m
+
k
E
B,φ
m−1
¸
+2 j κs ze
2 j β(m+1/2)z
¸
Sa(βz)
k
E
F,ψ
m
− j Sa
(βz/2)
k
E
F,φ
m
¸
+2 j γ s z
¸
k
C
m k
E
B,ψ
m
+
k
D
m k
E
B,φ
m
¸
,
(3.8)
P1: IML/FFX P2: IML
MOBK05803 MOBK058Sarris.cls December 15, 2006 7:45
54 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
with Sa(w) = sin w/w, Sa
(w) = sin
2
w/w and s = c t/n
0
z, where z is the MRTD
scaling cell. Furthermore,
k
A
m
= 
k
E
F,φ
m

2
+
k
E
F,ψ
m

2
+2

k
E
B,φ
m

2
+
k
E
B,ψ
m

2
,
k
B
m
= 1/2
¸

k
E
F,φ
m
+
k
E
F,ψ
m

2
−
k
E
F,φ
m
−
k
E
F,ψ
m

2
+ 2

k
E
B,φ
m
+
k
E
B,ψ
m

2
−
k
E
B,φ
m
−
k
E
B,ψ
m

2
¸
,
k
C
m
= 
k
E
B,φ
m

2
+
k
E
B,ψ
m

2
+2

k
E
F,φ
m

2
+
k
E
F,ψ
m

2
,
k
D
m
= 1/2
¸

k
E
B,φ
m
+
k
E
B,ψ
m

2
−
k
E
B,φ
m
−
k
E
B,ψ
m

2
+ 2

k
E
F,φ
m
+
k
E
F,ψ
m

2
−
k
E
F,φ
m
−
k
E
F,ψ
m

2
¸
.
(3.9)
It is worth mentioning that operations in (3.5)–(3.8) can be readily split in scaling and wavelet
related ones, thus facilitating their adaptive application.
3.3.2 Boundary Conditions
Assuming an optical ﬁber that extends from z = 0 to z = L, the following boundary conditions
are imposed [47]:
E
F
(0, t) =
A
2
(1 + j ) e
−t
2
/(2α
2
)
, E
B
(L, t) = 0, (3.10)
by local modiﬁcation of the update equations, based on the explicit enforcement of those
conditions (without extrapolations or interpolations as proposed in the past [14]). For example,
if z = 0 belongs to the mth scaling cell of the domain, and g
k
= E
F
(0, kt) is the discrete
time sample of the excitation function at the kth step, applied at the Mth equivalent grid point
(Fig. 3.6), then, by the wavelet transform:
k
E
F
M
=
1
√
2
k
E
F,φ
m
+
k
E
F,ψ
m
= g
k
/
√
2, (3.11)
k
E
F/B
M+1
=
1
√
2
k
E
F/B,φ
m
−
k
E
F/B,ψ
m
, (3.12)
k
E
F/B
M+2
=
1
√
2
k
E
F/B,φ
m+1
+
k
E
F/B,ψ
m+1
. (3.13)
P1: IML/FFX P2: IML
MOBK05803 MOBK058Sarris.cls December 15, 2006 7:45
EFFICIENTIMPLEMENTATIONOF ADAPTIVEMESHREFINEMENT 55
Optical fiber filter
z = 0 z = L
(z)
mcell (m + 1)cell (n − 1)cell ncell
M M + 1 M + 2
N + 1 N N − 1
: MRTD EGP
FIGURE3.6: Haar MRTD equivalent grid points at the beginning and the end of the ﬁber
Deﬁning an auxiliary variable: A
k
≡
√
2
k
E
F
M+1
, the following, explicit update equation is
derived:
A
k+1
= A
k−1
−2s
k
E
F,φ
m+1
+
k
E
F,ψ
m+1
− g
k
+2 j κs zSa (βz/2) e
−j β(M+3/2)z
k
E
B,φ
m
−
k
E
B,ψ
m
+2 j γ s z
¸
1
2
k
E
F,φ
m
−
k
E
F,ψ
m
2
+
k
E
B,φ
m
−
k
E
B,ψ
m
2
¸
(3.14)
Then, scaling and wavelet terms of the forward ﬁeld are updated at the excitation cell, as:
k+1
E
F,φ
m
=
1
2
(g
k
+ A
k
) ,
k+1
E
F,ψ
m
=
1
2
(g
k
− A
k
) (3.15)
Similarly, the hard boundary condition on the backward wave at the nth cell of the domain
(Fig. 3.6), at z = L is applied via the expressions:
B
k+1
= B
k−1
−2s
k
E
B,φ
n−1
+
k
E
B,ψ
n−1
+2 j κs zSa (βz/2) e
−j β(M+1/2)z
k
E
F,φ
n
−
k
E
F,ψ
n
+2 j γ s z
¸
1
2
k
E
B,φ
n
+
k
E
B,ψ
n
2
+
k
E
F,φ
m
+
k
E
F,ψ
m
2
¸
,
(3.16)
P1: IML/FFX P2: IML
MOBK05803 MOBK058Sarris.cls December 15, 2006 7:45
56 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
where B
k
=
√
2
k
E
B
N
. Then, the scaling and wavelet terms of the backward ﬁeld, at the nth
MRTD scaling cell, are updated according to the scheme:
k
E
B,φ
m
=
k
E
B,ψ
m
=
B
k
2
Finally, absorbing boundary conditions for backward and forward waves are imposed at
z = 0, L, via matched layer absorbers. The absorbers are implemented as in [47], by expanding
their quadratically varying conductivities in Haar scaling functions. A maximum conductivity
σ
max
= 0.1 S/m is used in all subsequent simulations.
3.3.3 Validation
For validation purposes, results of a nonadaptive Haar MRTD code are compared to FDTD.
The parameters of the ﬁber are normalized, with respect to the length L of the ﬁlter, as fol
lows: κL = 4, βL = 12, γ L = 2/3. The time step is t = 0.003125n
0
L/(10c ) and two
cases for the FDTD cell size are considered: z = 0.001L and z = 0.002L. Respectively,
scaling cell sizes for MRTD are chosen to be z = 0.002L and z = 0.004L, as the intro
duction of one wavelet level reﬁnes the resolution by a factor of two, thus allowing for the
reduction of the number of cells in half. Matched layers of 500 and 250 cells terminate the
FDTD and MRTD meshes respectively, with maximum conductivity σ
0
= 0.1 S/m. Simu
lation data for both aforementioned cases are provided in Table 3.1. Execution times for the
relevant FDTD and MRTD cases over 12,000 time steps (on a Sun Ultra 80 workstation at
500 MHz) reveal that the nonadaptive MRTD code is slower than FDTD by a factor of 11–
16%. The reason for this expected slowdown is that the factor of increase in operations per
cell between MRTD and FDTD is greater than the ratio of FDTD to MRTD cells (equal
to two). The two methods agree well on the peak of the transmitted intensity (which cor
responds to the forward wave intensity at the end of the ﬁber). The slight difference can be
attributed to the fact that in the MRTD absorber, the conductivity was assumed constant
within each scaling cell and therefore it varied less smoothly than the corresponding FDTD
TABLE3.1: Validation Data for Nonadaptive MRTD Code
PARAMETER FDTD, I MRTD, I FDTD, II MRTD, II
Cells 1000 500 500 250
CPU time (s) 43.25 48.23 25.55 29.56
max E
F
(z = L)
2
10.6042 10.6635 10.3541 10.4283
P1: IML/FFX P2: IML
MOBK05803 MOBK058Sarris.cls December 15, 2006 7:45
EFFICIENTIMPLEMENTATIONOF ADAPTIVEMESHREFINEMENT 57
2000 4000 6000 8000
0
4
8
12
Time step
F
o
r
w
a
r
d
f
i
e
l
d
i
n
t
e
n
s
i
t
y

E
F

2
z = 0 z = L / 2 z = L
: MRTD
: FDTD
FIGURE 3.7: MRTD and FDTD results for the forward ﬁeld intensity at the beginning, the middle,
and the end of the optical ﬁber ﬁlter (1000 grid points)
absorber conductivity. Moreover, Fig. ?? depicts the pulse compression along the ﬁber, since it
includes pulse waveforms (extracted via 1000 cell FDTD and 500 cell MRTD), as probed at
the beginning, the middle and the end of the ﬁber ﬁlter. Again, excellent correlation between
the two methods (FDTD and MRTD) is demonstrated. In consequence, the nonadaptive
MRTD results can be henceforth used as a measure for accuracy estimation of adaptive MRTD
algorithms.
Finally, Figs. ?? and ?? show snapshots of the forward ﬁeld intensity and the magnitude
squared of the forward ﬁeld wavelet coefﬁcients plotted in space–time coordinates. Evidently,
the evolution of both the wavelet coefﬁcients and the ﬁeld itself takes place along the char
acteristic line of forward ﬁeld propagation, z = c t. A similar pattern for wavelet behavior was
obtained in [15], where nonuniform multiconductor transmission line equations were solved
via a biorthogonal wavelet basis. In the following, this observation is utilized for the devel
opment of a computationally efﬁcient approach to the problem of thresholding of wavelet
coefﬁcients.
3.3.4 Implementation and Performance of Adaptive MRTD
Regarding the development of adaptive algorithms, two questions are addressed in this work:
First, the application of thresholding with as few operations (namely checks on wavelet
P1: IML/FFX P2: IML
MOBK05803 MOBK058Sarris.cls December 15, 2006 7:45
58 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
FIGURE3.8: Spatiotemporal propagation and compression of the nonlinear optical pulse (forward ﬁeld
intensity is shown)
FIGURE 3.9: Spatiotemporal evolution of forward ﬁeld wavelet coefﬁcients (square of magnitude of
wavelet coefﬁcients is shown)
P1: IML/FFX P2: IML
MOBK05803 MOBK058Sarris.cls December 15, 2006 7:45
EFFICIENTIMPLEMENTATIONOF ADAPTIVEMESHREFINEMENT 59
coefﬁcients being absolutely above or below a threshold ) as possible and second, the stable
exploitation of the compressed representation of the solution for the reduction in update op
erations at each time step. For this purpose, the following method is adopted [12, 15, 32], for
both forward and backward propagating wave arrays:
• All cells are initialized as being above threshold (referred to as active).
• Scaling coefﬁcients are updated throughout the mesh, along with active wavelet terms.
In all operations, only scaling and active wavelet terms are taken into account.
• At each time step, the magnitudes of wavelet terms at cells that are designated as active
are compared to an absolute threshold . The corresponding cells remain active only if
their wavelets are above the threshold.
• The region of the active cells is extended to all nearest neighbors of the latter. Since
only one wavelet level is used here, this implies cells that are immediately to the left or
to the right of cells on the border of active regions.
Hence, thresholding checks and update operations are limited to a subset of the ﬁeld coefﬁcients.
Note that the use of the Haar basis and a single wavelet level scheme, keeps the implementation
of this adaptive algorithm relatively simple and readily expandable to three dimensions. On the
other hand, the complexity of numerical solvers based on higherorder basis functions and mul
tiple wavelet levels has regularly undermined the potential of adaptivity to yield execution times
better than FDTD. Furthermore, the frequency of thresholding checks is implicitly dependent
on the CFL number s of the simulation, which effectively determines the maximum number of
cells that a front may move through in a single time step. Approximately, a thresholding check
window should be at most equal to 1/s , with the previously described procedure strictly corre
sponding to the value of s = 1. Physically, the thresholding algorithm, that was ﬁrst introduced
for the numerical solution of shock wave problems in [32], assumes the evolution of wavelet
coefﬁcients along wavefronts deﬁned from the characteristics of a given problem. Then, the
purpose of adding pivot elements, to extend the domain of active coefﬁcients, is actually the
tracking of these wavefronts as they move throughout the computational domain. The concept
of this algorithm is schematically explained in Fig. ??. As shown, adding the pivot elements at
the two edges of a wavefront, ensures the tracking of the direction of its movement. Indeed,
if that moves to the left, the left pivot element is activated at the “next” time step. Then, the
active wavelet region moves to the left, with the wavefront. Otherwise, it follows a movement
of the front to the right.
To this end, the second case that was presented in Section ?? is repeated here by using sev
eral absolute thresholds and CPU time measurements are carried out. Also, three thresholding
P1: IML/FFX P2: IML
MOBK05803 MOBK058Sarris.cls December 15, 2006 7:45
60 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
: active wavelet coefficients
: pivot elements
FIGURE3.10: Graphical presentation of the waveletbased, fronttracking algorithm
windows are investigated, with the outlined algorithm being applied every one, two, and four
time steps respectively. Although the latter was slightly larger than 1/s , it produced satisfac
tory (accuracywise) results. In all cases, thresholds up to 0.001 led to simulations that clearly
resolved the pulse compression and suffered from errors (with respect to the unthresholded
MRTD, which is used as a reference solution) of less than 1%.
In Figs. ?? and ??, forward ﬁeld intensity waveforms sampled at z = L, for thresholds 0.1,
10
−4
and 10
−7
are shown, along with the incident pulse. The last two are in good agreement both
with each other and with the previously presented FDTD and MRTD results, while the ﬁrst
suffers from signiﬁcant numerical errors, that demonstrate themselves as a ripple corrupting the
pattern of the waveform. More explicitly, CPUeconomy with respect to FDTDand error in the
peak transmitted forward ﬁeld intensity are plotted (with respect to the absolute threshold ),
in Figs. ?? and ??. It is thus shown, that the adaptive MRTD code can extract the solution to
this problem, at a CPU time reduced (compared to FDTD) by a factor close to 30%, with errors
limited at the order of 0.1%.
Thresholding operations in this problemrepresented a worstcase scenario for the compu
tational overhead that the adaptive algorithm may bring about, for the following reasons: First,
P1: IML/FFX P2: IML
MOBK05803 MOBK058Sarris.cls December 15, 2006 7:45
EFFICIENTIMPLEMENTATIONOF ADAPTIVEMESHREFINEMENT 61
1500 2500 3500 4500
0
2
4
6
8
10
Time step
I
n
c
i
d
e
n
t
f
o
r
w
a
r
d
f
i
e
l
d
i
n
t
e
n
s
i
t
y

E
F
(
z
=
0
)

2
4000 6000 8000 10000
0
2
4
6
8
10
12
Threshold = 10
−7
Time step
T
r
a
n
s
m
i
t
t
e
d
f
o
r
w
a
r
d
f
i
e
l
d
i
n
t
e
n
s
i
t
y

E
F
(
z
=
L
)

2
FIGURE3.11: Incident and transmitted forward ﬁeld intensity for a threshold of 10
−7
the geometry was a onedimensional one, a signiﬁcant part of which was almost throughout the
simulation occupied by the propagating pulse and second, these operations involved complex
numbers and nonlinear terms. Therefore, the fact that an accelerated performance of adaptive
MRTD (with respect to FDTD) was achieved is important and demonstrates the potential of
the algorithm for larger geometries.
4000 6000 8000 10000
0
2
4
6
8
10
12
Threshold = 10
−4
Time step
T
r
a
n
s
m
i
t
t
e
d
f
o
r
w
a
r
d
f
i
e
l
d
i
n
t
e
n
s
i
t
y

E
F
(
z
=
L
)

2
4000 6000 8000 10000
0
2
4
6
8
10
12
Threshold = 10
−1
Time step
T
r
a
n
s
m
i
t
t
e
d
f
o
r
w
a
r
d
f
i
e
l
d
i
n
t
e
n
s
i
t
y

E
F
(
z
=
L
)

2
FIGURE3.12: Transmitted forward ﬁeld intensity for thresholds of 10
−4
, 10
−1
P1: IML/FFX P2: IML
MOBK05803 MOBK058Sarris.cls December 15, 2006 7:45
62 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
10
−7
10
−6
10
−5
10
−4
10
−3
10
−2
10
−1
0
10
20
30
40
50
Threshold
E
c
o
n
o
m
y
[
%
]
i
n
C
P
U
t
i
m
e
Thres. window = 1
Thres. window = 2
Thres. window = 4
FIGURE3.13: CPU time economy for the adaptive MRTD code, with respect to FDTD, for absolute
thresholds from 10
−7
to 0.1 applied to Haar MRTD
10
−7
10
−6
10
−5
10
−4
10
−3
10
−2
10
−1
10
−2
10
−1
10
0
10
1
10
2
Threshold
R
e
l
a
t
i
v
e
e
r
r
o
r
i
n
p
e
a
k
i
n
t
e
n
s
i
t
y
[
%
]
Thres. window = 1
Thres. window = 2
Thres. window = 4
FIGURE 3.14: Relative error (%) in the peak intensity of the transmitted forward ﬁeld for adaptive
Haar MRTD
P1: IML/FFX P2: IML
MOBK05803 MOBK058Sarris.cls December 15, 2006 7:45
EFFICIENTIMPLEMENTATIONOF ADAPTIVEMESHREFINEMENT 63
3.4 CONCLUSIONS
Based on the study of a nonlinear pulse compression by an optical ﬁber ﬁlter, this chapter
demonstrated Haar waveletbased simulations with adaptive meshing that achieved better
thanFDTD execution times. To the extent of the author’s knowledge this is the ﬁrst time
when an adaptive, waveletbased code surpasses the efﬁciency of the conventional FDTD, not
only in terms of memory but also in terms of execution time requirements. The satisfactory
performance of the proposed technique stems from its relative simplicity that allows for an
efﬁcient implementation of its two components: thresholding tests of wavelet coefﬁcients and
operation savings while performing updates of ﬁeld arrays.
P1: IML/FFX P2: IML
MOBK05803 MOBK058Sarris.cls December 15, 2006 7:45
64
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05804 MOBK058Sarris.cls December 15, 2006 7:49
65
C H A P T E R 4
The Dynamically Adaptive Mesh
Reﬁnement (AMR)FDTD
Technique: Theory
The ﬁnitedifference timedomain(FDTD) method is combined withthe adaptive meshreﬁne
ment (AMR) technique of computational ﬂuid dynamics to achieve a fast, timedomain solver
for Maxwell’s equations, based on a threedimensional, adaptively reﬁnable moving cartesian
mesh. This combination allows the resulting technique to adapt to the problem at hand, opti
mally distributing computational resources in a given domain as needed, by recursively reﬁning
a coarse grid in regions of large gradient of electromagnetic ﬁeld energy. Thus, instead of setting
up a statically adaptive mesh, which is the subject of extensive previous research, the idea of a
dynamically adaptive mesh is pursued, because of the optimal suitability of the latter to the na
ture of timedomain electromagnetic simulations. However, the algorithm is still implemented
within the context of FDTD, with secondorder accurate update equations and is referred to as
dynamically AMRFDTD, to emphasize the dynamic evolution of the underlying mesh. This
chapter is aimed at providing the reader with a detailed description of the algorithm, ﬁlled with
all the information needed for its two and threedimensional implementation. The chapter is
concluded with a comparison between the dynamically AMRFDTD and the Haar MRTD
approach that was discussed in the last two chapters.
4.1 INTRODUCTION
The ﬁnitedifference timedomain technique [6] has been extensively employed in the modeling
of microwave integrated circuits [48]. It is especially suitable for wideband applications since it
allows for the characterization of a given structure in a broad frequency range, through a single
simulation. However, the FDTD stability and dispersion properties impose severe limitations
on the choice of the cell size and the time step of the method, rendering its application to
complex structures computationally expensive.
The challenge of accelerating FDTD simulations for practical geometries has been ad
dressed in the past with a variety of static subgridding techniques [10, 44]. According to those,
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05804 MOBK058Sarris.cls December 15, 2006 7:49
66 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
local mesh reﬁnement is pursued in a priori deﬁned regions of a computational domain, as dic
tated by physical considerations. For example, the presence of metallic edges, or high dielectric
permittivity inclusions, would call for a locally dense mesh, embedded in a coarser global one.
The use of local mesh reﬁnement typically results in signiﬁcant computational savings com
pared to the conventional FDTD, despite the fact that its implementation is associated with
additional interpolation and extrapolation operations in both space and time.
However, this approach ignores the dynamic nature of timedomain ﬁeld simulations.
In fact, techniques such as FDTD and TLM essentially register the history of a broadband
pulse propagating in a device under test, along with its multiple reﬂections from parts of the
latter. Hence, a sharp edge of a microstrip structure is not continuously illuminated by the pulse
excitation; on the contrary, it is so for a (potentially small) fraction of the total simulation time,
during which a local mesh reﬁnement around it is needed. Therefore, static mesh reﬁnement,
which is widely employed in frequencydomain simulations and has been incorporated in com
mercial ﬁniteelement tools, is only a suboptimal solution to the mesh reﬁnement problem in
the framework of timedomain analysis.
More recently, a movingwindow FDTD (MWFDTD) method was proposed for the
tracking of the forward propagating wave in the twodimensional terrain environment of a
wireless channel [49], similar to ideas previously proposed in [50, 51]. The single moving
window used by the method was characterized by ﬁxed size and velocity and therefore, it
could not track reﬂections (which were absorbed by terminating boundaries of the window).
As a result, the MWFDTD is not wellsuited for microwave circuit simulations, where the
modeling of phenomena as common as signal reﬂection and branching would require multiple
and potentially rotating windows.
Inthe context of computational ﬂuiddynamics, the technique of adaptive meshreﬁnement
(AMR) was introduced in [52], for the solution of hyperbolic partial differential equations. The
application of AMRis based on the use of a hierarchical mesh, recursively developed through the
reﬁnement of a coarse root mesh, which covers the entire computational domain. The regions
of the computational domain that need further mesh reﬁnement are detected via error estimates
or indicators such as gradients of the quantity to be solved for. There may also be dense mesh
regions, where the use of a dense mesh is not necessary after a certain time step. These can
then be coarsened, again in a recursive manner. Dense and coarse mesh regions are organized
via a clustering algorithm that is accompanied by regular checks (every certain time steps) of
the error estimates, which guide the process of migration of a cell from one level of resolution
to another. This procedure can be associated with the algorithm of the previous chapter, which
used wavelet ﬁeld expansions in order to track the spatiotemporal evolution of shockwave and
nonlinear optical pulse propagation problems, respectively. However, the generalization of such
waveletbased algorithms to three dimensions presents a signiﬁcant added complexity, while the
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05804 MOBK058Sarris.cls December 15, 2006 7:49
THEDYNAMICALLY ADAPTIVEMESHREFINEMENT(AMR)FDTDTECHNIQUE 67
implicit relation between the actual ﬁeld values and the wavelet expansion coefﬁcients renders
the application of boundary conditions, which are essential for the connection of nested meshes
of different resolutions, computationally burdensome.
Note that this chapter’s algorithm is a “multilevel” one, in the sense that it can support
multiple levels of mesh resolution, by reﬁning the Yee cells of an underlying root mesh by
a predeﬁned factor. In our terminology, a twolevel scheme would enclose Yee cells of size
x ×y ×z and x/N
s
×y/N
s
×z/N
s
, while an Nlevel scheme would enclose Yee
cells of size x/N
i
s
×y/N
i
s
×z/N
i
s
, for i = 0, 1, · · · N −1, where the factor N
s
(here
taken equal to 2) is the mesh reﬁnement factor.
In this chapter, the conventional FDTD is combined with the AMR method, in order to
formulate an efﬁcient AMRFDTD technique of superior performance [53]. Instead of using
ﬁxed subgrids, this method uses subgrids that are adaptively deﬁned, according to the evolution
of ﬁeld distributions in space and time. As an example, when a Gaussian pulse propagating
along a microstrip line is simulated, the adaptive mesh reﬁnement scheme successfully tracks
the movement of the pulse, thereby reﬁning only the region that surrounds the propagating
pulse. In this case, the AMR accuracy is comparable to that of a uniformly (throughout the
entire computational domain) dense mesh FDTD.
The chapter is organized as follows: The general structure of the AMRFDTDalgorithm
is presented in Section 4.2, while Section 4.3 refers to the update procedure of the tree of meshes
that the algorithm uses in order to adaptively track the ﬁeld evolution. Section 4.4 outlines
how cells that need further mesh reﬁnement or cells that can be removed from the mesh are
detected, clustered, and managed by the AMRFDTD scheme. Finally, Section 4.5 compares
AMRFDTD to the MRTD technique outlining their interesting similarities along with their
distinctive differences.
4.2 AMRFDTD: OVERVIEWOF THEALGORITHM
In general, a cartesian FDTD mesh occupies a rectangular region,
A =
¸
(x, y, z) , x
a
1
≤ x ≤ x
a
2
, y
a
1
≤ y ≤ y
a
2
, z
a
1
≤ z ≤ z
a
2
¸
,
terminatedby closedor absorbing boundaries. Inhomogeneous material properties canbe readily
assumed, by letting the dielectric permittivity and magnetic permeability of the structure be
generic functions of the space variables = (x, y, z), μ = μ(x, y, z), respectively. Let us
consider the subdivision of the domain A in N
x
× N
y
× N
z
Yee cells, indexed by a triplet
(i, j, k). Each cell occupies a volume x ×y ×z, where x, y, z denote the cell sizes
in x, y, and zdirections. The mesh of these N
x
× N
y
× N
z
cells is the coarse grid of region
A, that the algorithm under development will selectively and locally reﬁne.
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05804 MOBK058Sarris.cls December 15, 2006 7:49
68 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
Let us now assume that there are cells within the domain A whose reﬁnement is
necessary, according to certain accuracy criteria (the discussion of the latter is deferred to
Section 4.4). These cells are ﬁrst clustered together and then covered by rectangular subre
gions B
m
, m = 1, 2, · · · N which belong to A. Throughout the algorithmic development of the
AMRFDTD, it will be ensured that these subregions can share planar boundaries, yet they
cannot overlap. This is important in order to preserve the possibility of further reﬁnement of
these subregions independently from each other, as required by the evolution of the ﬁeld solu
tion. Each rectangular region B
n
is subdivided in Yee cells of dimensions: x/2, y/2, z/2.
Hence, a reﬁnement factor of 2 is used in every direction, reducing the Yee cell volume of the
initial mesh by a factor of 8.
To summarize, a coarse mesh has been deﬁned in the rectangular region A, enclosing ﬁner
meshes in rectangular subregions B
n
of the latter. The mesh of region A, henceforth referred to
as mesh A, will be called the root mesh, or level1 mesh. The meshes of regions B
m
, or meshes
B
m
, will be called child meshes of A, or level2 meshes. Recursively, each B
n
can be further
reﬁned to have its own child meshes, again reﬁning the cell sizes involved by a factor of 2. The
hierarchically deﬁned meshes, that the computational domain consists of in an AMRFDTD
simulation, can be assembled in the data structure of a “mesh tree.” The mesh tree is regenerated
every N
AMR
time steps through a recursive mesh reﬁnement procedure. An example is shown in
Fig. 4.1. There is only one level1 mesh, which is also called the root mesh and covers the
entire computational domain. Each levelm +1 mesh is created by reﬁning a subset of
the cells of a levelm mesh in a rectangular region by a factor N
s
(in each direction). Thus,
the two meshes form a child–parent relation. The child meshes of the same parent may share
an edge, but they may not overlap otherwise. In Fig. 4.1(b) a solid line corresponds to a child–
parent relation, while a dashed line corresponds to a boundary shared by meshes of the same
level.
C4
B1
C1
C2
C3
B3
C4
A
B2
(a) (b)
level 1
level 2
level 3 C3
A
B1 B2 B3
C1 C2
FIGURE4.1: (a) Geometry and (b) data structure of a threelevel mesh tree
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05804 MOBK058Sarris.cls December 15, 2006 7:49
THEDYNAMICALLY ADAPTIVEMESHREFINEMENT(AMR)FDTDTECHNIQUE 69
E
E
H
y
H
E
y
H
x
z
y
x
z
z
x
FIGURE4.2: Sampling points of electric and magnetic ﬁeld components in a Yee cell. Solid arrows are
for electric ﬁeld, whereas hollow arrows are for magnetic ﬁeld. The shaded areas represent reﬁned Yee
cells
According to the convention of the Yee cell in FDTD(Fig.4.2), electric ﬁeld components
are sampled at the center of the edges of each cell, while magnetic ﬁeld components are sampled
at face centers. The sampling points of a parent mesh may coincide or not with the sampling
points of its child mesh, depending on whether the reﬁnement factor is odd or even, respectively.
In this work, the reﬁnement factor is 2 (or powers of 2, with respect to the root mesh), hence the
grid points of child and parent meshes do not coincide. An alternative case, where the choice of
a reﬁnement factor of three renders the parent mesh sampling points also child mesh sampling
points can be found in [44].
The main difference between standard subgridded FDTD methods and AMRFDTD is
the dynamic mesh generation which is pursued in the latter, every N
AMR
time steps. To deﬁne
the time step of AMRFDTD, the following observations need to be made. For a levelM
mesh, Yee cell dimensions are: x/2
M−1
, y/2
M−1
, z/2
M−1
, where x, y, z are the
root mesh Yee cell dimensions. Furthermore, the Courant number is ﬁxed to a constant value
s in all meshes. This implies that the root mesh time step t (from now on referred to, as
AMRFDTD time step) is given as:
t = s
1
u
max
p
1
1
x
2
+
1
y
2
+
1
z
2
, (4.1)
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05804 MOBK058Sarris.cls December 15, 2006 7:49
70 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
where u
max
p
is the maximum phase velocity in the computational domain. Applying (4.1) for a
mesh of level M, keeping s ﬁxed, yields a time step t
M
for this mesh, equal to:
t
M
= t/2
M−1
. (4.2)
For example, level2 meshes are updated twice as many times as the root mesh. Thus, another
shortcoming of the conventional FDTD is addressed; the minimum time step of the algorithm
is only used for the update of regions of large ﬁeld variations, as opposed to the whole domain,
a salient feature that is also part of the ﬁxed subgridding algorithms of [10, 44].
The loop of the AMRFDTD operations is as follows:
1. Check the number of time steps executed. If it is an integer multiple of N
AMR
, perform
adaptive mesh reﬁnement to create a new mesh tree, and carry the ﬁeld values from the
old mesh tree to the new mesh tree.
2. Update ﬁelds of the root mesh.
3. Copy ﬁelds from the root mesh to the boundary of the child meshes. Update ﬁelds of
the child meshes 2
M−1
times, where M is the resolution level of the mesh. Copy ﬁelds
from child meshes back to the root mesh, for the time steps of the latter.
4. Check whether the maximum time step has been reached. If so, terminate the simula
tion, otherwise return to step 1.
The next two sections are aimed at explaining these steps in detail.
4.3 MESHTREEANDFIELDUPDATEPROCEDUREIN
AMRFDTD
AMRFDTD applies the wellknown ﬁeld update equations of FDTD for each mesh, yet the
interconnection of the different resolution meshes that march in time at different time steps
is an issue to be addressed explicitly. In this section, the types of interfaces that occur in an
AMRFDTD domain are presented, along with their treatment in the update process.
4.3.1 Categorization of Boundaries of Child Meshes
The different categories of child mesh boundaries that may practically occur are shown in
Fig. 4.3. To facilitate the presentation of the different cases, a twodimensional case is discussed
(readily extensible to three dimensions). Consider two child meshes B
1
and B
2
, embedded in a
root mesh A. Three separate cases of boundaries can be identiﬁed:
1. Segment e a: It is deﬁned here as a physical boundary (PB), including absorbing and/or
perfect electric conductor boundary conditions.
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05804 MOBK058Sarris.cls December 15, 2006 7:49
THEDYNAMICALLY ADAPTIVEMESHREFINEMENT(AMR)FDTDTECHNIQUE 71
a
c
d
b
e
B2
B1
A
FIGURE4.3: Types of boundaries of a child mesh
2. Segment cd: This is a boundary between “sibling” meshes (SB).
3. Segments ab, bc, de: These are boundaries between child and parent meshes (CPB).
Evidently, it is possible that a boundary may belong to more than one of the aforementioned
categories. Then, its classiﬁcation is based on the following hierarchy:
PB > SB > CPB.
4.3.2 Update of the Mesh Tree
In the following, the update of a twolevel mesh tree (with root and level1 meshes) is discussed.
The steps outlined here can be recursively extended to mesh trees of more levels, by considering,
for example, level2 meshes as roots for level 3meshes and so on. As before, t is the time step
of the root mesh, while the child level1 mesh uses a time step t/2. Let us consider the order of
updates, assuming that the timemarching procedure has reached the time t = nt. As a result,
it is assumed that the root mesh contains the updated values of the electric ﬁeld component
grid points at nt and those of the magnetic ﬁeld component grid points at (n −1/2) t. In
addition, the child meshes contain the updated electric and magnetic ﬁeld values at nt and
(n −1/4) t, respectively. Then, the following procedure is applied.
1. Backup magnetic ﬁeld components of the root mesh at (n −1/2) t. Obtain their
values at (n +1/2) t, by applying the FDTD update equations.
2. Backup the electric ﬁeld components of the root mesh at nt. Obtain their values at
(n +1) t, by applying the FDTD update equations.
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05804 MOBK058Sarris.cls December 15, 2006 7:49
72 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
3. For each child mesh, apply the update equations to obtain the magnetic ﬁeld values at
(n +1/4) t.
4. For each child mesh:
(a) Update the interior (nonboundary) electric ﬁeld grid points, to obtain their values
at (n +1/2) t.
(b) Update the boundary electric ﬁeld grid points, at boundaries of the PB, SB, CPB
type, to obtain their values at (n +1/2) t. Obviously, these updates are nontrivial,
since they invoke grid points of the root mesh, calculated at time steps of the child
mesh. Therefore, interpolation needs to be carried out, in a way that is analyzed in
the next subsection.
5. For each child mesh, backup the magnetic ﬁeld components at (n +1/4) t, and obtain
their values at (n +3/4) t, by applying the FDTD update equations.
6. Repeat steps 4a, 4b, to advance the electric ﬁeld values of the child meshes to (t +t).
7. For each child mesh:
(a) Put the spatially interpolated electric ﬁeld nodal values at (n +1) t back to the
parent mesh, excluding CPB grid points.
(b) Put the spatially interpolated magnetic ﬁeld nodal values at (n +1/2) t back to
the parent mesh.
This process can be readily extended to the multilevel case, where meshes of levels
1, 2, 3, . . . , N
L
are supported. If N
s
is the mesh reﬁnement factor, assuming that the level
1 (root) mesh uses a time step t, the levelL meshes use a time step t/N
(L−1)
s
. For levelL
meshes to advance one time step of theirs, level(L +1) meshes need to advance N
s
of their
own time steps to be synchronized with the former. It should be noted that all meshes of the
same level need to be updated for a certain time step before advancing to the next time step,
because their values are invoked in the update equations of neighboring meshes. The following
operations are recursively performed:
1. For each mesh of level L, the magnetic ﬁeld components are updated.
2. For each mesh of level L, the electric ﬁeld components are updated.
3. If L +1 ≤ N
L
, for each mesh of level L, for a time step counter k varying from 0 to
N
s
−1, the previous two sets of updates are performed.
4. For all the meshes except the root mesh, one can deﬁne a “local” time step variable
S, which enumerates the time steps executed within one time step of its parent mesh.
Hence, this counter will take on values from 0 to N
s
−1.
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05804 MOBK058Sarris.cls December 15, 2006 7:49
THEDYNAMICALLY ADAPTIVEMESHREFINEMENT(AMR)FDTDTECHNIQUE 73
(a) If S = Ns /2, the magnetic ﬁeld components of each mesh of level L, are trans
ferred to its parent mesh. Since the node points of parent and child meshes are not
necessarily collocated (they are evidently offset if N
s
= 2), the former are interpo
lated from the latter.
(b) If S = Ns −1, the electric ﬁeld components of each mesh of level L are transferred
to its parent mesh.
The aforementioned steps can be implemented in a function update Le vel (L, S), which is
recursively called to update all meshes with level greater or equal to L and a local time step S.
The main program just needs to call update Le vel (1, 0) for the update of the entire mesh tree
for one time step.
4.3.3 Field Updates on CPBType Boundaries
The electric ﬁeld components of a child mesh tangential to its CPBtype boundary (steps 4a,
4b) are obtained from its parent through trilinear interpolation in space and time, as shown in
Fig. 4.4. Since such a boundary is characterized by ﬁxing one spatial variable and letting the
other two vary, along with time, trilinear interpolation provides the expression employed in all
these updates. To interpolate a function f (η, ξ, τ) in the range 0 ≤ η, ξ < 1, 0 ≤ τ < 1, by
using values of the function at points (0, 0, 0), (0, 0, 1), (0, 1, 0), (0, 1, 1), (1, 0, 0), (1, 0, 1),
j y
(j+1) y
Δ
Δ
x Δ x (i + 1) i Δ
y
x
FIGURE 4.4: Sampling points for electric ﬁeld on CPBtype boundary. The solid arrows are for the
child mesh. The hollow arrows are for the parent mesh
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05804 MOBK058Sarris.cls December 15, 2006 7:49
74 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
(1, 1, 0), and (1, 1, 1), the following formula can be applied:
f (η, ξ, τ) = (1 −η) (1 −ξ) (1 −τ) f (0, 0, 0)
+η (1 −ξ) (1 −τ) f (1, 0, 0)
+ξ (1 −η) (1 −τ) f (0, 1, 0)
+ηξ (1 −τ) f (1, 1, 0)
+(1 −η) (1 −ξ) τ f (0, 0, 1)
+η (1 −ξ) τ f (1, 0, 1)
+ξ (1 −η) τ f (0, 1, 1)
+ξητ f (1, 1, 1) .
(4.3)
As an example, consider the E
x
ﬁeld on a CPBtype boundary located at z = 0. The par
ent mesh has E
x
((i +1/2) x, j y, 0, kt), whereas the child mesh needs E
x
(x, y, 0, t). If
(i −1/2)x ≤ x < (i +1/2)x, j y ≤ y < ( j +1)y, kt ≤ t < (k +1)t, with i, j, k
being integers. Then, η, ξ, and τ correspond to the following normalized spatial and temporal
variables:
x
x
−i −1/2,
y
y
− j,
t
t
−k,
which vary within the interval [0, 1). Since the child mesh regions B
n
belong to their root mesh
region A, the use of (4.3) allows for the determination of any sampling point of the child mesh
fromsampling points of its root mesh, enabling the transfer of data which is included in step 4b.
4.3.4 Field Updates on SBType Boundaries
First, note that boundaries between child meshes of different levels of resolution can be treated
as the CPB boundaries that were discussed above. Therefore, the treatment of SBtype bound
aries can be limited to boundaries between child meshes of the same level. Since, by Yee cell
convention, the electric ﬁeld is sampled at cell edge centers and the magnetic ﬁeld is sampled
at cell face centers, two meshes of the same level share tangential electric ﬁeld and normal
magnetic ﬁeld components at the interface between them (see, e.g., Fig. 4.5).
However, an inspection of the required grid points that the Yee’s algorithm invokes in
the update of the ﬁeld components indicated in Fig. 4.5, reveals that only the update of the
tangential electric ﬁeld component needs special handling. Referring to Fig. 4.5, which shows
two meshes interfaced at z = 0, the update of E
x
in mesh 1 is based on the retrieval of the
values of H
y
1
, H
y
2
, H
z
1
, H
z
2
. In particular, the index of H
y
1
in mesh 2 should be transparent to
mesh 1. For this purpose, the positions of the SBtype boundaries (between child meshes of the
same level) are recorded in a table, after each mesh reﬁnement (i.e., at each N
AMR
time steps).
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05804 MOBK058Sarris.cls December 15, 2006 7:49
THEDYNAMICALLY ADAPTIVEMESHREFINEMENT(AMR)FDTDTECHNIQUE 75
H
y1
H
y2
H
z2
H
z1
z
y
E
x
mesh1
mesh2
FIGURE4.5: Interface between two meshes of the same level
4.3.5 Field Updates at Junctions of SBType Boundaries
One speciﬁc difﬁculty arising in the update of an SBtype boundary comes from Tjunction
regions where three sibling meshes share one common edge, and crossjunction regions where
four sibling meshes share one common edge. Ideally, an SBtype boundary should be transparent
and meshes linked by SBtype boundaries (connecting meshes of the same resolution) should
e
b
c
d
E
A
C
D
B
a
FIGURE4.6: Typical SBtype boundaries. A, B, C, D, and E are meshes of the same level
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05804 MOBK058Sarris.cls December 15, 2006 7:49
76 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
behave like one larger mesh. Figure 4.6 shows some typical SBtype boundaries, including
junctions a, b, c , d, and e .
Astraightforwardsolutiontothe problemof handlingthese junctions, that indielectrically
homogeneous domains operates acceptably well, is to treat them as CPBtype boundaries.
This approach is employed in the microwave circuit examples of the Chapter 5. However, for
the analysis of structures such as the inhomogeneous waveguides of Chapter 6, the following
approach is proposed.
As shown in the ﬁgures, the update of an electric ﬁeld node (Enode) requires four
neighboring magnetic ﬁeld nodes (Hnodes), some of which may not belong to the same mesh.
Let us call those external Hnodes. It should be noted that an Enode on a boundary may be
shared by 2, 3, or 4 neighboring meshes of the same level. It only needs to be updated in one
mesh and then copied to all the other meshes. To avoid double updates, the following three
situations are considered:
1. If an Enode on an SBtype boundary is not at a corner of at least one neighboring
mesh, it should be updated in that mesh by using one external Hnode from the other
mesh, then copied to all other meshes. For example, in Fig. 4.6, the Enode a should
be updated in A then copied to B, b should be updated in B then copied to A, and d
should be updated in E then copied to A and D.
2. If an Enode on an SBtype boundary is at a corner of all the neighboring meshes, the
situation can be further divided in two subcases:
(a) There is at least one pair of meshes forming a crossjunction around the Enode.
For example, since meshes A and C form a crossjunction at the Enode c , c can
be updated in A by using two external Hnodes from C, or vice versa. Then, c is
copied to all the unupdated neighboring meshes.
(b) There is no pair of meshes forming a crossjunctionaround the Enode, for example,
Enode e . None of the neighboring meshes can obtain two external Hnodes,
therefore this Enode has to be treated as a CPBtype boundary and updated
accordingly.
4.3.6 Field Updates on PBType Boundaries
PBtype boundaries include absorbing and/or PEC boundary conditions. These conditions are
enforcedinbothroot andchildmeshes. For the applications that follow, Mur’s ﬁrstorder bound
ary condition [54] has been used. However, any other type of boundary condition can be readily
incorporated. Perfectly matched layer (PML)type conditions, would simply extend the com
putational domain by the number of the absorber cells. Since PMLs are terminated into PECs,
the only type of PB boundaries occurring in a PMLterminated domain are those of PECs.
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05804 MOBK058Sarris.cls December 15, 2006 7:49
THEDYNAMICALLY ADAPTIVEMESHREFINEMENT(AMR)FDTDTECHNIQUE 77
A
B
B
C
FIGURE4.7: Transition from old mesh tree to new mesh tree
4.3.7 Transition fromOld to NewMesh Tree
Each N
AMR
time steps of the root mesh, a new mesh tree is created, representing the adaptive
mesh regeneration. Field samples stored in the old mesh tree are transferred to the new mesh
tree. This is straightforward for the root mesh, since the new tree has the same root mesh as the
old tree. On the other hand, the possibility of an overlap between any child mesh of the newtree
and any child mesh of the old tree is checked. If there is such an overlap, the ﬁeld samples in the
overlapping region are transferred from the old to the new child mesh. For example, consider
the situation shown in Fig. 4.7, where A is the root mesh, B is an old child mesh, B
is a new
child mesh. B and B
overlap in region C. Fields of B within C will be copied to B
directly
and the rest of B
is initialized by interpolating ﬁelds of A. Note that if B
contains boundaries,
where conditions, such as Mur’s ﬁrstorder absorbing boundary condition, are applied, then
ﬁeld values of the current and the last time step are needed. These are maintained and kept
available, according to the proposed algorithm of Section 4.3.2. Again, trilinear interpolating
operations are employed to initialize the new mesh.
It is ﬁnally noted that source conditions are always enforced in the root mesh. If a child
mesh overlaps with the source region, the overlapping part of the source should also be enforced
in the child mesh.
4.4 ADAPTIVEMESHREFINEMENT
Up to this point, the features of AMRFDTDrelevant to the enforcement of a nonuniformgrid
and the implementation of multiple subgrids within a root mesh have been explained. Standard
interpolation operations, which are the common characteristic of any subgridding algorithm
have been proposed. What distinguishes AMRFDTD from previous subgridding approaches
is the adaptive mesh reﬁnement, which enables the adaptive movement of the subgrids. This
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05804 MOBK058Sarris.cls December 15, 2006 7:49
78 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
section explains the core of AMRFDTD, which is the detection of cells that need further
reﬁnement, the dilation of the detected region, to account for wave propagation during the
AMR interval and the clustering of the detected cells, to form child meshes.
4.4.1 Detection of Cells That Need Reﬁnement
Every N
AMR
steps, the following calculations are carried out. First, the energy of each (i, j, k)
cell of the root mesh is approximated as follows (assume that the computation is made at time
step n):
W
n
i, j,k
=
1
2
V
i, j,k
¸
E(r , nt)
2
+μ
H (r , nt)
2
¸
dv
≈
1
2
¸
i, j,k
E
n
i, j,k
2
+μ
i, j,k
H
n
i, j,k
2
¸
V
i, j,k
,
(4.4)
where V
i, j,k
is the volume of cell (i, j, k), W
n
i, j,k
is the electromagnetic energy in this cell at time
step n, and E
n
i, j,k
, H
n
i, j,k
are vector electric and magnetic ﬁeld values at the center of the cell
at time step n, which can be approximated by space/time averaging. Then, the gradient of the
energy is numerically approximated by a secondorder ﬁnitedifference expression, as:
∇W
n
i, j,k
=
W
n
i +1, j,k
− W
n
i −1, j,k
2x
ˆ x +
W
n
i, j +1,k
− W
n
i, j −1,k
2y
ˆ y
+
W
n
i, j,k+1
− W
n
i, j,k−1
2z
ˆ z.
(4.5)
Deﬁning thresholds θ
g
and θ
e
, a cell (i, j, k) is marked for reﬁnement if both of the following
criteria are met:
∇W
n
i, j,k
> θ
g
G
n
W
n
i, j,k
> θ
e
Q
n
,
(4.6)
where
G
n
= max
i, j,k
∇W
n
i, j,k
(4.7)
Q
n
= max
0≤m≤n
W
m
av
, (4.8)
W
m
av
=
1
N
x
N
y
N
z
¸
i, j,k
W
m
i, j,k
. (4.9)
Therefore, the adaptive mesh reﬁnement is executed in a cell when both an instantaneous and
a calculated over the whole simulated time threshold are exceeded. The ﬁrst criterion takes into
account the appearance of energy gradient peaks, while the second ensures that arbitrary ﬁeld
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05804 MOBK058Sarris.cls December 15, 2006 7:49
THEDYNAMICALLY ADAPTIVEMESHREFINEMENT(AMR)FDTDTECHNIQUE 79
ﬂuctuations, mainly stemming from numerical errors at the ﬁnal steps of the simulation, will
not create an unnecessary reﬁnement process.
On the other hand, as the AMRFDTD simulation begins, all the electric ﬁeld compo
nents assume zero values, except for the ones excited by the source. The detection of cells to be
reﬁned at this timemarching stage is difﬁcult, since the energy gradient is too small to surpass
the threshold set. To overcome this difﬁculty, in addition to the cells detected by thresholding,
the cells in the source region are also marked for reﬁnement, for a certain period of time. For
example, if a Gaussian excitation of the form:
exp
−
t −t
0
T
s
2
is used (with t
0
∼ 3T
s
), source region cells are reﬁned up to time t = 2t
0
.
For the multilevel case, the same algorithm is recursively applied at each mesh level.
Assuming again that N
L
is the maximum number of levels, the mesh regeneration (and, subse
quently, the update of the mesh tree) is pursued as follows:
1. For mesh level L varying from 1 to N
L
,
(a) The maximum average energy Q
n
L
and maximum gradient of energy G
n
L
for level L
is calculated.
(b) For each mesh of level L, the cells whose energy is greater than θ
e
Q
n
L
and gradient
of energy is greater than θ
g
G
n
L
are marked. If no cells are marked, we return to the
previous step. Otherwise, the marked cells are clustered into rectangular regions,
and for each rectangle, a level(L +1) mesh is created and added to the new mesh
tree.
2. For L varying from 2 to the number of levels of the new mesh tree, each levelL mesh
is initialized as follows: if it overlaps with any levelL mesh of the old mesh tree, the
electric and magnetic ﬁeld components are obtained from the overlapping region. For
the remainder of the mesh (which does not overlap with any other previous child mesh)
the electric and magnetic ﬁeld components are calculated by interpolating their values
from the corresponding parent mesh.
3. The previous mesh tree is then removed from memory.
The only mesh that is not subject to regeneration every N
AMR
time steps is the root mesh.
4.4.2 Modeling of Wave Propagation
The application of the previous criteria may result in a number of cells being marked for
reﬁnement. However, this reﬁnement process takes place every N
AMR
time steps (to avoid
loading every time step with the mesh reﬁnement operations). Within these time steps, wave
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05804 MOBK058Sarris.cls December 15, 2006 7:49
80 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
propagation within the computational domain can clearly generate the need for a denser mesh in
cells neighboring to the already reﬁned. To capture this ﬁeld movement, cells within a distance
D = σ
AMR
N
AMR
c t (4.10)
froma marked cell are also reﬁned, where c is the speed of light. The factor σ
AMR
is a predeﬁned,
greaterthanone positive real number, called the spreading factor. To facilitate computations,
the celltocell distances can be deﬁned in the sense of the Manhattan distance d(r
1
, r
2
) = x
1
−
x
2
 +y
1
− y
2
 +z
1
− z
2
 instead of the Euclidean distance d(r
1
, r
2
) = (x
1
− x
2
)
2
+(y
1
−
y
2
)
2
+(z
1
− z
2
)
2
. Physically, this distance D is the maximum distance that a wave can travel
within time equal to N
AMR
t, multiplied by the spreading factor. Note that no assumption is
being made as to the direction of the wave velocity, which in general is unknown.
The effect of the spreading factor, as well as the thresholds deﬁned in the previous section
will become evident in the numerical results of the next two chapters.
4.4.3 Clustering
Since the cells that are marked for reﬁnement would generally deﬁne irregularly shaped regions,
they are ﬁrst covered by a number of boxes, which are then divided in Yee cells. This procedure is
called clustering. To evaluate the quality of clustering, the box coverage efﬁciency is introduced,
deﬁned as the ratio of the total volume of the marked cells, covered by a box, to the volume of
the box.
For the implementation of this clustering procedure, the methodology proposed in [55]
is followed. There are three predeﬁned controlling parameters: the threshold for coverage efﬁ
ciency θ
c
, the minimum dimension of the box D
min
, and the maximum number of boxes N
max
.
At the beginning, the bounding box enclosing all the marked cells is found and its coverage
efﬁciency is calculated. If the coverage efﬁciency is greater than θ
c
or the dimension is less than
D
min
, the algorithm stops, otherwise the box is split in two boxes across a cut plane. Each of
the new boxes is shrunk to just cover the marked cells. Then, the coverage efﬁciency of each
box is calculated and compared with θ
c
. Again, either box with coverage efﬁciency less than θ
c
and dimension greater than D
min
will be split in two. This iterative process continues until the
maximum number of boxes is reached.
Figure. 4.8 illustrates the splitting and shrinking of box B. The black dots represent the
marked cells. Box A is the root mesh. Box B is split across the cut plane c c
and the resultant
two boxes are shrunk to obtain boxes C
1
and C
2
. The determination of the position of the cut
plane is detailed in [55].
4.4.4 Guidelines for Choosing AMRFDTDParameters
As discussed before, the accuracy and efﬁciency of AMRFDTD depends on ﬁve parameters;
θ
e
, θ
g
, θ
c
, σ
AMR
, and N
AMR
. Although a more systematic discussion of the dependence of
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05804 MOBK058Sarris.cls December 15, 2006 7:49
THEDYNAMICALLY ADAPTIVEMESHREFINEMENT(AMR)FDTDTECHNIQUE 81
C1
C2
A
c
c
B
A
split
and
shrink
FIGURE4.8: Splitting and shrinking of boxes (child meshes)
numerical errors of the algorithm on the values of these parameters is included in Section 6.6,
some empirical guidelines, qualitatively dictating their choice, can be provided here.
First, it should be noted that the best achievable accuracy by AMRFDTD is the one of
the FDTD scheme applied at a mesh as dense as the ﬁnest child mesh in the AMRFDTD
hierarchy of meshes. Therefore, the comments on the accuracy of the technique are meant to be
always referred to such an FDTD scheme. Let us also deﬁne the AMRcoverage as the ratio of
meshreﬁned regions to the total volume of the computational domain. In general, decreasing
the AMRcoverage will reduce the execution time of the code but will also reduce its accuracy,
since the mesh becomes coarser overall.
Furthermore, N
AMR
determines how frequently the AMR operations are performed. In
creasing this parameter leads to less AMRrelated operations. A subtle sideeffect of a large
N
AMR
is the following: newly generated meshreﬁned regions are always extended by a fac
tor proportional to N
AMR
(see (4.10)), to account for wave propagation between successive
mesh reﬁnements. Hence, the coverage increases and thus the execution time. In general, the
recommended values of N
AMR
are between 10 and 50.
The most important source of errors in any static or dynamic mesh reﬁnement scheme is
the reﬂections at a CPBtype boundary. The θ
e
parameter directly affects the wave amplitude
at such boundaries and as such it is the most important controlling parameter. On the other
hand, θ
g
ensures that the reﬁned regions enclose large energy gradient variations. In general,
θ
e
= 0.1 −0.5, θ
g
= 0.01 −0.02.
Finally, a value of σ
AMR
= 2 seems to yield satisfactory results in all cases, while the
coverage efﬁciency threshold θ
c
is chosen between 0.6 and 0.8. Note that a large value of
the latter enforces the generation of more smaller meshes, improving the AMRcoverage. At
the same time, it generates more CPB and SBtype boundaries, thus increasing the operations
related to their management.
These guidelines and the inherent tradeoffs in the choice of the parameters are further
illustrated in the numerical results of the next two chapters.
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05804 MOBK058Sarris.cls December 15, 2006 7:49
82 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
4.5 AMRFDTDANDMRTD: SIMILARITIES ANDDIFFERENCES
Essentially, the AMRFDTD technique belongs to the class of multiresolution timedomain
(MRTD) methods, although it does not implicate wavelet basis functions. However, it does
implement a multiresolution, space and time adaptive moving mesh in three dimensions, re
generated every certain time steps. However, the MRTD implementations, discussed in this
monograph, lack two important features of the AMRFDTDthat render the latter more ﬂexible
as a timedomain modeling tool:
1. While the adaptive MRTD operates on nonthresholded ﬁeld wavelet coefﬁcients, that
may be spatially and irregularly distributed, the AMRFDTD clusters the Yee cells
that need mesh reﬁnement and encloses them in rectangular subgrids. This leads to a
much more systematic mesh reﬁnement process.
2. In the MRTD implementations that are included in this monograph, as well as in the
vast majority of the MRTD literature, a uniform time step is set for the update of all
ﬁeld scaling and wavelet coefﬁcients, regardless of their resolution. On the other hand,
AMRFDTD employs asynchronous updates for the different mesh resolution levels.
However, both AMRFDTD features can be incorporated in the MRTD technique. An
important step toward this end is the work reported in [56]. In general, AMRFDTD has the
advantages of ﬂexibility and versatility and the disadvantages of only secondorder accuracy and
numerical dispersion that characterize FDTD. In fact, the realistic, two and threedimensional
applications presentedinthe next two chapters are aimedat highlighting the former two qualities
of the AMRFDTD. On the other hand, MRTD is a class of numerical techniques that can
achieve (with the use of higherorder basis functions, as opposed to the Haar basis) highorder
of accuracy, paying the price of complicated enforcement of source and boundary conditions.
Hence, despite its more recent introduction the dynamically AMRFDTDtechnique is closer to
being considered as a mature technique compared to MRTD, which is still related to important
and challenging questions when it comes to practical applications.
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05805 MOBK058Sarris.cls December 15, 2006 7:52
83
C H A P T E R 5
Dynamically Adaptive Mesh
Reﬁnement in FDTD: Microwave
Circuit Applications
The dynamically adaptive mesh reﬁnement FDTD technique of the previous chapter is now
applied to realistic microwave circuit geometries, namely a microstrip ﬁlter, a branch coupler,
and a spiral inductor. Through these applications, the salient properties of the technique along
with its excellent potential to dramatically reduce FDTD simulation times are shown. Trade
offs between speed and accuracy involved with the application of the algorithm to problems of
interest are also discussed.
5.1 INTRODUCTION
In this chapter, a number of microwave circuit analysis examples are presented, aimed at ad
dressing the following two questions that are naturally posed about the dynamically adaptive
mesh reﬁnement FDTDalgorithm. First, what is the effect of the various parameters, deﬁned in
the previous chapter on the performance and accuracy of the algorithm and second, whether the
overhead that AMRFDTD accumulates from the application of the mesh reﬁnement can still
leave some room for computational savings stemming from a reduced overall number of opera
tions. Both questions are negotiated through the application of the technique to three microwave
circuit geometries, namely, a microstrip lowpass ﬁlter, a branch coupler, and a spiral inductor.
In all cases, a maximum of two mesh levels is implemented (or a maximum mesh reﬁnement
factor of 8 for a Yee cell of the root mesh). In all these experiments, a twolevel AMRFDTD
is implemented, in order to facilitate the presentation of the effect of its parameters on accuracy
and execution time. All simulations were executed on an Intel Xeon 3.06 GHz machine.
5.2 MICROSTRIP LOWPASS FILTER
The ﬁrst example is a microstrip lowpass ﬁlter, shown in Fig. 5.1. The dimensions indicated in
the ﬁgure are A = 40 mm, B
1
= 2 mm, B
2
= 21 mm, W = 3 mm. The substrate thickness is
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05805 MOBK058Sarris.cls December 15, 2006 7:52
84 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
A
W
W
W
A
B B
P2
P1
A/2
1 2
FIGURE5.1: Microstrip lowpass ﬁlter geometry. P1 and P2 denote ports 1 and 2, respectively
0.8 mm and its dielectric constant is 2.2. The dimensions of the air box are 40 mm×40 mm×
4 mm. Two conventional FDTD simulations were performed for comparison: one using a
relatively coarse mesh (40 ×40 ×10 cells), and the second using a dense mesh (80 ×80 ×20
cells). AMRFDTD simulations use a 40 ×40 ×10 cell root mesh and different controlling
parameters for AMR. Both AMRFDTD and conventional FDTD use Mur’s ﬁrstorder
absorbing boundary condition [54]. A voltage source excitation is imposed at 3 mm from the
edges. In all simulations, a Courant number of 0.7 is used for determining the time step.
While AMRFDTD and the coarse FDTD technique are run for 4096 time steps, the dense
FDTD technique uses 8192 steps, since the time step t of the latter is equal to one half
the t of the former. The AMR parameters are: θ
g
= 0.02 , θ
e
= 0.1, θ
c
= 0.7, N
AMR
= 10,
σ
AMR
= 2.
To demonstrate the evolution of the child meshes over time, the ratio of the volume
of meshreﬁned areas (occupied by child meshes) to the total volume, referred to as AMR
coverage of the domain and the number of child meshes as a function of time steps are shown
in Fig. 5.2. In addition, Figs. 5.3–5.7 show the verticaltoground electric ﬁeld component E
z
and child meshes on the plane z = 0.4 mm at different time steps. The initial child mesh at
t = 0 is given in Fig. 5.3. As the wave propagates along the feed line, the coverage of reﬁne
ment increases, until about t = 250t. Between 250t and 700t, the coverage is relatively
large (40 to 60%) due to multiple reﬂections between the two open ends of the microstrip
line. As the ﬁelds impinge upon the absorbing boundaries of the structure, the ﬁeld values
in the working volume of the domain decrease. Consequently, the spatial ﬁeld variation be
comes smoother, which causes the AMRcoverage to decrease. In fact, after time t = 1000t,
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05805 MOBK058Sarris.cls December 15, 2006 7:52
DYNAMICALLY ADAPTIVEMESHREFINEMENTINFDTD 85
FIGURE5.2: AMR coverage and number of child meshes at different time steps for the lowpass ﬁlter
simulation
the coarse root mesh captures the ﬁeld solution sufﬁciently well, so that no child meshes are
necessary.
Figure 5.8 compares the vertical electrical ﬁeld at z = 0.4 mm and the center of the mi
crostrip line, 3 mmfromthe right edge, up to 40,000 time steps (32ns ). The timeseries deduced
via AMRFDTD cannot be distinguished from the one of the reference FDTD simulation,
FIGURE5.3: Vertical electric ﬁeldmagnitude at z = 0.4mmandt = 0. One childmesh; AMRcoverage
is 4.3%
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05805 MOBK058Sarris.cls December 15, 2006 7:52
86 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
FIGURE5.4: Vertical electric ﬁeld magnitude at z = 0.4 mm and t = 100t. One child mesh; AMR
coverage is 13.4%
whereas the result of FDTD using a coarser mesh has a signiﬁcant difference. The absence of
any latetime instability effects is also noted.
The calculated scattering (S) parameters and their differences are shown in Fig. 5.9. The
plot indicates the excellent approximation provided by AMRFDTD to the result of FDTD
using a dense mesh. This accuracy is quantiﬁed in Table ??, which employs the following
FIGURE5.5: Vertical electric ﬁeld magnitude at z = 0.4 mm and t = 200t. One child mesh; AMR
coverage is 52.5%
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05805 MOBK058Sarris.cls December 15, 2006 7:52
DYNAMICALLY ADAPTIVEMESHREFINEMENTINFDTD 87
FIGURE 5.6: Vertical electric ﬁeld magnitude at z = 0.4 mm and t = 500t. Three child meshes;
AMR coverage is 44.2%
Sparameter error metric:
E
S
=
k
l
m
S
l ,m
( f
k
) − S
ref
l ,m
( f
k
)
2
k
l
m
S
ref
l ,m
( f
k
)
2
, (5.1)
FIGURE 5.7: Vertical electric ﬁeld magnitude at z = 0.4 mm and t = 800t. Three child meshes;
AMR coverage is 28.0%
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05805 MOBK058Sarris.cls December 15, 2006 7:52
88 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
0 5 10 15 20 25 30 35
−0.2
0
0.2
0.4
0.6
0.8
1
Time (ns)
E
(
V
/
m
)
AMRFDTD
FDTD 40x40x10
FDTD 80x80x20
0.5 1 1.5
0
0.5
1
Time (ns)
E
(
V
/
m
)
AMRFDTD
FDTD 40x40x10
FDTD 80x80x20
z
z
FIGURE 5.8: Vertical electric ﬁeld magnitude at z = 0.4 mm and the center of the microstrip line,
3 mm from the right edge
where, f
k
is a discrete frequency within the modeled frequency band (up to 30 GHz), S
l ,m
the
(l , m)element of the scattering matrix of the simulated circuit, as determined by the AMR
FDTD or the coarsely meshed FDTD technique and S
ref
l ,m
is the same element, determined by
the densely meshed FDTD, which is used as a reference. From Table ??, it is concluded that
AMRFDTDcan closely followthe accuracy of the reference FDTDmethod, while consuming
only 5.4% of its total simulation time. These execution time savings are well above the savings
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05805 MOBK058Sarris.cls December 15, 2006 7:52
DYNAMICALLY ADAPTIVEMESHREFINEMENTINFDTD 89
0 5 10 15 20 25 30
−20
−15
−10
−5
0
Frequency (GHz)
A
m
p
l
i
t
u
d
e
o
f
S
1
1
(
d
B
)
AMRFDTD
FDTD 40x40x10
FDTD 80x80x20
0 5 10 15 20 25 30
0
0.1
0.2
0.3
0.4
0.5
Frequency (GHz)
A
b
s
o
l
u
t
e
v
a
l
u
e
o
f
e
r
r
o
r
o
f
S
1
1
AMRFDTD
FDTD 40x40x10
0 5 10 15 20 25 30
−40
−35
−30
−25
−20
−15
−10
−5
0
Frequency (GHz)
A
m
p
l
i
t
u
d
e
o
f
S
2
1
(
d
B
)
AMRFDTD
FDTD 40x40x10
FDTD 80x80x20
0 5 10 15 20 25 30
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
Frequency (GHz)
A
b
s
o
l
u
t
e
v
a
l
u
e
o
f
e
r
r
o
r
o
f
S
2
1
AMRFDTD
FDTD 40x40x10
FIGURE 5.9: Comparison of the scatteringparameters of the lowpassﬁlter of Fig. 5.1, obtained by
FDTD and AMRFDTD
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05805 MOBK058Sarris.cls December 15, 2006 7:52
90 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
TABLE 5.1: Computation Time and SParameter Error Metric E
S
for the Microstrip LowPass
Filter
TIME TOTAL E
s
METHOD MESH STEPS TIME(S) (%)
FDTD 40 ×40 ×10 4096 75 37.3
AMRFDTD 40 ×40 ×10 4096 518 2.4
FDTD 80 ×80 ×20 8192 9669 —
expected from static subgridding algorithms, previously reported in the literature and indicate
the potential of AMRFDTD.
It should be noted that the ratio between the FDTD simulation times for the ﬁne and
the coarse mesh is larger than 16:1, which is the ratio between the operations carried out in the
two cases. A careful study can show that for small meshes the FDTD simulation times in dense
and coarse meshes would tend to follow the 16:1 rule. However, once the mesh size exceeds
a certain limit, the simulation time increases faster than 16:1. This is due to the increase in
the memory access time for large meshes, which further extends their simulation time. For a
thorough investigation of memorycache related effects in the execution time of timediscrete
methods the reader is referred to [57].
The effect of the AMRFDTD controlling parameters on the simulation time and er
ror of this technique is studied next and results are shown in Figs. ??–??. As expected, the
decrease in θ
g
and θ
e
results in lower errors and longer simulation times. Essentially, as these
two thresholds are lowered, the AMRFDTD tends to become equivalent in operation and
FIGURE 5.10: Effect of threshold for gradient θ
g
. The rest of the AMR parameters are: θ
e
= 0.1,
θ
c
= 0.7, N
AMR
= 10, σ
AMR
= 2
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05805 MOBK058Sarris.cls December 15, 2006 7:52
DYNAMICALLY ADAPTIVEMESHREFINEMENTINFDTD 91
FIGURE 5.11: Effect of threshold for energy θ
e
. The rest of the AMR parameters are: θ
g
= 0.01,
θ
c
= 0.7, N
AMR
= 10, σ
AMR
= 2
FIGURE 5.12: Effect of threshold for coverage efﬁciency θ
c
. The rest of the AMR parameters are:
θ
g
= 0.01, θ
e
= 0.01, N
AMR
= 10, σ
AMR
= 2
FIGURE 5.13: Effect of interval of AMR N
AMR
. The rest of the AMR parameters are: θ
g
= 0.01,
θ
e
= 0.01, θ
c
= 0.7, σ
AMR
= 2
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05805 MOBK058Sarris.cls December 15, 2006 7:52
92 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
FIGURE 5.14: Effect of spreading factor of AMR σ
AMR
. The rest of the AMR parameters are: θ
g
=
0.01, θ
e
= 0.01, θ
c
= 0.7, N
AMR
= 10
performance to the dense FDTD method. On the contrary, increasing these thresholds can
reduce the overall computation time, without sacriﬁcing accuracy, up to some point. It should
be noted that the percentage error does not decrease monotonically as θ
g
and θ
e
decrease, because
certain child meshes generated by using smaller θ
g
or θ
e
may have larger reﬂections due to the
irregularity of the shape of the reﬁned regions. This problem can be alleviated by using higher
order interpolation schemes at the mesh boundaries to reduce the reﬂections [44–58]. It is also
noted that there is a sudden increase in the simulation time when θ
e
is below 0.01 without any
associated improvement in accuracy. At the late stage of the simulation, the ﬁeld components
assume some small values due to the reﬂections at the CPB andSBtype boundaries. If θ
e
is very
small, a large number of child meshes can be generated as a result. Subsequently, the simulation
time increases, without any improvement in accuracy. An appropriate choice of θ
e
can essentially
eliminate this problem. Also, when θ
c
reaches 0.8, both the error and the computation time
increase. The reason is that the numerical error triggers the automatic generation of multiple
spurious child meshes that are clustered independently. As a result, their management by the
algorithm adds an overhead that completely eliminates any savings due to the mesh reﬁnement.
Tradeoff effects related to the choice of N
AMR
and σ
AMR
are also evident in Figs. ?? and ??.
The appropriate values for N
AMR
and σ
AMR
, indicated by these plots, have been used in all the
aforementioned numerical experiments. The choice of the AMRparameters is further discussed
in this and the next chapter.
5.3 MICROSTRIP BRANCHCOUPLER
The microstrip branch coupler geometry of Fig. ?? is analyzed next. The geometric parameters
indicated in the ﬁgure are as follows: A = 40 mm, w
1
= 2 mm, w
2
= 3 mm, B
1
= 7 mm,
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05805 MOBK058Sarris.cls December 15, 2006 7:52
DYNAMICALLY ADAPTIVEMESHREFINEMENTINFDTD 93
A
A
w
w
B
w
B
P2 P4
P1 P3
B
2
2
3
1
1
1
FIGURE5.15: Microstrip branch coupler geometry. P1, P2, P3, and P4 denote port 1, 2, 3, and 4
B
2
= 11 mm, B
3
= 9 mm. The thickness of the substrate is 0.8 mmand its dielectric constant is
2.2. The dimensions of the computational domainenclosing the structure are 40 mm×40 mm×
4 mm. The AMRFDTD method uses a 40 ×40 ×10 mesh and takes 4096 time steps. A
reference FDTD simulation of an 80 ×80 ×20mesh is used for comparison. The AMR
controlling parameters are θ
g
= 0.01, θ
e
= 0.1, θ
c
= 0.7, N
AMR
= 10, σ
AMR
= 2. Figures ??–
?? show the Sparameters and their errors, as determined by the AMRFDTD, plain FDTD
executed on the AMR root mesh, and the reference FDTD simulation. Comparisons regarding
computation times and numerical errors of the AMRFDTD are included in Table ??. The
data demonstrate the capability of AMRFDTD to deduce the dense FDTD results at a
greatly reduced computational cost, which is reﬂected on an execution time reduction by a
factor of 20. Timedomain ﬁeld waveforms up to 40,000 time steps are shown in Fig. ??. Again,
the AMRFDTD and densemesh FDTD waveforms coincide, without any sign of latetime
instability.
5.4 MICROSTRIP SPIRAL INDUCTOR
As a last example, the geometry of a spiral inductor of Fig. ?? is analyzed. The parameters
of this geometry are: A
1
= 60 mm, A
2
= 40 mm, w
1
= w
2
= 2 mm, B
1
= 24 mm, B
2
=
20 mm, B
3
= 18 mm, B
4
= 4 mm. The thickness of the substrate is 0.8 mm and its dielectric
constant is 2.2. The dimensions of the computational domain enclosing the structure are 60
mm × 40 mm × 4 mm. The air bridge is 0.8 mm above the substrate. The AMRFDTD
method uses a 60 ×40 ×10 mesh and 8192 time steps. A reference FDTD simulation of a
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05805 MOBK058Sarris.cls December 15, 2006 7:52
94 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
0 5 10 15 20 25 30
−30
−25
−20
−15
−10
−5
0
Frequency (GHz)
A
m
p
l
i
t
u
d
e
o
f
S
1
1
(
d
B
)
AMRFDTD
FDTD 40x40x10
FDTD 80x80x20
0 5 10 15 20 25 30
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
Frequency (GHz)
A
b
s
o
l
u
t
e
v
a
l
u
e
o
f
e
r
r
o
r
o
f
S
2
1
AMRFDTD
FDTD 40x40x10
FIGURE5.16: S
11
for the microstrip branch coupler geometry of Fig. ??, as determined by FDTD and
AMRFDTD
120 ×80 ×20 mesh is used for comparison. The AMR controlling parameters are θ
g
= 0.001,
θ
e
= 0.1, θ
c
= 0.6, N
AMR
= 50, σ
AMR
= 2. The mesh reﬁnement process is illustrated in Fig.
??, which shows the effective vertical electric ﬁeld wavefront tracking achieved by the algorithm.
Moreover, Fig. ?? shows the Sparameters and their errors, as determined by the three methods
under comparison, while error and execution time data are shown in Table ??. Timedomain
results are also shown in Fig. ??.
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05805 MOBK058Sarris.cls December 15, 2006 7:52
DYNAMICALLY ADAPTIVEMESHREFINEMENTINFDTD 95
0 5 10 15 20 25 30
−40
−35
−30
−25
−20
−15
−10
−5
0
Frequency (GHz)
A
m
p
l
i
t
u
d
e
o
f
S
2
1
(
d
B
)
AMRFDTD
FDTD 40x40x10
FDTD 80x80x20
0 5 10 15 20 25 30
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
Frequency (GHz)
A
b
s
o
l
u
t
e
v
a
l
u
e
o
f
e
r
r
o
r
o
f
S
2
1
AMRFDTD
FDTD 40x40x10
FIGURE5.17: S
21
for the microstrip branch coupler geometry of Fig. ??, as determined by FDTD and
AMRFDTD
Compared to the lowpass ﬁlter and the branch coupler, the execution time savings
achieved by AMRFDTD over the conventional FDTD are smaller (yet large, of about 80%),
while the associated errors are larger. These effects stem from the highly resonant nature of
the spiral inductor, which necessitates the use of a signiﬁcant number of time steps for the
extraction of the Sparameters. At late stages of the simulation, relatively small ﬁeld values are
easily contaminated by spurious reﬂections at the parent–child mesh interfaces. Note that in
this case, the error of the coarse mesh FDTD becomes excessively large as well.
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05805 MOBK058Sarris.cls December 15, 2006 7:52
96 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
0 5 10 15 20 25 30
−30
−25
−20
−15
−10
−5
0
Frequency (GHz)
A
m
p
l
i
t
u
d
e
o
f
S
3
1
(
d
B
)
AMRFDTD
FDTD 40x40x10
FDTD 80x80x20
0 5 10 15 20 25 30
0
0.05
0.1
0.15
0.2
0.25
Frequency (GHz)
A
b
s
o
l
u
t
e
v
a
l
u
e
o
f
e
r
r
o
r
o
f
S
3
1
AMRFDTD
FDTD 40x40x10
FIGURE5.18: S
31
for the microstrip branch coupler geometry of Fig. ??, as determined by FDTD and
AMRFDTD
5.5 DISCUSSION: STABILITY ANDACCURACY
OF AMRFDTDRESULTS
Based on the previous examples, two further comments are in order. First, the timedomain
results accompanying the three numerical experiments (Figs. 5.8, ??, ??), demonstrate the ab
sence of latetime instability in the AMRFDTD. In fact, the convergence of the number of
AMRFDTD child meshes to zero over time, implies that only the root mesh is still present
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05805 MOBK058Sarris.cls December 15, 2006 7:52
0 5 10 15 20 25 30
−10
−8
−6
−4
−2
0
Frequency (GHz)
A
m
p
l
i
t
u
d
e
o
f
S
4
1
(
d
B
)
AMRFDTD
FDTD 40x40x10
FDTD 80x80x20
0 5 10 15 20 25 30
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Frequency (GHz)
A
b
s
o
l
u
t
e
v
a
l
u
e
o
f
e
r
r
o
r
o
f
S
4
1
AMRFDTD
FDTD 40x40x10
FIGURE5.19: S
41
for the microstrip branch coupler geometry of Fig. ??, as determined by FDTD and
AMRFDTD
TABLE 5.2: Computation Time and SParameter Error Metric E
S
for the Microstrip Branch
Coupler
TIME TOTAL E
s
METHOD MESH STEPS TIME(S) (%)
FDTD 40 ×40 ×10 4096 262 34.6
AMRFDTD 40 ×40 ×10 4096 1021 2.43
FDTD 80 ×80 ×20 8192 19765 —
97
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05805 MOBK058Sarris.cls December 15, 2006 7:52
98 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
0 5 10 15 20 25 30 35
−0.2
0
0.2
0.4
0.6
0.8
1
Time (ns)
E
(
V
/
m
)
AMRFDTD
FDTD 40x40x10
FDTD 80x80x20
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
−0.2
0
0.2
0.4
0.6
0.8
Time (ns)
AMRFDTD
FDTD 40x40x10
FDTD 80x80x20
z
E
(
V
/
m
)
z
FIGURE5.20: Vertical electric ﬁeld magnitude at z = 0.6 mm and the center of the microstrip line of
port 3, 6 mm from the edge. The excitation is imposed at 3 mm from the edge of port 1
A
w
B
B
B
B
P1 P2 A
w
microstrip
via
air bridge
4
3
2
2
2
1
1
1
FIGURE5.21: Spiral inductor geometry. P1 and P2 denote port 1 and port 2, respectively
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05805 MOBK058Sarris.cls December 15, 2006 7:52
DYNAMICALLY ADAPTIVEMESHREFINEMENTINFDTD 99
FIGURE5.22: Vertical ﬁeld evolution and associated mesh reﬁnement in the microstrip spiral inductor,
simulated by a twolevel dynamic AMRFDTD
at a late stage of the code. Therefore, no spatial or temporal interpolation operations, which are
the primary sources of instabilities in adaptive mesh FDTD codes [44], are applied then. This
is an additional advantage of using a dynamically adaptive instead of a statically adaptive mesh
in timedomain simulations.
Another aspect of the AMRFDTD accuracy is associated with the reﬂections at
dense/coarse mesh interfaces. A pulse propagating at a statically reﬁned mesh will be reﬂected
off such an interface, creating numerical errors. Note that in AMRFDTD, such a pulse would
always be enclosed in a dense mesh, while its retroreﬂections might encounter dense/coarse grid
TABLE 5.3: Computation Time and SParameter Error Metric E
S
for the Microstrip Spiral
Inductor
TIME TOTAL E
s
METHOD MESH STEPS TIME(S) (%)
FDTD 60 ×40 ×10 8192 411 52.1
AMRFDTD 60 ×40 ×10 8192 5276 3.74
FDTD 120 ×80 ×20 16384 29154 —
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05805 MOBK058Sarris.cls December 15, 2006 7:52
100 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
0 5 10 15 20 25 30
−30
−25
−20
−15
−10
−5
0
Frequency (GHz)
A
m
p
l
i
t
u
d
e
o
f
S
1
1
(
d
B
)
AMRFDTD
FDTD 60x40x10
FDTD 120x80x20
0 5 10 15 20 25 30
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Frequency (GHz)
A
b
s
o
l
u
t
e
v
a
l
u
e
o
f
e
r
r
o
r
o
f
S
2
1
AMRFDTD
FDTD 40x40x10
0 5 10 15 20 25 30
−40
−35
−30
−25
−20
−15
−10
−5
0
Frequency (GHz)
A
m
p
l
i
t
u
d
e
o
f
S
2
1
(
d
B
)
AMRFDTD
FDTD 60x40x10
FDTD 120x80x20
0 5 10 15 20 25 30
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
Frequency (GHz)
A
b
s
o
l
u
t
e
v
a
l
u
e
o
f
e
r
r
o
r
o
f
S
2
1
AMRFDTD
FDTD 40x40x10
FIGURE5.23: Sparameters for the spiral inductor geometry of Fig. ??, as determined by FDTD and
AMRFDTD
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05805 MOBK058Sarris.cls December 15, 2006 7:52
DYNAMICALLY ADAPTIVEMESHREFINEMENTINFDTD 101
0 5 10 15 20 25 30 35
−0.2
−0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Time (ns)
E
(
V
/
m
)
AMRFDTD
FDTD 60x40x10
FDTD 120x80x20
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
−0.2
−0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Time (ns)
E
(
V
/
m
)
AMRFDTD
FDTD 60x40x10
FDTD 120x80x20
z
z
FIGURE 5.24: Vertical electric ﬁeld magnitude at z = 0.6 mm and the center of the microstrip line,
6 mm from the right edge. The excitation is imposed at 3 mm from the left edge of the microstrip line
interfaces before the AMR algorithm creates new meshes for them. This latter case pro
duces errors in AMRFDTD, which are evidently smaller than those arising in a static
subgrid.
However, the application of interpolations in space and time generates a degradation in
the FDTD stability factor. This wellknown effect, also discussed in [44], can be alleviated by
using higherorder interpolation schemes. For the simple trilinear interpolations employed in
these computations, the AMRFDTD stability limit is observed to be 0.9 of the corresponding
FDTD one in twodimensional cases, and 0.85 of the corresponding FDTD one in three
dimensional cases.
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05805 MOBK058Sarris.cls December 15, 2006 7:52
102 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
Finally, the comparison between coarse/dense FDTD results and AMRFDTD results
in the frequency domain, reveals the standard pattern of the results being in a relatively good
agreement with each other up to the middle of the simulated frequency band and diverging
afterwards. In these simulations, the coarse grid Yee cell size is about λ
min
/7.5 is each dimension.
Therefore, up to the frequency f
max
/2, the socalled coarse mesh uses a sampling rate of at least
λ( f
max
/2)/15, which is dense enough to determine the Sparameters accurately. From that
point on, the effect of the FDTD numerical dispersion becomes more severe, leading to the
large errors shown in the ﬁgures.
5.6 CONCLUSION
In this chapter, the dynamically AMRFDTD technique was applied to realistic microwave
circuit and optical waveguide geometries. The purpose of the applications that were shown,
was to demonstrate whether the mesh adaptation overhead of AMRFDTD can still allow
for important computational savings. The conclusion is that it does, because the method is
optimally suited to the nature of timedomain simulations. The latter are characterized by
spatially and temporally localized phenomena that call for a dense mesh not throughout space
and time, but only at the certain time and space they happen. Although microwavecircuit
examples were presented, the technique evidently holds a great promise for largescale optical
structure modeling. This direction is further investigated in the next chapter.
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05806 MOBK058Sarris.cls December 15, 2006 7:53
103
C H A P T E R 6
Dynamically Adaptive Mesh
Reﬁnement in FDTD: Optical
Applications and Error Estimates
After presenting microwave circuit applications, the dynamically adaptive mesh reﬁnement
FDTD is employed for the modeling of twodimensional optical waveguide structures,
including power splitters, junctions, and ring resonators. The conventional FDTD often
requires extremely long simulation times for the characterization of this type of problems;
on the contrary, the use of an adaptively moving mesh is shown to result in dramatically
accelerated simulations. In all the examples that follow, a multilevel adaptive mesh reﬁnement
is pursued. Hence, while in the previous chapter, a maximum to minimum Yee cell edge ratio
of two was used, in the following examples this ratio goes up to 64. The use of multiple levels
also allows us to compare the efﬁciency of the mesh reﬁnement algorithm as a function of the
number of levels used. The chapter is concluded by extracting error estimates that allow for an
educated choice of the AMR parameters.
6.1 MULTILEVEL AMRFDTD
To demonstrate the application of the AMRFDTDalgorithmwith multiple resolution levels, a
2DTEmode (with an E
z
electric ﬁeld component only) optical waveguide (shown in Fig. 6.1)
is simulated [59]. Its width is 0.3 μm and its dielectric constant is 10.24. The computational
domain is 6 μm×6 μm and a 1 μm thick matched absorber is used to truncate it. Figure 6.2
shows a snapshot of the electric ﬁeld obtained by AMRFDTD with a maximum number of
levels equal to 4. The excitation is imposed at 1 μm from the left edge of the waveguide and
the electric ﬁeld is recorded at the center of waveguide and 1.5 μm from the right edge. The
excitation is a modulated Gaussian pulse of the form:
exp
−
y − y
0
W
2
−
t −t
0
T
s
2
sin (2π f t) ,
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05806 MOBK058Sarris.cls December 15, 2006 7:53
104 MESH REFINEMENTFORTIMEDOMAIN NUMERICAL ELECTROMAGNETICS
y
absorber
waveguide
6 μm
6 μm 0.3 μm
1 μm
x
FIGURE6.1: Optical waveguide geometry
where T
s
= 0.02 ps, t
0
= 3T
s
, f = 200 THz, W = 0.7 μm, y
0
= 3 μm. Table 6.1 compares the
accuracy and computation time of AMRFDTD using a root mesh of 120 ×120 and 2–4 mesh
levels, to a reference FDTD simulation using a 1920 ×1920 mesh and several coarser FDTD
schemes. The accuracy is quantiﬁed by employing the following timedomain error metric:
E
t
=
¸
k
f (t
k
) − f
ref
(t
k
)
2
¸
k
f
ref
(t
k
)
2
, (6.1)
where t
k
is discrete time within the modeled time range (up to 16 ps), f is the sampled
electric ﬁeld as determined by AMRFDTD or the coarse mesh FDTD techniques, and f
ref
is the sampled electric ﬁeld determined by the reference FDTD simulation. All AMRFDTD
FIGURE 6.2: Vertical electric ﬁeld magnitude distribution across the optical waveguide of Fig. 6.1, at
t = 100t. One level2 child mesh and one level3 child mesh are shown
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05806 MOBK058Sarris.cls December 15, 2006 7:53
DYNAMICALLY ADAPTIVE MESHREFINEMENTIN FDTD 105
TABLE 6.1: Computation Time and TimeDomain Error Metric E
t
for the Optical Waveguide
of Fig. 6.1
NUMBER TIME TOTAL E
t
METHOD MESH OF LEVELS STEPS TIME(S) (%)
FDTD 120 ×120 — 2000 6.4 74.9
FDTD 240 ×240 — 4000 66.7 17.2
FDTD 480 ×480 — 8000 635 3.64
FDTD 960 ×960 — 16000 13794 0.60
FDTD 1920 ×1920 — 32000 181997 —
AMRFDTD 120 ×120 2 2000 31.4 17.0
AMRFDTD 120 ×120 3 2000 133 3.57
AMRFDTD 120 ×120 4 2000 641 0.56
simulations use a reﬁnement factor of 2 for successive levels, and θ
e
= 0.1, θ
g
= 0, θ
c
= 0.8,
N
AMR
= 10, σ
AMR
= 2. Figure 6.2 shows that AMRFDTD can reach the accuracy of the
reference FDTD method in a greatly reduced computation time. The time domain simulation
results of the AMRFDTD with four levels and the reference FDTD code are compared in
Fig. 6.3. The AMRFDTDwaveformis not visually discernible fromthe reference FDTDone.
0.4 0.6 0.8 1 1.2 1.4 1.6
x 10
−4
−4
−3
−2
−1
0
1
2
3
4
5
Time (ns)
E
(
V
/
m
)
AMRFDTD
FDTD 120x120
FDTD 1920x1920
z
FIGURE 6.3: Temporal waveform of the vertical electric ﬁeld magnitude at the center of the optical
waveguide of Fig. 6.1, 1.5 μm from the right edge
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05806 MOBK058Sarris.cls December 15, 2006 7:53
106 MESH REFINEMENTFORTIMEDOMAIN NUMERICAL ELECTROMAGNETICS
6.2 DIELECTRIC WAVEGUIDEWITHACORRUGATED
PERMITTIVITY PROFILE
A dielectric waveguide with a corrugated permittivity proﬁle, similar to the one presented in
[60], is simulated by FDTD and a multilevel dynamic AMRFDTD [59]. This geometry is
a good benchmark application for the AMRFDTD technique, since the multiple reﬂections
created by the dielectric corrugations enforce strong interactions between spatially distributed
meshes, testing the stability and accuracy of the mesh reﬁnement scheme. All interpolations
employed are ﬁrstorder polynomial ones.
The geometry of the waveguide is shown in Fig. 6.4. Its width is 2 μm and the dielectric
constant of the host medium is 9. The dielectric constant of the corrugations is 7.9976. The
computational domain is 12 μm×8 μm and is truncated by a 1 μm thick matched absorber.
The excitation is imposed 1 μm from the left edge of the waveguide and the electric ﬁeld is
recorded at the center of the waveguide and 1.2 μm from the right edge. The excitation is a
modulated pulse of the form:
(u(t) −u(t − T)) sin
2
(πt
/
T) sin (2π f t) e
−(y−y
0
)
2
/W
2
, (6.2)
where u(t) is the unit step function, f = 193 THz, T = 5/f , W = 0.7 μm, and h = 2y
0
is
the height of the waveguide (hence, the pulse is centered in the middle of the guide, along
the ydirection). Table 6.2 compares the accuracy and computation time of AMRFDTD
using a root mesh of 240 ×160 and 2–4 mesh levels, to a reference FDTD simulation using
a 3840 ×2560 mesh and several coarser FDTD schemes. The simulated timewindow in all
methods remains constant. Therefore, if one method has a time step t and another t/2,
the latter will be run for twice the total number of time steps of the former. The accuracy is
quantiﬁed by employing the timedomain waveformerror metric of Eq. (6.1). All AMRFDTD
1 1 1 1 1
y
x absorber
12
2
1
3
P2 8 P1
FIGURE6.4: Geometry of the dielectric waveguide with a corrugated permittivity proﬁle. Dimensions
are given in units of μm. P1 and P2 denote ports 1 and 2, respectively
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05806 MOBK058Sarris.cls December 15, 2006 7:53
DYNAMICALLY ADAPTIVE MESHREFINEMENTIN FDTD 107
TABLE6.2: Computation Time and TimeDomain Error Metric E
t
for the Dielectric Waveguide
of Fig. 6.4
NUMBER TIME TOTAL E
t
METHOD MESH OF LEVELS STEPS TIME(S) (%)
FDTD 240 ×160 — 2000 20.07 157.1
FDTD 480 ×320 — 4000 183.9 41.89
FDTD 960 ×640 — 8000 3544 8.54
FDTD 1920 ×1280 — 16000 50230 0.71
FDTD 3840 ×2560 — 32000 559714 —
AMRFDTD 240 ×160 2 2000 91.8 41.2
AMRFDTD 240 ×160 3 2000 246 8.43
AMRFDTD 240 ×160 4 2000 1253 0.91
AMRFDTD 240 ×160 5 2000 11337 1.39
simulations use a reﬁnement factor N
s
of 2 for successive levels, and θ
e
= 0.01, θ
g
= 0, θ
c
= 0.8,
N
AMR
= 10, σ
AMR
= 2. Table 6.2 shows that the dynamic AMRFDTDcan reach the accuracy
of the reference FDTD method in a greatly reduced computation time. It should be noted that
the error no longer decreases once the number of levels of the mesh tree reaches 5, while
the execution time also starts increasing then. This trend continues as the number of levels
increases and reﬂects a saturation point beyond which the beneﬁts of the AMR algorithm are
compensated for by the errors accumulated through the interpolating operations, as well as
the time for their execution. These two problems are interconnected, since the reduction of
interpolation errors can be achieved by applying a higherorder polynomial interpolation [44],
which in turn increases the operations needed. Therefore, the conclusion of this study is that
four levels, combined with ﬁrstorder interpolations can achieve an impressive reduction in the
execution time (by a factor of 446.7), with less than 1% error in the timedomain.
Figure 6.5 compares the timedomain waveforms of AMR and conventional FDTD
schemes, while Fig. 6.6 demonstrates the latetime stability of the multilevel dynamic AMR
FDTD. In both ﬁgures, the AMRFDTD uses four levels and a 240 ×160 mesh. The result
of Fig. 6.6 can be explained by the following note regarding the stability of the dynamic AMR
FDTD, which is also its fundamental difference from static subgridding schemes. Assume
a pulse propagating through a static interface between a coarse and a dense grid. Then, the
interpolating operations needed for the update equations would involve large ﬁeld values. As a
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05806 MOBK058Sarris.cls December 15, 2006 7:53
108 MESH REFINEMENTFORTIMEDOMAIN NUMERICAL ELECTROMAGNETICS
0.95 1 1.05 1.1 1.15 1.2 1.25 1.3 1.35 1.4 1.45
x 10
−4
−3
−2
−1
0
1
2
3
Time (ns)
E
(
V
/
m
)
AMRFDTD
FDTD 240x160
FDTD 3840x2560
z
FIGURE6.5: Electric ﬁeld at the center of the dielectric waveguide of Fig. 6.4, 1.2 μm from the right
edge
result, the associated errors are large, contributing to the wellknown latetime instability effects
in static subgridding schemes. On the other hand, such an interface would not exist under the
dynamic AMR scheme, the reason being that this algorithm tracks the propagating wavefronts.
Therefore, the pulse of this thought example would be enclosed in a moving dense mesh.
0 0.5 1 1.5 2 2.5 3 3.5
x 10
−3
−3
−2
−1
0
1
2
3
Time (ns)
E
(
V
/
m
)
z
FIGURE6.6: Electric ﬁeld at the center of the dielectric waveguide of Fig. 6.4, 1.2 μm from the right
edge, obtained with the dynamic AMRFDTD, for 40,000 time steps, indicating the absence of latetime
instability
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05806 MOBK058Sarris.cls December 15, 2006 7:53
DYNAMICALLY ADAPTIVE MESHREFINEMENTIN FDTD 109
6.3 DIELECTRIC WAVEGUIDEPOWERSPLITTER
To demonstrate the efﬁciency of the multilevel AMRFDTD for the simulation of structures
with nonconformal dielectric constant distribution, a dielectric waveguide power splitter, shown
in Fig. 6.7 [60] is simulated [59]. The dielectric constant of the crossshaped waveguide is 9,
whereas the dielectric constant of the slanted splitter (lens) is 4. It should be noted that due to the
slanted orientation of the lens, its dielectric constant proﬁle does not conform to a rectangular
mesh. In FDTD simulations using different meshes, the dielectric constant distribution is
discretized independently, according to an optimal staircase approximation. The multilevel
AMRFDTDsimilarly rediscretizes the dielectric constant distribution at different levels. The
simulations use the same excitation as (6.2), which is imposed at (1.2 μm, 9 μm), and the electric
ﬁeld is recorded at 1 μm from each port. All AMRFDTD simulations use a reﬁnement factor
of 2 for successive levels, and θ
e
= 0.01, θ
g
= 0, θ
c
= 0.8, N
AMR
= 10, σ
AMR
= 2. Table 6.3
shows that AMRFDTD can reach the accuracy of the reference FDTD method, in a much
smaller execution time. In addition, the timedomain results of Figs. 6.8–6.10 demonstrate the
latetime stability and accuracy of the method.
Figures 6.11–6.13 showthe evolution of the ﬁeld and the mesh tree, which uses a 60 ×60
root mesh and a maximum number of three levels. The level1 and level2 meshes are drawn
in the ﬁgures, while the regions without grid lines are covered by level3 meshes (not drawn
for visualization purposes). From these ﬁgures, the effective wavefront tracking achieved by
AMRFDTD is demonstrated.
y
x
absorber
waveguide P1
P2
P3
P4
18
18
0.5
2
2
1
l
e
n
s
FIGURE 6.7: Geometry of the dielectric waveguide power splitter. Dimensions are given in units of
μm. P1, P2, P3, and P4 denote ports 1, 2, 3, and 4, respectively
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05806 MOBK058Sarris.cls December 15, 2006 7:53
110 MESH REFINEMENTFORTIMEDOMAIN NUMERICAL ELECTROMAGNETICS
TABLE6.3: Computation Time and TimeDomain Error Metric E
t
for the Dielectric Waveguide
Power Splitter for Fields Recorded at Port 2
NUMBER TIME TOTAL E
t
METHOD MESH OF LEVELS STEPS TIME(S) (%)
FDTD 360 ×360 — 4000 148.92 168
FDTD 720 ×720 — 8000 2313 41.0
FDTD 1440 ×1440 — 16000 39181 2.85
FDTD 2880 ×2880 — 32000 447497 —
AMRFDTD 360 ×360 2 4000 394.6 41.0
AMRFDTD 360 ×360 3 4000 816.6 2.98
AMRFDTD 360 ×360 4 4000 3683 1.07
6.4 DIELECTRIC WAVEGUIDEYJUNCTION
Adielectric waveguide Yjunction, shown in Fig. 6.14 is simulated [59]. The dielectric constant
of the waveguide is 9 and the dielectric constant of the surrounding medium is 1. The compu
tational domain is 25 μm×10 μm and a 1 μm thick matched absorber is used to truncate it.
The excitation is imposed at 1.2 μm from the edge of port 1 and the electric ﬁeld is recorded at
the center of the waveguide, 2 μm from the edge of port 2. The excitation is a modulated pulse
1.4 1.6 1.8 2 2.2 2.4
x 10
−4
−3
−2
−1
0
1
2
3
Time (ns)
E
(
V
/
m
)
AMRFDTD
FDTD 360x360
FDTD 2880x2880
z
FIGURE6.8: Electric ﬁeld at 1 μm from the edge of port 2 of the power splitter of Fig. 6.7
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05806 MOBK058Sarris.cls December 15, 2006 7:53
DYNAMICALLY ADAPTIVE MESHREFINEMENTIN FDTD 111
1.4 1.6 1.8 2 2.2 2.4
x 10
−4
−1
−0.5
0
0.5
1
Time (ns)
E
(
V
/
m
)
AMRFDTD
FDTD 360x360
FDTD 2880x2880
z
FIGURE6.9: Electric ﬁeld at 1 μm from the edge of port 3 of the power splitter of Fig. 6.7
of the form:
e
−(y−y
0
)
2
/W
2
−(t−t
0
)
2
/T
2
sin (2π f t) , (6.3)
where f = 200 THz, T = 0.02 ps, t
0
= 3T, W = 1 μm, and y
0
corresponds to the center
of the waveguide. Table 6.4 compares the accuracy and computation time of AMRFDTD
using a root mesh of 250 ×100 and 2–4 mesh levels, to a reference FDTD simulation using a
0 0.5 1 1.5 2 2.5 3 3.5
x 10
−3
−3
−2
−1
0
1
2
3
Time (ns)
E
(
V
/
m
)
z
FIGURE 6.10: Electric ﬁeld obtained with the dynamic AMRFDTD at 1 μm from the edge of port
2 of the power splitter of Fig. 6.7, obtained with the dynamic AMRFDTD, for 40,000 time steps,
indicating the absence of latetime instability
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05806 MOBK058Sarris.cls December 15, 2006 7:53
112 MESH REFINEMENTFORTIMEDOMAIN NUMERICAL ELECTROMAGNETICS
FIGURE6.11: Electric ﬁeld magnitude distribution across the power splitter of Fig. 6.7, at t = 100t.
There are three meshes in total: one level1, one level2, and one level3 mesh
FIGURE6.12: Electric ﬁeld magnitude distribution across the power splitter of Fig. 6.7, at t = 300t.
There are three meshes in total: one level1, one level2, and one level3 mesh
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05806 MOBK058Sarris.cls December 15, 2006 7:53
DYNAMICALLY ADAPTIVE MESHREFINEMENTIN FDTD 113
FIGURE6.13: Electric ﬁeld magnitude distribution across the power splitter of Fig. 6.7, at t = 400t.
There are eight meshes in total: one level1, three level2, and four level3 meshes (tracking transmitted
and reﬂected wavefronts)
2000 ×800 mesh and several coarser FDTD schemes. The AMR parameters are: θ
e
= 0.001,
θ
g
= 0, θ
c
= 0.8, N
AMR
= 10, σ
AMR
= 2. All AMRFDTDsimulations use a reﬁnement factor
of 2 for their successive mesh levels. Table 6.4 shows that AMRFDTDcan achieve the accuracy
of the reference FDTD method in a dramatically reduced computation time.
Field waveforms in time, extracted by a fourlevel AMRFDTD, with a 250 ×100 root
mesh, and the reference FDTDsimulation are shown and compared in Fig. 6.15, demonstrating
y
absorber
P2
P3
x
25
1
waveguide P1
10
4
2
2
2
11 7
2
FIGURE6.14: Geometry of the dielectric waveguide Yjunction. The dielectric constant of the waveg
uide is 9 and the dielectric constant of the surrounding medium is 1. Dimensions are given in units of
μm. P1, P2, and P3 denote ports 1, 2, and 3, respectively
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05806 MOBK058Sarris.cls December 15, 2006 7:53
114 MESH REFINEMENTFORTIMEDOMAIN NUMERICAL ELECTROMAGNETICS
TABLE6.4: Computation Time and TimeDomain Error Metric E
t
for the Dielectric Waveguide
YJunction for Fields Recorded at Port 2
NUMBER TIME TOTAL E
t
METHOD MESH OF LEVELS STEPS TIME(S) (%)
FDTD 250 ×100 — 2500 15.7 166.7
FDTD 500 ×200 — 5000 155.0 174.4
FDTD 1000 ×400 — 10000 1296 77.8
FDTD 2000 ×800 — 20000 23418 —
AMRFDTD 250 ×100 2 2500 184.7 174.5
AMRFDTD 250 ×100 3 2500 646.1 78.4
AMRFDTD 250 ×100 4 2500 2665 0.55
the excellent agreement of the AMRFDTDapproach to the reference data, despite its reduced
numerical cost. A waveform extracted by a conventional FDTD implemented on the root mesh
of the AMRFDTDdomain alone is appended, to indicate the accuracy improvement, brought
about by the mesh reﬁnement. Moreover, Fig. 6.16 shows that the multilevel AMRFDTD is
free from the late instability problems that characterize static subgridding techniques.
Figures 6.17–6.20 show the evolution of the electric ﬁeld and the mesh tree, which uses
a 125 ×50 root mesh and maximum level of 3. The level1 mesh is drawn in the ﬁgures and
2 3 4
x 10
−4
−2
−1
0
1
2
Time (ns)
E
(
V
/
m
)
AMRFDTD
FDTD 250x100
FDTD 2000x800
z
FIGURE 6.15: Electric ﬁeld at the center of the dielectric waveguide Yjunction of Fig. 6.14, 2 μm
from the edge of port 2
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05806 MOBK058Sarris.cls December 15, 2006 7:53
DYNAMICALLY ADAPTIVE MESHREFINEMENTIN FDTD 115
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
x 10
−3
−2
−1
0
1
2
Time (ns)
E
(
V
/
m
)
z
FIGURE 6.16: Electric ﬁeld at the center of the dielectric waveguide Yjunction of Fig. 6.14, 2 μm
from the edge of port 2, obtained with the dynamic AMRFDTD, for 25,000 time steps, indicating the
absence of latetime instability
FIGURE 6.17: Electric ﬁeld magnitude distribution across the Yjunction of Fig. 6.14, at t = 100t.
There are three meshes in total: one level1, one level2, and one level3 mesh
FIGURE 6.18: Electric ﬁeld magnitude distribution across the Yjunction of Fig. 6.14, at t = 400t.
There are three meshes in total: one level1, one level2, and one level3 mesh (enclosed in the black
solid line)
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05806 MOBK058Sarris.cls December 15, 2006 7:53
116 MESH REFINEMENTFORTIMEDOMAIN NUMERICAL ELECTROMAGNETICS
FIGURE 6.19: Electric ﬁeld magnitude distribution across the Yjunction of Fig. 6.14, at t = 600t.
There are three meshes in total: one level1, one level2, and one level3 mesh (enclosed in the black
solid line). Note that level2 and level3 meshes track the pulse into the junction
FIGURE 6.20: Electric ﬁeld magnitude distribution across the Yjunction of Fig. 6.14, at t = 800t.
There are four meshes in total: one level1, two level2, and one level3 mesh (enclosed in the black solid
line)
the regions without grid lines are covered by level2 and level3 meshes, which are not drawn
because they are too dense. The rectangular boxes embedded in the solid gray regions indicate
the boundaries of level3 meshes. These ﬁgures indicate the wavefront tracking achieved by the
dynamic AMRFDTD.
6.5 DIELECTRIC RING RESONATOR
A highly resonant THz structure of a dielectric ring resonator coupled to two dielectric waveg
uides, previously modeled in [61] and shown in Fig. 6.21, is simulated [59]. The width of the
waveguide and the ring is 0.3 μm, and their dielectric constant is 10.24. The external radius of
the ring is 2.5 μm. The edgetoedge distance between the ring and the waveguide is 0.232 μm.
The computational domain is 12 μm×12 μm and it is truncated by a 1 μm thick matched
absorber. The excitation is a modulated pulse of the form:
e
−(y−y
0
)
2
/W
2
−(t−t
0
)
2
/T
2
sin (2π f t) , (6.4)
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05806 MOBK058Sarris.cls December 15, 2006 7:53
DYNAMICALLY ADAPTIVE MESHREFINEMENTIN FDTD 117
x
y
1
0.3
0.3
0.232
2.5
12
12
P2
P4 P3
P1
FIGURE6.21: Geometry of the dielectric ring resonator coupled to two dielectric waveguides. Dimen
sions are in units of μm. P1, P2, P3, and P4 denote ports 1, 2, 3, and 4, respectively
where f = 200 THz, T = 0.02 ps, t
0
= 3T, W = 0.7 μm, and y
0
corresponds to the midsection
of the waveguide. The excitation is imposed 1.1 μm from the end of the lower left waveguide
port, and the electric ﬁeld is recorded at 2 μm from the edge of each port. All AMRFDTD
simulations use a reﬁnement factor of 2 for successive levels, and θ
e
= 0.01, θ
g
= 0, θ
c
= 0.8,
N
AMR
= 10, σ
AMR
= 2. Table 6.5 compares accuracy and execution times of AMRFDTDand
conventional FDTDfor this largescale computational problem. Again, a fourlevel scheme can
reproduce the results of the reference FDTD simulation within a smaller execution time (by a
factor of 33), at a relative timedomain error of ﬁelds recorded at port 2 of 1.56%.
TABLE6.5: Computation Time and TimeDomain Error Metric E
t
for the Dielectric Ring Res
onator for Fields Recorded at Port 2
NUMBER TIME TOTAL E
t
METHOD MESH OF LEVELS STEPS TIME(S) (%)
FDTD 120 ×120 — 20000 60.71 187.0
FDTD 240 ×240 — 40000 703.1 163.6
FDTD 480 ×480 — 80000 6904.4 37.2
FDTD 960 ×960 — 160000 133993 —
AMRFDTD 120 ×120 2 20000 466.5 163.6
AMRFDTD 120 ×120 3 20000 1080 36.7
AMRFDTD 120 ×120 4 20000 4096 1.56
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05806 MOBK058Sarris.cls December 15, 2006 7:53
118 MESH REFINEMENTFORTIMEDOMAIN NUMERICAL ELECTROMAGNETICS
The timedomain results of Figs. 6.22–6.24 demonstrate the latetime stability and ac
curacy of the method. It should be noted that the ﬁrst Gaussian pulse arrives at port 3 at about
0.2 ps. Before that, the ﬁeld at port 3 is due to the disturbance caused by the excitation, which
propagates through free space. Since this disturbance is very small compared with the main pulse
propagating along the waveguide, it is not tracked by the ﬁnest mesh and therefore it induces
a relatively large error. However, the pulses circulating the ring resonator and coupled to the
waveguides are tracked by the ﬁnest mesh and therefore their amplitude and group velocity are
very accurately resolved even at very late stages of the simulation.
1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8 3
x 10
−4
−5
0
5
Time (ns)
E
(
V
/
m
)
AMRFDTD
FDTD 120x120
FDTD 960x960
z
2.9 2.95 3 3.05 3.1 3.15 3.2
x 10
−3
−0.05
0
0.05
Time (ns)
E
(
V
/
m
)
AMRFDTD
FDTD 120x120
FDTD 960x960
z
FIGURE6.22: Electric ﬁeld at 2 μm from port 2 of the ring resonator of Fig. 6.21
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05806 MOBK058Sarris.cls December 15, 2006 7:53
DYNAMICALLY ADAPTIVE MESHREFINEMENTIN FDTD 119
0 1 2
x 10
−4
−0.02
−0.015
−0.01
−0.005
0
0.005
0.01
0.015
0.02
Time (ns)
E
(
V
/
m
)
AMRFDTD
FDTD 120x120
FDTD 960x960
2 3 4
x 10
−4
−0.1
−0.05
0
0.05
0.1
Time (ns)
E
(
V
/
m
)
AMRFDTD
FDTD 120x120
FDTD 960x960
3 3.1 3.2 3.3
x 10
−3
−0.05
0
0.05
Time (ns)
E
(
V
/
m
)
AMRFDTD
FDTD 120x120
FDTD 960x960
z
z
z
FIGURE6.23: Electric ﬁeld at 2 μm from port 2 of the ring resonator of Fig. 6.21
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05806 MOBK058Sarris.cls December 15, 2006 7:53
120 MESH REFINEMENTFORTIMEDOMAIN NUMERICAL ELECTROMAGNETICS
0 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016
−5
0
5
Time (ns)
E
(
V
/
m
)
z
FIGURE 6.24: Electric ﬁeld at 2 μm from port 2 of the ring resonator of Fig. 6.21, obtained with the
dynamic AMRFDTD, for 100,000 time steps, indicating the absence of latetime instability
185 190 195 200 205 210
−10
−8
−6
−4
−2
0
Frequency (THz)
A
m
p
l
i
t
u
d
e
o
f
S
2
1
(
d
B
)
FIGURE6.25: S
21
obtained with the fourlevel AMRFDTD
Finally, the S
21
obtained by using AMRFDTDis shown in Fig. 6.25, which agrees quite
well with the results of [61], extracted by FDTD. The AMRFDTD uses a 120 ×120 root
mesh and four levels. Since the structure is highly resonant, 100,000 time steps are executed.
The resonant frequencies of the present method and [61] are compared in Table 6.6, which
indicates a very good agreement between the two sets of results.
6.6 NUMERICAL ERRORESTIMATION ANDCONTROL
There are ﬁve parameters controlling the accuracy and computation time of AMRFDTD, as
described by the previous sections. To use AMRFDTD as a CAD technique, a clear set of
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05806 MOBK058Sarris.cls December 15, 2006 7:53
DYNAMICALLY ADAPTIVE MESHREFINEMENTIN FDTD 121
TABLE 6.6: Resonant Frequencies (in THz) of the
Dielectric Ring Resonator
ORDER AMRFDTD FDTD[61]
25 186.01 185.85
26 192.03 191.88
27 198.06 197.90
28 204.08 203.92
29 210.11 209.94
guidelines for the choice of the ﬁve parameters is needed. The main objective of this section is
to produce such guidelines, that would allow a user to determine the AMRFDTD parameters
for a given error tolerance [59].
The numerical error whose dependence on the AMRFDTD parameters is sought for is
deﬁnedas follows, for a 2DTEcase withﬁeldcomponents (E
z
, H
x
, H
y
). Givena computational
domain, an array of probes is considered, where ﬁeld values are recorded at each time step. The
reference solution, that AMRFDTD is compared to, is provided by applying the FDTD
method in a uniform mesh, at a Yee cell resolution equal to the maximum achievable resolution
of the AMRFDTD. This is chosen to be dense enough to guarantee convergence. Then, if
a ﬁeld component recorded at the mth probe at time t
k
is denoted by f
m
(t
k
) and f
ref
m
(t
k
) the
corresponding reference ﬁeld value, the error is deﬁned as:
E =
¸
k,m
f
m
(t
k
) − f
ref
m
(t
k
)
2
¸
k,m
f
ref
m
(t
k
)
2
(6.5)
Note that different probes encounter different ﬁeldwaveforms, but alsomeshconﬁgurations, due
to the dynamic nature of mesh evolution in AMRFDTD. Therefore, this metric is much more
effective in capturing generic error effects, than a normthat would be based on ﬁeld sampling at
a port of a device, or the error in a frequency domain quantity (such as characteristic impedance
or propagation constant). This approach is analogous to a MonteCarlotype simulation and
is aimed at rendering the obtained error less dependent on the device under test. Furthermore,
such an approach is the only feasible, given the nature of the problem at hand. Contrary to
static subgridding methods, the extraction of analytical error bounds seems to be impossible,
due to the arbitrary distribution of submeshes in a domain. Finally, the same methodology can
be easily extended to 2D TM cases and 3D cases.
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05806 MOBK058Sarris.cls December 15, 2006 7:53
122 MESH REFINEMENTFORTIMEDOMAIN NUMERICAL ELECTROMAGNETICS
While the AMRFDTDerror performance on a number of structures was studied, results
obtainedfromsimulating three dielectric waveguide ones are shown. These are the power splitter,
corrugated permittivity proﬁle waveguide and the Yjunction, presented and simulated in the
previous section.
First, the effect of N
AMR
, σ
AMR
, and θ
c
on the accuracy and simulation time is inves
tigated. Figure 6.26 includes errorsimulation time curves, deduced by changing one of these
3 3.2 3.4 3.6 3.8 4
−3
−2.5
−2
−1.5
log
10
(Simulation time)
l
o
g
1
0
ε
10
20
40
10
20
40
10
20
40
splitter
waveguide
Yjunction
(a)
3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8
−2.5
−2
−1.5
−1
log
10
(Simulation time)
l
o
g
1
0
ε
1
2
3
4 1
2
3
4
1
2
3
4
splitter
waveguide
Yjunction
(b)
3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9
−2.2
−2
−1.8
−1.6
−1.4
log
10
(Simulation time)
l
o
g
1
0
ε
0.8
0.7
0.5
0.3
0.8 0.7
0.5
0.3
0.8
0.7
0.5
0.3
splitter
waveguide
Yjunction
(c)
FIGURE6.26: Error versus simulation time for variable (a) N
AMR
, (b) σ
AMR
, (c) θ
c
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05806 MOBK058Sarris.cls December 15, 2006 7:53
DYNAMICALLY ADAPTIVE MESHREFINEMENTIN FDTD 123
three parameters and ﬁxing the rest. All ﬁgures refer to fourlevel AMRFDTDsimulations, but
twoand threelevel simulation results exhibit a similar behavior. The general pattern of these
curves suggests that although increasing N
AMR
and σ
AMR
or decreasing θ
c
may improve accu
racy (e.g., as N
AMR
is increased the AMRFDTD tends to become equivalent to the reference
FDTD technique), this improvement reaches a plateau beyond which, any change in the pa
rameters merely increases the execution time. In a number of versatile situations, the choice of
N
AMR
= 10, σ
AMR
= 2, and θ
c
= 0.7 balances accuracy and efﬁciency well. This conclusion
is also supported by Fig. 6.26. Essentially, these parameters do not explicitly control the mesh
reﬁnement procedure, which is why their choice is rather straightforward. This is not the case
for θ
e
, θ
g
that are studied next.
Figure 6.27 shows the dependence of the error E on θ
e
and θ
g
for all the cases studied. As
expected, the error decreases as θ
e
or θ
g
decreases. For very small values of θ
g
, the error is mainly
a function of θ
e
and vice versa, since marking a cell for reﬁnement depends on both thresholds.
Based on this data E(θ
e
, θ
g
) can be approximated as:
˜
E(θ
e
, θ
g
) = C
1
E(0, θ
g
) arctan
θ
g
θ
e
C
2
+C
3
E(θ
e
, 0) arctan
θ
e
θ
g
C
4
,
(6.6)
0
.9
0.9 0.9
1
1
1
1
1
1
.
1
1
.
1
1.1
1.1
1.1
1
.
2
1
.
2
1.2
1.2
1
.
3
1
.
3
1.3
1.3
1
.
4
1
.
4
1.4
1
.
5
1
.5
1.5
1
.
6
1
.6
1.6
1
.
7
1
.7
1
.
8
1.8
1
.
9
1.9
2
2
2
.1
2
.2
log
10
θ
e
l
o
g
1
0
θ
g
−4 −3.5 −3 −2.5 −2 −1.5 −1 −0.5 0
−5
−4.5
−4
−3.5
−3
−2.5
−2
−1.5
−1
FIGURE6.27: Dependence of −l og
10
E on θ
e
and θ
g
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05806 MOBK058Sarris.cls December 15, 2006 7:53
124 MESH REFINEMENTFORTIMEDOMAIN NUMERICAL ELECTROMAGNETICS
0
.
8
0
.
9
0.9
0.9 0.9
1
1
1
1
1
.
1
1
.
1
1.1
1.1
1
.
2
1
.
2
1
.2
1.2
1
.
3
1
.
3
1
.3
1.3
1
.
4
1
.4
1.4
1
.
5
1
.
5
1.5
1
.
6
1
.
6
1.6
1
.
7
1.7
1
.
8
1.8
1
.
9
1.9
2
2
2
.
1
log
10
θ
e
l
o
g
1
0
θ
g
−4 −3.5 −3 −2.5 −2 −1.5 −1 −0.5 0
−5
−4.5
−4
−3.5
−3
−2.5
−2
−1.5
−1
FIGURE6.28: Curveﬁtted dependence of −l og
10
˜
E on θ
e
and θ
g
based on (6.6)
where C
1
, C
2
, C
3
, and C
4
are ﬁnetuned to minimize the difference between
˜
E and E. As
an example, Fig. 6.28 shows the curveﬁtted error obtained by (6.6), where C
1
= C
3
= 1.02,
C
2
= 0.05, C
4
= 0.01, which depicts a very good agreement with Fig. 6.27. Therefore, this
approximation, requiring only E(θ
e
, 0) and E(0, θ
g
), is useful as sufﬁciently accurate.
By simulating typical optical waveguide structures, two empirical error bound functions
are obtained for the special cases θ
e
= 0 and θ
g
= 0, respectively,
E
bg
(θ
e
= 0, θ
g
) = 0.4θ
0.35
g
, (6.7)
E
be
(θ
e
, θ
g
= 0) = 0.15θ
0.37
e
. (6.8)
Following (6.6), we deﬁne a general error bound as
E
b
(θ
e
, θ
g
) = C
1
E
bg
(θ
e
= 0, θ
g
) arctan
θ
g
θ
e
C
2
+C
3
E
be
(θ
e
, θ
g
= 0) arctan
θ
e
θ
g
C
4
(6.9)
where C
1
= C
3
= 1.02, C
2
= 0.05, C
4
= 0.01. This general error bound is shown in Fig. 6.29
and can be used to determine θ
e
and θ
g
given the desired accuracy.
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05806 MOBK058Sarris.cls December 15, 2006 7:53
DYNAMICALLY ADAPTIVE MESHREFINEMENTIN FDTD 125
0
.
7
0
.
8
0
.
8
0.8
0.8
0
.
9
0
.
9
0
.9
0.9
0.9
1
1
1
1
1
.
1
1
.
1
1
.1
1.1
1
.
2
1
.2
1.2
1
.
3
1
.
3
1.3
1
.
4
1.4
1
.
5
1
.5
1
.6
1
.7
log
10
θ
e
l
o
g
1
0
θ
g
−4 −3.5 −3 −2.5 −2 −1.5 −1 −0.5 0
−4
−3.5
−3
−2.5
−2
−1.5
−1
FIGURE6.29: Curveﬁtted dependence of error bound −l og
10
E
b
on θ
e
and θ
g
based on (6.6)
While the validity of these bounds can be conﬁrmed through all previously published
examples of the dynamic AMRFDTD, the case of a dielectric waveguide directional coupler
(Fig. 6.30), similar to that of [62], is also discussed here. In this example, letting N
AMR
= 10,
σ
AMR
= 2 and θ
c
= 0.7, θ
g
and θ
e
were chosen, requiring that the error be less than 3%. Then,
from Fig. 6.29, θ
e
= 0.001, θ
g
= 0.0003 satisﬁed this criterion. The excitation was a sinewave
y
absorber
P1 P2
P4 P3
x
waveguide
2.9
0.5
0.35
16
6
0.25
2.82
FIGURE 6.30: Geometry of a directional coupler (not drawn to scale). Dimensions are given in μm.
The surrounding medium is air. The dielectric constant of the waveguide and the center piece is 3.24
and 4.41, respectively
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05806 MOBK058Sarris.cls December 15, 2006 7:53
126 MESH REFINEMENTFORTIMEDOMAIN NUMERICAL ELECTROMAGNETICS
modulated Gaussian pulse with a center frequency of 230 THz and a pulse width of 20 fs.
The excitation was imposed at 1.2 μm from port 1. For AMRFDTD, four levels of mesh
resolution were used; the size of the coarsest mesh was 320 ×120. The time step for each
mesh was determined by letting all mesh Courant numbers be equal to 0.7. Fields and errors
were recorded over 3200 time steps of the root mesh. An array of 14 ×8 probes was used to
record the E
z
ﬁeld throughout the computation domain. The reference result was obtained by
applying FDTD in a 2560 ×960 uniform mesh. The actual error was 0.46%, which satisﬁed
the requirement, although it indicated that the deduced bound was rather conservative. Finally,
Fig. 6.31 compares the ﬁelds obtained by AMRFDTD and FDTD in ports 3 and 4 of the
coupler. Evidently, AMRFDTD cannot be distinguished from the reference solution, even
80 100 120 140 160 180
−1
−0.5
0
0.5
1
Time (fs)
E
(
V
/
m
)
FDTD 2560x960
AMRFDTD
FDTD 320x120
(a) Port 3
50 100 150 200 250
−0.1
−0.05
0
0.05
0.1
Time (fs)
E
(
V
/
m
)
FDTD 2560x960
AMRFDTD
FDTD 320x120
(b) Port 4
z
z
FIGURE6.31: Electric ﬁeld at port 3 and 4 of the directional coupler
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05806 MOBK058Sarris.cls December 15, 2006 7:53
DYNAMICALLY ADAPTIVE MESHREFINEMENTIN FDTD 127
for the weak ﬁeld at port 4. For that waveform, the coarse mesh FDTD result suffers from a
signiﬁcant numerical dispersion induced error. The AMRFDTD simulation lasts 2.8 hours,
whereas the reference FDTD simulation 19.3 hours.
6.7 CONCLUSION
This chapter discussed the application of the multilevel, dynamically AMRFDTD technique
to optical waveguide structure analysis. Numerical examples demonstrated the efﬁciency and
latetime stability of the AMRFDTD technique. Large speedup factors (ranging from 30 to
100) compared to the conventional FDTD were achieved, while all errors involved remained
small, typically of the order of 1% or less. Furthermore, the numerical error associated with the
application of the AMRFDTD technique was studied, and its dependence on the parame
ters controlling the accuracy and performance of this technique was outlined. Guidelines for
the a priori choice of these parameters were provided, thus fulﬁlling an important condition
for AMRFDTD to be considered as a CADoriented tool for electromagnetic and optical
applications.
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK05806 MOBK058Sarris.cls December 15, 2006 7:53
128
P1: IML/FFX P2: IML
MOBK058BIB MOBK058Sarris.cls December 15, 2006 17:39
129
Bibliography
[1] J. B. Keller, “Geometrical theory of diffraction,” J. Opt. Soc. Am., vol. 52, no. 2,
pp. 116–130, Feb. 1962.
[2] R. Mittra and S. W. Lee, Analytical Techniques in the Theory of Guided Waves. New York:
Macmillan, 1971.
[3] R. F. Harrington, Field Computation by Moment Methods. New York: Macmillan, 1968.
[4] A. Sommerfeld, Partial Differential Equations in Physics. New York: Academic Press,
1964.
[5] J. Jin, The Finite Element Method in Electromagnetics. New York: J. W. Wiley & Sons,
1993.
[6] K. S. Yee, “Numerical solution of initial boundary value problems involving Maxwell’s
equations in isotropic media,” IEEETrans. Antennas Propagat., vol. AP14, pp. 302–307,
Mar. 1966.
[7] W. J. R. Hoefer, “The transmissionline matrix method—Theory and applications,”
IEEE Trans. Microwave Theory Tech., vol. MTT33, no. 10, pp. 882–893, Oct. 1985.
[8] E. Chiprout and M. S. Nakhla, Asymptotic Waveform Evaluation and Moment Matching
for Interconnect Analysis. Norwell, MA: Kluwer, 1994.
[9] K. L. Schlager et al., “Relative accuracy of several ﬁnitedifference timedomain meth
ods in two and three dimensions,” IEEE Trans. Antennas Propagat., vol. 41, no. 12,
pp. 1732–1737, Dec. 1993.doi:10.1109/8.273296
[10] I. S. Kim and W. J. R. Hoefer, “A local mesh reﬁnement algorithm for the time domain
ﬁnite difference method using Maxwell’s curl equations,” IEEETrans. Microwave Theory
Tech., vol. MTT38, no. 6, pp. 812–815, June 1990.doi:10.1109/22.130985
[11] M. Krumpholz and L. P. B. Katehi, “MRTD: New time domain schemes based on
multiresolution analysis,” IEEE Trans. Microwave Theory Tech., vol. MTT44, no. 4,
pp. 555–561, Apr. 1996.doi:10.1109/22.491023
[12] W. Werthen and I. Wolff, “A novel wavelet based time domain approach,” IEEE Mi
crowave Guided Wave Lett., vol. 6, pp. 438–440, Dec. 1996.doi:10.1109/75.544542
[13] M. Aidam and P. Russer, “Application of biorthogonal B—spline wavelets to Telegra
pher’s equations,” Proc. Annu. Rev. Prog. Appl. Comput. Electromagn., pp. 983–990, Mar.
1998.
P1: IML/FFX P2: IML
MOBK058BIB MOBK058Sarris.cls December 15, 2006 17:39
130 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
[14] M. Fujii and W. J. R. Hoefer, “Athreedimensional Haar waveletbased multiresolution
analysis similar to the 3D FDTD method—derivation and application,” IEEE Trans.
Microwave Theory Tech., vol. MTT46, pp. 2463–2475, Dec. 1998.
[15] S. GrivetTalocia, “Adaptive transient solution of nonuniform multiconductor trans
mission lines using wavelets,” IEEE Trans. Antennas Propagat., vol. 48, no. 10,
pp. 1563–1573, Oct. 2000.doi:10.1109/8.899673
[16] X. Zhu and L. Carin, “Multiresolution timedomain analysis of planewave scattering
from general threedimensional surface and subsurface dielectric targets,” IEEE Trans.
Antennas Propagat., vol. 49, no. 11, pp. 1568–1578, Nov. 2001.doi:10.1109/8.964093
[17] E. M. Tentzeris et al., “Space and timeadaptive gridding using MRTD technique,”
IEEE MTTS Int. Microwave Symp. Dig., vol. 1, pp. 337–340, 1997.
[18] M. Krumpholz, E. M. Tentzeris, and L. P. B. Katehi, “Application of MRTD to
printed transmission lines,” IEEE MTTS Int. Microwave Symp. Dig., pp. 573–576,
1996.
[19] E. M. Tentzeris, L. Roselli, and L. P. B. Katehi, “Nonlinear circuit characterization using
a multiresolutiontime domaintechnique (MRTD),” IEEEMTTSInt. Microwave Symp.
Dig., vol. 3, pp. 1397–1400, 1998.
[20] E. M. Tentzeris, R. L. Robertson, and L. P. B. Katehi, “MRTD applied to complex
airdielectric boundary structures,” IEEE MTTS Int. Microwave Symp. Dig., vol. 4,
pp. 1463–1466, 1999.
[21] K. Goverdhanam et al., “Applications of a multiresolution based FDTD multigrid,”
IEEE MTTS Int. Microwave Symp. Dig., vol. 1, pp. 333–336, 1997.
[22] S. G. Mallat, A Wavelet Tour of Signal Processing. New York: Academic Press, 1999.
[23] F. Arndt and J. Ritter, “A generalized 3D subgrid technique for the FDTD method,”
IEEE MTTS Int. Microwave Symp. Dig., vol. 3, pp. 1563–1566, 1997.
[24] I. Wolff and M. Walter, “An algorithm for realizing Yee’s FDTD method in the wavelet
domain,” IEEE MTTS Int. Microwave Symp. Dig., vol. 3, pp. 1301–1304, 1999.
[25] R. G. Plumb, Z. Huang, and K. Demarest, “An FDTD/MoM hybrid technique for
modeling complex antennas in the presence of heterogeneous grounds,” IEEE Trans.
Geosci. Remote Sensing, vol. 37, no. 6, pp. 2692–2698, Nov. 1999.doi:10.1109/36.803416
[26] T. D. Tsiboukis, T. I. Kosmanis, and N. V. Kantartzis, “A hybrid FDTDwavelet
galerkin technique for the numerical analysis of ﬁeld singularities inside waveguides,”
IEEE Trans. Magn., vol. 36, no. 4, pp. 902–906, July 2000.doi:10.1109/20.877589
[27] M. Fujii et al., “A 2D TLM and Haar MRTD real time hybrid connection algorithm,”
Proc. Annu. Rev. Prog. Appl. Comput. Electromagn., pp. 1013–1020, 2000.
[28] H. Kim, S. Ju, and D.H. Bae, “The modeling of lumped elements using the Haar
wavelet multiresolution time domain technique,” Proc. IEEE APS Symp., 2000.
P1: IML/FFX P2: IML
MOBK058BIB MOBK058Sarris.cls December 15, 2006 17:39
BIBLIOGRAPHY 131
[29] C. D. Sarris and L. P. B. Katehi, “An efﬁcient numerical interface between Haar MRTD
andFDTD: Formulationandapplications,” IEEETrans. Microwave Theory Tech., vol. 51,
no. 4, pp. 1146–1156, Apr. 2003.doi:10.1109/TMTT.2003.809620
[30] I. Daubechies, Ten Lectures on Wavelets. Philadelphia, PA: Society for Industrial and
Applied Mathematics, 1992.
[31] H. L. Royden, Real Analysis, 3rd edn. Englewood Cliffs, NJ: Prentice Hall, 1988.
[32] G. Papanicolaou, E. Bacry, and S. Mallat, “A wavelet based spacetime adaptive nu
merical method for partial differential equations,” Math. Modeling Numer. Anal., vol. 26,
no. 7, p. 793, 1992.
[33] A. Haar, “Zur Theorie der Orthogonalen Funktionensysteme,” Math. Ann., vol. 69,
pp. 331–371, 1910.doi:10.1007/BF01456326
[34] G. Battle, “A block spin construction of ondelettes, Part I : Lemarie functions,” Comm.
Math. Phys., vol. 110, pp. 601–615, 1987.doi:10.1007/BF01205550
[35] C. Huber, M. Krumpholz, and P. Russer, “A ﬁeld theoretical comparison of FDTD
and TLM,” IEEE Trans. Microwave Theory Tech., vol. 43, pp. 1935–1950, Aug. 1995.
doi:10.1109/22.402284
[36] J. C. Strikwerda, Finite Difference Schemes and Partial Differential Equations. Boca Raton,
FL: CRC Press, 1989.
[37] T. Dogaru and L. Carin, “Application of Haarwaveletbased multiresolution time
domain schemes to electromagnetic scattering problems,” IEEE Trans. Antennas Prop
agat., vol. 50, no. 6, pp. 774–784, June 2002.doi:10.1109/TAP.2002.1017657
[38] C. D. Sarris and L. P. B. Katehi, “Fundamental gridding related dispersion effects in
MRTD schemes,” IEEE Trans. Microwave Theory Tech., vol. 49, no. 12, pp. 2248–2257,
Dec. 2001.doi:10.1109/22.971607
[39] A. Taﬂove, Computational Electrodynamics: The FiniteDifference TimeDomain Method.
Norwood, MA: Artech House, 1995.
[40] E. M. Tentzeris et al., “Multiresolution timedomain (MRTD) adaptive schemes using
arbitrary resolutions of wavelets,” IEEE Trans. Microwave Theory Tech., vol. 50, no. 2,
pp. 501–516, Feb. 2002.doi:10.1109/22.982230
[41] K. Goverdhanam et al., “A perfectly matched layer formulation for Haar wavelet based
MRTD,” Proc. Eur. Microwave Conf., pp. 243–246, 1999.
[42] S. D. Gedney, “An anisotropic perfectly matched layerabsorbing mediumfor the trunca
tion of FDTDlattices,” IEEETrans. Antennas Propagat., vol. 44, no. 12, pp. 1360–1369,
Dec. 1996.doi:10.1109/8.546249
[43] J. Citerne, G. Carat, R. Gillard, and J. Wiart, “An efﬁcient analysis of planar microwave
circuits using a DWTbased Haar MRTD scheme,” IEEE Trans. Microwave Theory
Tech., vol. 48, no. 12, pp. 2261–2270, Dec. 2000.doi:10.1109/22.898973
P1: IML/FFX P2: IML
MOBK058BIB MOBK058Sarris.cls December 15, 2006 17:39
132 MESHREFINEMENTFORTIMEDOMAINNUMERICAL ELECTROMAGNETICS
[44] M. A. Stuchly, M. Okoniewski, and E. Okoniewska, “Three dimensional subgridding
algorithm for FDTD,” IEEE Trans. Antennas Propagat., vol. 45, no. 3, pp. 422–429,
Mar. 1997.doi:10.1109/8.558657
[45] M. Kuzyk, D. Sullivan, and J. Liu, “Threedimensional optical pulse simulation using
the FDTDmethod,” IEEETrans. Microwave Theory Tech., vol. 48, no. 7, pp. 1127–1133,
July 2000.doi:10.1109/22.848495
[46] H. G. Winful, “Pulse compression in optical ﬁber ﬁlters,” Appl. Phys. Lett., vol. 46,
pp. 527–529, Mar. 1985.doi:10.1063/1.95580
[47] H. G. Winful, M. Krumpholz, and Linda P. B. Katehi, “Nonlinear time domain mod
eling by multiresolution time domain (MRTD),” IEEE Trans. Microwave Theory Tech.,
vol. 45, no. 3, pp. 385–393, Mar. 1997.doi:10.1109/22.563337
[48] D. M. Sheen, S. M. Ali, M. Abouzahra, and J. A. Kong, “Application of the three
dimensional ﬁnitedifference timedomain method to the analysis of planar microstrip
circuits,” IEEE Trans. Microwave Theory Tech., vol. 38, no. 7, pp. 849–857, July 1990.
doi:10.1109/22.55775
[49] R. Luebbers, J. Schuster, and K. Wu, “Application of moving window FDTD to pre
diction of path loss over irregular terrain,” Proc. IEEE APS Symp., vol. 2, pp. 610–613,
June 2003.
[50] R. Kastner, B. Fidel, E. Heyman, and R. W. Ziolkowski, “Hybrid rayFDTD moving
frame approach to pulse propagation,” Proc. IEEE APS Symp., vol. 3, pp. 1414–1417,
1994.
[51] R. Kastner, B. Fidel, E. Heyman, and R. W. Ziolkowski, “Absorbing boundary con
ditions in the context of the hybrid rayFDTD moving window solution,” Proc. IEEE
APS Symp., vol. 2, pp. 1006–1009, 1997.
[52] M. Berger and J. R. Oliger, “Adaptive mesh reﬁnement for hyperbolic partial dif
ferential equation,” J. Computat. Phys., vol. 53, pp. 484–512, 1984.doi:10.1016/0021
9991(84)900731
[53] Y. Liu and C. D. Sarris, “Efﬁcient modeling of microwave integrated circuit ge
ometries via a dynamically adaptive mesh reﬁnement (AMR)FDTD technique,”
IEEE Trans. Microwave Theory Tech., vol. 54, no. 2, pp. 689–703, Feb. 2006.
doi:10.1109/TMTT.2005.862660
[54] G. Mur, “Absorbing boundary conditions for the ﬁnitedifference approximation of
the timedomain electromagnetic ﬁeld equations,” IEEE Trans. Electromagn. Comput.,
vol. EMC23, pp. 377–382, 1981.
[55] M. Berger and I. Rigoutsos, “An algorithm for point clustering and grid generation,”
IEEE Trans. Syst., Man. Cybern., vol. 21, no. 5, pp. 1278–1286, Sept./Oct. 1991.
doi:10.1109/21.120081
P1: IML/FFX P2: IML
MOBK058BIB MOBK058Sarris.cls December 15, 2006 17:39
BIBLIOGRAPHY 133
[56] A. Rennings, V. Toni, P. Waldow, Y. Qian, T. Itoh, and I. Wolff, “A novel adaptivity
for em time domain methods: Scale adaptive time steps (SATS),” IEEE MTTS Int.
Microwave Symp. Dig., vol. 2, pp. 757–760, May 2001.
[57] T. Mangold, J. Rebel, W. J. R. Hoefer, P. P. M. So, and P. Russer, “What determines the
speed of timediscrete algorithms,” Proc. Annu. Rev. Prog. Appl. Comput. Electromagn.,
vol. 2, pp. 594–601, 2000.
[58] S. S. Zivanovic, K. S. Yee, and K. K. Mei, “A subgridding method for the timedomain
ﬁnitedifference method to solve Maxwells equations,” IEEE Trans. Microwave Theory
Tech., vol. 39, no. 3, pp. 471–479, Mar. 1991.doi:10.1109/22.75289
[59] Y. Liu and C. D. Sarris, “Fast timedomain simulation of optical waveguide structures
with a multilevel dynamically adaptive mesh reﬁnement FDTDapproach,” IEEE/OSAJ.
Lightwave Tech., vol. 24, no. 8, pp. 3235–3247, Aug. 2006.doi:10.1109/JLT.2006.876899
[60] M. Fujii and W. J. R. Hoefer, “Interpolating wavelet collocation method of time depen
dent Maxwells equations: Characterization of electrically large optical waveguide discon
tinuities,” J. Computat. Phys., vol. 186, no. 2, pp. 666–689, Apr. 2003.doi:10.1016/S0021
9991(03)000913
[61] S. C. Hagness, D. Raﬁzadeh, S. T. Ho, and A. Taﬂove, “FDTD microcavity simu
lations: Design and experimental realization of waveguide coupled singlemode ring
and whisperinggallerymode disk resonators,” IEEE/OSA J. Lightwave Tech., vol. 15,
no. 11, pp. 2154–2165, Nov. 1997.doi:10.1109/50.641537
[62] S. T. Chu and S. K. Chaudhuri, “A ﬁnitedifference timedomain method for the design
and analysis of guidedwave optical structures,” IEEE/OSA J. Lightwave Tech., vol. 7,
no. 12, pp. 2033–2038, Dec. 1989.doi:10.1109/50.41625
P1: IML/FFX P2: IML
MOBK058BIB MOBK058Sarris.cls December 15, 2006 17:39
134
P1: OIY
MOBK058BIO MOBK058Sarris.cls December 15, 2006 7:54
135
Author Biography
CostasD. Sarris received a Ph.D. and a M.Sc. in Electrical Engineering, and a M.Sc. in Applied
Mathematics from the University of Michigan, Ann Arbor, in 2002, 1998 and 2002, respec
tively. He also received a Diploma in Electrical and Computer Engineering (with distinction)
from the National Technical University of Athens (NTUA), Greece, in 1997. In November
2002, he joined the Edward S. Rogers Sr. Department of Electrical and Computer Engineer
ing (ECE), University of Toronto, Toronto, ON, Canada, where he is currently an Assistant
Professor. His research interests are in the area of computational electromagnetics, with em
phasis in highorder, meshadaptive techniques. He is currently involved with basic research in
novel numerical techniques, as well as applications of timedomain analysis to wireless chan
nel modeling, wavepropagation in complex media and metamaterials and electromagnetic
compatibility/interference (EMI/EMC) problems.
Prof. Sarris has received a number of scholarship distinctions, including Hellenic Fel
lowship Foundation (1993–1997) and Technical Chamber of Greece (1994–1997) awards for
academic excellence and an NTUA1997 class bronze medal. He received a student paper award
in the 2001 International Microwave Symposium for his work on a hybrid FDTD/MRTD nu
merical scheme, a Canada Foundation for Innovation New Opportunities Fund Award in 2004
and an award for excellence in undergraduate teaching from the Department of Electrical and
Computer Engineering, University of Toronto.
P1: OIY
MOBK058BIO MOBK058Sarris.cls December 15, 2006 7:54
136
Copyright © 2007 by Morgan & Claypool All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted in any form or by any means—electronic, mechanical, photocopy, recording, or any other except for brief quotations in printed reviews, without the prior permission of the publisher. Adaptive Mesh Reﬁnement for TimeDomain Numerical Electromagnetics Costas D. Sarris www.morganclaypool.com
ISBN: 1598290789
paperback
ISBN: 9781598290783 paperback ISBN: 1598290797 ebook
ISBN: 9781598290790 ebook
DOI:
10.2200/S00052ED1V01Y200609CEM011
A Publication in the Morgan & Claypool Publishers series SYNTHESIS LECTURES ON COMPUTATIONAL ELECTROMAGNETICS #11 Series Editor: Constantine A. Balanis, Arizona State University ISSN: 19321252 (Print) ISSN: 19321716 (Electronic) First Edition 10 9 8 7 6 5 4 3 2 1
Adaptive Mesh Reﬁnement for TimeDomain Numerical Electromagnetics
Costas D. Sarris
University of Toronto Ontario, Canada
SYNTHESIS LECTURES ON COMPUTATIONAL ELECTROMAGNETICS #11
M &C
Mor gan
& Cl aypool
Publishers
the dynamically adaptive mesh reﬁnement (AMR) FDTD technique. . the Haar wavelet basis. along with challenges related to its application to the analysis of reallife microwave and optical structures. KEYWORDS Computational electromagentics. couplers. the most popular electromagnetic ﬁeld solver of the timedomain form of Maxwell’s equations. engineers.iv ABSTRACT This monograph is a comprehensive presentation of stateoftheart methodologies that can dramatically enhance the efﬁciency of the ﬁnitedifference timedomain (FDTD) technique. Therefore. ﬁnite difference time domain (FDTD). The speciﬁc techniques presented in this book are characterized by stability and robustness. optical waveguides. employed in a wide range of exciting applications in microwave/millimeterwave and optical engineering. a number of applications are presented. and graduate students. Optical power splitters. cavities). It is designed to provide a solid reference for highly efﬁcient timedomain solvers. known to reduce simulation times by orders of magnitude. multiresolution time domain (MRTD). It then proceeds to developing an adaptive mesh reﬁnement method based on the use of multiresolution analysis and. While mesh adaptation is an extremely desirable FDTD feature. is introduced and explained in detail. which amount to varying the total number of unknown ﬁeld quantities in the course of the simulation to adapt to temporally or spatially localized ﬁeld features. this book presents advanced concepts and cuttingedge modeling techniques in an intuitive way for programmers. These methodologies are aimed at optimally tailoring the computational resources needed for the wideband simulation of microwave and optical structures to their geometry. microwave simulation. Furthermore. Building on ﬁrst principles of timedomain electromagnetic simulations. and couplers Optical ring resonators Nonlinear optical waveguides. spiral inductors. they are excellent computer analysis and design (CAD) tools. it is not always robust. a new method to embed a moving adaptive mesh in FDTD. as well as the nature of the ﬁeld solutions they support. The book starts by introducing the FDTD technique. Yjunctions. more speciﬁcally. To highlight the properties of the theoretical tools developed in the text. including: • • • • Microwave integrated circuits (microstrip ﬁlters. That is achieved by the development of robust “adaptive meshing” approaches. adaptive mesh reﬁnement.
. . . . . .1 Metal FinLoaded Cavity . . . .3 Adaptive Haar Wavelet Simulation of Pulse Compression in an Optical Fiber Filter . . . . . . . . . . . . . . . . . . . . . FDTD Equations . . . . . . . . . . . . . . . . . 57 3. . . . . . . . . . . . . . . . . . . . . . . . .1 Introduction . .2 Connection Algorithm . . . . 7 2. .2 Wavelet Bases . . . . . . . . . . . . . . . . . . .2 Multiresolution Analysis: A brief Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . .1 Introduction . . . . . . . . . . . . . . . .1 Formulation of Haar MRTD Scheme . . . . . . . . 63 3. . . . . 32 2. . . . . . . . . . . . . . . 47 3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2 Mesh Truncation . . . . . . . . . . . . . . . . . . . . . .3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 A Numerical Interface Between FDTD and Haar MRTD: Formulation and Applications . . . . . .6 Numerical Results: Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .v Contents 1. . . . . . . .1 MRTD. . . . . . . . . . . . . . . . 20 2. . . . . . . . . . . . . . . . . . . . . . Introduction . . . . .4 Conclusions . . . . . . 20 2. . . . . . . . . . . . . 51 3. . . . . 9 2. . . . . . . . . . . . . . . . . . . . . . . . .2. . . . .4 TwoDimensional Hybrid ArbitraryOrder Haar MRTD/FDTD Scheme: Formulation .4 Implementation and Performance of Adaptive MRTD . . . . . . . . . . . . . . . . . 9 2. . . . . . . . . . . . . . . . . . . . . .6. . 7 2. . . . . . . . . . . . . . . . . . . . . . . .3 Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4. . . . . . .3 Derivation of TimeDomain Schemes by the Method of Moments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2 WaveletBased FrontTracking . . . . . . . . . . . . . . . . . . . . . .3. . . . . . . . . . . . . . .4. . . . . . . . . . . . . . . .7 Conclusions . . . . . 48 3. . . . . . . . . . . . . . . . . 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 2. . . . . . . . . . . . . . . . . . . . .6. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54 3.2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52 3. . . . . . . . 56 3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3. . . . . . . . . . .1 General . . . . . . . . . . . . . . . . 44 Efﬁcient Implementation of Adaptive Mesh Reﬁnement in the Haar Waveletbased MRTD Technique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 2. . . . . . . . . . . . . 27 2. . . . . . .3.2 Boundary Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47 3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5 Numerical Results: Validation . . 30 2. . . . . . . . . .
. . . . . . . .1 Introduction . . . . . . . . . . . . . . . . . . . . . . . .4. . . . . 82 Dynamically Adaptive Mesh Reﬁnement in FDTD: Microwave Circuit Applications . . . . . . . . . . . . . . . . . . .3. 110 6. . . . . . . . . 109 6. . . . . . . 106 6.5 Dielectric Ring Resonator . . 83 5. . . . . . . . . . . . . . . . . . . .5 Discussion: Stability and Accuracy of AMRFDTD results . . The Dynamically Adaptive Mesh Reﬁnement (AMR)FDTD Technique: Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93 5. . . . . . . . . . . . . . .4 Adaptive Mesh Reﬁnement . . . . . 6. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .80 4. .3 Field Updates on CPBType Boundaries . . . . 79 4. . . 65 4. .4. . . . . . . . . . . .2 AMRFDTD: Overview of the Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . .2 Microstrip LowPass Filter . . . . . . . . . . . . . . . . . . .1 Detection of Cells That Need Reﬁnement .3. . . . . . . . . . . . . . . . . . . . . . . . . .3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73 4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127 5. . . . . . . . . .6 Conclusion . . . . . . . . . . . . . . . . . . . . . . .7 Transition from Old to New Mesh Tree . . . . . .3 Mesh Tree and Field Update Procedure in AMRFDTD . . . . . . . . . . . . . . 71 4. . 83 5. . .1 Introduction . . . . . . . . . 103 6. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78 4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116 6. . . . . . . . . . . .3 Microstrip Branch Coupler . . . . . . . . . . . . . . .7 Conclusion . . .2 Modeling of Wave Propagation . . . 77 4. . . . . . . . . . 74 4. 77 4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4 Microstrip Spiral Inductor . . . . . . . . . . . . . . . . . . . . . . . . . . .6 Field Updates on PBType Boundaries . . . . . . . . . . . . . . . . . . . . . . . 102 Dynamically Adaptive Mesh Reﬁnement in FDTD: Optical Applications and Error Estimates . . . . . . . . . . . . . . . .6 Numerical Error Estimation and Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5 AMRFDTD and MRTD: Similarities and Differences . . . . . . . . . 75 4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103 6. . . . . . . . . .3 Dielectric Waveguide Power Splitter . . . . . . . . . . . . . . . . . . . . . . . . 96 5. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80 4. . . . . . . . . . . .vi CONTENTS 4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1 Multilevel AMRFDTD . 70 4. . . . . . . . . . . . . . . . . . . . . . . . 65 4. . 67 4. . . . . . . .4 Field Updates on SBType Boundaries . . .3. . . . . . . . . . . . . . . . . . . . . . .5 Field Updates at Junctions of SBType Boundaries . . . . 92 5. . . . .4 Dielectric Waveguide YJunction . . . . . . . . . . . . . . . . . . . . . .3. . . . . . . .1 Categorization of Boundaries of Child Meshes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2 Update of the Mesh Tree . . . . . . . . . . . . . . . .4. . . . . . . .3 Clustering. 83 5. . . . 70 4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4. . . . . . . . .2 Dielectric Waveguide with a Corrugated Permittivity Proﬁle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4 Guidelines for Choosing AMRFDTD Parameters . . . . . . . . . . . . . . . . . . . . . . . .3. . . . . . . . . . . 120 6. 76 4. . . . .3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Ann Arbor. Liu was partially supported by the Ontario Centers of Excellence through a Collaborative Research Grant and NSERC through a Collaborative Research and Development (CRD) Grant. Barry Perlman deserving a special acknowledgment as well). Constantine Balanis. University of Toronto. to whom the author is ever indebted for her mentorship and example of achievement through excellence. Rogers Sr. Department of Electrical and Computer Engineering. September 2006 .vii Acknowledgments First and foremost. Sarris. for asking me to contribute to this series. Costas D. and to Joel Claypool for his support to this project. Dr. It was primarily supported by the Natural Sciences and Engineering Research Council of Canada (NSERC) through a Discovery Grant to the author. Toronto.S. Harvey should be personally acknowledged for his continued interest) and the U. Army CECOM (with Dr. Yaxun Liu at the Edward S. I would like to express my gratitude to Prof. Linda Katehi. James F. The research included in the part of this monograph that refers to the MRTD technique was conducted at the University of Michigan. this work was supported by the Army Research Ofﬁce (and Dr. Financially. under the inspiring guidance of Prof. Ontario. The work on the dynamically adaptive mesh reﬁnement ﬁnitedifference timedomain (AMRFDTD) technique is the result of a productive collaboration of the author with Dr. All this generous support is gratefully acknowledged.
.
. . . . . . . . 19 Validation Data for Nonadaptive MRTD Code .4. . . . . . . . . . . . . . . . . . . . . . . .2 6. . . . . . . . . . . . . . . . . . . . 110 Computation Time and TimeDomain Error Metric Et for the Dielectric Waveguide YJunction for Fields Recorded at Port 2. . . . . 6. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6 Position of and Field Values at the MRTD Equivalent Grid Points of Fig. . . . . . . . 107 Computation Time and TimeDomain Error Metric Et for the Dielectric Waveguide Power Splitter for Fields Recorded at Port 2. . . .1 3. . . . . . . . . . . . . . . . . . . 99 Computation Time and TimeDomain Error Metric Et for the Optical Waveguide of Fig. . .1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3 6. . . . . . . . . . . . .ix List of Tables 2. . . . 121 . . . . .3 6. . 90 Computation Time and SParameter Error Metric E S for the Microstrip Branch Coupler . . . . . . . . . . . . . . . . . . . . . . .10. . . . . . . . . . . . . .1 6. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1 5. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56 Computation Time and SParameter Error Metric E S for the Microstrip LowPass Filter . . . . . . 105 Computation Time and TimeDomain Error Metric Et for the Dielectric Waveguide of Fig. . . . . . . . . . . . . . . . 117 Resonant Frequencies (in THz) of the Dielectric Ring Resonator. . . . . . . . . . .2 5. . . . . . . . . . . .5 6. . . . . . . . . . . . 97 Computation Time and SParameter Error Metric E S for the Microstrip Spiral Inductor . . . . . . . . . . . . . 114 Computation Time and TimeDomain Error Metric Et for the Dielectric Ring Resonator for Fields Recorded at Port 2. . . . . . . . . . . . . . . . . . . . . . . . . . 6. . . .1 5. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2. . . . . . . . . . .4 6. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.6 2. . . . .15 2. . . . . . . . . . . . . . . . .xi List of Figures 2.14 Battle–Lemarie scaling and mother wavelet functions in the Fourier domain . . . . . . . . . . . V2 . . . . . . 15 Explanation of deﬁnition (2. . 12 Demonstration of the multiresolution principle as implemented by the Haar basis . . . . . . . 13 Haar scaling and mother wavelet functions in the Fourier domain . . . . . . . . . .) within a scaling cell as introduced by a ﬁrstorder wavelet expansion in x. . . . . .16 2. . . . . .13 2. . .17 2. . . . . . . . . . . . . . . . . . .11 Coarse rectangular mesh for a layered coplanar waveguide structure. . . . . V1 . . .by 3order MRTD/FDTD mesh conﬁguration (dimensions are given in FDTD cells. . . . . . . . . . . 18 Electric/magnetic ﬁeld equivalent grid points for zeroorder Haar MRTD in one dimension. . . . . . . . . .12 2. . . . . . . . . . .by 4order MRTD/FDTD mesh conﬁguration (dimensions are given in FDTD cells. . . . 13 Battle–Lemarie scaling and mother wavelet function. . . . . 31 Time and frequency domain patterns of electric ﬁeld (Ey ) sampled within the cavity of Fig. . . . . . . . . . 23 Onedimensional interface between FDTD and a ﬁrstorder Haar MRTD scheme .2 2. . . . . . . 10 Haar scaling and mother wavelet function . . . . . . . . . . . . . . . . . . . . . . . . .18 Mesh reﬁnement in the Haar basis . . . . .3 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . .p. . . . . . . . . . . . . . . MRTD mesh is 2 × 2) . . .4 2. . . . . . . . . . . . . 2. . . . . . . . . . . . . . 29 MRTD update from FDTD data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .17. . .22) for the Haar basis. . . . . . . . . . . . . .19 . . .10 2. . . . . . . . . . . . . . . . . . . . 28 Reﬂectionless propagation through a onedimensional FDTD/fourthorder MRTD interface . . . . .14 2. . . . . .9 2. . . . . . . .1 2. . . . . . . . . . . . . . . . . . as obtained by FDTD and the hybrid scheme . . . . . . . . . . . . . . . . . . . with cells containing variable dielectric permittivity and electric conductivity proﬁles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 Empty rectangular cavity geometry and interface of a 3. . . . . . . . 30 Empty rectangular cavity geometry and interface of a 3. . . . . . . . . . . . . . . . .5 2. . . . . . . . . . . . . MRTD mesh is 1 × 1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . with electric/magnetic scaling cell offset chosen in consistence with (2. . . . . . . . . . .18 2. under the common convention of half a cell offset between electric/magnetic scaling cells . . .8 2. . 32 2. 8 Projections of a function f on three successive approximation spaces V0 . . . . 29 FDTD update from MRTD data . . . . . .g. . . . .7 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 Electric/magnetic ﬁeld equivalent grid points for zeroorder Haar MRTD in one dimension. . . . . . . . . zdirections . . . . . . . 14 Equivalent grid points (e. . . . . . .33) . . . . . . . . . . . . . . . . . . . . . .
the FDTD region terminated into the ABC should at least extend over one scaling MRTD cell . .21 Empty square cavity geometry and interface of a 4. . . . . . . . . . 39 2. . . . . . . . . .xii MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS 2. . . . . . . 2. . . .18. . . . . . .32 FDTDUPML interfaced with a 0. . .24 Metal ﬁnloaded cavity geometry and interface of a 2. . . . in the case of onedimensional wave propagation (Fig. . . . . . . 38 2. MRTD mesh is 1 × 1) . . . . . . .by 4order MRTD/FDTD mesh conﬁguration (dimensions are given in FDTD cells. . . . . . . . . . . . . . . . . .22 Electric ﬁeld pattern for the T E22 mode of the empty cavity of Fig. . . . .39 Concept of ABC termination of an MRTD mesh. .31 at time steps 20. . . . . . . . . . . . . . sampled in front of the dielectric slab of Fig. . 2. . . . 8.000 . . . . .000–20. . . . 44 . . . . . . . . . . . . . . . . . . .39). . . . . . . . . . . .by 2order MRTD (MRTD mesh is 8 × 8) . . . . . . . . . . 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2. . . . . . as obtained by FDTD and the hybrid scheme . when a 4. .34 FDTDUPML interfaced with a 4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .34 . . . . . . . . . . . . . .36 Electric ﬁeld at point B of Fig. . . . . . . . . . . . . . 42 2. . . .25 Electric ﬁeld distribution in the metal ﬁnloaded cavity (dominant TEmode) . . . . .38 Spatial ﬁeld distribution in the interfaceterminated domain of Fig. . . .43 2. . . . . . for MRTD orders 0–3 . . . . . . . . . . . . . . . . . . . .33 FDTDUPML interfaced with a 2. . . . MRTD mesh is 4 × 4) . . . . . . . . . . . . . . . . . . . . . 34 2. . . . .by 2order MRTD/FDTD mesh conﬁguration (dimensions are given in FDTD cells. . . . . . . . . . . . . .by 0order MRTD (MRTD mesh is 32 × 32) . . . . . . . . . . . . . . . . . . . . . . . . 2. . . . . .35 Comparison of FDTD and MRTD/FDTD interface results for the three case studies of Figs. . . . . . . . . . . . . 42 2. . . . . . . . . . . . . . . . .37 Spatial ﬁeld distribution in the interfaceterminated domain of Fig. 16 FDTD UPML is used to terminate the MRTD mesh. . . . . . . . . . . . .28 Slab waveguide geometry . . . . .31 Case study for the interfacebased termination of a twodimensional MRTD domain . . . . . . . 41 2. . . . . . . . . . .30 Numerical and theoretical S11 for the slab geometry . . .by 4order MRTD (MRTD mesh is 2 × 2) . . . .33. . . . . . . . . . . . . . . . . 2. . . . . . . . 35 2. . . . . . . . . . . . . . .21 (4 by 4 MRTD) . .27 Concept of MRTD mesh termination via an FDTD/MRTD interface . . An FDTD solution (with a 32cell UPML) is appended for comparison . . . . . . . . .20 Time and frequency domain patterns of electric ﬁeld (Ey ) sampled within the cavity of Fig. . . . . . . . . . . . 2. . . . . . . . . . . .32–2. . 40 2. . . . . . .40 Incident and reﬂected ﬁeld waveforms. . 41 2. . . . . . . . . . . . . .26 Electric ﬁeld sampled within the ﬁnloaded cavity as a function of time for time steps 18. . . 2. .34 2. . . . . . . . . . . . 40 2. . . . . . . . . . . . . . . . . . . . . .29 Waveforms of incident and reﬂected electric ﬁeld (scaling and wavelet terms of order 0–2). . . . . . . 35 2. . 36 2. . . . . . . . 39 2. . . . . . . . . 43 2. . . . . . . . . . . . . . . . . . . . . . . . .28 . . . . . . . .23 Electric ﬁeld pattern for the T E22 mode of the empty cavity of Fig. . . . 600. . . . . . . . . . . . . . . . . . . . . . . . . . . .21 (4 by 4 MRTD). . . 33 2. . . . . . . . 2. . . . . 33 2. 50 . . . . . . . . . . . 37 2. . . . . . . . . . for Mur’s ﬁrstorder ABC (implemented via an FDTD/MRTD interface). . . . . 38 2.31 at time steps 80. . . . . . . . . . .
. . . . . . . . . . . . . . . .39) . . . . and the end of the optical ﬁber ﬁlter (1000 grid points) . . . . . . . . . derived by adaptive MRTD with = 10−5 . . . . . 57 3.1 (a) Geometry and (b) data structure of a threelevel mesh tree. . . . . . . 10−1 . . 2. 2. . . . D. . . . . . . .28 . .7 MRTD and FDTD results for the forward ﬁeld intensity at the beginning. .2 Sampling points of electric and magnetic ﬁeld components in a Yee cell. . . . . . . . . . . .5 Interface between two meshes of the same level. . 55 3. . . . .7 Transition from old mesh tree to new mesh tree. . . . . .1 applied to Haar MRTD . . . . . 2. . . . . . for a fourthorder Haar MRTD simulation of the problem of Fig. . . . . . . . . . . .9 Spatiotemporal evolution of forward ﬁeld wavelet coefﬁcients (square of magnitude of wavelet coefﬁcients is shown) . . . . 61 3. . . . . . . . . . . . . .28 . . . . . . The hollow arrows are for the parent mesh. . . . . . .11 Incident and transmitted forward ﬁeld intensity for a threshold of 10−7 . in the case of onedimensional wave propagation (Fig. . . . . . . . . . . . . . . . . . . .13 CPU time economy for the adaptive MRTD code. . . . . . . . . . . 51 3. . . .6 Haar MRTD equivalent grid points at the beginning and the end of the ﬁber . . . . . . . . . . . . . . . A. . . . . . . for a fourthorder Haar MRTD simulation of the problem of Fig. . . . . . . . .14 Relative error (%) in the peak intensity of the transmitted forward ﬁeld for adaptive Haar MRTD . . . . . . . . . The shaded areas represent reﬁned Yee cells. . . . .1 Unthresholded to total number (%) of wavelet coefﬁcients of orders 0. . . . . . . . . . . . . . . 77 . . . . . . . a threshold = 10−4 is used) . .10 Graphical presentation of the waveletbased. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . with respect to FDTD. . . . . . . . . . . . . . . . 3. .6 Typical SBtype boundaries. . . . .28) and regions of unthresholded wavelet coefﬁcients (boundaries are deﬁned by blue dots. 71 4. . . . . . . . . . . . . 2. . . . . . . . . . 52 3. . . . 75 4. 62 4. .4 Sampling points for electric ﬁeld on CPBtype boundary. . . . . . .3 Types of boundaries of a child mesh. 58 3. . . . . . . . and E are meshes of the same level. . . . . . . . . . . . . . . . . . . . . . 2. . . . . . . . . 10−4 for the dielectric slab incidence problem of Fig. . . . . . . . . . . . . . . . . . . . . . . .5 Evolution of wavefronts in the dielectric slab simulation (continued) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58 3. . . . . . . . 60 3. . . . . . . . . . . . . . . . . . . . 62 3. . 1. . . . . . . . . . . C. . . . 50 3. . . . . . . . . 45 3. 68 4. . . . . . . . . . . . .41 Reﬂection coefﬁcient for Mur’s ﬁrstorder ABC (implemented via an FDTD/MRTD interface). . . . . . . . . . . whereas hollow arrows are for magnetic ﬁeld. . . . . . 75 4. . . . . . . . . . . . . . . . . . . .3 Comparison of S11 . . . . . . . . . . . . . . . . . . . for absolute thresholds from 10−7 to 0. . . . . . . . . . The solid arrows are for the child mesh. . . . . . . . . . . . . Solid arrows are for electric ﬁeld. .12 Transmitted forward ﬁeld intensity for thresholds of 10−4 . . B. . . . 69 4. . . . . .8 Spatiotemporal propagation and compression of the nonlinear optical pulse (forward ﬁeld intensity is shown) . . . . . . . . 61 3. . . . . . . . . . . . . . . . . . . . . .LIST OF FIGURES xiii 2. . 49 3. . . . fronttracking algorithm . the middle. . . . . . . . . . . . 73 4. . . .4 Evolution of wavefronts in the dielectric slab simulation (Fig. . . . . . .28 . . . .2 Unthresholded to total number (%) of wavelet coefﬁcients of orders 2. . . . . . . 49 3. . . . . . . . . . . . . . . . . . . . . . .
. . . σAMR = 2 . .1 5. . . . . . . . . . . . . . . . . . . . . . . . . . 87 Vertical electric ﬁeld magnitude at z = 0. . . . 5. . . . . . . . . . NAMR = 10. . . . . . . . . . . . obtained by FDTD and AMRFDTD . . . . . . . . . . . . . . . . . . . . . . . . .4 mm and the center of the microstrip line. . . . . . 84 AMR coverage and number of child meshes at different time steps for the lowpass ﬁlter simulation . . . . . . . 92 Microstrip branch coupler geometry. . . . . . . as determined by FDTD and AMRFDTD .01.8 5. . . . . . . . . 96 . . . . . . . . . . . . . . . . . . . P1 and P2 denote ports 1 and 2. . . . . . . . . .11 5. . .4 mm and t = 100 t. . . . . . . . . . . . . 5. . .13 5. . . . . . . . 3 mm from the right edge . . . . . . . . . . . . . 5. . . . . . . . . . . . . . . . . . . .2 5. . . . . . . .6 5. 91 Effect of spreading factor of AMR σAMR . . . . NAMR = 10 . . . . . . . . . . . . . . . . .17 5. . . . . . 91 Effect of threshold for coverage efﬁciency θc . . . . . AMR coverage is 13. . . The rest of the AMR parameters are: θg = 0. . . θc = 0. . . . .4% . . . . P2. . . . . . NAMR = 10. . . .9 5. . . .3 5. . .5 5. . . . . . . .0% . . . . . . . . . . . . . . . P1. . θc = 0. . . . . . . . . . . . . . . . σAMR = 2 . . . . . . . . . . . . . . . . . 89 Effect of threshold for gradient θg . . . . . . . . 94 S21 for the microstrip branch coupler geometry of Fig. . AMR coverage is 52. . . . . . . . . . . . . . . .7. . . . . .15.01. . . . . . . . . . . . . . . . . . . . . NAMR = 10. . . . . .5% . . . . . . . . . . 86 Vertical electric ﬁeld magnitude at z = 0. . . . .4 mm and t = 0. . . 85 Vertical electric ﬁeld magnitude at z = 0. . . The rest of the AMR parameters are: θg = 0. . . .1. The rest of the AMR parameters are: θg = 0. . . . . . . . . . . . Three child meshes. . . . . . . . . . . . . . . . . . . . . .7. . . as determined by FDTD and AMRFDTD . . . . . . . . . . AMR coverage is 44. .15. . . . . . . .1. . . . . . . .14 5. 2. . . . . . . . Three child meshes. . respectively . . . . . . . . . . . . . . . . . . . . . . . . 81 Microstrip lowpass ﬁlter geometry. . .01. . . . . . . . . . . θc = 0. . . θe = 0. . . 5. and P4 denote port 1. . . 95 S31 for the microstrip branch coupler geometry of Fig. . . 90 Effect of threshold for energy θe . . . . . . . .7. . . .16 5. . . . . . . . . . .18 Splitting and shrinking of boxes (child meshes). .01. . . .4 mm and t = 200 t. . . . . . . . . 86 Vertical electric ﬁeld magnitude at z = 0. . . . . . . . . . . . .12 5. 3. . θe = 0. . The rest of the AMR parameters are: θg = 0. . . . . . and 4 . . . . . . . . . . . . . .2% . AMR coverage is 4. . . .4 mm and t = 800 t.4 mm and t = 500 t. .8 5. .3% . . . 91 Effect of interval of AMR NAMR . . . . . One child mesh. . . . . .15 5. . . . . as determined by FDTD and AMRFDTD . . . . . . . . . . . . . . . . . .01. . . . . . . . . . .01. . . . . . . . . The rest of the AMR parameters are: θe = 0. . . . . . σAMR = 2 . σAMR = 2 . .15. . . . . . .10 5. . . . . P3. . . . . . . . . . . . . . . . . . . . . .xiv MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS 4. . . 87 Vertical electric ﬁeld magnitude at z = 0. 93 S11 for the microstrip branch coupler geometry of Fig. . . . . . . . . . . . . . . . . . . . .4 5. .01. . . . . . . . . . . . . . . . . . . . . . . . . . . . . AMR coverage is 28. . . . . . . . . . . . . . . . One child mesh. . . . . . . . 85 Vertical electric ﬁeld magnitude at z = 0. . . . . . . . . . . . .7 5. . . . . . . . . θe = 0. . . . . . . . . . . . . . . . . . . . . . . . . 88 Comparison of the scatteringparameters of the lowpassﬁlter of Fig. . θc = 0. . . . .7. . . One child mesh. .
5. . .7 . . . . . . . . . . . and 4. at t = 100 t. .7 . .6 mm and the center of the microstrip line of port 3. Dimensions are given in units of μm. . . . . . . . . . . . . . . . . . . . . . . . . 6. . . . . 1. . . 6. .5 Electric ﬁeld at the center of the dielectric waveguide of Fig. . . . . . . . . . . . P2. Dimensions are given in units of μm. . . . . . . respectively . . . . . . . . .108 6. . . . . . . . . . . . . . . . . . indicating the absence of latetime instability . . .15. . . . . . . . . . 6. .4 Geometry of the dielectric waveguide with a corrugated permittivity proﬁle. . . . . .3 Temporal waveform of the vertical electric ﬁeld magnitude at the center of the optical waveguide of Fig. . . . . .20 Vertical electric ﬁeld magnitude at z = 0. . . . . . . .6 mm and the center of the microstrip line. . .9 Electric ﬁeld at 1 μm from the edge of port 3 of the power splitter of Fig. . . . . . . . . . . . . . . . . . . . 1. . . . . obtained with the dynamic AMRFDTD. . . . . . . . . .2 μm from the right edge. . . . . 6 mm from the right edge. . . . . for 40. . . . . . . . . . . .1 Optical waveguide geometry . .2 Vertical electric ﬁeld magnitude distribution across the optical waveguide of Fig. . . . . . indicating the absence of latetime instability . . . . . . . . . The excitation is imposed at 3 mm from the edge of port 1 . . . . . . . 101 6. . . . . . . . . P1. . . . . .19 S41 for the microstrip branch coupler geometry of Fig. . . .2 μm from the right edge. . . . . . . . .000 time steps. . . .6 Electric ﬁeld at the center of the dielectric waveguide of Fig.000 time steps. . . . . . . One level2 child mesh and one level3 child mesh are shown . . . . . . . as determined by FDTD and AMRFDTD . . . simulated by a twolevel dynamic AMRFDTD .LIST OF FIGURES xv 5. . . . . . . . . . . . . 2. . . . . . . . . . . . . . 3. . . . . . 97 5. . . 6. .4. . . . . . . . . . . . . . . . . . . . . . for 40. . . . . . . . . . . P1 and P2 denote ports 1 and 2. . . . . . . . . . . . . . . . . . . . 108 6. . . . P3. . . 99 5. . . . . . . . . . . . . . . . . . . . . . 111 6. . . . . . . .1. . . . . . .5 μm from the right edge . . . . . . . . . . . . . . . . . . . The excitation is imposed at 3 mm from the left edge of the microstrip line . . . . . . . 1. . .23 Sparameters for the spiral inductor geometry of Fig. . . . . . . . . . . . . . . 100 5. . . . . . . 109 6. . . . . . . . . . . . . . . . . . .4. . . 6. . . . . . . . . . . . . . . . . . . . 104 6. . . . 106 6. . . . . . . . . . . . . . . . .21. . .21 Spiral inductor geometry. . . . . . .24 Vertical electric ﬁeld magnitude at z = 0. . . . . . . 110 6. . . . . . . . . . . . . . . . . . . . . . 104 6. . .7 Geometry of the dielectric waveguide power splitter. . . . . . . . . . . . . . 6. . . 5. obtained with the dynamic AMRFDTD. . 6. as determined by FDTD and AMRFDTD . . . . . . . 105 6. . .7. . . . . . . . . . . . . . . . . . . 111 . . . . . .1. . . . . . . . . . .8 Electric ﬁeld at 1 μm from the edge of port 2 of the power splitter of Fig. . P1 and P2 denote port 1 and port 2.10 Electric ﬁeld obtained with the dynamic AMRFDTD at 1 μm from the edge of port 2 of the power splitter of Fig. . . . . . . . . . . . . . . . . . . . . respectively . . . . . . . . 6 mm from the edge. . . . . . . . and P4 denote ports 1. 98 5. . . . . . . . . . . . . .22 Vertical ﬁeld evolution and associated mesh reﬁnement in the microstrip spiral inductor. . . . . . . . . . . . . . . . . . respectively . . . 98 5. . . . . . . . .
. There are three meshes in total: one level1. . . one level2.21 Geometry of the dielectric ring resonator coupled to two dielectric waveguides. . indicating the absence of latetime instability . . . obtained with the dynamic AMRFDTD. . . . . 114 6. . 115 6. . . . . . .14. . P1. . . .22 Electric ﬁeld at 2 μm from port 2 of the ring resonator of Fig.11 Electric ﬁeld magnitude distribution across the power splitter of Fig. and one level3 mesh . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .21 . . . . . . . .xvi MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS 6. . . . . . . . . . . . . . . . 112 6. . . . . .7. 3. . . . . . . . . . . P2. 2. . . . 6. . Dimensions are given in units of μm. . . . . . P3. .14. . . . . . . . . . . . . . . 6. . . . . . . . . . . and one level3 mesh .12 Electric ﬁeld magnitude distribution across the power splitter of Fig. at t = 100 t. . . .14 Geometry of the dielectric waveguide Yjunction. . . The dielectric constant of the waveguide is 9 and the dielectric constant of the surrounding medium is 1. . . . . . . . . . . . . . . . .14. . one level2. . . . . . . . . . . . . . . . 115 6. . . and one level3 mesh . . . . . . . 6. . . . . . .16 Electric ﬁeld at the center of the dielectric waveguide Yjunction of Fig. . . . . . . . . 6. . . . . . . . . . . .21 . . . respectively . . . . . . . . .20 Electric ﬁeld magnitude distribution across the Yjunction of Fig. . . P2.14. . . . . . . . 2. .7. . . . . P1. . 6. . . . . . . . . . . . . . . . . . and 3.14. . 113 6. . . and 4. . . . . . . . . . There are three meshes in total: one level1. . . . . . . . . There are three meshes in total: one level1. . . . . . . 6. at t = 600 t. . and P4 denote ports 1. . . . . and one level3 mesh (enclosed in the black solid line). respectively . . . . . . . 6. . 2 μm from the edge of port 2 . . . . . 116 6. . . . . at t = 800 t. and four level3 meshes (tracking transmitted and reﬂected wavefronts) . . . . . . .18 Electric ﬁeld magnitude distribution across the Yjunction of Fig. . . . . . . . . . . 6. . Dimensions are in units of μm. . . . . . . . . 6. . There are three meshes in total: one level1. . . . .000 time steps. for 25. . 113 6. . . 116 6. . . . . 112 6. . 6. . and one level3 mesh (enclosed in the black solid line) . There are three meshes in total: one level1. three level2. . . . . . . . . . . . . .13 Electric ﬁeld magnitude distribution across the power splitter of Fig. . . . one level2. . There are four meshes in total: one level1. .23 Electric ﬁeld at 2 μm from port 2 of the ring resonator of Fig. .14. . . . . 2 μm from the edge of port 2. . . . . . . . . one level2. . . two level2. . . at t = 100 t. . and P3 denote ports 1. . . . . and one level3 mesh (enclosed in the black solid line) . . . . . . . . . . one level2. . . . 6. . . . . . . . . . . . . .17 Electric ﬁeld magnitude distribution across the Yjunction of Fig. 119 . . . . . . Note that level2 and level3 meshes track the pulse into the junction . . . . 117 6. . . . . . . . . . . . . . at t = 300 t. . . . . . . . . . . . .19 Electric ﬁeld magnitude distribution across the Yjunction of Fig. at t = 400 t. . . . . 118 6. . . . . . . . . . . . There are eight meshes in total: one level1.7. . . . . . . 115 6. . . . . . . . . . . . . . . . at t = 400 t. . . . . . . . . .15 Electric ﬁeld at the center of the dielectric waveguide Yjunction of Fig. . . .
. . . .25 S21 obtained with the fourlevel AMRFDTD . . . . . . . . . . . obtained with the dynamic AMRFDTD.31 Electric ﬁeld at port 3 and 4 of the directional coupler . . . . The dielectric constant of the waveguide and thecenter piece is 3. . . . . . . . . . . . . .LIST OF FIGURES xvii 6. . . . . . . . . .24 and 4. . . . . . .21. The surrounding medium is air. . . . . . . . . . . . . . .24 Electric ﬁeld at 2 μm from port 2 of the ring resonator of Fig. . . 122 6. . . . . . . . (b) σAMR . . . . .26 Error versus simulation time for variable (a) NAMR . . . . . . . . . . . . . . . . . . . for 100. . . . . . . . . . . . . 123 ˜ 6. . 125 6. . . . . . . . 120 6.27 Dependence of −lo g 10 E on θe and θg . .29 Curveﬁtted dependence of error bound −lo g 10 Eb on θe and θg based on (6. . . . . . . 126 . .000 time steps. . . . . . . . . . . . . . . . . . . . . . . . . .30 Geometry of a directional coupler (not drawn to scale). . . . 124 6. . . . . respectively . 125 6. . . . . . . . . . . . 120 6. . . 6. . . . . . . . . . . . . . . . . . . . . . . . . Dimensions are given in μm. . . . . . . . . . . .6) . . . . . . . . .41.6) . . . . . . . . . . . . (c) θc .28 Curveﬁtted dependence of −lo g 10 E on θe and θg based on (6. . . . indicating the absence of latetime instability . . .
.
which are based on representing multielement structures as an assembly of cascaded. numerical schemes are developed by formulating an operator problem and subsequently pursuing its solution by computational (as opposed to analytical) means. proceed to the formulation of a solution to a given boundary value electromagnetic problem. classiﬁes accordingly the related scheme. which assumes in general the form of a dyad G(r . multiplexers). unconventionally shaped objects [1]. that serve the need of the wireless industry for light and small transceiver structures well. densely integrated architectures. r ). for high power/lowloss. In addition. These provide lowcost. As a result. The question of the operator itself being either an integral or a differential one. approximate circuit design approaches. Yet. with no assumptions on the shape of the wavefronts involved. The profound need to model and eventually mitigate such effects fuels the research on fullwave techniques of computational electromagnetics (CEM).. tactical and commercial communication systems. individually functioning blocks. the latter produces parasitic effects such as crosstalk and signal distortion. The next steps beyond the current stateoftheart include the monolithic integration of silicon/silicon germanium (SiGe) circuits. ﬁlters. Electromagnetic problems are cast in the form of an integral equation by ¯ ¯ use of a Green function. do not constitute a reliable simulation tool for cuttingedge applications.g. Despite the fact . In essence. microelectromechanical switch (MEMS) devices. circuit miniaturization allows for the implementation of systemonachip concepts. highfrequency asymptotics assume the reduction of wavefronts to rays. Methods that are referred to as such. facilitating the application of optics principles to problems of scattering from electrically large. and digital CMOS signal processing units. lowloss but bulky components. lightweight and reproducible alternatives to conventional. that generate signal integrity issues with an adverse impact on the overall performance of wireless communication systems. due to their failure to account for the inevitable electromagnetic coupling of closely spaced components.1 CHAPTER 1 Introduction Recent advances in solidstate device technology have enabled the development of planar monolithic microwave integrated circuits (MMICs) in multilayer. micromachined components (e. On the other hand. resonators.
since Yee’s scheme [6] is secondorder accurate only. Timedomain differential methods are becoming increasingly popular among the electromagnetic community as a consequence of their versatility and their ability to provide simulation results that are intuitively meaningful to circuit designers and microwave engineers. in realistic cases. In addition. have enhanced the capability of frequencydomain techniques to produce broadband results in certain cases. memory and execution time consuming calculation. when narrowband simulations are pursued. the dominant integral equation solver is arguably the numerical method of moments [3]. Therefore. However. aimed at the discretization of electromagnetic structures at rates that . translate into largescale algebraic systems. and sensitive to numerical dispersion. but usually twentyﬁve points per wavelength is necessary for the extraction of a convergent solution. if tractable at all. that trade versatility with computational effort. To this end. such as a Gaussian pulse. frequency sweep techniques. reasonably accurate techniques. the ﬁnitedifference timedomain (FDTD) technique offers a mathematically straightforward analysis method. with computational requirements that can be met by nowadays computer performance. [7]) method that solve partial differential equations in a mesh. a dense discretization of at least ten. suitable for arbitrary electromagnetic geometries. such as the asymptotic waveform evaluation (AWE. This fact naturally poses the signiﬁcant question of deriving new. the FDTD treatment of either electrically large geometries or ﬁne detail structures typically results in a computationally intensive. In fact. is injected as an initial condition to the computational domain. In addition. naturally giving way to differential methods. Then. until the steady state is reached. frequencydomain techniques extract ﬁeld information at a given frequency point. suffering though from general issues of accuracy and efﬁciency. [5]). [6]). [8]). the WienerHopf method [2]. it inherently accounts for Sommerfeld’s radiation condition [4]. A Green function incorporates an intuitively rich description of the physics of a given problem. Clearly. In particular. Nevertheless. a marchingintime procedure is followed to calculate the evolution of its wavefront. or the transmission line matrix (TLM. thus prompting an everincreasing research interest in those.2 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS that special forms of integral equations can be treated by analytical techniques. a broadband impulse excitation. As an alternative to the conventional FDTD. among others. techniques such as the ﬁnite element method (FEM. several higher order numerical techniques have been developed [9]. On the contrary. recent advances in highperformance and parallel computing have made these methods a feasible alternative. This monograph focuses on the solution of Maxwell’s equations in the timedomain. the latter are preferred to the former. rendering the use of absorbing boundaries unnecessary. such as. the ﬁnite difference timedomain (FDTD. However. the use of Green dyadics becomes extremely complex.
pulse functions are used for the temporal ﬁeld expansion. The concept of adaptivity implies the possibility of dynamic mesh reﬁnement at regions of the computational domain that are electromagnetically active. In [11]. that render the rigorous enforcement of the divergence free nature of the magnetic ﬁeld and the continuity conditions a rather subtle issue. It is worth noting though. It is noted that spatial variations of ﬁeld components are reﬂected on the magnitude of their respective wavelet coefﬁcients. that effectively deﬁne a discrete mesh in space. Daubechies. According to MRTD. beyond the limits that a static subgridding scheme can provide for. Furthermore. Signiﬁcant memory economy was shown to be associated to this method. that these computational gains were strictly associated to the Battle–Lemarie scaling function rather than the Battle–Lemarie wavelet function basis. and Haar wavelets have been presented. As the timedomain characterization of planar structures typically registers the history of a wideband (narrow) pulse propagation. Battle–Lemarie scaling functions enhanced by one wavelet level (zeroorder wavelet functions) were employed for the formulation of the socalled WMRTD scheme. several types of MRTD schemes with distinct numerical properties can be derived. space adaptivity via an a posteriori . biorthogonal. a third. standard subgridding techniques involve spatiotemporal interpolations or extrapolations at the boundaries of different resolution parts of the numerical grid. but equally important one.INTRODUCTION 3 may even approach the Nyquist limit. In fact. that allowed for the modeling of electromagnetic structures at discretization rates that approached the Nyquist limit (λ/2). These ﬁnite difference equations are solved by “marching in time. depending on the wavelet basis that is employed for the discretization of Maxwell’s equations. adaptively imposing dense gridding conditions only in and around the pulse and the products of its retroreﬂections can further extend the efﬁciency of a technique. Upon substitution of these expansions into Maxwell’s equations. henceforth referred to as multiresolution timedomain (MRTD). in [11–16] and references therein. from the size of the smallest geometric detail that is contained in the latter.” in the FDTD sense. because of its highly linear dispersion properties. for example. Galerkin method is applied to derive ﬁnite difference equations with respect to the ﬁeld scaling and wavelet spatiotemporal coefﬁcients. Furthermore. which are either stored in computer memory or removed according to certain criteria (thresholding). Yet. the incorporation of subgridding algorithms in the FDTD scheme [10] has indicated the potential of disconnecting the mean cell size within a domain. Recently. employing Battle–Lemarie. is adaptivity. If allowing for a relatively coarse mesh and taking into account localized geometric details are two challenges that novel numerical schemes are expected to meet. the electromagnetic ﬁeld components are spatially expanded in a basis composed of scaling and wavelet functions. In addition. waveletbased timedomain methods.
This monograph is aimed at presenting two efﬁcient solutions to the problem of establishing mesh adaptive timedomain numerical electromagnetic techniques. even when electrically small geometric features are to be included. by employing a highorder basis (such as cubic splines). seriously compromises their potential applicability to stateoftheart devices. includes printed transmission line analysis [18]. and antenna geometries [20]. Because of their relative simplicity. It is worth noting that no interpolations or extrapolations (spatial or temporal) are necessary for the establishment of this interface. Then. in Chapter 2. In addition. Since Haar MRTD constitutes the “wavelet extension” of the FDTD technique. no stability issues are associated with its application. the range of WMRTD applications that have been demonstrated so far. First. an inherent contradiction related to MRTD schemes is that although their efﬁciency is expected to increase with the order of the multiresolution expansion at hand for homogeneous domains. by adding wavelets and using adaptivity algorithms. wavefronttracking ideas are introduced and applied to onedimensional linear and nonlinear optical wave propagation problems. which is an attractive feature of an algorithm especially when it is combined with parallelization/domain decomposition techniques. Moreover. applicationoriented question. the popularity of Haar MRTD has recently grown in the computational electromagnetics community. also increasing with wavelet order. complex airdielectric boundary problems. a straightforward implementation of adaptive gridding is obtained. It is worth noting that wavelet methods meet all aforementioned challenges for timedomain numerical schemes. dielectric. a great deal of FDTD algorithms can be modiﬁed in order to resolve Haar MRTD modeling issues. Then. a numerical interface between FDTD and the Haar waveletbased MRTD is presented in detail. ﬁnite difference equations in Haar MRTD involve only “nearest neighbor” interactions. Additionally. 21]. Chapter 3 deals with the most fundamental. This complexity is associated with the incorporation of highorder wavelets into the ﬁnite difference expression of the constitutive relations for dielectrically inhomogeneous or conducting media.4 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS thresholding of wavelet coefﬁcients was pursued in [17]. the complexity of conductor. a coarse mesh can be established. First. nonlinear circuit modeling [19]. related to wavelet methods: How to efﬁciently implement the widely advertised mesh adaptivity. It is shown that the interface can greatly facilitate the application of both localized hard boundary conditions and the implementation perfectly matched layer (PML) absorbers for mesh truncation. similar to the ones that have been demonstrated in signal processing applications [22]. To this end. Also. Haar MRTD schemes present an attractive alternative to their entire domain counterparts and have therefore become the subject of several studies [14. without defeating the purpose of obtaining operation savings. and boundary modeling that they present. . For these reasons.
Its exceptional performance is demonstrated through microwave and optical applications in Chapters 5. . while fully preserving the versatility of FDTD. This technique. Chapter 4 presents a computationally ﬂexible and efﬁcient method that can achieve mesh adaptivity. essentially translates the AMR technique of computational ﬂuid dynamics to threedimensional electromagnetic problems.INTRODUCTION 5 Following up to the technique of Chapter 3. called the dynamically adaptive mesh reﬁnement (AMR)FDTD. 6.
.
7
CHAPTER 2
A Numerical Interface Between FDTD and Haar MRTD: Formulation and Applications
A numerically stable interface between the Haar waveletbased MRTD technique and FDTD is presented in this section. Such a hybridization of MRTD facilitates the application of the method to open structures and inhomogeneous circuit geometries, where the use of highorder wavelets signiﬁcantly complicates the formulation of a pure MRTD scheme. Furthermore, it allows for the straightforward enforcement of localized and perfectly matched layer (PML)type absorbing boundary conditions. The fact that the implementation of the proposed interface involves no spatial or temporal interpolations indicates its potential to efﬁciently connect FDTD and Haar MRTD.
2.1
INTRODUCTION
The timedomain characterization of microwave structures, often encountered in wireless frontend applications such as ﬁlter, resonator or feed components, usually includes the modeling of ﬁne detail complex boundaries and regions of dynamically varying ﬁeld distributions. For this purpose, dense gridding conditions are necessary for the extraction of an accurate solution to the problem at hand. Nevertheless, the implementation of a uniformly dense mesh for an electrically large structure, translates to computationally burdensome simulations, as a result of both the size of the domain (in cells) and the stabilityrelated restriction on the time step of a ﬁnitedifference timedomain (FDTD)type of scheme, imposed by the small dimensions of the unit cell. Hence, the incorporation of local mesh reﬁnement techniques into conventional timedomain solvers is motivated, as a means of alleviating the overall cost of modeling structures with localized geometric details. Typically, subgridding techniques are implemented via spatiotemporal interpolations or extrapolations. Furthermore, a higherorder static subgridded FDTD algorithm, that needs no interpolatory operations has been recently proposed in [23]. However, waveletbased numerical
8
MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS
FIGURE 2.1: Coarse rectangular mesh for a layered coplanar waveguide structure, with cells containing variable dielectric permittivity and electric conductivity proﬁles
algorithms provide a natural framework for the implementation of dynamically adaptive subgridding, which holds the promise of signiﬁcantly accelerating conventional CAD tools, even those with local subgridding features. For homogeneous domains, the computational efﬁciency of waveletbased schemes, such as MRTD, is expected to increase with the order of the multiresolution expansion [24]. Yet, the complexity of conductor, dielectric, and boundary modeling that they present, also increasing with wavelet order, seriously compromises their potential applicability to stateoftheart devices. In particular, whenever a single cell includes variable material properties (such as dielectric and/or conducting layers), direct MRTD update equations are replaced by matrix expressions, resulting from the discretization of constitutive relations [11, 14]. Such cases are encountered in typical microwave devices, such as microstrip lines and coplanar waveguides (Fig. 2.1). In addition, the most effective mesh truncation technique nowadays, the perfectly matched layer (PML) absorber, is itself an inhomogeneous, uniaxially anisotropic (both electrically and magnetically) material. The aforementioned contradiction is addressed here by means of a hybrid approach that connects the FDTD technique with the Haar waveletbased MRTD, via a numerical interface. The purpose of this approach is to establish an efﬁcient algorithm, allowing for the combination of the versatility of FDTD with the adaptivity of MRTD, employing the ﬁrst in geometrically complex parts of the domain and the second in homogeneous regions. In the past, interfaces between different numerical methods have been developed in order to address speciﬁc problems of interest. In [25], the frequencydomain method of moments (MoM) was coupled to FDTD, for the analysis of ground penetrating radar (GPR) problems, where GPR antenna operation was modeled by MoM, while the inhomogeneous ground (including dielectric stratiﬁcation) was incorporated in an FDTD mesh. In [26], a Daubechies scaling functionbased Wavelet–Galerkin scheme was coupled to FDTD for the modeling of
A NUMERICAL INTERFACE BETWEEN FDTD AND HAAR MRTD
9
wedgeloaded twodimensional waveguide structures, apparently leading to signiﬁcant computational savings compared to its pure FDTD counterpart. Still, the general applicability and efﬁciency of the scheme was questionable, given the largely different dispersion characteristics of the two methods. Moreover, a hybridization of the transmission line matrix (TLM) method and MRTD was pursued in [27], where the issue of dealing with disparate dispersion methods was explicitly addressed via a conditionally stable space–time interpolation scheme. Finally, [28] demonstrated a technique for including FDTDmodeled lumped elements in a threedimensional domain simulated by the Haar waveletbased MRTD formulation that was introduced in [14] and was restricted to one wavelet level. This chapter proposes an interface between a Haar wavelet MRTD scheme of arbitrary number of wavelet levels and FDTD [29]. Under certain conditions, the two schemes are characterized by the same dispersion properties, a fact that is utilized in order to couple them with no interpolations or extrapolations and with absolutely no spurious reﬂections at their interface. The structure of this chapter is as follows: First, a brief overview of multiresolution analysis (that the MRTD technique is based upon) is provided, along with a presentation of how the technique can be systematically formulated. Then, the formulation of a Haar MRTD scheme (supporting an arbitrarily high spatial resolution) is presented. Noting that the dispersion equation for the latter coincides with the dispersion equation of an FDTD scheme of the same spatial resolution, a simple connection algorithm between the two is proposed. Emphasis is placed on the implementation problems that arise in the application of update equations at the boundaries of the two domains. Validation studies include the analysis of twodimensional dielectric cavity structures, under various FDTD/MRTD mesh conﬁgurations. Finally, the method is applied for mesh truncation in open domain problems and the modeling of structures with metal inserts such as a ﬁnloaded cavity.
2.2
MULTIRESOLUTION ANALYSIS: A BRIEF OVERVIEW
The theoretical foundation of waveletbased numerical techniques is the socalled multiresolution analysis theory, that is often encountered under the headline of wavelets. A brief overview of this subject is provided here for the purpose of completeness. A detailed study of the topic is contained in [30].
2.2.1
General
Wavelet bases are a mathematical tool for hierarchical decomposition of functions in an orthogonal expansion, according to the general scheme: f (ξ ) =
k
c k φ(ξ − k) +
j k
d j,k ψ(2 j ξ − k)
(2.1)
3) 1 Hilbert space of square integrable functions. 2. V2 where f is assumed to belong to L2 (R). The resolution of V0 is successively reﬁned by the second sum.1). The notion of a projection of f onto approximation spaces at successive levels of resolution is explained in Fig. the ﬁrst sum represents the projection of f (ξ ) onto a subspace V0 . each one being spanned by a wavelet basis ψ(2 j ξ − k) . that corresponds to an approximation of f at a “coarse” level of resolution. W0 is V0 W0 = V1 (2. It is noted that a basic property of the subspaces V0 .1 In (2. Recursively. The basis of V0 is generated by orthogonal translations of φ(ξ ) which is called the scaling function.2: Projections of a function f on three successive approximation spaces V0 . V0 W0 W1 · · · Wk−1 = Vk (2.k are simply produced by dilations (for adjustment of resolution) and translations of ψ. V1 . . This means that scaling and zero wavelet functions form a basis of orthogonal functions that spans V1 . which consists of projections of f onto the subspaces W j . The function ψ(ξ ) is called the mother wavelet. since all other wavelet basis functions ψ j.2.10 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS FIGURE 2.2) where V1 corresponds to a level of resolution twice that of V0 .
Thresholding of a wavelet coefﬁcient implies its elimination from subsequent operations.2 Wavelet Bases The simplest (and oldest) orthogonal wavelet system is the Haar basis [33]. g > implies inner product.2 Hence. 3 The notation < f. 2. 2 (2.n >  = O 2−(α+1/2) j . 0 ≤ ξ < 1/2 ⎪ ⎨ ψ(ξ ) = − 1. The function f (x) is uniformly Lipschitz of order α over the interval [a. it is said that f (h) = O(g (h)). wavelet coefﬁcients assume large values near discontinuities or singularities. If f (h) and g (h) are two functions of h. as h → 0. hence its study is also practically interesting. by a factor of two. once a numerical analysis of a problem is performed by expanding the unknown function in a wavelet basis.5) (2. due to its simplicity.2. b). is equivalent to improving the resolution of an approximation. modifying the dilation factor for the generation of the wavelet basis from the mother wavelet can yield other types of MRA. or—equivalently—it’s being treated as if it were equal to zero. By this procedure. that can be thresholded to zero. (2. otherwise (2. typically assume insigniﬁcant values. Its study is useful from a theoretical point of view because it offers an intuitive understanding of many multiresolution properties.3 Therefore. highorder wavelet coefﬁcients. any desired resolution of approximation can be achieved by adding an appropriate number of wavelet levels. such that:  f (h) < Cg (h). like the one in Eq. . if there is some constant C. b] if and only if for any n ∈ Z and j ∈ Z such that 2− j n ∈ (a. In particular. 1/2 ≤ ξ < 1 ⎪ ⎪ ⎩ 0. 0 ≤ ξ < 1 0. which are located away from discontinuities or singularities.A NUMERICAL INTERFACE BETWEEN FDTD AND HAAR MRTD 11 In other words. ψ j. for h sufﬁciently small.  < f. an adaptive. Indeed. The advantage of this approach lies in the dependence of the magnitude of wavelet coefﬁcients on the local regularity of f . this basis is widely employed in a series of applications. otherwise. temporally moving mesh is implemented.1). In general. Furthermore. The Haar scaling function is deﬁned as a pulse of unit length: φ(ξ ) = 1. an explicit relationship between the magnitude and the “Lipschitz uniformity” of a square integrable function [31] is given by the following theorem [32]: Let 0 < α < n be a real number that is not an integer.4) and the scaling function basis is produced by orthogonal translations of the latter: φk (ξ ) = φ(ξ − k) The Haar mother wavelet function is deﬁned as: ⎧ ⎪ 1.6) Dyadic multiresolution analyses are considered here. adding one wavelet level.
2 0. (2. Battle–Lemarie scaling and wavelets have an excellent localization both in space and Fourier domains (Fig.5 0 −0.7).2 scaling function 0 0. A basis that has also been used in applications.8 1 1 0. 2. is qualitatively explained in Fig. 1 0.6) are entire domain functions and therefore schemes that are developed in this basis have to be truncated with respect to space. if φ or ψ functions represent a variation along the xdirection of a computational mesh with a cell size x.7) It is noted that the same deﬁnition holds for the derivation of any dyadic wavelet basis.2 0. The Battle–Lemarie scaling and mother wavelet (Fig.4 0. In all these deﬁnitions ξ is a normalized. then simply: ξ = x/ x. and the mechanism.4 0. dimensionless variable.8 1 Normalized space variable (e.2 mother wavelet 0 0.6 0. from its generating mother wavelet. The Haar scaling and mother wavelet functions are depicted in Fig. as shown in Fig. 2. (2. is the cubic spline Battle–Lemarie basis [34]. For example. in which multiresolution analysis is brought about by employing this basis (in the sense of Eq. a feature that permits an a priori estimate of the necessary levels of resolution for correct ﬁeld modeling.6 0.5 0 −0. as shown in the deﬁnition: ψ j. Therefore. However.k (ξ ) = 2 j/2 ψ(2 j ξ − k).5 −0.3)).12 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS The wavelet basis is then produced by both translations and dilations.5.g. 2.3: Haar scaling and mother wavelet function .5 −1 −0. x/Δ x) FIGURE 2. The Haar basis has an excellent localization in space (or time) domain and poor localization in the corresponding Fourier domain. 2. 2.3. the Haar basis does not share the typical wavelet property of combining good localization in both domains (within the limits of the uncertainty principle [22]).4.
4 0.A NUMERICAL INTERFACE BETWEEN FDTD AND HAAR MRTD V0 13 W0 V1 V0 W0 V2 W1 V0 W0 W1 FIGURE 2.8 0.5: Haar scaling and mother wavelet functions in the Fourier domain .2 0 −40 −30 −20 −10 0 10 20 30 40 Normalized wavenumber variable (e.4: Demonstration of the multiresolution principle as implemented by the Haar basis 1 F {φ}(k) F {ψ}(k) 0.6 0.g. kx Δ x) FIGURE 2.
14 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS Normalized space variable (e.7 also provide an intuitive explanation of the successive approximation procedure that is connected to multiresolution analysis. scaling functions alone provide a description of “low frequency” characteristics of a signal.5 and 2.8 0. while the addition of wavelets enhances the capability of 1.6: Battle–Lemarie scaling and mother wavelet function Figures 2.g. kx Δ x) −5 −5 −15 FIGURE 2.2 F{φ}(k) F{ψ}(k) 0. while the mother wavelet is by the same token a band—pass ﬁlter. x/Δ x) FIGURE 2. Hence. Evidently.4 0 −15 Normalized wavenumber variable (e.g. the scaling function has the Fourier domain pattern of a low—pass ﬁlter.7: Battle–Lemarie scaling and mother wavelet functions in the Fourier domain .
a systematic way to formulate timedomain numerical schemes for Maxwell’s equations is provided. zdirections a scheme to describe the highfrequency content of a signal. This argument also explains the adaptivity property of waveletbased numerical schemes: wavelet coefﬁcients correspond to highfrequency signal variations. which is also employed in the present monograph for the derivation of MRTD schemes. the value of these coefﬁcients decays to insigniﬁcant levels.p. t) 1 ∂ E(z. The FDTD scheme for this equation is derived by approximating the partial derivatives of each side of the . a discrete space–time mesh is introduced and the values of the electric ﬁeld E(m z.3 DERIVATION OF TIMEDOMAIN SCHEMES BY THE METHOD OF MOMENTS In [35].8) For the numerical solution of this equation. k t) = k Em are sought by marching in time. in the absence of which.8: Equivalent grid points (e.) within a scaling cell as introduced by a ﬁrstorder wavelet expansion in x. This technique. t) = ∂z c ∂t (2. is explained here.g. the following onedimensional wave equation example: ∂ E(z. 2. by considering ﬁrst.A NUMERICAL INTERFACE BETWEEN FDTD AND HAAR MRTD 15 (m1) cell mth cell φm1 φm ψm ψm1 PEC k E M 3 k E M 2 k E M 1 kEM =0 FIGURE 2.
n−1 ) dx 2 h n (x) h n (x) d x = δn . the electric ﬁeld E(z. For this scheme to be stable. the CFL number has to be less than one. for the discretization of (2. (k − 1) t) = + O( t 2 ) ∂t 2 t k+1 Em − k−1 Em ≈ 2 t (2. t) around m z and k t. m 1.15) Then.17) . (2. n x < x < (n + 1) x otherwise (2.8).9) (2. k t) − E ((m − 1) z.8): k+1 Em = = k−1 Em k−1 Em t ( k Em+1 − k Em−1 ) z + s ( k Em+1 − k Em−1 ) + c (2.16) (2. k t).16 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS equation at a space–time mesh point (m z.10) The previous expressions stem from a reformulation of a Taylor series expansion for E(z. t) = k.n x (2.12) where s = c t/ z is the CFL (Courant–Friedrichs–Levy) number [36]. h n (t) in space and time: E(z.14) d h m (z) 1 h k (t) = dz c k Em k. Alternatively. m k Em h m (z) h k (t) (2. k t) E ((m + 1) z. 0.n+1 − δn . the following integrals are employed: +∞ −∞ h n (x) +∞ −∞ 1 d h n (x) d x = (δn .11) (2. m h m (z) d h k (t) dt (2.15) is sampled in space and time using the complex conjugate of the basis functions themselves as testing functions (Galerkin method). Then. (k + 1) t) − E (m z. simple algebraic manipulations lead to the following secondorder accurate approximation of (2. k t) = + O( z 2 ) ∂z 2 z k Em+1 − k Em−1 ≈ 2 z ∂ E(m z. To carry out the testing procedure.8) via the method of moments. t) is expanded in pulse basis functions (or equivalently: Haar scaling functions) h n (z) .13) with h n (x) = Substituting into (2. k t) E (m z. by centered differencing [36]: ∂ E(m z. with respect to space and time. yields: k Em k.
m+1 − δm .m k Em δm .m−1 ) δk . z φm (z) = φ −m z is a scaling basis produced by a scaling function φ and: r ψm. are shown to be equivalent. .A NUMERICAL INTERFACE BETWEEN FDTD AND HAAR MRTD 17 where δn . testing the left handside of (2. . In this case. p (z) = 2r/2 ψ 2r (2.n = 1. 1.m k Em (δm . the ﬁeld expansion in (2. . 2. a slightly modiﬁed deﬁnition with respect to (2. .m (δk . (2.8) in a discrete space.9. 2 t 2 (2.n is Kronecker’s delta: δn .7) is used for the wavelet basis here. 2c Thus.21) z −m − p z (2.22) with p = 0. p k Em r ψm. p (z) h k (t).k+1 − δk . The introduction of wavelets gives rise to a reﬁnement of the scaling functionbased approximation of E(z.k−1 ) k−1 Em ) z 2c (2. (2.12) is again derived and therefore the two methods for approximating (2.13) is written as: E(z. t) = c ∂t k . yields: +∞ −∞ +∞ −∞ d z d t h m (z) h k (t) ∂ E(z.15): +∞ −∞ +∞ −∞ d z d t h m (z) h k (t) 1 ∂ E(z. t) by a factor 2 Rmax +1 . if n = n 0. In a similar manner. .20) where Rmax is the maximum waveletorder that has been introduced. t) = k.19) = ( k+1 Em − z . Hence. it is easily concluded that Eq.15). if n = n Hence. testing the right handside of (2. m φ k Em φm (z) h k (t) + Rmax 2r −1 r =0 p=0 ψr. in order to keep a correspondence between the cell index m of the scaling and the wavelet functions. To explain the mechanics of this mesh reﬁnement. This deﬁnition is schematically explained for the case of the Haar basis in Fig. t) = ∂z k . Also. 2r − 1.k t .18) = ( k Em+1 − k Em−1 ) Similarly. expanding the electric and magnetic ﬁelds in scaling and wavelet functions and applying the method of moments for the discretization of Maxwell’s equations (as indicated in the previous section). the MRTD scheme is derived.
1 0 ψm. 2. Linear combinations of their coefﬁcients with the scaling coefﬁcient. The electric ﬁeld values at the equivalent grid points and the positions of the n . deﬁned by pulse functions are shown.10.10: Mesh reﬁnement in the Haar basis . ψn.0 .p.1 . ψn. In Fig. the 0 1 1 following wavelet terms are included within the nth cell: ψn. Assuming the use of a ﬁrstorder scheme (two wavelet levels).0 .MRTD cell (n + 1)  MRTD cell xn × × x n+1 × (x) xn+2 : MRTD equivalent grid points x : Scaling cell limits FIGURE 2. produce four equivalent grid points (e.) within the cell.18 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS φm 1 ψm.9: Explanation of deﬁnition (2. 0 1 ψm.g.22) for the Haar basis the simple example of the Haar basis is employed. 0 (z) mΔ z (Δ m+1) z FIGURE 2. two scaling cells.
0 − 2En 1.23) The matrix on the right handside of (2.1: Position of and Field Values at the MRTD Equivalent Grid Points of Fig.625 xn + 0.0 ⎥ ˆ ⎢ E ⎥ ⎢ +1 −1 0 + 2 ⎥ ⎢ En ⎥ ⎣ n⎦ ⎣ ⎦ √ ⎦ ⎣ ψ ˆ4 +1 −1 0 − 2 En En 1.1 √ ψ ψ φ En − En 0.0 √ ψ ψ φ En + En 0.1.n In. . p (z) d φn (z) dz dz r d ψn .0 ⎥ ⎢ E 2 ⎥ ⎢ +1 +1 − 2 ˆ 0 ⎥ ⎢ En ⎥ ⎢ n⎥ ⎢ ⎥ √ ⎥·⎢ ⎢ 3⎥ = ⎢ ⎥ ⎢ ψ1. The modiﬁed ﬁnite difference equations for the MRTD approximation of (2.375 xn + 0.10 POSITION xn + 0. This formulation will be later on employed in the explicit enforcement of localized boundary conditions and the development of a numerical interface between the FDTD method and MRTD. A mathematically elegant and compact way to express the relationship between ﬁeld nodal values and MRA expansion coefﬁcients is the following matrix formulation: √ ⎤ ⎡ ⎤ ⎡ φ ⎤ ˆ1 +1 +1 + 2 0 En En √ ⎢ ⎥ ⎢ ⎥ ⎢ ψ0.0 + 2En 1. p (z) dz dz .23) is the onedimensional wavelet transform matrix for a twolevel wavelet scheme.n φ.24) ψr . p (z) φ.n ψr . ψ r p dz dz (2. φ = = = = +∞ −∞ +∞ −∞ +∞ −∞ +∞ −∞ φn (z) φn (z) d φn (z) dz dz r d ψn . p (z) p In. φ p In.0 − 2En 1.1 latter within the nth cell are provided in Table 2.0 √ ψ ψ φ En − En 0.875 x x x x FIELD NODAL VALUE √ ψ ψ φ En + En 0.8) are now dependent on the form that the following integrals assume in a given wavelet basis: In.n r ψn.1 ⎡ (2.125 xn + 0. 2.0 + 2En 1.A NUMERICAL INTERFACE BETWEEN FDTD AND HAAR MRTD 19 TABLE 2. ψr p r ψn.
the variation of the index l of the inﬁnite sums is restricted within a ﬁnite range [−N. if wavelets up to order n are introduced. N]. is called secondorder accurate. because it implies a numerical approximation of the spatial and temporal derivatives of (2. The numerical discretization of (2. with an error proportional to z2 and t 2 . MRTD is applied to the following system of twodimensional T Ez Maxwell’s equations: ∂ Ey 1 (ρ.24) can be carried out either analytically or computationally. if entire domain basis functions are involved. or simply approximated as ﬁnite ones.10). t) = ∂t ∂ Hx ∂ Hz (ρ. depending on the integrands themselves. the following set of update equations for scaling and wavelet terms. This fact becomes evident by pure inspection of expressions (2. is deduced: φ k+1 Em = = φ k−1 Em 2c t + z 2c t + z +∞ l=−∞ +∞ l=−∞ φ. FDTD Equations In this section.8). p p Im. 2. the clariﬁcation of the use of the term “order” of an MRTD scheme.. ψ r ψ (2. for completeness: The support of a function f is the closure of the set {x : f (x) = 0}. The computation of the integrals of (2. as opposed to its common use in numerical methods. It is noted that wavelets are solely connected to mesh reﬁnement and do not affect the order of accuracy of the underlying scheme. respectively. As a ﬁnal note.25). t) − (ρ. p k−1 Em ψr. an MRTD scheme is of nth order.4 2.m+l k Em+lp ψr.8). given in (2. Therefore.25) ψ r .26) . In the case of inﬁnitely supported basis functions (e. On the other hand. or—simply— a secondorder scheme. if ﬁnite domain basis functions are involved. t) ∂z ∂x (2.20 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS Letting l = m − m. φ φ Im.m+l k Em+l The inﬁnite sums with respect to the position index in (2. The deﬁnition of this term is provided here. the Battle–Lemarie basis employed in [11]). p Im. or {x : f (x) = 0} [31]. where N is called the stencil of the method. Fourier calculus has to be employed for the extraction of their value.g. cast in a generic form. are either reduced to ﬁnite ones. that wavelet functions are deﬁned in the Fourier domain. It is often the case. as a formal substitute for the concept of entire or ﬁnite domain functions is that of support. p . is necessary.ψ r ψr .m+l k Em+l φ. p k+1 Em ψr. rather than directly.4.12).1 TWODIMENSIONAL HYBRID ARBITRARYORDER HAAR MRTD/FDTD SCHEME: FORMULATION MRTD. A term widely used in the wavelet literature.m+l k Em+l + + Rmax 2r −1 r =0 p=0 Rmax 2r −1 r =0 p =0 p r. φ φ Im.
m n r z. Given these deﬁnitions. the wavelet functions of order r that recursively reﬁne the resolution of φm are deﬁned r as: ψm. p = 2r/2 ψ (2r (x/ x − m) − p).max . In the subsequent development. Based on the method outlined in [11].m x. p z (z) (2.max in x. φ φ φi (x) φm (z) + + rx . A basis of pulse functions h n (t) = h(t/ t − n) (deﬁned as in [11]). as they are the most commonly used for the purpose of adaptively solving partial differential equations. assuming a spatial expansion of electromagnetic ﬁeld components in scaling and wavelet functions of an arbitrary basis and up to arbitrary orders r x. φm (ξ ) = φ (ξ/ ξ − m) denotes the mth scaling function in ξ = x. p x (x) n Ei. p z r x ψi. t) = n h n (t) i.m φi (x) φm (z) + r z.28) 1 ∂ Ey ∂ Hz (ρ. p x φ r x ψi. p z (z) r ψi. ψr x . p x r z. the discretization of a twodimensional domain (in which ﬁeld solutions are sought) in cells of x by z is pursued. t) = ∂t 1 ∂ Ey (ρ. where p = 0. ψr x . t) ∂t μ ∂x ˆ with ρ = x x + zˆ .m x. p z r r ψi. where t denotes the time step.m φm (z) rz ψm. Then. p x ψr z . t) μ ∂z (2. However.xp x (x) ψmz . p z (z) + rx . zdirections respectively. p z y. r z. zdirections.A NUMERICAL INTERFACE BETWEEN FDTD AND HAAR MRTD 21 ∂ Hx (ρ. p z (z) + r x .m x.27) (2.29) ¯ Hx (ρ.m y. update equations are derived by z the method of moments. p z y. px y. px Hi. p z n Hi. ψr x . p z φi (x) n Ei. φ φ n Ei.xp x (x) φm (z) n Hi. limited by the choice of x and z through the stability condition.m x. · · · 2r − 1. Ey .30) . Accordingly. Hz are expressed in the form of the following orthogonal expansions: ¯ Ey (ρ.m (2. is employed for ﬁeld expansion in time. Hx . p x r z. φ ψr z .m rz ψm. p z n Hi. p z r φi (x) ψmz . by means of a wavelet basis. our analysis is restricted to dyadic wavelet transforms. t) = n h n (t) i. deﬁned by the scaling function φ and the socalled mother wavelet ψ [30]. φ ψr z . ψr x . p x (x) n Ei. t) = − (ρ. p x φ + r x . p x ψr z .
(2.m r x . Nξ .30) can be cast in the next equivalent form [30]: ¯ Ey (ρ. since each wavelet level successively doubles the resolution of the underlying approximation. including Ey and Hz grid points (which are necessary . p z rz ψir x.32) via the wavelet transform.max .and zdirections is left as a parameter under investigation. by introducing the notion of equivalent MRTD grid points. p z n Hi . t) = n h n (t) i. utilizing Haar scaling and zeroorder wavelet functions (Fig. 1 · · · 2r x. 2. p x ψir x. ψ ψr z . the introduction of wavelets of orders r = 0. r ξ.m r x . Thus. in one dimension.max + 1 and φn (ξ ) = 2 R/2 φ(2 R(ξ/ ξ ) − n) is the scaling basis that produces by itself an approximation of the same resolution as (2. the latter corresponds to the hierarchical multiresolution decomposition of the former and the coefﬁcients in (2.max + 1. m = m + s z. p z n (2. This waveletinduced mesh reﬁnement can also be perceived as a procedure of generating equivalent grid points that give rise to a mesh of cell sizes x/ρx . .m z. . 2. φ ψr z . p x z. in Fig. p z (z) z. . the offset of electric and magnetic scaling cells in x. p z n Hi . p z (z) z.m n r z. linear combinations of scaling and wavelet functions yield electric ﬁeld values at points: i + ( p + 0.3).m φi (x) φm (z) rz φi (x) ψm. 2. .11 and 2. with the notable exception of [37]. based on the FDTD practice. This argument is demonstrated for a zeroorder Haar MRTD scheme. Most MRTD studies.max +1 .max +1 − 1. Rz = r z. φ φ + + rx . 1. . . these ﬁgures represent zcuts of the mesh.12.31).31). i = 1. i = i + s x .5)/2r x.31) where n = n + 1/2. t) = n h n (t) i. In mathematical terms.30) can directly be deduced from the ones in (2. z/ρz. reﬁnes the mesh in ξ direction by a factor of ρξ = 2r ξ. px Hi . z) : i · ξ . It is noted that in general. p x (x) φm (z) Hi . Similar observations hold for (2. . the Ey expansion of (2. a systematic way of determining these offsets is set forth.max +1 x. with p = 0. ψ φ + r x . p x r z. which depict the equivalent grid points that are generated in both cases. In our twodimensional example. while half a time step offset between the update of electric and magnetic ﬁeld terms is kept (as in FDTD).m R φiRx (x) φm z (z) . φ Rx φ Rz n Ei. Assume that a certain scaling function basis generates electric ﬁeld grid points in one dimension (ξ = x. In this section. Then. . In fact.m y. (2.22 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS ¯ Hz(ρ.32) R where Rx = r x. p x (x) ψm. adopt the choice of s x = s z = 1/2.30).
12: Electric/magnetic ﬁeld equivalent grid points for zeroorder Haar MRTD in one dimension. the ﬁeld values at equivalent grid points within each cell. Accordingly.max +1 x. the purpose (i − 1)−cell i−cell (i + 1)−cell E 2i1 E 2i E 2i+1 E 2i+2 (x) H2i1 H2i H 2i+1 H 2i+2 (x) : Electric field equivalent grid point : Magnetic field equivalent grid point FIGURE 2. respectively. However. Hz equivalent grid points are located at i + s x + ( p + 0.33) . Up to a normalization multiplicative constant. under the common convention of half a cell offset between electric/magnetic scaling cells for the approximation of xpartial derivatives involved in Ey updates).A NUMERICAL INTERFACE BETWEEN FDTD AND HAAR MRTD (i − 1) − cell i − cell (i + 1) − cell φ E 2i1 E 2i E 2i+1 E 2i+2 23 (x) ψ H 2i1 H 2i H 2i+1 H 2i+2 (x) : Electric field equivalent grid point : Magnetic field equivalent grid point FIGURE 2. with electric/magnetic scaling cell offset chosen in consistence with (2. are computed as the sum and the difference of scaling and zeroorder wavelet terms.11: Electric/magnetic ﬁeld equivalent grid points for zeroorder Haar MRTD in one dimension.5)/2r x.
φψ r.11 shows the node arrangement in zeroorder Haar MRTD.m + − n y.φφ n+1 Ei. p. The scaling cell dimensions are x × z.33) Figure 2. p z .24 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS of using wavelets is to implement in this “new” mesh of equivalent grid points.φψ x. p x and η = ψr z.ψ φ (2. p q . Galerkin’s method is applied for the derivation of ﬁeld update equations.m−1−1 z.12). under the common convention of half a cell offset between electric/magnetic scaling cells.φφ n h i. p − z. p z.φψr. p n+1 Ei.2r −1 h i . the method produced by scaling functions only. (2.m = n Ei.2 h i. Apparently. Also.φφ r .m + r 2r/2 n .φψr.φψ x. h i −1. Hence.m + y.5ρz = 2−r z. Finally. m = m + s z. r . ζ = ψr x .φφ n h i. D1 are given in Section 2.m −1 −1 −n h i.35) .φφ y. In the following.p r.ψ q ψr.max −2 s z = 0.m r.8.2 −n h i −1. at a resolution that is ﬁner (in xdirection) by the wavelet reﬁnement factor ρx . p t 2r/2 δ p. i = i + s x .26)– (2. Upon substitution of the MRTD expansions of all ﬁeld components into Eq.m − x.28).2 h i .φφ n h i.φφ 2r/2 n r.m r x. the update equations for Ey . p − z.0 x z x. p ) n h i. p (2.m −1.φφ h i −1. p ) n h i.34) y. the coefﬁcients D0 .m − x. the wavelet augmented method has to deﬁne half an e q uivale nt cell offset between the equivalent electric/magnetic grid points in this direction. r . this latter approach leads to MRTD schemes with consistent numerical dispersion properties (as opposed to the former approach) and therefore.p D2 (r. h x = Hx x and h z = Hz z coefﬁcients are provided.33). p h i −1.φψr h i. which in this case yields an offset of one quarter of a cell (Fig.m−1 z. it is adopted in the following derivations.mp + r + D3 (r. p.5ρx = 2−r x. n = n + 1/2.max −2 (2.m − + z. equivalent electric and magnetic equivalent nodes are now collocated.ψ q ψr. As discussed in [38].ψ n φ x. if this method deﬁnes half a scaling cell offset between electric and magnetic nodes. The choice of s x and s z that is consistent with this requirement is given below: s x = 0. On the other hand. Ey update equations t x z n z.mr. if the separation of electric and magnetic scaling cells is chosen in consistence with (2.m −1 + r z.m h i .φψr. 2. the equivalent grid points are leapfrogged in space and correctly correspond to the mesh of an FDTD scheme of cell size x/2.m − n h i .m = n Ei.φψ n n 2q /2 q n q .m −1 x.
m y.p n n q .ψr. r z.w − 2r/2 δ p.2r z −1 y.ψ n n h i.ψ x. p D1 (r z. q .2q −1 x. w. p ψq . r . p + q .ψr.φψ .0 (2.0 z.39) n h i.m+1 + y. p ψ D2 (q .m+1 + y.φφ n Ei.0 n Ei.φφ x. p q .m q .φφ n Ei.ψ n h i . p −1 x. w ) n h i.36) z.m = x.m r.ψ ψq .m − n Ei. p z. p z y.m r.38) n h i. p z y. p − − z.2 −1 ψ q −1 h i. p x. p z) n Ei.ζ φ n −1 h i. p.p r . p ) n h i .m + y.m + y.ζ ψr z .ζ φ n Ei. p φ − 2r/2 δ p. r .ζ ψr.φψr.m = x. p −1 ψq .w h i −1.ψ φ n h i.p D2 (r.φη n −1 h i.w Hx update equations h i.m + t x μ z + 2r z/2 δ p z.m h i .ζ ψr.m r.φφ D0 (r z.w r . p. p. q .ψr. p z n Ei.m (2. r . w.w y.ζ ψr . p ) n h i . r . p ψq n h i.0 n Ei.ζ φ D0 (r z.37) x.max y.φφ 2r/2 0≤r ≤r z.m−1 z. p + + q y.ψ φ n h i.ψ n x.m = n Ei.ψ φ 2q /2 h i.ψ φ t x z x. p z. p.max y.m r.φη x. p ) n h i −1.ζ φ x.m z z (2. p + D3 (r.2r z −1 y.φφ n −1 h i.m r.m+1 − n Ei.m − n Ei.A NUMERICAL INTERFACE BETWEEN FDTD AND HAAR MRTD y. r z.m 25 = n Ei. p z y. p z D1 (r z.0 x z x.m + t x μ z + 2r z/2 δ p z. p z.m y.w − z.m+1 r z. p z. p z r z.0 (2. p z.mr .m n + t x μ z y.m r.ψ ψq .m r. p z) n Ei.p + D3 (r.40) n h i. r z.w + D3 (q .m r.φψr z .mr .m y. p φ n+1 Ei.ζ φ 2r/2 0≤r ≤r z. p ) n h i .m+1 r z.m+1 − n Ei.m −1 q .ζ η n −1 h i.ψ ψq .m = x.41) . p z r z.m − φ z.ζ φ n Ei.ζ η x. w ) n h i.m = x.w + t 2q /2 δw.m+1 − n Ei. r z.0 + r .φψr z .m+1 − n Ei.φψ D2 (r.m z.φψr.m (2.φφ n h i −1.ψr. p z n Ei.w n+1 Ei.φφ + r .ψ φ (2. p z y.m + t x μ z y.
m+ 1 − n+ 1 h i. px y.m + y.m − n Ei+1.m + y.m − n Ei.m r.m = z. p x η n Ei+1.φη z. An FDTD cell size of δx × δz and time step δt is assumed.ψr x .26 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS Hzupdate equations n h i .m 2 2 δt x x n+ 1 h i. p x φ n Ei+1.m 2 2 (2.m 2 2 μδz δtδz y y z = n− 1 h i+ 1 .m − n Ei.0 φ n Ei+1.ψr x .m − t z r x /2 2 δ p x .m + y.m − n Ei.φφ D0 (r x .2r x −1 μ x + rx .φφ n Ei.0 y.φφ n −1 h i .45) A numerical dispersion analysis of this system of equations yields the following expression [38]: ω t sin up t 2 1 2 = + k x xeff 1 sin xeff 2 1 k z zeff sin zeff 2 2 2 (2.46) .max y.ζ φ n −1 h i .m x n+ 1 h i.m 2 2 μδx = n Ei. px y. p x .m = z.m+ 1 + n Ei.m − t z μ x y.m D1 (r x .φη n Ei.ζ φ z.m D1 (r x .φη 2r/2 0≤r ≤r x.ζ η n −1 h i . px rx .φη D0 (r x .m (2.φη n Ei+1. p x φ y. px rx .φφ z. p x .m 2 2 + z n+ 1 h i− 1 . p x .φφ n Ei+1.m − n Ei+1.ψr.0 η n Ei+1. p x ) n Ei.ψr x .47) .φφ 2r/2 0≤r ≤r x.ψr. px y.44) n h i .42) t z μ x n h i .ψ η (2.m+ 1 2 2 z n+ 1 h i+ 1 .φη n −1 h i .m + y.m r. the wellknown FDTD equations for T E z waves are provided below.m (2.m − n Ei+1.m − y.max y.m − n Ei.2r x −1 μ x + rx . rx .0 y. px y.m + y z n+ 1 h i+ 1 . r x .m − t z r x /2 2 δ p x . r x .ψ φ (2. p x η y. Then: y n+1 Ei.43) n h i .ζ η z.ψr x . p x . p x ) n Ei.m − n Ei. rx .m− 1 − 2 2 2 2 δxδz δtδx y y x = n− 1 h i. For completeness.m = z.m = z.m+1 − n Ei.
For the update of the electric ﬁeld ψ ψ ψ φ MRTD coefﬁcients E5 . a twodimensional FDTD/MRTD connection algorithm is presented.13. E5 . In the following. (2. z eff ≡ δz. that allows for MRTD domain truncation via existing absorber codes. are necessary. leads to the following expression [39]: 1 ωδt sin u p δt 2 2 = 1 k x δx sin δx 2 2 1 k z δz + sin δz 2 2 . to derive FDTD coefﬁcients (nodal ﬁeld values) from MRTD terms. (2. For their calculation. while the MRTD scaling cell size is 76. the FDTD nodal values of the magnetic ﬁeld H1 . an inverse fast wavelet transform (IFWT) is applied. at the FDTD/MRTD boundary. (2. H4 are used as an input to a recursive fast wavelet transform (FWT). It is noted that both FWT and IFWT are characterized by an optimal complexity of O(N). it is readily concluded that under the condition: xeff ≡ δx. In the case of an FDTD to MRTD transition.49) the FDTD and Haar MRTD schemes present the same dispersion properties. Hence.46) and (2.8 mm. Smooth and stable transition from the MRTD to FDTD region is observed. a stable time step for the MRTD region is also stable for FDTD. 2. Similar concepts can be employed for interfaces of other types. the wavelet decomposition of the magnetic ﬁeld at that cell is deduced and employed for the electric ﬁeld updates of MRTD. The time step is 0.4 mm. their computational implementation is relatively simple. An application of interest is the termination of an MRTD mesh in an FDTD perfectly matched layer absorber. The interface for the electric ﬁeld nodes is located at cell 1440. As a computational validation of these interface principles.4. the propagation of a 0–5 GHz Gaussian pulse through an FDTD/fourthorder MRTD interface is shown in Fig. 2. 2. E5 10 . The cell size for the FDTD region (and effective cell size for MRTD) is 2. magnetic scaling and wavelet terms one cell to the left. within the FDTD region. onedimensional example of an interface between a ﬁrstorder Haar MRTD and the corresponding FDTD scheme is shown in Fig. In addition.48). As a result. since the same time step is chosen for both methods.48) By inspection of Eq. H3 . Time marching is carried out concurrently in the FDTD and MRTD regions. E5 11 .9 of the Courant limit.A NUMERICAL INTERFACE BETWEEN FDTD AND HAAR MRTD 27 A dispersion analysis for the system of FDTD equations (2. Thus. The common dispersion equation of the two schemes also implies common stability properties.14.2 Connection Algorithm A simple. for the case where an FDTD region encloses an MRTD one. H2 .47). as expected. a reﬂectionless interface between the two methods can be established. .
16 schematically explains this procedure. the independent solution of this region becomes possible. via a twodimensional FWT routine.equivalent grid points FIGURE 2.13: Onedimensional interface between FDTD and a ﬁrstorder Haar MRTD scheme To begin with. Conversely. the complete determination of the tangential electromagnetic ﬁeld components along the boundary of the FDTD domain is allowed for. 2. since a perfect conductorbacked absorber is typically used for the simulation of an open boundary.grid points E Efield : MRTD E . For the update of all FDTD grid points whose stencil extends into the MRTD region. It is noted that no FDTD grid points are sought for at the (shaded) corner . the problem is reduced to calculating nodal ﬁeld values along the boundary of the MRTD region. which is exactly the function of IFWT. Figure 2. In consequence. for a case where r x.max = 1.equivalent grid points 5 H1 H2 H3 H4 Hfield : FDTD H .grid points : MRTD H . If this scaling cell extends beyond the domain.15). for the update of the MRTD boundary magnetic ﬁeld coefﬁcients. FDTD electric ﬁeld nodes extending over one scaling cell within the FDTD region are wavelet transformed. since the MRTD calculated tangential ﬁelds are used as boundary conditions for FDTD. data transfer from MRTD to FDTD is addressed.28 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS E ψ 10 5 E ψ 11 5 FDTD MRTD E φ 5 x x: x x x ψ FDTD E . Then. Thus. This is possible and physically correct for both closed and open domain problems. it sufﬁces to retrieve the nodal ﬁeld values of the electric ﬁeld one FDTD cell within the MRTD region (Fig. zero ﬁeld values are used.max = r z.
5 0 1250 1300 1350 1400 1450 1500 1550 1600 MRTD cells FDTD cells FIGURE 2.5 0 1250 1 1300 1350 1400 1450 1500 1550 1600 t = 650 Δ t 0.A NUMERICAL INTERFACE BETWEEN FDTD AND HAAR MRTD 1 29 t = 550 Δ t 0.14: Reﬂectionless propagation through a onedimensional FDTD/fourthorder MRTD interface 2D Yee’s cell FDTD/MRTD boundary : E y FDTD field node : H x FDTD field node : H z FDTD field node MRTD region (x) FDTD region (z) : E y MRTD equivalent grid point FIGURE 2.5 0 1250 1 1300 1350 1400 1450 1500 1550 1600 t = 770 Δ t 0.15: FDTD update from MRTD data .
the same principles lead to interfaces between wavelet schemes of an arbitrary basis and the ones that are formulated by the corresponding scaling functions only. The choice of those conﬁgurations was made in order to demonstrate that the stability and accuracy of the algorithm was preserved under any gridding conditions. these being either related to the order of the MRTD .5 NUMERICAL RESULTS: VALIDATION Several twodimensional airﬁlled square resonators have been analyzed in various FDTD/MRTD mesh conﬁgurations for validation purposes. uses a crossshaped stencil for the update of all grid points. all operations that implement this connection algorithm are performed at the same time step. This is due to the fact that MRTD. provided that the effective resolutions in the two regions are kept the same. just as FDTD. determines again the tangential electromagnetic ﬁeld components along the boundary of the MRTD region. during the update of the electric ﬁeld coefﬁcients in both regions.16. It is also noted that the extension of the interface algorithm to three dimensions is accomplished by treating each face of Yee’s cell according to the method that has been set forth in this work. in an absolutely stable fashion. due to the matching of the dispersion properties of the two schemes. is sufﬁcient for its independent. Computing the electric ﬁeld MRTD coefﬁcients from the FDTD data via an FWT and updating the tangential magnetic ﬁeld component coefﬁcients via the standard MRTD ﬁnite difference equations. Furthermore. shown in Fig. 2. Evidently. in turn. 2. This.30 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS FDTD/MRTD boundary Δx MRTD scaling cell : Ey MRTD equivalent y grid point : Ey FDTD field node y (necessary for the MRTD update) (x) (z) Δz MRTD region FDTD region FIGURE 2.16: MRTD update from FDTD data region. MRTDbased solution.
an MRTD scheme of orders 3 and 4 in x. In the ﬁrst case (Fig.18 depict two case studies for MRTD/FDTD mesh conﬁgurations.and zdirections forms a mesh of 2 by 2 scaling cells corresponding to 32 × 32 FDTD cells. enclosing the MRTD region in an FDTD one. On the other hand. A pure MRTD scheme would model these electric walls by means of image theory.by 3order MRTD/FDTD mesh conﬁguration (dimensions are given in FDTD cells. Similarly. 5 2 32 2 16 (x) 2 (z) : FDTD/MRTD interface FIGURE 2.17 and 2. The dimensions of the cavity structures in all ﬁgures are given in FDTD cells (or MRTD equivalent grid points). The whole domain corresponds to 36 × 40 FDTD cells of dimension 1 cm × 1 cm. an MRTD scheme of order 3 in both x. whose FDTD modeling amounts to setting the tangential to PEC electric ﬁeld nodal values equal to zero.17).17: Empty rectangular cavity geometry and interface of a 3.A NUMERICAL INTERFACE BETWEEN FDTD AND HAAR MRTD 40 31 36 (x) 2 7 (z) : FDTD/MRTD interface FIGURE 2. 2.by 4order MRTD/FDTD mesh conﬁguration (dimensions are given in FDTD cells.and zdirections respectively forms a 1 by 1 scaling mesh and is interfaced with FDTD of cell size 1 cm × 1 cm. facilitates the treatment of these hard boundaries. MRTD mesh is 1 × 1) . that is terminated into the metal boundaries of the cavity. asymmetrically placed within an FDTD region. in the second case.18: Empty rectangular cavity geometry and interface of a 3. MRTD mesh is 2 × 2) scheme or the proximity of the MRTD region to the hard boundary of the domain. necessitating the introduction of several images of highorder wavelet terms [11]. Figures 2.
2. This observation is also in agreement with the Haar MRTD dispersion analysis of [38]. 2.9 1.17.19: Time and frequency domain patterns of electric ﬁeld (Ey ) sampled within the cavity of Fig.6. are shown in Figs.85 1.8 1.9 of the Courant stability limit. In that case. In all cases the time step was set equal to 0.95 x 10 2 4 Time step Ey [Spectral density] 10 0 Interface FDTD 10 −1 10 −2 0 0. whenever a scaling cell greater than the ﬁn dimensions is chosen.20.5 2 2.6 2. the MRTD region is symmetrically placed within the FDTD mesh and only two FDTD cells away from the boundary. restricts the order of the MRTD scheme that can be employed for its analysis. caused by the scaling function deﬁning the ﬁn cell (or wavelets extending beyond the ﬁn .1 NUMERICAL RESULTS: APPLICATIONS Metal FinLoaded Cavity The method of this chapter is applied for the simulation of a metal ﬁnloaded cavity.22 and 2. The patterns of the T E21 . Cavity resonances that are derived via a pure FDTD scheme and the proposed interfacebased method agree well (in both time and frequency domain).23. 2.5 3 3. A third validation case is shown in Fig. 2. similar to the one presented in [27]. The MRTD scheme is of order 4 in both directions.and T E22 modes.21.19 and 2.9 of the Courant limit. This structure is chosen for the reason that the presence of the metal ﬁn within the domain. utmost care is necessary for the compensation of the unphysical coupling of the regions below and above the ﬁn. The smooth patterns conﬁrm the correctness of the scheme and the absence of any type of spurious effects that would corrupt its performance.5 1 1. 2. as shown in Figs.5 Frequency [GHz] FIGURE 2. as obtained by FDTD and the hybrid scheme The total domain corresponds to 20 × 39 FDTD cells. The effective cell size is again 1 cm × 1 cm and the time step equal to 0. In particular. derived via the interface algorithm.32 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS Ey 2 Interface FDTD 0 Electric field −2 1. hence giving rise to a single scaling cell mesh in the MRTD region.
36 32 Haar MRTD (x) (z) FDTD FIGURE 2. as obtained by FDTD and the hybrid scheme limits). MRTD mesh is 1 × 1) . However.5 1 1.85 1. Furthermore.95 x 10 2 4 Time step Ey [Spectral density] 10 0 Interface FDTD 10 −1 10 −2 0.5 2 2. domain split [40]). result in a local increase of operations and consumption of computational resources. the strategies that are followed in this case (e.9 1.21: Empty square cavity geometry and interface of a 4.g. they bring about practical implementation problems.8 1.by 4order MRTD/FDTD mesh conﬁguration (dimensions are given in FDTD cells. 2.20: Time and frequency domain patterns of electric ﬁeld (Ey ) sampled within the cavity of Fig.A NUMERICAL INTERFACE BETWEEN FDTD AND HAAR MRTD 2 33 Electric field Ey Interface FDTD 0 − 2 1. especially when one is interested in developing generic waveletbased CAD tools.18..5 Frequency [GHz] FIGURE 2.
5 Electric field 0 −0. Under these conditions. f c being the cutoff frequency of the T E11 mode of the cavity.5 −1 40 40 20 30 20 0 10 0 xaxis zaxis FIGURE 2. As an excitation. a second.by secondorder scheme is employed (4 by 4 scaling cells).34 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS 1 Ey 0. with its 3 dB frequency chosen to be equal to f c /2.21 (4 by 4 MRTD) For the interfacebased solution of the problem.5 0 −0.23: Electric ﬁeld pattern for the T E22 mode of the empty cavity of Fig.21 (4 by 4 MRTD) . 2. 2. the gridding conditions are given in Fig. In the MRTD region.8 of the Courant limit and FDTD cells of 1 cm × 1 cm are used.22: Electric ﬁeld pattern for the T E22 mode of the empty cavity of Fig. a Gaussian pulse. an absolutely stable performance of the code 1 Electric field E y 0. 2.5 −1 40 30 20 10 30 20 10 0 0 40 xaxis zaxis FIGURE 2. The time step is set at 0. is applied in the FDTD region (at the plane z = 5 cm).24.
55 cm.000 time steps. The deduced electric ﬁeld spatial distribution is shown in Fig.25 0 40 30 20 10 10 0 0 20 40 30 xaxis zaxis FIGURE 2.25: Electric ﬁeld distribution in the metal ﬁnloaded cavity (dominant TEmode) . sampled at the point (x = 1.by 2order MRTD/FDTD mesh conﬁguration (dimensions are given in FDTD cells. MRTD mesh is 4 × 4) was obtained. It is noted that no latetime instabilities were observed over as many as 100. z = 2.A NUMERICAL INTERFACE BETWEEN FDTD AND HAAR MRTD 35 40 36 (x) 2 7 : FDTD/MRTD interface (z) FIGURE 2.75 0.5 0.35 cm).26.25.24: Metal ﬁnloaded cavity geometry and interface of a 2. in order to demonstrate the stability of the solution. in Fig. Moreover.000 time steps. is plotted as a function of time for an arbitrary interval of 18. 2. 1 Electric field E y 0. the electric ﬁeld.000– 20. 2.
ﬁeld wavelet coefﬁcients at the absorber are small enough and the numerical error produced by their omission is controllable. under the condition that the absorber has been placed far enough from the source.5 1.2. Furthermore. For example. in [37].000 2.2 Mesh Truncation 2. this would lead to absorber regions that are much longer than the ones typically used in FDTD applications. a formulation for a matched layer absorber that was proposed in [41] highlighted the difﬁculty of such efforts. For highorder Haar wavelet schemes. terminating an MRTD with ﬁve wavelet levels (fourthorder scheme) in just a single cell of it. In this case.26: Electric ﬁeld sampled within the ﬁnloaded cavity as a function of time for time steps 18.36 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS 1. is equivalent to using an FDTD absorber of 64 grid points per direction.85 1. optimized FDTDuniaxial perfectly .95 x 10 2 4 Time step FIGURE 2. On the other hand. this method is not applicable to open guiding structures because of the signiﬁcant ﬁeld values that typically impinge upon the absorber region of their computational domain.5 0 −0.6. excellent performance of 8–16 cell. Nevertheless. the obvious compromise in accuracy that is related to this approach has negligible effect only in radiation problems.5 1 Electric field Ey 0. the approach of omitting wavelet operations in the absorber region was adopted. However. that essentially stems from the sampling of varying electric and magnetic conductivities by a multilevel basis.6.1 PML Absorber A typical problem related to higherorder and multiresolution formulations of ﬁnite difference schemes is the inherently complex modeling of boundary and material conditions. However.000–20. it may seem necessary that the absorber region of an MRTD mesh be discretized by at least the degrees of freedom of a single MRTD cell. for highorder schemes.8 1. On the contrary.5 −1 −1.9 1.
Hence.27. as expected. while the ﬁve wavelet levels successively reﬁne the resolution of the scheme to λmin /25. In Fig. the concept of the FDTD/MRTD interface is employed for the implementation of a UPML termination of an MRTD domain. 2. is solved by a fourthorder MRTD scheme.25 of a scaling cell. also presented in [17]. In case the FDTD region corresponds to a fraction of a scaling cell.28 × λmin . waveforms of incident and reﬂected electric ﬁelds. A 0–30 GHz Gaussian pulse excitation is used. the waveguide structure of Fig. and the reﬂection coefﬁcient from the slab is calculated. 2. The latter are simply zeroed out and fed back as such to the FWT routine.27: Concept of MRTD mesh termination via an FDTD/MRTD interface .1875 and 0. 2. 2. terminated into a fourcell FDTD PML (half a scaling cell). the application of MRTD update equations needs FDTD grid points beyond the conductor that backs the PML. which in this case is 8 mm. Using this method.30 depicts Zero field grid points MRTD cell P MRTD region FDTDPML region E C 2ndorder MRTD equivalent grid points FDTD grid points FIGURE 2. in Fig. truncated with a 6 and 8 grid point PML corresponding to 0.A NUMERICAL INTERFACE BETWEEN FDTD AND HAAR MRTD 37 matched layer (UPML) absorbers has been recently demonstrated [42]. the scaling cell is 1. prompting the quest for MRTD absorbers that may extend over a fraction of an MRTD scaling cell. an FDTDUPML region encloses an MRTD one and the interface algorithm is applied for the reﬂectionless connection of the two. This concept is explained for a onedimensional case of a secondorder MRTD scheme (with eight equivalent grid points per cell). Their correctness is readily veriﬁed. Additionally. by noticing that they assume the form of spatial ﬁeld derivatives. In this work. analyzed in their scaling and wavelet constituents are shown.29. Fig.28. In particular.
a twodimensional. Evidently. 8 cellUPML with theoretical reﬂection coefﬁcient R = exp(−16) and fourthorder polynomial conductivity variation is used to terminate MRTD meshes that correspond to a 64×64 FDTD domain (Fig.28 . along with the theoretical S11 form derived by transmission line theory. The three case studies are 1 Incident 0.2 0 −0.31). 2.28: Slab waveguide geometry comparative plots of the numerical results derived by the two termination types. Then.56 air PML P E C Excitation plane MRTD/FDTD PML interface (z) (x) FIGURE 2. 2.5 scal 0−wav 1−wav 2−wav 0 600 650 700 750 800 850 0. sampled in front of the dielectric slab of Fig.2 2800 2850 2900 2950 3000 3050 3100 3150 3200 scal 0−wav 1−wav 2−wav 3250 3300 Time step FIGURE 2.38 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS 8 mm air ε r = 2. all three sets of results are in good correlation with each other.4 Reflected 0.29: Waveforms of incident and reﬂected electric ﬁeld (scaling and wavelet terms of order 0–2).
5 ps . B.50) with t0 = 1/(4π f 0 ) and f 0 =1 GHz. is applied in the middle of the domain.31: Case study for the interfacebased termination of a twodimensional MRTD domain .5 mm × 2.34 and correspond to MRTD orders 0 by 0. C in the three ﬁgures. Timedomain ﬁeld waveforms are then sampled at the points indicated as A. 2.32–2.5 mm and the time step is t = 4. and 4 by 4. The FDTD cell size is set equal to 2. An electric current excitation of the form: J y (t) = 1 4 (t/t0 )3 − (t/t0 )4 exp(−t/t0 ) x z (2. 2 by 2.A NUMERICAL INTERFACE BETWEEN FDTD AND HAAR MRTD 39 −10 [dBs] S −30 11 −50 8 point PML 6 point PML TL theory −70 0 10 20 30 Frequency [GHz] FIGURE 2. respectively.30: Numerical and theoretical S11 for the slab geometry depicted in Figs. corresponding 8 8 8 64 FDTD UPML MRTD 64 FDTD UPML Jy FDTD UPML (x) 8 FDTD UPML (z) FIGURE 2.
applied B (x) (z) FIGURE 2. 31). and 16 FDTDUPML cells (0. and 0. The same waveform is determined via the FDTD method. 1.5. 2. (1. shown in Fig. 8. 2. and 2 MRTD cells respectively).by 2order MRTD (MRTD mesh is 8 × 8) .35. (17. 1. In addition. demonstrate an excellent agreement between the pure FDTD scheme and the FDTDUPML terminated MRTD.36 depicts the electric ﬁeld waveform sampled at point B of the MRTD domain of Fig.40 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS A (x) (z) FIGURE 2.by 0order MRTD (MRTD mesh is 32 × 32) to FDTD cells (31. The results. It is noted that the UPML regions in these three examples extend over 4. when terminated at 4.33: FDTDUPML interfaced with a 2. 2.33.32: FDTDUPML interfaced with a 0. demonstrating the ability of the interface to provide MRTD absorbing boundary conditions with efﬁciency and complexity that are independent of the order of the underlying wavelet expansion. Fig. 1) of the 64 × 64 domain.25 MRTD cells respectively. 17).
shows the absence of any signiﬁcant retroreﬂections from the absorbers in all three termination schemes. its propagation toward and its absorption from the 1 0 −1 x 10 4 MRTD+Interf.32–2.34: FDTDUPML interfaced with a 4. the broad time window over which the results are given. Moreover.35: Comparison of FDTD and MRTD/FDTD interface results for the three case studies of Figs. Finally. all four curves agree well.34 .37 and 2. Evidently.by 4order MRTD (MRTD mesh is 2 × 2) at a 64×64 mesh. FDTD Electric field Ey 0 4000 2000 0 −2000 −4000 0 2000 0 −2000 −4000 0 50 100 150 200 250 300 350 50 100 150 200 250 300 350 50 100 150 200 250 300 350 Time step FIGURE 2. 2. terminated at 32 UPML cells. Figs. 2.38 depict the spatial evolution of the pulse.A NUMERICAL INTERFACE BETWEEN FDTD AND HAAR MRTD 41 C (x) (z) FIGURE 2.
since it employs existing UPML implementations. 16 FDTD UPML is used to terminate the MRTD mesh. essentially represents the level of errors due to artiﬁcial reﬂections from the absorber.31 at time steps 20. FIGURE 2. one can consider the proposed method as an effective alternative for MRTD mesh termination.) 2000 Ey [V/m] 0 −2000 −4000 −6000 0 100 200 300 400 500 600 700 800 Time step FIGURE 2. when a 4. the spatial ﬁeld distribution at time step 600. An FDTD solution (with a 32cell UPML) is appended for comparison FDTDUPML boundaries. linking them to MRTD codes via the simple connection algorithm that was earlier described. 8.37: Spatial ﬁeld distribution in the interfaceterminated domain of Fig. In fact. 50 . that are interfaced to the MRTD domain. 2.36: Electric ﬁeld at point B of Fig.42 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS 4000 4 PML cells 8 PML cells 16 PML cells FDTD (ref. 2. This level appears to be at −80 dB. In conclusion.33.
2. even the application of simple absorbing boundary conditions (ABCs) assumes signiﬁcant complexity in the context of multiresolution techniques.2 Absorbing Boundary Conditions Evidently. while the boundary condition itself takes a single cell.6. The reason for that is that ABCs impose mathematical conditions at planes that include only a fraction of the equivalent grid points contained within a terminal cell. 600 2.39: Concept of ABC termination of an MRTD mesh. ABC termination MRTD/FDTD interface MRTD region Excitation plane FIGURE 2. 2. Thus. an FDTD region—no shorter than a single MRTD scaling cell—is introduced. as in the case of a PECbacked PML absorber.31 at time steps 80. Then. the FDTD region terminated into the ABC should at least extend over one scaling MRTD cell . The concept of an interfacebased solution to this question is depicted in Fig. now the FDTD region needs to extend over the degrees of freedom of an MRTD scaling cell—not a fraction of it. For the purpose of applying an absorbing boundary condition.A NUMERICAL INTERFACE BETWEEN FDTD AND HAAR MRTD 43 FIGURE 2.2. FDTD region is terminated into a classically implemented Mur’s ABC.38: Spatial ﬁeld distribution in the interfaceterminated domain of Fig. this mesh truncation method is more suitable to loworder MRTD schemes. Because of this tradeoff.39.
.40. 0. in the case of onedimensional wave propagation (Fig. in a TEM waveguide mode. respectively.9 of their Courant limit.7 CONCLUSIONS A numerical interface between an arbitraryorder Haar MRTD and FDTD was developed and applied in this chapter.40 and 2.41 presents the reﬂection coefﬁcient of each of the four terminations. The FDTD region occupies two MRTD scaling cells for each scheme.39). the thickness of the FDTD region for each case study is: 0. 2.40: Incident and reﬂected ﬁeld waveforms.75 Incident 0.25 0 50 −4 100 150 200 250 2 x 10 Reflected 1 Order=0 Order=1 Order=2 Order=3 0 −1 1950 2000 2050 2100 2150 2200 Time step FIGURE 2. 2.41. 2.0096.0192. A 0–5 GHz Gaussian pulse propagation. Obviously. 2.0384. The two unique features of the proposed technique are that ﬁrst. 0. are terminated into Mur’s ﬁrstorder ABC. for MRTD orders 0–3 Results from a onedimensional implementation of these concepts are shown in Figs.5 0. the obtained ABC performance is similar in all MRTD cases and conﬁrms the capability of the method to provide an efﬁcient means of implementing an ABCtruncation of wavelet grids. Therefore.44 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS 1 0. That is also the reason for the phase difference of the reﬂected waveforms in Fig. with effective cell size of 0. Fig. for Mur’s ﬁrstorder ABC (implemented via an FDTD/MRTD interface). 2. is considered. Four Haar MRTD schemes of orders 0–3.0024 m and time step equal to 0. 0.0768 m. Finally.
while s R = 1/2 R+2 is the shift itself. p (t) d t = 2r/2 (δk .k+1 − δk .1) (A. p d φk r (t)ψk+s R. when applied to largescale problems of current microwave technology structures. in the case of onedimensional wave propagation (Fig.A NUMERICAL INTERFACE BETWEEN FDTD AND HAAR MRTD −60 MRTD order = 0 MRTD order = 1 MRTD order = 2 MRTD order = 3 45 −80 S11 [dBs] −100 −120 −140 0 1 2 3 Frequency [GHz] FIGURE 2.k − δk .39) it does not employ any interpolations or extrapolations and second.k dt dt (t)φk+s R (t) d t = 2r/2 (δk .2) r d ψk . the developed algorithm constitutes a computationally efﬁcient tool for jointly exploiting the advantages of FDTD and MRTD.41: Reﬂection coefﬁcient for Mur’s ﬁrstorder ABC (implemented via an FDTD/MRTD interface). it is applicable for any MRTD order.0 (A.k ) δ p. which is critical for the acceleration of timedomain schemes. 2.k+1 − δk . Hence. The maximum waveletorder that determines the mutual shift of electric/magnetic ﬁeld nodes in axial directions is denoted as R. +∞ −∞ +∞ −∞ +∞ −∞ d φk (t)φk+s R (t) d t = δk .2r −1 dt .k+1 ) δ p. APPENDIX: COEFFICIENTS FOR HAAR MRTD UPDATE EQUATIONS Analytical expressions of integrals that are encountered in the application of the method of moments for the derivation of Haar MRTD update equations are provided below.
λ is the wellknown Kronecker delta and ψ the Haar mother wavelet function. This is done at the preprocessing stage of an MRTD code. p. p.k − δk . 2r − p − 1. (A.5)/2r − 1/2 R+2 − p ξ3 = 2r 1 + ( p + 1)/2r − 1/2 R+2 − p ξ1 = 2r p /2r − 1/2 R+2 − p ξ2 = 2r ( p + 0. r .k+1 dt +∞ −∞ +∞ +D1 (r.k d φk +s R (t)φk (t) d t = δk . r . r .k−1 dt dt (t)φk (t) d t = 2r/2 (δk .k r d ψk +s R.k − δk . r . p ) = −D0 (r. r .k ) δ p. r . 2r − p − 1. 2r − p − 1) and ξ1 = 2r 1 + p /2r − 1/2 R+2 − p ξ2 = 2r 1 + ( p + 0. p )δk . p ) = 2(r +r )/2 ψ(ξ1 ) − 2ψ(ξ2 ) + ψ(ξ3 ) D1 (r. p r (t)ψk+s R.k−1 with D0 (r.0 dt dt r (t)ψk. p. p (t) d t = 2r/2 (δk .k−1 ) δ p. r . p. p (t) d t = D2 (r. r . p ) = 2(r +r )/2 ψ(ξ1 ) − 2ψ(ξ2 ) + ψ(ξ3 ) D2 (r. r .46 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS +∞ −∞ r d ψk . p )δk . D1 . p )δk .k−1 − δk . p. p. D3 and omitting operations that involve multiplications by zero in the main timestepping loop. p. 2r − p − 1) D3 (r.5) (A. p. p +D3 (r. p )δk . r . code efﬁciency is greatly enhanced by a priori recognizing the nonzero coefﬁcients D0 . p ) = −D1 (r.3) r d ψk +s R. p −∞ +∞ −∞ +∞ −∞ d φk +s R r (t)ψk.4) . The previous expressions are readily programmable and allow for the development of arbitraryorder Haar wavelet MRTD codes.2r −1 (A.5)/2r − 1/2 R+2 − p ξ3 = 2r ( p + 1)/2r − 1/2 R+2 − p Finally. D2 . Yet. p (t) d t = D0 (r. δκ.
signiﬁcant wavelet values are expected at space–time regions. This chapter presents an efﬁcient implementation of thresholding applied to a nonlinear problem and reports signiﬁcant execution time savings compared to the conventional FDTD technique. Several approaches to waveletbased mesh reﬁnement have been presented in the literature. at a limited compromise of accuracy. it can be proven that the decay of wavelet expansion coefﬁcients of a square integrable function depends on the local smoothness of the latter [32]. timedomain techniques is the simple implementation of adaptive meshing. In [37. However. dynamically following local variations and singularities of solutions to partial differential equations. while allowing for a denser mesh at parts of the domain where the solution varies less smoothly. In particular.1 INTRODUCTION Multiresolution analysis (MRA) concepts have already been employed in a wide range of applications [11. 3. sparing the arithmetic operations on wavelet coefﬁcients below a certain threshold—small enough according to accuracy requirements—amounts to imposing coarse gridding conditions at those regions.47 CHAPTER 3 Efﬁcient Implementation of Adaptive Mesh Reﬁnement in the Haar Waveletbased MRTD Technique The main attractive feature of waveletbased. 43].37]. Hence. In this sense. A motivating force for this research activity is the fact that waveletbased methods provide the most natural framework for the implementation of adaptive grids. through the application of a thresholding procedure to eliminate wavelet coefﬁcients that attain relatively insigniﬁcant values. Haar wavelets were used to selectively reﬁne the resolution of an underlying .15.14. that the application of the proposed method has led to. little attention has been devoted so far to the investigation of computational costs and accuracy tradeoffs in order to obtain thresholdingrelated operation savings. where high variations in the numerical solution evolve.
the number of wavelet coefﬁcients that are above threshold is computed at each time step. since it was not enhanced by the interpolatory operations. for wavelets of orders 0–3 are shown in Figs. = 10−6 and 10−4 .48 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS FDTD scheme at parts of the domain where this seemed to be a priori necessary. typically employed in subgridded FDTD [44]. In general. . unthresholded wavelet coefﬁcients are doubled. For two different threshold values. On the contrary. the necessity to incorporate wavelets into a timedomain simulation technique is always related to dynamic rather than static subgridding. 2. Section 2. waveletbased meshing was introduced in [32] and applied to electromagnetic structures in [12–15]. This choice is due to the fact that the maximum absolute value of the basis function ψ r p (x) is 2r/2 . The resultant plots. Then. 3. that generates an additional (reﬂected) wavefront.1) is revisited.28. based on ideas originally related to shockwave problems of computational ﬂuid dynamics. Thresholding of Haar wavelet coefﬁcients is applied. the Haar basis and attempting a simple and explicit implementation of an algorithm for the thresholding of wavelet coefﬁcients.2. Execution time measurements for the algorithm as applied to a nonlinear optics problem. through the use of a normalization term 2−r/2 . symmetrically generated by the source in the middle of the domain. j. since the latter—having been extensively studied in the literature—can be nowadays efﬁciently implemented by several existing engines. thresholding amounts to omitting wavelet terms whose maximum contribution to the ﬁeld value is less than . the construction of waveletbased discretizations with a robustness and ﬂexibility comparable to FDTD is still a challenging task” [15]. In both cases. 3. based on the condition: ψr. show that this procedure can actually lead to fasterthanFDTD simulations.6. Then. adaptive. as a result of the pulse incidence on the dielectric slab.2.1 and 3. Hence. the amplitude of the threshold is adapted to the order r of each wavelet coefﬁcient. the resulting method was strictly equivalent to a subgridded FDTD and presented obvious accuracy disadvantages.2 WAVELETBASED FRONTTRACKING In this section. However. the PML absorption of the two wavefronts. the case study of the incidence of a 0–30 GHz Gaussian pulse on a dielectric slab (Fig. where is a predeﬁned threshold. . always in conjunction with highorder basis functions. Evidently. causes a stepwise decrease . all four curves have a similar pattern: Around time step 2000. The current work meets this challenge by building up a twolevel scheme on the simplest wavelet basis. despite small differences between them. p n Ej ≤ 2−r/2 . in a static sense. these being either Daubechies or Bsplines. the complexity of the proposed scheme made clear that “.
1: Unthresholded to total number (%) of wavelet coefﬁcients of orders 0. for a fourthorder Haar MRTD simulation of the problem of Fig.2: Unthresholded to total number (%) of wavelet coefﬁcients of orders 2.6% (in 8192 time steps). thresholding of wavelet coefﬁcients yields a compression in memory requirements by 64.28 . 3. 1. Finally. Fig. Overall.28 in the number of active wavelet coefﬁcients. 3.EFFICIENT IMPLEMENTATION OF ADAPTIVE MESH REFINEMENT 49 FIGURE 3. 2. for a fourthorder Haar MRTD simulation of the problem of Fig.3 depicts the reﬂection coefﬁcient Thres = 10−6 Thres = 10−6 Thres = 10−4 Secondorder wavelets Thres = 10−4 Time step Thirdorder wavelets Time step FIGURE 3. the absorption of the reﬂected wavefront signals the end of the simulation and the decay of the number of unthresholded coefﬁcients to almost zero. Also. 2.
Figs. where ﬁeld values at all time steps for a certain domain need to be stored in memory. wavelets follow successfully the ﬁeld wavefronts. Two observations are now in order: First. one can interpret the Courant stability condition (c t ≤ x). At each time step.and rightpropagating waves. This task is limited to tracking the direction of the velocity of those boundaries. solved with a onelevel Haar MRTD scheme (under same gridding conditions as before). as computed via adaptive MRTD. derived by adaptive MRTD with slab incidence problem of Fig. the regions of unthresholded wavelet coefﬁcients (for = 10−4 ) are found and their boundaries are plotted (in blue dots). The source is located in the middle of the domain and produces two symmetric. demonstrates that wavelet thresholding has a limited effect on the accuracy of the algorithm. . Yet. Moreover. 2. Second. this aspect of wavelet adaptivity will not be further investigated in this work. in order to keep track of the evolution of wavelet coefﬁcients. presupposes the application of memory operations (reshufﬂe of ﬁeld arrays) that have an adverse effect on execution time. as follows: Wavefront position can change at each time step at most by one cell.28 = 10−5 .4 and 3. left. throughout the domain. The excellent comparison of the numerical results to transmission line theory. 10−4 for the dielectric S11 for the dielectric slab. these observations are exploited for the efﬁcient application of thresholding in MRTD. A posteriori processing those and wavelet thresholding them.5 depict six snapshots of electric ﬁeld distribution in space. for the same problem. 3. Indeed. can lead to signiﬁcant memory economy [45]. that there are applications.50 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS 0 −10 −20 S11 [dBs] −30 −40 S11 [dBs] 0 −10 −20 −30 −40 −50 −60 −70 0 5 10 15 20 Frequency [GHz] −50 −60 MRTD TLtheory 25 30 −70 −80 0 5 10 15 20 Frequency [GHz] MRTD TLtheory 25 30 FIGURE 3. It is noted though. one has to essentially study the movement of the boundaries of the unthresholded wavelet regions. In the following.3: Comparison of S11 . because taking advantage of wavelet compression at runtime.
3 ADAPTIVE HAAR WAVELET SIMULATION OF PULSE COMPRESSION IN AN OPTICAL FIBER FILTER As a vehicle for the demonstration of the algorithms developed in this study. the propagation of an optical pulse through a ﬁber ﬁlter and its gradual compression is simulated.5 0 0 200 400 600 800 1000 1200 1400 1600 1800 2000 200 400 600 800 1000 1200 1400 1600 1800 2000 Scaling cell Scaling cell FIGURE 3.5 0 −0.5 −0.5 0. 2.5 1 1 0.5 1.5 51 1.5 1.5 0.5 0 0 200 400 600 800 1000 1200 1400 1600 1800 2000 Scaling cell 1.5 0.5 0.EFFICIENT IMPLEMENTATION OF ADAPTIVE MESH REFINEMENT 1.5 1 1 0. a threshold = 10−4 is used) 3.4: Evolution of wavefronts in the dielectric slab simulation (Fig.5 0 500 1000 1500 2000 2500 3000 3500 4000 MRTD equivalent grid point 2 MRTD equivalent grid point 2 Front boundaries 1. This nonlinear phenomenon is based on a selfphase modulation (SPM) induced negative dispersion that the pulse experiences along the ﬁber and was originally reported and studied in [46].5 Electric field [V/m] 1 Electric field [V/m] 0 500 1000 1500 2000 2500 3000 3500 4000 1 0. It is noted that .28) and regions of unthresholded wavelet coefﬁcients (boundaries are deﬁned by blue dots.5 Front boundaries 200 400 600 800 1000 1200 1400 1600 1800 2000 1.5 0 −0.5 Scaling cell Electric field [V/m] Electric field [V/m] 1 1 0.5 0 0 −0.5 0 500 1000 1500 2000 2500 3000 3500 4000 0 500 1000 1500 2000 2500 3000 3500 4000 MRTD equivalent grid point 2 2 MRTD equivalent grid point Front boundaries Front boundaries 1.
3.5 −0.5 0.1 Formulation of Haar MRTD Scheme E(z.2) βz (3.3) . t) = E F (z. by means of the Battle–Lemarie cubic splinebased SMRTD technique.5 Electric field [V/m] 1 0. Numerical modeling of this case was also pursued in [47].5: Evolution of wavefronts in the dielectric slab simulation (continued) the refractive index in the ﬁber assumes the form: n(z) = n0 + n1 cos(2β0 z) + n2 E2 .5 0 500 1000 1500 2000 2500 3000 3500 4000 MRTD equivalent grid point 2 MRTD equivalent grid point 2 Front boundaries 1. (3. where the cosinusoidal term is due to a periodic structure (grating) written within the core of the ﬁber.52 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS 1. t) e − j (βz+ωt) . In this work.5 1 1 0. (3. a one waveletlevel Haar MRTD scheme is used.5 0.5. the following system of equations is deduced: ∂ EF n0 ∂ EF + = j κ E B e −2 j ∂z c ∂t n0 ∂ E B ∂ EB − = − j κ EF e 2 j ∂z c ∂t βz + j γ E F 2 + 2E B 2 E F − j γ E B 2 + 2E F 2 E B . resulting in execution time that was reported to be larger than FDTD by a factor of 1.5 Electric field [V/m] 0 500 1000 1500 2000 2500 3000 3500 4000 0 0 −0.1) The electric ﬁeld in the ﬁber is decomposed in forward and backward propagating waves as: Substituting this expression into Maxwell’s equations and discarding terms with spatial variation faster than e j 2βz (slowly varying envelope approximation). 3. for the purpose of estimating the improvement in computational performance that exclusively originates from the dynamic adaptivity of MRTD rather than the highorder of an underlying scaling basis.5 Front boundaries 200 400 600 800 1000 1200 1400 1600 1800 2000 1. t) e j (βz−ωt) + E B (z.5 0 0 200 400 600 800 1000 1200 1400 1600 1800 2000 Scaling cell Scaling cell FIGURE 3.
φ = B.3) can be retrieved from [47]. E F. by ﬁrst expanding both forward and backward ﬁelds in terms of Haar scaling φm (z) = φ(z/ z − m) and wavelet ψm (z) = ψ(z/ z − m) functions in space and pulse functions h k (t) = h(t/ t − k) in time. ψ k−1 Em −s B. Upon substitution of (3. B.2).ψ B.3) and application of the method of moments. − 2 k Em B.φ k Am k Em + k Bm k Em F.ψ . − 2 k Em + k Em−1 + j Sa ( β z/2) k Em B.φ F. (3.ψ k Em+1 F. The FDTD update equations for the system of (3. κ = π n1 /λ and β(ω) = n0 (ω)ω/c − 2πn0 /λ0 and λ0 the freespace wavelength that satisﬁes the Bragg condition for the grating.ψ F. (2. the system of Haar MRTD update equations is formulated.5) F. φ k+1 Em = F.2).φ = F.φ (3.4). (3. Haar MRTD equations are derived. ψ k+1 Em β(m+1/2) z Sa( β z) k Em − j Sa ( β z/2) k Em B.φ B.φ k Em+1 − k Em−1 + F. Field expansions read: Ex (z. In this work.6). deﬁnitions (2.φ F. E B. φ k−1 Em −s F. (3. φ k+1 Em β(m+1/2) z Sa( β z) k Em F.ψ + 2 j κs ze −2 j + 2 jγ s z F. ψ k−1 Em +s F.φ k Em x.ψ k Am k Em + k Bm k Em B.φ k Em+1 F.ψ k Cm k Em + k Dm k Em . update equations for the mth cell values of E F.ψ k Em+1 B.4) into (3.φ k Em+1 − k Em−1 + B.ψ .φ . In particular. For these functions. φ k−1 Em +s B.φ + k Em−1 F.φ F. (2.k=−∞ where x = F.φ k Em+1 + 2 j κs ze −2 j + 2 jγ s z B.ψ .ψ at the kth time step read: F.EFFICIENT IMPLEMENTATION OF ADAPTIVE MESH REFINEMENT 53 with γ = π n2 /λ.φ k Cm k Em + k Dm k Em − k Em−1 + B.4) m.ψ (3.φ B.ψ B.ψ k Em+1 − 2 k Em F.7) B.ψ (3.φ .14) are followed.φ . . ψ k+1 Em β(m+1/2) z Sa( β z) k Em + j Sa ( β z/2) k Em F.8) B.ψ + 2 j κs ze 2 j + 2 jγ s z B.ψ + 2 j κs ze 2 j + 2 jγ s z β(m+1/2) z Sa( β z) k Em B. t) = ∞ x.ψ k Em+1 − k Em−1 + F.ψ + k Em−1 B. E B.φ (3.ψ = B.6) F.φ B.ψ φm (z) + k Em ψm (z) h k (t). − 2 k Em + k Em−1 − j Sa ( β z/2) k Em F.
ψ .φ F.10) by local modiﬁcation of the update equations.φ F.5)–(3.ψ z is the MRTD =  k Em 2 +  k Em 2 + 2  k Em 2 +  k Em 2 .ψ B. (3.φ F.ψ F.φ B.ψ = 1/2  k Em + k Em 2 −  k Em − k Em 2 + 2  k Em + k Em 2 −  k Em − k Em 2 F. then. 3.ψ F.9) . 3.11) F/B k E M+1 1 =√ 2 1 =√ 2 F/B.φ F. It is worth mentioning that operations in (3.2 Boundary Conditions Assuming an optical ﬁber that extends from z = 0 to z = L. and g k = E F (0. k t) is the discretetime sample of the excitation function at the kth step.12) F/B k E M+2 F/B. where B. k Cm k Dm = B.ψ . − k Em (3.φ 2 k Em  + B.φ B.ψ t/n0 z. thus facilitating their adaptive application.φ k Em √ F. t) = 0.ψ . applied at the Mth equivalent grid point (Fig.φ B.13) .and waveletrelated ones. t) = A 2 2 (1 + j ) e −t /(2α ) .ψ 2 k Em  +2  F.φ 2 k Em  B.φ k Em (3. = 1/2  k Em + k Em 2 −  k Em − k Em 2 + 2  k Em + k Em 2 −  k Em − k Em 2 B. by the wavelet transform: F k EM 1 =√ 2 F.8) can be readily split in scaling.ψ + k Em = g k / 2.φ k Em+1 + k Em+1 F/B. 2 E B (L. For example.ψ F.3. k Am k Bm F. (3. Sa (w) = sin2 w/w and s = c scaling cell.φ + F.6).ψ 2 k Em  B. F/B. Furthermore.φ F. if z = 0 belongs to the mth scaling cell of the domain. (3.54 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS with Sa(w) = sin w/w.ψ .φ B. the following boundary conditions are imposed [47]: E F (0. based on the explicit enforcement of those conditions (without extrapolations or interpolations as proposed in the past [14]).ψ B.
φ k+1 Em = 1 (g k + Ak ) . B.φ k En (3. as: F.φ k Em B.16) .φ k En F.ψ k+1 Em = 1 (g k − Ak ) 2 (3.φ k Em F.ψ + 2 j κs zSa ( β z/2) e − j + 2 jγ s z 1 2 F.φ k Em (3.ψ 2 k Em F. at z = L is applied via the expressions: Bk+1 = Bk−1 − 2s B. explicit update equation is F.15) Similarly.14) − F. scaling and wavelet terms of the forward ﬁeld are updated at the excitation cell.φ k Em+1 + k Em+1 − g k β(M+3/2) z B.6).φ k Em − Then.EFFICIENT IMPLEMENTATION OF ADAPTIVE MESH REFINEMENT 55 Optical fiber filter mcell (m + 1)cell (n − 1)cell ncell M M+1 M+2 N−1 N (z) N+1 z=0 z=L : MRTD EGP FIGURE 3.ψ − k Em B.ψ + k Em . the following.6: Haar MRTD equivalent grid points at the beginning and the end of the ﬁber Deﬁning an auxiliary variable: Ak ≡ derived: Ak+1 = Ak−1 − 2s √ F 2 k EM+1 .ψ + k En + F. 3.φ k En−1 + k En−1 B.ψ β(M+1/2) z 2 F. the hard boundary condition on the backward wave at the nth cell of the domain (Fig.ψ − k En 2 + 2 j κs zSa ( β z/2) e − j + 2 jγ s z 1 2 B. 2 F.ψ 2 k Em + B.
A maximum conductivity σmax = 0.56 √ B where Bk = 2 k EN .φ k Em B. L. scaling cell sizes for MRTD are chosen to be z = 0.23 10. at the nth MRTD scaling cell.1: Validation Data for Nonadaptive MRTD Code PARAMETER Cells CPU time (s) max E F (z = L) 2 FDTD.6042 MRTD. as the introduction of one wavelet level reﬁnes the resolution by a factor of two. absorbing boundary conditions for backward and forward waves are imposed at z = 0. The reason for this expected slowdown is that the factor of increase in operations per cell between MRTD and FDTD is greater than the ratio of FDTD to MRTD cells (equal to two). The time step is t = 0.6635 FDTD. as follows: κ L = 4.002L and z = 0. Then. 3.3541 MRTD. the conductivity was assumed constant within each scaling cell and therefore it varied less smoothly than the corresponding FDTD TABLE 3.1 S/m is used in all subsequent simulations.25 10. Matched layers of 500 and 250 cells terminate the FDTD and MRTD meshes respectively.3. The slight difference can be attributed to the fact that in the MRTD absorber. Respectively. thus allowing for the reduction of the number of cells in half. the scaling and wavelet terms of the backward ﬁeld.004L. Simulation data for both aforementioned cases are provided in Table 3. The absorbers are implemented as in [47]. β L = 12.56 10. Execution times for the relevant FDTD and MRTD cases over 12. The two methods agree well on the peak of the transmitted intensity (which corresponds to the forward wave intensity at the end of the ﬁber). I 500 48. via matched layer absorbers.000 time steps (on a Sun Ultra 80 workstation at 500 MHz) reveal that the nonadaptive MRTD code is slower than FDTD by a factor of 11– 16%. results of a nonadaptive Haar MRTD code are compared to FDTD.1 S/m. II 250 29. with respect to the length L of the ﬁlter.003125n0 L/(10c ) and two cases for the FDTD cell size are considered: z = 0. by expanding their quadratically varying conductivities in Haar scaling functions. II 500 25.002L. The parameters of the ﬁber are normalized.1.55 10. I 1000 43. with maximum conductivity σ0 = 0. are updated according to the scheme: B.001L and z = 0.3 Validation For validation purposes. γ L = 2/3.ψ = k Em = MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS Bk 2 Finally.4283 .
this observation is utilized for the development of a computationally efﬁcient approach to the problem of thresholding of wavelet coefﬁcients. In consequence. Finally. the middle. Evidently. Moreover. since it includes pulse waveforms (extracted via 1000 cell FDTD and 500 cell MRTD).EFFICIENT IMPLEMENTATION OF ADAPTIVE MESH REFINEMENT 12 57 : MRTD : FDTD Forward field intensity E 2 F 8 z =0 z =L/2 z =L 4 0 2000 4000 6000 8000 Time step FIGURE 3. Figs. as probed at the beginning. z = c t. In the following. the application of thresholding with as few operations (namely checks on wavelet . two questions are addressed in this work: First. Fig. A similar pattern for wavelet behavior was obtained in [15].7: MRTD and FDTD results for the forward ﬁeld intensity at the beginning.3. the nonadaptive MRTD results can be henceforth used as a measure for accuracy estimation of adaptive MRTD algorithms. the evolution of both the wavelet coefﬁcients and the ﬁeld itself takes place along the characteristic line of forward ﬁeld propagation. and the end of the optical ﬁber ﬁlter (1000 grid points) absorber conductivity. 3. ?? depicts the pulse compression along the ﬁber. the middle and the end of the ﬁber ﬁlter. excellent correlation between the two methods (FDTD and MRTD) is demonstrated. ?? and ?? show snapshots of the forward ﬁeld intensity and the magnitude squared of the forward ﬁeld wavelet coefﬁcients plotted in space–time coordinates.4 Implementation and Performance of Adaptive MRTD Regarding the development of adaptive algorithms. where nonuniform multiconductor transmission line equations were solved via a biorthogonal wavelet basis. Again.
9: Spatiotemporal evolution of forward ﬁeld wavelet coefﬁcients (square of magnitude of wavelet coefﬁcients is shown) .8: Spatiotemporal propagation and compression of the nonlinear optical pulse (forward ﬁeld intensity is shown) FIGURE 3.58 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS FIGURE 3.
EFFICIENT IMPLEMENTATION OF ADAPTIVE MESH REFINEMENT
59
coefﬁcients being absolutely above or below a threshold ) as possible and second, the stable exploitation of the compressed representation of the solution for the reduction in update operations at each time step. For this purpose, the following method is adopted [12, 15, 32], for both forward and backward propagating wave arrays:
• • •
All cells are initialized as being above threshold (referred to as active). Scaling coefﬁcients are updated throughout the mesh, along with active wavelet terms. In all operations, only scaling and active wavelet terms are taken into account. At each time step, the magnitudes of wavelet terms at cells that are designated as active are compared to an absolute threshold . The corresponding cells remain active only if their wavelets are above the threshold. The region of the active cells is extended to all nearest neighbors of the latter. Since only one wavelet level is used here, this implies cells that are immediately to the left or to the right of cells on the border of active regions.
•
Hence, thresholding checks and update operations are limited to a subset of the ﬁeld coefﬁcients. Note that the use of the Haar basis and a single wavelet level scheme, keeps the implementation of this adaptive algorithm relatively simple and readily expandable to three dimensions. On the other hand, the complexity of numerical solvers based on higherorder basis functions and multiple wavelet levels has regularly undermined the potential of adaptivity to yield execution times better than FDTD. Furthermore, the frequency of thresholding checks is implicitly dependent on the CFL number s of the simulation, which effectively determines the maximum number of cells that a front may move through in a single time step. Approximately, a thresholding check window should be at most equal to 1/s , with the previously described procedure strictly corresponding to the value of s = 1. Physically, the thresholding algorithm, that was ﬁrst introduced for the numerical solution of shock wave problems in [32], assumes the evolution of wavelet coefﬁcients along wavefronts deﬁned from the characteristics of a given problem. Then, the purpose of adding pivot elements, to extend the domain of active coefﬁcients, is actually the tracking of these wavefronts as they move throughout the computational domain. The concept of this algorithm is schematically explained in Fig. ??. As shown, adding the pivot elements at the two edges of a wavefront, ensures the tracking of the direction of its movement. Indeed, if that moves to the left, the left pivot element is activated at the “next” time step. Then, the active wavelet region moves to the left, with the wavefront. Otherwise, it follows a movement of the front to the right. To this end, the second case that was presented in Section ?? is repeated here by using several absolute thresholds and CPU time measurements are carried out. Also, three thresholding
60
MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS
: active wavelet coefficients : pivot elements
FIGURE 3.10: Graphical presentation of the waveletbased, fronttracking algorithm
windows are investigated, with the outlined algorithm being applied every one, two, and four time steps respectively. Although the latter was slightly larger than 1/s , it produced satisfactory (accuracywise) results. In all cases, thresholds up to 0.001 led to simulations that clearly resolved the pulse compression and suffered from errors (with respect to the unthresholded MRTD, which is used as a reference solution) of less than 1%. In Figs. ?? and ??, forward ﬁeld intensity waveforms sampled at z = L, for thresholds 0.1, −4 10 and 10−7 are shown, along with the incident pulse. The last two are in good agreement both with each other and with the previously presented FDTD and MRTD results, while the ﬁrst suffers from signiﬁcant numerical errors, that demonstrate themselves as a ripple corrupting the pattern of the waveform. More explicitly, CPU economy with respect to FDTD and error in the peak transmitted forward ﬁeld intensity are plotted (with respect to the absolute threshold ), in Figs. ?? and ??. It is thus shown, that the adaptive MRTD code can extract the solution to this problem, at a CPU time reduced (compared to FDTD) by a factor close to 30%, with errors limited at the order of 0.1%. Thresholding operations in this problem represented a worstcase scenario for the computational overhead that the adaptive algorithm may bring about, for the following reasons: First,
EFFICIENT IMPLEMENTATION OF ADAPTIVE MESH REFINEMENT
61
Threshold = 10 −7
Incident forward field intensity E (z=0) 2
10
2
12
8
Transmitted forward field intensity EF (z=L)
2500 3500 4500
10
F
8
6
6
4
4
2
2
0 1500
0 4000
6000
8000
10000
Time step
Time step
FIGURE 3.11: Incident and transmitted forward ﬁeld intensity for a threshold of 10−7
the geometry was a onedimensional one, a signiﬁcant part of which was almost throughout the simulation occupied by the propagating pulse and second, these operations involved complex numbers and nonlinear terms. Therefore, the fact that an accelerated performance of adaptive MRTD (with respect to FDTD) was achieved is important and demonstrates the potential of the algorithm for larger geometries.
Threshold = 10−4
12
Transmitted forward field intensity E (z=L)
Transmitted forward field intensity E (z=L)
Threshold = 10−1
12
2
F
2 F
10
10
8
8
6
6
4
4
2
2
0 4000
6000
8000
10000
0 4000
6000
8000
10000
Time step
Time step
FIGURE 3.12: Transmitted forward ﬁeld intensity for thresholds of 10−4 , 10−1
with respect to FDTD.14: Relative error (%) in the peak intensity of the transmitted forward ﬁeld for adaptive Haar MRTD . window = 2 Thres. for absolute thresholds from 10−7 to 0. window = 4 10 1 10 0 10 −1 10 −2 10 −7 10 −6 10 −5 10 −4 10 −3 10 −2 10 −1 Threshold FIGURE 3. window = 1 Thres. window = 2 Thres.1 applied to Haar MRTD 2 10 Relative error in peak intensity [%] Thres. window = 1 Thres.62 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS 50 40 Thres. window = 4 Economy[%] in CPU time 30 20 10 0 −7 10 10 −6 10 −5 10 −4 10 −3 10 −2 10 −1 Threshold FIGURE 3.13: CPU time economy for the adaptive MRTD code.
waveletbased code surpasses the efﬁciency of the conventional FDTD. this chapter demonstrated Haar waveletbased simulations with adaptive meshing that achieved betterthanFDTD execution times. The satisfactory performance of the proposed technique stems from its relative simplicity that allows for an efﬁcient implementation of its two components: thresholding tests of wavelet coefﬁcients and operation savings while performing updates of ﬁeld arrays. To the extent of the author’s knowledge this is the ﬁrst time when an adaptive.EFFICIENT IMPLEMENTATION OF ADAPTIVE MESH REFINEMENT 63 3. .4 CONCLUSIONS Based on the study of a nonlinear pulse compression by an optical ﬁber ﬁlter. not only in terms of memory but also in terms of execution time requirements.
.
the FDTD stability and dispersion properties impose severe limitations on the choice of the cell size and the time step of the method. to emphasize the dynamic evolution of the underlying mesh. Thus. According to those. timedomain solver for Maxwell’s equations. The challenge of accelerating FDTD simulations for practical geometries has been addressed in the past with a variety of static subgridding techniques [10. .and threedimensional implementation. However.65 CHAPTER 4 The Dynamically Adaptive Mesh Reﬁnement (AMR)FDTD Technique: Theory The ﬁnitedifference timedomain (FDTD) method is combined with the adaptive mesh reﬁnement (AMR) technique of computational ﬂuid dynamics to achieve a fast. instead of setting up a statically adaptive mesh. rendering its application to complex structures computationally expensive. It is especially suitable for wideband applications since it allows for the characterization of a given structure in a broad frequency range. through a single simulation. This chapter is aimed at providing the reader with a detailed description of the algorithm. This combination allows the resulting technique to adapt to the problem at hand.1 INTRODUCTION The ﬁnitedifference timedomain technique [6] has been extensively employed in the modeling of microwave integrated circuits [48]. by recursively reﬁning a coarse grid in regions of large gradient of electromagnetic ﬁeld energy. which is the subject of extensive previous research. 4. adaptively reﬁnable moving cartesian mesh. the idea of a dynamically adaptive mesh is pursued. the algorithm is still implemented within the context of FDTD. optimally distributing computational resources in a given domain as needed. However. based on a threedimensional. 44]. because of the optimal suitability of the latter to the nature of timedomain electromagnetic simulations. with secondorder accurate update equations and is referred to as dynamically AMRFDTD. ﬁlled with all the information needed for its two. The chapter is concluded with a comparison between the dynamically AMRFDTD and the Haar MRTD approach that was discussed in the last two chapters.
while the . In the context of computational ﬂuid dynamics. the presence of metallic edges. Therefore. is only a suboptimal solution to the mesh reﬁnement problem in the framework of timedomain analysis. techniques such as FDTD and TLM essentially register the history of a broadband pulse propagating in a device under test. static mesh reﬁnement. where the use of a dense mesh is not necessary after a certain time step. However. a movingwindow FDTD (MWFDTD) method was proposed for the tracking of the forward propagating wave in the twodimensional terrain environment of a wireless channel [49]. it is so for a (potentially small) fraction of the total simulation time. As a result. However. The application of AMR is based on the use of a hierarchical mesh. There may also be dense mesh regions. embedded in a coarser global one. Hence. on the contrary. where the modeling of phenomena as common as signal reﬂection and branching would require multiple and potentially rotating windows. this approach ignores the dynamic nature of timedomain ﬁeld simulations. as dictated by physical considerations. again in a recursive manner. a sharp edge of a microstrip structure is not continuously illuminated by the pulse excitation. respectively. it could not track reﬂections (which were absorbed by terminating boundaries of the window). The regions of the computational domain that need further mesh reﬁnement are detected via error estimates or indicators such as gradients of the quantity to be solved for. For example. despite the fact that its implementation is associated with additional interpolation and extrapolation operations in both space and time. These can then be coarsened. The single moving window used by the method was characterized by ﬁxed size and velocity and therefore. which covers the entire computational domain. the generalization of such waveletbased algorithms to three dimensions presents a signiﬁcant added complexity. This procedure can be associated with the algorithm of the previous chapter. In fact. similar to ideas previously proposed in [50. The use of local mesh reﬁnement typically results in signiﬁcant computational savings compared to the conventional FDTD. More recently. Dense and coarse mesh regions are organized via a clustering algorithm that is accompanied by regular checks (every certain time steps) of the error estimates. which used wavelet ﬁeld expansions in order to track the spatiotemporal evolution of shockwave and nonlinear optical pulse propagation problems. the technique of adaptive mesh reﬁnement (AMR) was introduced in [52]. which guide the process of migration of a cell from one level of resolution to another. which is widely employed in frequencydomain simulations and has been incorporated in commercial ﬁniteelement tools.66 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS local mesh reﬁnement is pursued in a priori deﬁned regions of a computational domain. would call for a locally dense mesh. or high dielectric permittivity inclusions. during which a local mesh reﬁnement around it is needed. for the solution of hyperbolic partial differential equations. 51]. along with its multiple reﬂections from parts of the latter. the MWFDTD is not wellsuited for microwave circuit simulations. recursively developed through the reﬁnement of a coarse root mesh.
In this chapter. · · · N − 1. Each cell occupies a volume x × y × z. In general. Instead of using ﬁxed subgrids.5 compares AMRFDTD to the MRTD technique outlining their interesting similarities along with their distinctive differences. the conventional FDTD is combined with the AMR method. y. z denote the cell sizes in x. by reﬁning the Yee cells of an underlying root mesh by a predeﬁned factor. where x. in order to formulate an efﬁcient AMRFDTD technique of superior performance [53]. y. while Section 4.4 outlines how cells that need further mesh reﬁnement or cells that can be removed from the mesh are detected. Inhomogeneous material properties can be readily assumed. by letting the dielectric permittivity and magnetic permeability of the structure be generic functions of the space variables = (x. Let us consider the subdivision of the domain A in Nx × Ny × Nz Yee cells. za 1 ≤ z ≤ za 2 . in the sense that it can support multiple levels of mesh resolution. terminated by closed or absorbing boundaries. clustered. according to the evolution of ﬁeld distributions in space and time. when a Gaussian pulse propagating along a microstrip line is simulated. z). thereby reﬁning only the region that surrounds the propagating pulse. where the factor Ns (here taken equal to 2) is the mesh reﬁnement factor.THE DYNAMICALLY ADAPTIVE MESH REFINEMENT (AMR)FDTD TECHNIQUE 67 implicit relation between the actual ﬁeld values and the wavelet expansion coefﬁcients renders the application of boundary conditions. y. which are essential for the connection of nested meshes of different resolutions. a cartesian FDTD mesh occupies a rectangular region. a twolevel scheme would enclose Yee cells of size x × y × z and x/Ns × y/Ns × z/Ns . In our terminology. j. and managed by the AMRFDTD scheme. and zdirections. the AMR accuracy is comparable to that of a uniformly (throughout the entire computational domain) dense mesh FDTD. As an example. Note that this chapter’s algorithm is a “multilevel” one. while an Nlevel scheme would enclose Yee cells of size x/Nsi × y/Nsi × z/Nsi . respectively. y. 4. Section 4. computationally burdensome.2. k).3 refers to the update procedure of the tree of meshes that the algorithm uses in order to adaptively track the ﬁeld evolution. indexed by a triplet (i.2 AMRFDTD: OVERVIEW OF THE ALGORITHM A = (x. the adaptive mesh reﬁnement scheme successfully tracks the movement of the pulse. 1. y. μ = μ (x. In this case. that the algorithm under development will selectively and locally reﬁne. ya 1 ≤ y ≤ ya 2 . The mesh of these Nx × Ny × Nz cells is the coarse grid of region A. xa 1 ≤ x ≤ xa 2 . for i = 0. Section 4. Finally. z) . this method uses subgrids that are adaptively deﬁned. z). The chapter is organized as follows: The general structure of the AMRFDTD algorithm is presented in Section 4. .
y/2. 4.1: (a) Geometry and (b) data structure of a threelevel mesh tree . or level1 mesh. a coarse mesh has been deﬁned in the rectangular region A.68 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS Let us now assume that there are cells within the domain A whose reﬁnement is necessary. which is also called the root mesh and covers the entire computational domain. The child meshes of the same parent may share an edge. will be called the root mesh. Throughout the algorithmic development of the AMRFDTD. henceforth referred to as mesh A. · · · N which belong to A. while a dashed line corresponds to a boundary shared by meshes of the same level. again reﬁning the cell sizes involved by a factor of 2. There is only one level1 mesh.1(b) a solid line corresponds to a child– parent relation. reducing the Yee cell volume of the initial mesh by a factor of 8. but they may not overlap otherwise. C1 B1 C2 C3 B2 C4 B3 B1 A level 1 B2 B3 level 2 C1 C2 C3 C4 level 3 A (a) (b) FIGURE 4. Each rectangular region Bn is subdivided in Yee cells of dimensions: x/2. enclosing ﬁner meshes in rectangular subregions Bn of the latter. or meshes Bm . The hierarchically deﬁned meshes. To summarize. Thus. The mesh of region A. m = 1.” The mesh tree is regenerated every NAMR time steps through a recursive mesh reﬁnement procedure. Hence. Each levelm + 1 mesh is created by reﬁning a subset of the cells of a levelm mesh in a rectangular region by a factor Ns (in each direction). can be assembled in the data structure of a “mesh tree. Recursively. z/2. These cells are ﬁrst clustered together and then covered by rectangular subregions Bm . it will be ensured that these subregions can share planar boundaries.1. the two meshes form a child–parent relation. or level2 meshes.4). that the computational domain consists of in an AMRFDTD simulation. according to certain accuracy criteria (the discussion of the latter is deferred to Section 4. as required by the evolution of the ﬁeld solution. In Fig. This is important in order to preserve the possibility of further reﬁnement of these subregions independently from each other. 4. 2. yet they cannot overlap. each Bn can be further reﬁned to have its own child meshes. The meshes of regions Bm . An example is shown in Fig. will be called child meshes of A. a reﬁnement factor of 2 is used in every direction.
y. Solid arrows are for electric ﬁeld. Furthermore. Yee cell dimensions are: x/2 M−1 . To deﬁne the time step of AMRFDTD.4. y/2 M−1 . In this work. (4. depending on whether the reﬁnement factor is odd or even. The shaded areas represent reﬁned Yee cells According to the convention of the Yee cell in FDTD (Fig. the following observations need to be made. An alternative case. This implies that the root mesh time step t (from now on referred to. while magnetic ﬁeld components are sampled at face centers. where x. For a levelM mesh. the Courant number is ﬁxed to a constant value s in all meshes. The sampling points of a parent mesh may coincide or not with the sampling points of its child mesh.THE DYNAMICALLY ADAPTIVE MESH REFINEMENT (AMR)FDTD TECHNIQUE z 69 Hz Ey Ex Hy Hx Ez x y FIGURE 4. every NAMR time steps.1) . The main difference between standard subgridded FDTD methods and AMRFDTD is the dynamic mesh generation which is pursued in the latter. hence the grid points of child and parent meshes do not coincide. with respect to the root mesh). the reﬁnement factor is 2 (or powers of 2. where the choice of a reﬁnement factor of three renders the parent mesh sampling points also child mesh sampling points can be found in [44].2). electric ﬁeld components are sampled at the center of the edges of each cell. as AMRFDTD time step) is given as: t=s 1 u max p 1 1 1 1 + + 2 2 x y z2 . z are the root mesh Yee cell dimensions.2: Sampling points of electric and magnetic ﬁeld components in a Yee cell. whereas hollow arrows are for magnetic ﬁeld. z/2 M−1 . respectively.
4. 4. If it is an integer multiple of NAMR . the types of interfaces that occur in an AMRFDTD domain are presented. (4. the minimum time step of the algorithm is only used for the update of regions of large ﬁeld variations. Three separate cases of boundaries can be identiﬁed: 1. level2 meshes are updated twice as many times as the root mesh. Consider two child meshes B1 and B2 .3. In this section. including absorbing and/or perfect electric conductor boundary conditions. keeping s ﬁxed.2) For example. Update ﬁelds of the child meshes 2 M−1 times. 44]. yet the interconnection of the different resolution meshes that march in time at different time steps is an issue to be addressed explicitly. The next two sections are aimed at explaining these steps in detail. Segment e a: It is deﬁned here as a physical boundary (PB). otherwise return to step 1. Copy ﬁelds from child meshes back to the root mesh. Thus. 2. Check whether the maximum time step has been reached. a salient feature that is also part of the ﬁxed subgridding algorithms of [10. 3. embedded in a root mesh A.1) for a p mesh of level M. for the time steps of the latter. perform adaptive mesh reﬁnement to create a new mesh tree. Update ﬁelds of the root mesh. If so. a twodimensional case is discussed (readily extensible to three dimensions). another shortcoming of the conventional FDTD is addressed. along with their treatment in the update process. Applying (4. 4. 4. where M is the resolution level of the mesh. and carry the ﬁeld values from the old mesh tree to the new mesh tree. To facilitate the presentation of the different cases.3. yields a time step tM for this mesh. . The loop of the AMRFDTD operations is as follows: 1. equal to: tM = t/2 M−1 . Check the number of time steps executed.3 MESH TREE AND FIELD UPDATE PROCEDURE IN AMRFDTD AMRFDTD applies the wellknown ﬁeld update equations of FDTD for each mesh. as opposed to the whole domain.70 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS where u max is the maximum phase velocity in the computational domain. Copy ﬁelds from the root mesh to the boundary of the child meshes. terminate the simulation.1 Categorization of Boundaries of Child Meshes The different categories of child mesh boundaries that may practically occur are shown in Fig.
THE DYNAMICALLY ADAPTIVE MESH REFINEMENT (AMR)FDTD TECHNIQUE
71
e B1
d c
B2
a
b
A
FIGURE 4.3: Types of boundaries of a child mesh
2. Segment cd: This is a boundary between “sibling” meshes (SB). 3. Segments ab, bc, de: These are boundaries between child and parent meshes (CPB). Evidently, it is possible that a boundary may belong to more than one of the aforementioned categories. Then, its classiﬁcation is based on the following hierarchy: PB > SB > CPB.
4.3.2
Update of the Mesh Tree
In the following, the update of a twolevel mesh tree (with root and level1 meshes) is discussed. The steps outlined here can be recursively extended to mesh trees of more levels, by considering, for example, level2 meshes as roots for level 3meshes and so on. As before, t is the time step of the root mesh, while the child level1 mesh uses a time step t/2. Let us consider the order of updates, assuming that the timemarching procedure has reached the time t = n t. As a result, it is assumed that the root mesh contains the updated values of the electric ﬁeld component grid points at n t and those of the magnetic ﬁeld component grid points at (n − 1/2) t. In addition, the child meshes contain the updated electric and magnetic ﬁeld values at n t and (n − 1/4) t, respectively. Then, the following procedure is applied. 1. Backup magnetic ﬁeld components of the root mesh at (n − 1/2) t. Obtain their values at (n + 1/2) t, by applying the FDTD update equations. 2. Backup the electric ﬁeld components of the root mesh at n t. Obtain their values at (n + 1) t, by applying the FDTD update equations.
72
MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS
3. For each child mesh, apply the update equations to obtain the magnetic ﬁeld values at (n + 1/4) t. 4. For each child mesh: (a) Update the interior (nonboundary) electric ﬁeld grid points, to obtain their values at (n + 1/2) t. (b) Update the boundary electric ﬁeld grid points, at boundaries of the PB, SB, CPBtype, to obtain their values at (n + 1/2) t. Obviously, these updates are nontrivial, since they invoke grid points of the root mesh, calculated at time steps of the child mesh. Therefore, interpolation needs to be carried out, in a way that is analyzed in the next subsection. 5. For each child mesh, backup the magnetic ﬁeld components at (n + 1/4) t, and obtain their values at (n + 3/4) t, by applying the FDTD update equations. 6. Repeat steps 4a, 4b, to advance the electric ﬁeld values of the child meshes to (t + t). 7. For each child mesh: (a) Put the spatially interpolated electric ﬁeld nodal values at (n + 1) t back to the parent mesh, excluding CPB grid points. (b) Put the spatially interpolated magnetic ﬁeld nodal values at (n + 1/2) t back to the parent mesh. This process can be readily extended to the multilevel case, where meshes of levels 1, 2, 3, . . . , NL are supported. If Ns is the mesh reﬁnement factor, assuming that the level1 (root) mesh uses a time step t, the levelL meshes use a time step t/Ns(L−1) . For levelL meshes to advance one time step of theirs, level(L + 1) meshes need to advance Ns of their own time steps to be synchronized with the former. It should be noted that all meshes of the same level need to be updated for a certain time step before advancing to the next time step, because their values are invoked in the update equations of neighboring meshes. The following operations are recursively performed: 1. For each mesh of level L, the magnetic ﬁeld components are updated. 2. For each mesh of level L, the electric ﬁeld components are updated. 3. If L + 1 ≤ NL , for each mesh of level L, for a time step counter k varying from 0 to Ns − 1, the previous two sets of updates are performed. 4. For all the meshes except the root mesh, one can deﬁne a “local” time step variable S, which enumerates the time steps executed within one time step of its parent mesh. Hence, this counter will take on values from 0 to Ns − 1.
THE DYNAMICALLY ADAPTIVE MESH REFINEMENT (AMR)FDTD TECHNIQUE
73
(a) If S = Ns /2 , the magnetic ﬁeld components of each mesh of level L, are transferred to its parent mesh. Since the node points of parent and child meshes are not necessarily collocated (they are evidently offset if Ns = 2), the former are interpolated from the latter. (b) If S = Ns − 1, the electric ﬁeld components of each mesh of level L are transferred to its parent mesh. The aforementioned steps can be implemented in a function upd ate Le vel(L, S), which is recursively called to update all meshes with level greater or equal to L and a local time step S. The main program just needs to call upd ate Le vel(1, 0) for the update of the entire mesh tree for one time step.
4.3.3
Field Updates on CPBType Boundaries
The electric ﬁeld components of a child mesh tangential to its CPBtype boundary (steps 4a, 4b) are obtained from its parent through trilinear interpolation in space and time, as shown in Fig. 4.4. Since such a boundary is characterized by ﬁxing one spatial variable and letting the other two vary, along with time, trilinear interpolation provides the expression employed in all these updates. To interpolate a function f (η, ξ, τ ) in the range 0 ≤ η, ξ < 1, 0 ≤ τ < 1, by using values of the function at points (0, 0, 0), (0, 0, 1), (0, 1, 0), (0, 1, 1), (1, 0, 0), (1, 0, 1),
y
(j+1) Δ y
jΔ y x iΔx (i + 1) Δ x
FIGURE 4.4: Sampling points for electric ﬁeld on CPBtype boundary. The solid arrows are for the child mesh. The hollow arrows are for the parent mesh
3. In particular. 0. k t ≤ t < (k + 1) t.3) As an example. 0) + (1 − η) (1 − ξ ) τ f (0. j y. . y. at each NAMR time steps).3) allows for the determination of any sampling point of the child mesh from sampling points of its root mesh. ξ .5..e. 1. 1).g. For this purpose. the electric ﬁeld is sampled at cell edge centers and the magnetic ﬁeld is sampled at cell face centers. reveals that only the update of the tangential electric ﬁeld component needs special handling. 4. j. 1. e. k being integers. ξ. 1.4 Field Updates on SBType Boundaries First. k t). 1. τ ) = (1 − η) (1 − ξ ) (1 − τ ) f (0. 1. which shows two meshes interfaced at z = 0.. 0) +ξ (1 − η) (1 − τ ) f (0. 0) +η (1 − ξ ) (1 − τ ) f (1. and (1. Therefore. the use of (4. 1) +η (1 − ξ ) τ f (1. 1) . Fig. 0) +ηξ (1 − τ ) f (1. 0. 1) +ξ (1 − η) τ f (0. consider the Ex ﬁeld on a CPBtype boundary located at z = 0.5. and τ correspond to the following normalized spatial and temporal variables: x y t − i − 1/2. 0. Since. Hz2 . Since the child mesh regions Bn belong to their root mesh region A.5). 4. the index of Hy1 in mesh 2 should be transparent to mesh 1. by Yee cell convention. after each mesh reﬁnement (i. with i. enabling the transfer of data which is included in step 4b. the positions of the SBtype boundaries (between child meshes of the same level) are recorded in a table. − j. 1) +ξ ητ f (1. t). (4. Hy2 . − k. Referring to Fig. whereas the child mesh needs Ex (x. 1. 0. However. the following formula can be applied: f (η. If (i − 1/2) x ≤ x < (i + 1/2) x.74 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS (1. η. 4. Then. 0. 1). 0). 4. the update of Ex in mesh 1 is based on the retrieval of the values of Hy1 . note that boundaries between child meshes of different levels of resolution can be treated as the CPB boundaries that were discussed above. 0. two meshes of the same level share tangential electric ﬁeld and normal magnetic ﬁeld components at the interface between them (see. The parent mesh has Ex ((i + 1/2) x. Hz1 . the treatment of SBtype boundaries can be limited to boundaries between child meshes of the same level. j y ≤ y < ( j + 1) y. x y t which vary within the interval [0. an inspection of the required grid points that the Yee’s algorithm invokes in the update of the ﬁeld components indicated in Fig.
A. and E are meshes of the same level c A d D .THE DYNAMICALLY ADAPTIVE MESH REFINEMENT (AMR)FDTD TECHNIQUE 75 z mesh1 Ex mesh2 Hz1 Hy2 y Hz2 Hy1 FIGURE 4.5: Interface between two meshes of the same level 4. D.5 Field Updates at Junctions of SBType Boundaries One speciﬁc difﬁculty arising in the update of an SBtype boundary comes from Tjunction regions where three sibling meshes share one common edge.6: Typical SBtype boundaries. C.3. and crossjunction regions where four sibling meshes share one common edge. Ideally. B. an SBtype boundary should be transparent and meshes linked by SBtype boundaries (connecting meshes of the same resolution) should C B b a e E FIGURE 4.
b should be updated in B then copied to A. These conditions are enforced in both root and child meshes. Mur’s ﬁrstorder boundary condition [54] has been used. any other type of boundary condition can be readily incorporated. None of the neighboring meshes can obtain two external Hnodes. Then. some of which may not belong to the same mesh. 4. A straightforward solution to the problem of handling these junctions. (b) There is no pair of meshes forming a crossjunction around the Enode. c . However. For example. 2. the situation can be further divided in two subcases: (a) There is at least one pair of meshes forming a crossjunction around the Enode. in Fig. therefore this Enode has to be treated as a CPBtype boundary and updated accordingly. If an Enode on an SBtype boundary is not at a corner of at least one neighboring mesh. is to treat them as CPBtype boundaries. c is copied to all the unupdated neighboring meshes. 4. This approach is employed in the microwave circuit examples of the Chapter 5. Since PMLs are terminated into PECs. d . c can be updated in A by using two external Hnodes from C. Figure 4.6 shows some typical SBtype boundaries. It should be noted that an Enode on a boundary may be shared by 2. the update of an electric ﬁeld node (Enode) requires four neighboring magnetic ﬁeld nodes (Hnodes). that in dielectrically homogeneous domains operates acceptably well. including junctions a. Enode e . the following approach is proposed. the Enode a should be updated in A then copied to B. and e . then copied to all other meshes. the following three situations are considered: 1.76 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS behave like one larger mesh. for example. b. As shown in the ﬁgures.6. To avoid double updates. or 4 neighboring meshes of the same level. it should be updated in that mesh by using one external Hnode from the other mesh. and d should be updated in E then copied to A and D. since meshes A and C form a crossjunction at the Enode c .3. would simply extend the computational domain by the number of the absorber cells. Let us call those external Hnodes. 3. for the analysis of structures such as the inhomogeneous waveguides of Chapter 6. For example. However. . the only type of PB boundaries occurring in a PMLterminated domain are those of PECs.6 Field Updates on PBType Boundaries PBtype boundaries include absorbing and/or PEC boundary conditions. Perfectly matched layer (PML)type conditions. If an Enode on an SBtype boundary is at a corner of all the neighboring meshes. It only needs to be updated in one mesh and then copied to all the other meshes. or vice versa. For the applications that follow.
These are maintained and kept available. a new mesh tree is created. 4.3. Standard interpolation operations. representing the adaptive mesh regeneration. B and B overlap in region C. Again. the overlapping part of the source should also be enforced in the child mesh.7 Transition from Old to New Mesh Tree Each NAMR time steps of the root mesh. where A is the root mesh.3.THE DYNAMICALLY ADAPTIVE MESH REFINEMENT (AMR)FDTD TECHNIQUE 77 A B C B FIGURE 4. the features of AMRFDTD relevant to the enforcement of a nonuniform grid and the implementation of multiple subgrids within a root mesh have been explained. If a child mesh overlaps with the source region. B is an old child mesh. are applied. B is a new child mesh.4 ADAPTIVE MESH REFINEMENT Up to this point. What distinguishes AMRFDTD from previous subgridding approaches is the adaptive mesh reﬁnement. On the other hand. For example. Fields of B within C will be copied to B directly and the rest of B is initialized by interpolating ﬁelds of A. which are the common characteristic of any subgridding algorithm have been proposed. Note that if B contains boundaries.7: Transition from old mesh tree to new mesh tree 4. which enables the adaptive movement of the subgrids. 4. trilinear interpolating operations are employed to initialize the new mesh. It is ﬁnally noted that source conditions are always enforced in the root mesh. since the new tree has the same root mesh as the old tree.7. the possibility of an overlap between any child mesh of the new tree and any child mesh of the old tree is checked.2. If there is such an overlap. Field samples stored in the old mesh tree are transferred to the new mesh tree. This . consider the situation shown in Fig. the ﬁeld samples in the overlapping region are transferred from the old to the new child mesh. This is straightforward for the root mesh. such as Mur’s ﬁrstorder absorbing boundary condition. where conditions. then ﬁeld values of the current and the last time step are needed. according to the proposed algorithm of Section 4.
the gradient of the energy is numerically approximated by a secondorder ﬁnitedifference expression. j. j.k − Wi.k 2 2 dv (4. k) cell of the root mesh is approximated as follows (assume that the computation is made at time step n): n Wi.k+1 − Wi. the dilation of the detected region. Wi. j. j.k Therefore.k i. n t) Vi. j. the adaptive mesh reﬁnement is executed in a cell when both an instantaneous and a calculated over the whole simulated time threshold are exceeded.4.k .k .k = 1 2 E (r .78 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS section explains the core of AMRFDTD. j. j.k H i.k is the volume of cell (i.k > θg G n n Wi.k−1 2 z ˆ x+ ˆ z. j +1. to account for wave propagation during the AMR interval and the clustering of the detected cells.k E i. the following calculations are carried out. which is the detection of cells that need further reﬁnement. j. The ﬁrst criterion takes into account the appearance of energy gradient peaks.4) 1 ≈ 2 + 2 n μi. k). Then.1 Detection of Cells That Need Reﬁnement Every NAMR steps.6) where n G n = max ∇Wi. to form child meshes. 0≤m≤n m Wav = (4. n (4.k . the energy of each (i.k = n n Wi+1. j.k are vector electric and magnetic ﬁeld values at the center of the cell at time step n.8) m Wi. j.7) (4. j. j. n where Vi. while the second ensures that arbitrary ﬁeld . j.k 2 n i. H i.5) Deﬁning thresholds θg and θe . as: n ∇Wi.k + 2 x n n Wi.k − Wi−1. n t) + μ H (r . j. which can be approximated by space/time averaging. 1 Nx Ny Nz (4. j.k is the electromagnetic energy in this cell at time n n step n. j. First. k) is marked for reﬁnement if both of the following criteria are met: n ∇Wi. j.k Vi. j.9) i. j.k 2 y ˆ y (4.k m Q n = max Wav . j. j. j. and Ei. n n Wi. j.k > θe Q . j −1. j. j. a cell (i. 4.
the cells in the source region are also marked for reﬁnement. source region cells are reﬁned up to time t = 2t0 . the marked cells are clustered into rectangular regions. the mesh regeneration (and. except for the ones excited by the source. wave . the cells whose energy is greater than θe Q n and gradient L of energy is greater than θg G n are marked. since the energy gradient is too small to surpass the threshold set. However. we return to the L previous step. 3. For the remainder of the mesh (which does not overlap with any other previous child mesh) the electric and magnetic ﬁeld components are calculated by interpolating their values from the corresponding parent mesh. 2. all the electric ﬁeld components assume zero values. 4.4. For L varying from 2 to the number of levels of the new mesh tree. subsequently. a level(L + 1) mesh is created and added to the new mesh tree. in addition to the cells detected by thresholding. (b) For each mesh of level L. The detection of cells to be reﬁned at this timemarching stage is difﬁcult. On the other hand. for a certain period of time. Within these time steps. For the multilevel case. the update of the mesh tree) is pursued as follows: 1. For example. The previous mesh tree is then removed from memory. mainly stemming from numerical errors at the ﬁnal steps of the simulation.2 Modeling of Wave Propagation The application of the previous criteria may result in a number of cells being marked for reﬁnement. the electric and magnetic ﬁeld components are obtained from the overlapping region. If no cells are marked. and for each rectangle. this reﬁnement process takes place every NAMR time steps (to avoid loading every time step with the mesh reﬁnement operations). if a Gaussian excitation of the form: exp − t − t0 Ts 2 is used (with t0 ∼ 3Ts ). the same algorithm is recursively applied at each mesh level. will not create an unnecessary reﬁnement process.THE DYNAMICALLY ADAPTIVE MESH REFINEMENT (AMR)FDTD TECHNIQUE 79 ﬂuctuations. as the AMRFDTD simulation begins. each levelL mesh is initialized as follows: if it overlaps with any levelL mesh of the old mesh tree. Otherwise. Assuming again that NL is the maximum number of levels. The only mesh that is not subject to regeneration every NAMR time steps is the root mesh. To overcome this difﬁculty. (a) The maximum average energy Q n and maximum gradient of energy G n for level L L L is calculated. For mesh level L varying from 1 to NL .
called the spreading factor. r 2 ) = x1 − x2  + y1 − y2  + z1 − z2  instead of the Euclidean distance d (r 1 . Box A is the root mesh.10) from a marked cell are also reﬁned. There are three predeﬁned controlling parameters: the threshold for coverage efﬁciency θc . which are then divided in Yee cells. the box coverage efﬁciency is introduced. this distance D is the maximum distance that a wave can travel within time equal to NAMR t.4. θe . multiplied by the spreading factor. the coverage efﬁciency of each box is calculated and compared with θc . and NAMR . Note that no assumption is being made as to the direction of the wave velocity. θc . to the volume of the box. otherwise the box is split in two boxes across a cut plane. The factor σAMR is a predeﬁned. the accuracy and efﬁciency of AMRFDTD depends on ﬁve parameters. The black dots represent the marked cells. This procedure is called clustering. Physically. r 2 ) = (x1 − x2 )2 + (y1 − y2 )2 + (z1 − z2 )2 . deﬁned as the ratio of the total volume of the marked cells. the algorithm stops. θg . where c is the speed of light. Again. The effect of the spreading factor. cells within a distance D = σAMR NAMR c t (4. which in general is unknown. σAMR . For the implementation of this clustering procedure. This iterative process continues until the maximum number of boxes is reached.8 illustrates the splitting and shrinking of box B. At the beginning. If the coverage efﬁciency is greater than θc or the dimension is less than Dmin . the bounding box enclosing all the marked cells is found and its coverage efﬁciency is calculated. covered by a box. 4. either box with coverage efﬁciency less than θc and dimension greater than Dmin will be split in two. To facilitate computations. 4. as well as the thresholds deﬁned in the previous section will become evident in the numerical results of the next two chapters. To evaluate the quality of clustering. Box B is split across the cut plane c c and the resultant two boxes are shrunk to obtain boxes C1 and C2 . 4. Although a more systematic discussion of the dependence of .3 Clustering Since the cells that are marked for reﬁnement would generally deﬁne irregularly shaped regions.4 Guidelines for Choosing AMRFDTD Parameters As discussed before.80 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS propagation within the computational domain can clearly generate the need for a denser mesh in cells neighboring to the already reﬁned. the minimum dimension of the box Dmin . Then. the methodology proposed in [55] is followed.4. the celltocell distances can be deﬁned in the sense of the Manhattan distance d (r 1 . they are ﬁrst covered by a number of boxes. and the maximum number of boxes Nmax . To capture this ﬁeld movement. The determination of the position of the cut plane is detailed in [55]. Each of the new boxes is shrunk to just cover the marked cells. Figure. greaterthanone positive real number.
qualitatively dictating their choice. the comments on the accuracy of the technique are meant to be always referred to such an FDTD scheme.8.5. while the coverage efﬁciency threshold θc is chosen between 0. the recommended values of NAMR are between 10 and 50. First.1 − 0. θe = 0. A subtle sideeffect of a large NAMR is the following: newly generated meshreﬁned regions are always extended by a factor proportional to NAMR (see (4. The most important source of errors in any static or dynamic mesh reﬁnement scheme is the reﬂections at a CPBtype boundary.8: Splitting and shrinking of boxes (child meshes) numerical errors of the algorithm on the values of these parameters is included in Section 6. The θe parameter directly affects the wave amplitude at such boundaries and as such it is the most important controlling parameter. Furthermore. Hence.02.6. Note that a large value of the latter enforces the generation of more smaller meshes.THE DYNAMICALLY ADAPTIVE MESH REFINEMENT (AMR)FDTD TECHNIQUE c split and shrink B A c A C1 81 C2 FIGURE 4. decreasing the AMRcoverage will reduce the execution time of the code but will also reduce its accuracy. improving the AMRcoverage. some empirical guidelines.10)).6 and 0. Increasing this parameter leads to less AMRrelated operations. In general. At the same time. Finally. θg ensures that the reﬁned regions enclose large energy gradient variations. it generates more CPB. to account for wave propagation between successive mesh reﬁnements. . the coverage increases and thus the execution time. Therefore. These guidelines and the inherent tradeoffs in the choice of the parameters are further illustrated in the numerical results of the next two chapters.01 − 0.and SBtype boundaries. NAMR determines how frequently the AMR operations are performed. can be provided here. On the other hand. thus increasing the operations related to their management. In general. it should be noted that the best achievable accuracy by AMRFDTD is the one of the FDTD scheme applied at a mesh as dense as the ﬁnest child mesh in the AMRFDTD hierarchy of meshes. a value of σAMR = 2 seems to yield satisfactory results in all cases. θg = 0. In general. since the mesh becomes coarser overall. Let us also deﬁne the AMRcoverage as the ratio of meshreﬁned regions to the total volume of the computational domain.
although it does not implicate wavelet basis functions. the realistic.and threedimensional applications presented in the next two chapters are aimed at highlighting the former two qualities of the AMRFDTD. the AMRFDTD technique belongs to the class of multiresolution timedomain (MRTD) methods. On the other hand. regenerated every certain time steps. both AMRFDTD features can be incorporated in the MRTD technique. despite its more recent introduction the dynamically AMRFDTD technique is closer to being considered as a mature technique compared to MRTD. AMRFDTD has the advantages of ﬂexibility and versatility and the disadvantages of only secondorder accuracy and numerical dispersion that characterize FDTD. a uniform time step is set for the update of all ﬁeld scaling and wavelet coefﬁcients. On the other hand. In fact.82 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS 4. AMRFDTD employs asynchronous updates for the different mesh resolution levels. In the MRTD implementations that are included in this monograph.5 AMRFDTD AND MRTD: SIMILARITIES AND DIFFERENCES Essentially. An important step toward this end is the work reported in [56]. discussed in this monograph. the AMRFDTD clusters the Yee cells that need mesh reﬁnement and encloses them in rectangular subgrids. that may be spatially and irregularly distributed. However. In general. lack two important features of the AMRFDTD that render the latter more ﬂexible as a timedomain modeling tool: 1. regardless of their resolution. However. While the adaptive MRTD operates on nonthresholded ﬁeld wavelet coefﬁcients. as opposed to the Haar basis) highorder of accuracy. Hence. This leads to a much more systematic mesh reﬁnement process. . 2. as well as in the vast majority of the MRTD literature. the MRTD implementations. it does implement a multiresolution. However. two. space and time adaptive moving mesh in three dimensions. which is still related to important and challenging questions when it comes to practical applications. paying the price of complicated enforcement of source and boundary conditions. MRTD is a class of numerical techniques that can achieve (with the use of higherorder basis functions.
a branch coupler. Tradeoffs between speed and accuracy involved with the application of the algorithm to problems of interest are also discussed. 5. All simulations were executed on an Intel Xeon 3. a branch coupler. and a spiral inductor. a maximum of two mesh levels is implemented (or a maximum mesh reﬁnement factor of 8 for a Yee cell of the root mesh). Both questions are negotiated through the application of the technique to three microwave circuit geometries.1. B2 = 21 mm. W = 3 mm. In all these experiments. namely a microstrip ﬁlter. In all cases. the salient properties of the technique along with its excellent potential to dramatically reduce FDTD simulation times are shown.1 INTRODUCTION In this chapter.83 CHAPTER 5 Dynamically Adaptive Mesh Reﬁnement in FDTD: Microwave Circuit Applications The dynamically adaptive mesh reﬁnement FDTD technique of the previous chapter is now applied to realistic microwave circuit geometries. and a spiral inductor. what is the effect of the various parameters. in order to facilitate the presentation of the effect of its parameters on accuracy and execution time. namely. The substrate thickness is . aimed at addressing the following two questions that are naturally posed about the dynamically adaptive mesh reﬁnement FDTD algorithm.2 MICROSTRIP LOWPASS FILTER The ﬁrst example is a microstrip lowpass ﬁlter. a twolevel AMRFDTD is implemented. 5. Through these applications. whether the overhead that AMRFDTD accumulates from the application of the mesh reﬁnement can still leave some room for computational savings stemming from a reduced overall number of operations. 5. a number of microwave circuit analysis examples are presented. shown in Fig. The dimensions indicated in the ﬁgure are A = 40 mm.06 GHz machine. deﬁned in the previous chapter on the performance and accuracy of the algorithm and second. First. B1 = 2 mm. a microstrip lowpass ﬁlter.
σAMR = 2.4 mm at different time steps. Two conventional FDTD simulations were performed for comparison: one using a relatively coarse mesh (40 × 40 × 10 cells). respectively 0. 5. The dimensions of the air box are 40 mm × 40 mm × 4 mm.2. θe = 0. the ﬁeld values in the working volume of the domain decrease. To demonstrate the evolution of the child meshes over time. the coverage of reﬁnement increases. Figs. the coverage is relatively large (40 to 60%) due to multiple reﬂections between the two open ends of the microstrip line. P1 and P2 denote ports 1 and 2. until about t = 250 t. after time t = 1000 t. the dense FDTD technique uses 8192 steps. which causes the AMRcoverage to decrease.7. 5. and the second using a dense mesh (80 × 80 × 20 cells). A voltage source excitation is imposed at 3 mm from the edges. 5.3. While AMRFDTD and the coarse FDTD technique are run for 4096 time steps.1. AMRFDTD simulations use a 40 × 40 × 10 cell root mesh and different controlling parameters for AMR. a Courant number of 0. Consequently. . Between 250 t and 700 t.2. In all simulations. In fact. The initial child mesh at t = 0 is given in Fig. since the time step t of the latter is equal to one half the t of the former. In addition.3–5.7 is used for determining the time step.7 show the verticaltoground electric ﬁeld component Ez and child meshes on the plane z = 0. referred to as AMRcoverage of the domain and the number of child meshes as a function of time steps are shown in Fig. NAMR = 10. the spatial ﬁeld variation becomes smoother. As the ﬁelds impinge upon the absorbing boundaries of the structure. As the wave propagates along the feed line. Both AMRFDTD and conventional FDTD use Mur’s ﬁrstorder absorbing boundary condition [54]. The AMR parameters are: θg = 0.1: Microstrip lowpass ﬁlter geometry. θc = 0.8 mm and its dielectric constant is 2.84 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS A W W A P1 W B1 B2 P2 A/2 FIGURE 5. the ratio of the volume of meshreﬁned areas (occupied by child meshes) to the total volume.02 .
Figure 5.DYNAMICALLY ADAPTIVE MESH REFINEMENT IN FDTD 85 FIGURE 5. so that no child meshes are necessary.8 compares the vertical electrical ﬁeld at z = 0. The timeseries deduced via AMRFDTD cannot be distinguished from the one of the reference FDTD simulation. One child mesh. up to 40. AMR coverage is 4.3% .000 time steps (32ns ).4 mm and the center of the microstrip line.3: Vertical electric ﬁeld magnitude at z = 0.4 mm and t = 0.2: AMR coverage and number of child meshes at different time steps for the lowpass ﬁlter simulation the coarse root mesh captures the ﬁeld solution sufﬁciently well. FIGURE5. 3 mm from the right edge.
5. AMR coverage is 52. The calculated scattering (S) parameters and their differences are shown in Fig.4: Vertical electric ﬁeld magnitude at z = 0. AMR coverage is 13.4 mm and t = 100 t. One child mesh. The plot indicates the excellent approximation provided by AMRFDTD to the result of FDTD using a dense mesh. The absence of any latetime instability effects is also noted.86 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS FIGURE 5.9.4% whereas the result of FDTD using a coarser mesh has a signiﬁcant difference. This accuracy is quantiﬁed in Table ??. which employs the following FIGURE 5. One child mesh.4 mm and t = 200 t.5: Vertical electric ﬁeld magnitude at z = 0.5% .
1) FIGURE 5.0% . (5.2% Sparameter error metric: ref Sl.m ( f k ) − Sl.m ( f k ) k l m 2 . Three child meshes. AMR coverage is 44.4 mm and t = 800 t.6: Vertical electric ﬁeld magnitude at z = 0. Three child meshes.DYNAMICALLY ADAPTIVE MESH REFINEMENT IN FDTD 87 FIGURE 5.4 mm and t = 500 t. AMR coverage is 28.7: Vertical electric ﬁeld magnitude at z = 0.m ( f k ) 2 ES = k l m ref Sl.
which is used as a reference. f k is a discrete frequency within the modeled frequency band (up to 30 GHz). determined by the densely meshed FDTD.8 0. From Table ??.2 0 −0.m the (l.5 Time (ns) FIGURE 5. it is concluded that AMRFDTD can closely follow the accuracy of the reference FDTD method. while consuming only 5.5 1 1.8: Vertical electric ﬁeld magnitude at z = 0. Sl. m)element of the scattering matrix of the simulated circuit. as determined by the AMRref FDTD or the coarsely meshed FDTD technique and Sl.4 mm and the center of the microstrip line.88 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS 1 0.6 AMRFDTD FDTD 40x40x10 FDTD 80x80x20 Ez (V/m) 0.2 0 5 10 15 20 25 30 35 Time (ns) 1 AMRFDTD FDTD 40x40x10 FDTD 80x80x20 Ez (V/m) 0. These execution time savings are well above the savings .m is the same element.4% of its total simulation time. 3 mm from the right edge where.4 0.5 0 0.
DYNAMICALLY ADAPTIVE MESH REFINEMENT IN FDTD 0 89 Amplitude of S11 (dB) −5 −10 AMRFDTD FDTD 40x40x10 FDTD 80x80x20 −15 −20 0 5 10 15 Frequency (GHz) 20 25 30 AMRFDTD FDTD 40x40x10 Absolute value of error of S11 0.2 0.15 0.4 0.3 0. 5.05 0 0 5 10 15 Frequency (GHz) 20 25 30 AMRFDTD FDTD 40x40x10 5 10 15 Frequency (GHz) 20 25 30 AMRFDTD FDTD 40x40x10 FDTD 80x80x20 5 10 15 Frequency (GHz) 20 25 30 FIGURE 5.1.25 0.3 0.2 0.5 0. obtained by FDTD and AMRFDTD .1 0 0 0 −5 Amplitude of S21 (dB) −10 −15 −20 −25 −30 −35 −40 0 0.9: Comparison of the scatteringparameters of the lowpassﬁlter of Fig.35 Absolute value of error of S21 0.1 0.4 0.
the decrease in θg and θe results in lower errors and longer simulation times. As expected. ??–??. σAMR = 2 . once the mesh size exceeds a certain limit. previously reported in the literature and indicate the potential of AMRFDTD. However.1: Computation Time and SParameter Error Metric E S for the Microstrip LowPass Filter METHOD FDTD AMRFDTD FDTD MESH 40 × 40 × 10 40 × 40 × 10 80 × 80 × 20 TIME STEPS 4096 4096 8192 TOTAL TIME (S) 75 518 9669 Es (%) 37. The rest of the AMR parameters are: θe = 0.3 2. A careful study can show that for small meshes the FDTD simulation times in dense and coarse meshes would tend to follow the 16:1 rule. θc = 0. The effect of the AMRFDTD controlling parameters on the simulation time and error of this technique is studied next and results are shown in Figs. which is the ratio between the operations carried out in the two cases.90 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS TABLE 5.1. as these two thresholds are lowered. This is due to the increase in the memory access time for large meshes. For a thorough investigation of memorycache related effects in the execution time of timediscrete methods the reader is referred to [57].4 — expected from static subgridding algorithms.7. the AMRFDTD tends to become equivalent in operation and FIGURE 5. Essentially.10: Effect of threshold for gradient θg . the simulation time increases faster than 16:1. NAMR = 10. It should be noted that the ratio between the FDTD simulation times for the ﬁne and the coarse mesh is larger than 16:1. which further extends their simulation time.
θe = 0. The rest of the AMR parameters are: θg = 0. θc = 0.7.01. σAMR = 2 FIGURE 5.01. NAMR = 10.7.01. σAMR = 2 FIGURE 5.13: Effect of interval of AMR NAMR .DYNAMICALLY ADAPTIVE MESH REFINEMENT IN FDTD 91 FIGURE 5.11: Effect of threshold for energy θe . θc = 0.01. The rest of the AMR parameters are: θg = 0.01. NAMR = 10. σAMR = 2 . The rest of the AMR parameters are: θg = 0.12: Effect of threshold for coverage efﬁciency θc . θe = 0.
7. The choice of the AMR parameters is further discussed in this and the next chapter. B1 = 7 mm. when θc reaches 0.8.14: Effect of spreading factor of AMR σAMR . a large number of child meshes can be generated as a result. This problem can be alleviated by using higherorder interpolation schemes at the mesh boundaries to reduce the reﬂections [44–58]. Tradeoff effects related to the choice of NAMR and σAMR are also evident in Figs. θe = 0. ?? is analyzed next. 5. have been used in all the aforementioned numerical experiments. . indicated by these plots. the simulation time increases. Also. The geometric parameters indicated in the ﬁgure are as follows: A = 40 mm. ?? and ??. without sacriﬁcing accuracy. An appropriate choice of θe can essentially eliminate this problem.3 MICROSTRIP BRANCH COUPLER The microstrip branch coupler geometry of Fig. without any improvement in accuracy.01. the ﬁeld components assume some small values due to the reﬂections at the CPB. NAMR = 10 performance to the dense FDTD method. If θe is very small.92 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS FIGURE 5. It should be noted that the percentage error does not decrease monotonically as θg and θe decrease. As a result. both the error and the computation time increase. On the contrary.01 without any associated improvement in accuracy.and SBtype boundaries. w2 = 3 mm.01. w1 = 2 mm. up to some point. because certain child meshes generated by using smaller θg or θe may have larger reﬂections due to the irregularity of the shape of the reﬁned regions. The reason is that the numerical error triggers the automatic generation of multiple spurious child meshes that are clustered independently. The appropriate values for NAMR and σAMR . their management by the algorithm adds an overhead that completely eliminates any savings due to the mesh reﬁnement. The rest of the AMR parameters are: θg = 0. It is also noted that there is a sudden increase in the simulation time when θe is below 0. Subsequently. increasing these thresholds can reduce the overall computation time. At the late stage of the simulation. θc = 0.
Comparisons regarding computation times and numerical errors of the AMRFDTD are included in Table ??.8 mm above the substrate. B3 = 9 mm. The AMR controlling parameters are θg = 0. The parameters of this geometry are: A1 = 60 mm. w1 = w2 = 2 mm. P3.8 mm and its dielectric constant is 2. B1 = 24 mm. P1. NAMR = 10.1.DYNAMICALLY ADAPTIVE MESH REFINEMENT IN FDTD 93 A w2 P4 w1 B1 P1 B3 FIGURE 5. The data demonstrate the capability of AMRFDTD to deduce the dense FDTD results at a greatly reduced computational cost. θe = 0. ??. σAMR = 2. The AMRFDTD method uses a 40 × 40 × 10 mesh and takes 4096 time steps. and 4 P2 A w1 B2 P3 B2 = 11 mm. Figures ??– ?? show the Sparameters and their errors. B2 = 20 mm.4 MICROSTRIP SPIRAL INDUCTOR As a last example. θc = 0. A reference FDTD simulation of an 80 × 80 × 20mesh is used for comparison. as determined by the AMRFDTD. The AMRFDTD method uses a 60 × 40 × 10 mesh and 8192 time steps. the AMRFDTD and densemesh FDTD waveforms coincide. Again. The dimensions of the computational domain enclosing the structure are 60 mm × 40 mm × 4 mm. P2. without any sign of latetime instability. 2. The dimensions of the computational domain enclosing the structure are 40 mm × 40 mm × 4 mm. which is reﬂected on an execution time reduction by a factor of 20.2.8 mm and its dielectric constant is 2. The air bridge is 0. The thickness of the substrate is 0.15: Microstrip branch coupler geometry. Timedomain ﬁeld waveforms up to 40. and the reference FDTD simulation. and P4 denote port 1. the geometry of a spiral inductor of Fig.7. ?? is analyzed. plain FDTD executed on the AMR root mesh. A reference FDTD simulation of a . 5. B4 = 4 mm. 3. A2 = 40 mm.2.000 time steps are shown in Fig.01. B3 = 18 mm. The thickness of the substrate is 0.
??.2 0.001. Moreover.05 0 AMRFDTD FDTD 40x40x10 0 5 10 15 Frequency (GHz) 20 25 30 FIGURE 5.35 Absolute value of error of S21 0.4 0. ??.15 0. as determined by the three methods under comparison.25 0.1.6. ?? shows the Sparameters and their errors. σAMR = 2.94 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS 0 −5 Amplitude of S11 (dB) −10 −15 −20 −25 −30 0 AMRFDTD FDTD 40x40x10 FDTD 80x80x20 5 10 15 Frequency (GHz) 20 25 30 0. ??. as determined by FDTD and AMRFDTD 120 × 80 × 20 mesh is used for comparison. Fig. NAMR = 50. The mesh reﬁnement process is illustrated in Fig. . The AMR controlling parameters are θg = 0.16: S11 for the microstrip branch coupler geometry of Fig.1 0. θe = 0. θc = 0.3 0. while error and execution time data are shown in Table ??. which shows the effective vertical electric ﬁeld wavefront tracking achieved by the algorithm. Timedomain results are also shown in Fig.
??. .25 0. which necessitates the use of a signiﬁcant number of time steps for the extraction of the Sparameters.3 0.15 0. Note that in this case.35 Absolute value of error of S21 0.05 0 AMRFDTD FDTD 40x40x10 95 AMRFDTD FDTD 40x40x10 FDTD 80x80x20 5 10 15 Frequency (GHz) 20 25 30 0 5 10 15 Frequency (GHz) 20 25 30 FIGURE 5. while the associated errors are larger. At late stages of the simulation. relatively small ﬁeld values are easily contaminated by spurious reﬂections at the parent–child mesh interfaces. as determined by FDTD and AMRFDTD Compared to the lowpass ﬁlter and the branch coupler. These effects stem from the highly resonant nature of the spiral inductor. the execution time savings achieved by AMRFDTD over the conventional FDTD are smaller (yet large.4 0.2 0.DYNAMICALLY ADAPTIVE MESH REFINEMENT IN FDTD 0 −5 Amplitude of S21 (dB) −10 −15 −20 −25 −30 −35 −40 0 0. the error of the coarse mesh FDTD becomes excessively large as well.1 0. of about 80%).17: S21 for the microstrip branch coupler geometry of Fig.
??. two further comments are in order. implies that only the root mesh is still present .1 0.25 Absolute value of error of S31 0. First. In fact.05 0 0 5 10 15 Frequency (GHz) 20 25 30 FIGURE 5. the timedomain results accompanying the three numerical experiments (Figs. the convergence of the number of AMRFDTD child meshes to zero over time.2 0. ??.5 DISCUSSION: STABILITY AND ACCURACY OF AMRFDTD RESULTS Based on the previous examples.18: S31 for the microstrip branch coupler geometry of Fig. as determined by FDTD and AMRFDTD 5. demonstrate the absence of latetime instability in the AMRFDTD.8.96 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS 0 −5 Amplitude of S31 (dB) −10 −15 −20 −25 −30 0 AMRFDTD FDTD 40x40x10 FDTD 80x80x20 5 10 15 Frequency (GHz) 20 25 30 AMRFDTD FDTD 40x40x10 0. ??). 5.15 0.
as determined by FDTD and AMRFDTD TABLE 5.3 0.7 Absolute value of error of S41 0.5 0.6 0.8 0.0 −2 Amplitude of S41 (dB) −4 AMRFDTD FDTD 40x40x10 FDTD 80x80x20 −6 −8 −10 0 0.1 0 5 10 15 Frequency (GHz) 20 25 30 AMRFDTD FDTD 40x40x10 0 5 10 15 Frequency (GHz) 20 25 30 FIGURE 5. ??.2 0.6 2.2: Computation Time and SParameter Error Metric E S for the Microstrip Branch Coupler METHOD FDTD AMRFDTD FDTD MESH 40 × 40 × 10 40 × 40 × 10 80 × 80 × 20 TIME STEPS 4096 4096 8192 TOTAL TIME (S) 262 1021 19765 Es (%) 34.43 — .4 0.19: S41 for the microstrip branch coupler geometry of Fig.
6 AMRFDTD FDTD 40x40x10 FDTD 80x80x20 Ez (V/m) 0.8 0.20: Vertical electric ﬁeld magnitude at z = 0.8 AMRFDTD FDTD 40x40x10 FDTD 80x80x20 0.6 mm and the center of the microstrip line of port 3.6 Ez (V/m) 0. 6 mm from the edge.5 0.2 0.2 0 5 10 15 Time (ns) 20 25 30 35 0.98 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS 1 0.21: Spiral inductor geometry.8 FIGURE 5.2 0 −0. P1 and P2 denote port 1 and port 2.3 0.4 0.4 Time (ns) 0.6 0.4 0.7 0.2 0 −0.2 0.1 0. respectively . The excitation is imposed at 3 mm from the edge of port 1 A1 B1 w2 A 2 P1 w1 B2 microstrip via air bridge B3 B4 P2 FIGURE 5.
Note that in AMRFDTD.74 — .22: Vertical ﬁeld evolution and associated mesh reﬁnement in the microstrip spiral inductor. Another aspect of the AMRFDTD accuracy is associated with the reﬂections at dense/coarse mesh interfaces.1 3. Therefore. simulated by a twolevel dynamic AMRFDTD at a late stage of the code.DYNAMICALLY ADAPTIVE MESH REFINEMENT IN FDTD 99 FIGURE 5. while its retroreﬂections might encounter dense/coarse grid TABLE 5. no spatial or temporal interpolation operations. This is an additional advantage of using a dynamically adaptive instead of a statically adaptive mesh in timedomain simulations.3: Computation Time and SParameter Error Metric E S for the Microstrip Spiral Inductor METHOD FDTD AMRFDTD FDTD MESH 60 × 40 × 10 60 × 40 × 10 120 × 80 × 20 TIME STEPS 8192 8192 16384 TOTAL TIME (S) 411 5276 29154 Es (%) 52. creating numerical errors. such a pulse would always be enclosed in a dense mesh. A pulse propagating at a statically reﬁned mesh will be reﬂected off such an interface. are applied then. which are the primary sources of instabilities in adaptive mesh FDTD codes [44].
3 0.25 0.5 0.8 0.100 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS 0 −5 Amplitude of S11 (dB) −10 −15 −20 −25 −30 0 0.35 0.2 0.5 0.4 0.1 0.2 0.15 0.1 0 AMRFDTD FDTD 40x40x10 AMRFDTD FDTD 60x40x10 FDTD 120x80x20 5 10 15 Frequency (GHz) 20 25 30 0 5 10 15 Frequency (GHz) 20 25 30 0 −5 Amplitude of S21 (dB) −10 −15 −20 −25 −30 −35 −40 0 0.05 0 0 5 10 15 Frequency (GHz) 20 25 30 AMRFDTD FDTD 40x40x10 AMRFDTD FDTD 60x40x10 FDTD 120x80x20 5 10 15 Frequency (GHz) 20 25 30 FIGURE 5.6 0.3 0. as determined by FDTD and AMRFDTD .45 Absolute value of error of S21 0. ??.7 Absolute value of error of S21 0.23: Sparameters for the spiral inductor geometry of Fig.4 0.
5 0.6 0.1 0 −0. the AMRFDTD stability limit is observed to be 0.2 0. the application of interpolations in space and time generates a degradation in the FDTD stability factor.3 0.24: Vertical electric ﬁeld magnitude at z = 0.8 0.8 0.6 mm and the center of the microstrip line.1 −0. For the simple trilinear interpolations employed in these computations.6 1. and 0.7 0.DYNAMICALLY ADAPTIVE MESH REFINEMENT IN FDTD 0.8 2 FIGURE 5.4 1. This wellknown effect. The excitation is imposed at 3 mm from the left edge of the microstrip line interfaces before the AMR algorithm creates new meshes for them.6 0.2 0. .4 0. This latter case produces errors in AMRFDTD.6 0.7 0. can be alleviated by using higherorder interpolation schemes.85 of the corresponding FDTD one in threedimensional cases.3 0.2 0 5 10 15 Time (ns) 20 25 30 35 0.2 Time (ns) 1.1 −0.4 AMRFDTD FDTD 60x40x10 FDTD 120x80x20 Ez (V/m) 0. 6 mm from the right edge.5 0. However.2 0.9 of the corresponding FDTD one in twodimensional cases. also discussed in [44].8 1 1.2 0.4 AMRFDTD FDTD 60x40x10 FDTD 120x80x20 101 Ez (V/m) 0.1 0 −0. which are evidently smaller than those arising in a static subgrid.
102
MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS
Finally, the comparison between coarse/dense FDTD results and AMRFDTD results in the frequency domain, reveals the standard pattern of the results being in a relatively good agreement with each other up to the middle of the simulated frequency band and diverging afterwards. In these simulations, the coarse grid Yee cell size is about λmin /7.5 is each dimension. Therefore, up to the frequency f max /2, the socalled coarse mesh uses a sampling rate of at least λ( f max /2)/15, which is dense enough to determine the Sparameters accurately. From that point on, the effect of the FDTD numerical dispersion becomes more severe, leading to the large errors shown in the ﬁgures.
5.6
CONCLUSION
In this chapter, the dynamically AMRFDTD technique was applied to realistic microwave circuit and optical waveguide geometries. The purpose of the applications that were shown, was to demonstrate whether the mesh adaptation overhead of AMRFDTD can still allow for important computational savings. The conclusion is that it does, because the method is optimally suited to the nature of timedomain simulations. The latter are characterized by spatially and temporally localized phenomena that call for a dense mesh not throughout space and time, but only at the certain time and space they happen. Although microwavecircuit examples were presented, the technique evidently holds a great promise for largescale optical structure modeling. This direction is further investigated in the next chapter.
103
CHAPTER 6
Dynamically Adaptive Mesh Reﬁnement in FDTD: Optical Applications and Error Estimates
After presenting microwave circuit applications, the dynamically adaptive mesh reﬁnement FDTD is employed for the modeling of twodimensional optical waveguide structures, including power splitters, junctions, and ring resonators. The conventional FDTD often requires extremely long simulation times for the characterization of this type of problems; on the contrary, the use of an adaptively moving mesh is shown to result in dramatically accelerated simulations. In all the examples that follow, a multilevel adaptive mesh reﬁnement is pursued. Hence, while in the previous chapter, a maximum to minimum Yee cell edge ratio of two was used, in the following examples this ratio goes up to 64. The use of multiple levels also allows us to compare the efﬁciency of the mesh reﬁnement algorithm as a function of the number of levels used. The chapter is concluded by extracting error estimates that allow for an educated choice of the AMR parameters.
6.1
MULTILEVEL AMRFDTD
To demonstrate the application of the AMRFDTD algorithm with multiple resolution levels, a 2D TEmode (with an Ez electric ﬁeld component only) optical waveguide (shown in Fig. 6.1) is simulated [59]. Its width is 0.3 μm and its dielectric constant is 10.24. The computational domain is 6 μm × 6 μm and a 1 μm thick matched absorber is used to truncate it. Figure 6.2 shows a snapshot of the electric ﬁeld obtained by AMRFDTD with a maximum number of levels equal to 4. The excitation is imposed at 1 μm from the left edge of the waveguide and the electric ﬁeld is recorded at the center of waveguide and 1.5 μm from the right edge. The excitation is a modulated Gaussian pulse of the form: exp − y − y0 W
2
−
t − t0 Ts
2
sin (2π f t) ,
104
MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS
y 1 μm
6 μm
waveguide 0.3 μm
absorber 6 μm
FIGURE 6.1: Optical waveguide geometry
x
where Ts = 0.02 ps, t0 = 3Ts , f = 200 THz, W = 0.7 μm, y0 = 3 μm. Table 6.1 compares the accuracy and computation time of AMRFDTD using a root mesh of 120 × 120 and 2–4 mesh levels, to a reference FDTD simulation using a 1920 × 1920 mesh and several coarser FDTD schemes. The accuracy is quantiﬁed by employing the following timedomain error metric: f (tk ) − f ref (tk ) Et =
k 2
f ref (tk )
k
2
,
(6.1)
where tk is discrete time within the modeled time range (up to 16 ps), f is the sampled electric ﬁeld as determined by AMRFDTD or the coarse mesh FDTD techniques, and f ref is the sampled electric ﬁeld determined by the reference FDTD simulation. All AMRFDTD
FIGURE 6.2: Vertical electric ﬁeld magnitude distribution across the optical waveguide of Fig. 6.1, at t = 100 t. One level2 child mesh and one level3 child mesh are shown
8 1 Time (ns) 1.DYNAMICALLY ADAPTIVE MESH REFINEMENT IN FDTD 105 TABLE 6.60 — 17. 5 4 3 2 AMRFDTD FDTD 120x120 FDTD 1920x1920 Ez (V/m) 1 0 −1 −2 −3 −4 0. θc = 0. 6.9 17.0 3.4 0.8.1.2 shows that AMRFDTD can reach the accuracy of the reference FDTD method in a greatly reduced computation time.4 133 641 Et (%) 74. 6. NAMR = 10.56 simulations use a reﬁnement factor of 2 for successive levels. 6.5 μm from the right edge .3.6 x 10 −4 FIGURE 6. and θe = 0.1: Computation Time and TimeDomain Error Metric Et for the Optical Waveguide of Fig.4 66.7 635 13794 181997 31. 1. Figure 6.2 3.1 METHOD FDTD FDTD FDTD FDTD FDTD AMRFDTD AMRFDTD AMRFDTD MESH 120 × 120 240 × 240 480 × 480 960 × 960 1920 × 1920 120 × 120 120 × 120 120 × 120 NUMBER OF LEVELS — — — — — 2 3 4 TIME STEPS 2000 4000 8000 16000 32000 2000 2000 2000 TOTAL TIME (S) 6.1.4 1.57 0.64 0.2 1.6 0. The AMRFDTD waveform is not visually discernible from the reference FDTD one. θg = 0. σAMR = 2. The time domain simulation results of the AMRFDTD with four levels and the reference FDTD code are compared in Fig.3: Temporal waveform of the vertical electric ﬁeld magnitude at the center of the optical waveguide of Fig.
106 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS 6. The excitation is imposed 1 μm from the left edge of the waveguide and the electric ﬁeld is recorded at the center of the waveguide and 1. testing the stability and accuracy of the mesh reﬁnement scheme.7 μm. along the ydirection).9976. 6. This geometry is a good benchmark application for the AMRFDTD technique.4: Geometry of the dielectric waveguide with a corrugated permittivity proﬁle. The accuracy is quantiﬁed by employing the timedomain waveform error metric of Eq. All AMRFDTD y 1 1 1 1 1 1 8 P1 3 absorber 12 2 P2 x FIGURE 6.1). Dimensions are given in units of μm. The dielectric constant of the corrugations is 7. (6. Its width is 2 μm and the dielectric constant of the host medium is 9.2) where u(t) is the unit step function. The geometry of the waveguide is shown in Fig. W = 0.4. to a reference FDTD simulation using a 3840 × 2560 mesh and several coarser FDTD schemes. 2 2 (6. P1 and P2 denote ports 1 and 2. the pulse is centered in the middle of the guide. respectively . is simulated by FDTD and a multilevel dynamic AMRFDTD [59]. since the multiple reﬂections created by the dielectric corrugations enforce strong interactions between spatially distributed meshes. Table 6. and h = 2y0 is the height of the waveguide (hence. Therefore. All interpolations employed are ﬁrstorder polynomial ones.2 DIELECTRIC WAVEGUIDE WITH A CORRUGATED PERMITTIVITY PROFILE A dielectric waveguide with a corrugated permittivity proﬁle.2 μm from the right edge. The excitation is a modulated pulse of the form: (u(t) − u(t − T)) sin2 (πt / T) sin (2π f t) e −(y−y0 ) /W . if one method has a time step t and another t/2. similar to the one presented in [60].2 compares the accuracy and computation time of AMRFDTD using a root mesh of 240 × 160 and 2–4 mesh levels. the latter will be run for twice the total number of time steps of the former. f = 193 THz. The computational domain is 12 μm × 8 μm and is truncated by a 1 μm thick matched absorber. The simulated timewindow in all methods remains constant. T = 5/ f .
while the execution time also starts increasing then. as well as the time for their execution.9 3544 50230 559714 91. the conclusion of this study is that four levels. since the reduction of interpolation errors can be achieved by applying a higherorder polynomial interpolation [44].01.54 0.2 shows that the dynamic AMRFDTD can reach the accuracy of the reference FDTD method in a greatly reduced computation time.91 1.4 METHOD FDTD FDTD FDTD FDTD FDTD AMRFDTD AMRFDTD AMRFDTD AMRFDTD MESH 240 × 160 480 × 320 960 × 640 1920 × 1280 3840 × 2560 240 × 160 240 × 160 240 × 160 240 × 160 NUMBER OF LEVELS — — — — — 2 3 4 5 TIME STEPS 2000 4000 8000 16000 32000 2000 2000 2000 2000 TOTAL TIME (S) 20.43 0. while Fig.07 183. which in turn increases the operations needed.89 8. Therefore. The result of Fig. Figure 6.DYNAMICALLY ADAPTIVE MESH REFINEMENT IN FDTD 107 TABLE 6.1 41. σAMR = 2. It should be noted that the error no longer decreases once the number of levels of the mesh tree reaches 5. This trend continues as the number of levels increases and reﬂects a saturation point beyond which the beneﬁts of the AMR algorithm are compensated for by the errors accumulated through the interpolating operations. 6. 6. θc = 0.71 — 41. the AMRFDTD uses four levels and a 240 × 160 mesh.5 compares the timedomain waveforms of AMR and conventional FDTD schemes.8. θg = 0.2: Computation Time and TimeDomain Error Metric Et for the Dielectric Waveguide of Fig.6 demonstrates the latetime stability of the multilevel dynamic AMRFDTD. combined with ﬁrstorder interpolations can achieve an impressive reduction in the execution time (by a factor of 446.39 simulations use a reﬁnement factor Ns of 2 for successive levels. and θe = 0.2 8. Then. These two problems are interconnected. Table 6. which is also its fundamental difference from static subgridding schemes. the interpolating operations needed for the update equations would involve large ﬁeld values. Assume a pulse propagating through a static interface between a coarse and a dense grid. In both ﬁgures.6 can be explained by the following note regarding the stability of the dynamic AMRFDTD. NAMR = 10. As a .7).8 246 1253 11337 Et (%) 157. 6. with less than 1% error in the timedomain.
5 1 1. the reason being that this algorithm tracks the propagating wavefronts. such an interface would not exist under the dynamic AMR scheme.4.2 1. contributing to the wellknown latetime instability effects in static subgridding schemes.3 1.1 1.5: Electric ﬁeld at the center of the dielectric waveguide of Fig. indicating the absence of latetime instability .2 μm from the right edge result.15 1. the pulse of this thought example would be enclosed in a moving dense mesh.5 2 Time (ns) 2.95 1 1.45 x 10 −4 FIGURE 6.4 1.6: Electric ﬁeld at the center of the dielectric waveguide of Fig. On the other hand. 3 2 1 Ez (V/m) 0 −1 −2 −3 0 0. 6.108 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS 3 2 1 Ez (V/m) 0 −1 −2 −3 AMRFDTD FDTD 240x160 FDTD 3840x2560 0.000 time steps.25 Time (ns) 1. for 40.5 3 x 10 3.4. 1.05 1. the associated errors are large. Therefore. obtained with the dynamic AMRFDTD. 6.2 μm from the right edge.35 1. 1.5 −3 FIGURE 6.
8. 2. its dielectric constant proﬁle does not conform to a rectangular mesh.11–6. the timedomain results of Figs. The multilevel AMRFDTD similarly rediscretizes the dielectric constant distribution at different levels. In FDTD simulations using different meshes. a dielectric waveguide power splitter.13 show the evolution of the ﬁeld and the mesh tree. y P4 1 0.7: Geometry of the dielectric waveguide power splitter. whereas the dielectric constant of the slanted splitter (lens) is 4. the dielectric constant distribution is discretized independently. shown in Fig.3 DIELECTRIC WAVEGUIDE POWER SPLITTER To demonstrate the efﬁciency of the multilevel AMRFDTD for the simulation of structures with nonconformal dielectric constant distribution. The level1 and level2 meshes are drawn in the ﬁgures.10 demonstrate the latetime stability and accuracy of the method. In addition. 9 μm). which is imposed at (1. 6. The dielectric constant of the crossshaped waveguide is 9. which uses a 60 × 60 root mesh and a maximum number of three levels. From these ﬁgures. according to an optimal staircase approximation. Table 6. P3.2). and the electric ﬁeld is recorded at 1 μm from each port.5 18 P1 waveguide n le s 2 P3 2 absorber P2 18 x FIGURE 6.3 shows that AMRFDTD can reach the accuracy of the reference FDTD method. P2. θg = 0. 3. The simulations use the same excitation as (6. 6. and P4 denote ports 1. in a much smaller execution time. All AMRFDTD simulations use a reﬁnement factor of 2 for successive levels. Dimensions are given in units of μm. and 4. θc = 0. the effective wavefront tracking achieved by AMRFDTD is demonstrated. It should be noted that due to the slanted orientation of the lens.DYNAMICALLY ADAPTIVE MESH REFINEMENT IN FDTD 109 6.8–6. while the regions without grid lines are covered by level3 meshes (not drawn for visualization purposes). σAMR = 2.01. Figures 6. NAMR = 10.2 μm. respectively . P1.7 [60] is simulated [59]. and θe = 0.
6. shown in Fig.6 1.8: Electric ﬁeld at 1 μm from the edge of port 2 of the power splitter of Fig. The dielectric constant of the waveguide is 9 and the dielectric constant of the surrounding medium is 1. 6.4 1.6 816.0 2.8 Time (ns) 2 2. The computational domain is 25 μm × 10 μm and a 1 μm thick matched absorber is used to truncate it.2 μm from the edge of port 1 and the electric ﬁeld is recorded at the center of the waveguide.110 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS TABLE 6.3: Computation Time and TimeDomain Error Metric Et for the Dielectric Waveguide Power Splitter for Fields Recorded at Port 2 METHOD FDTD FDTD FDTD FDTD AMRFDTD AMRFDTD AMRFDTD MESH 360 × 360 720 × 720 1440 × 1440 2880 × 2880 360 × 360 360 × 360 360 × 360 NUMBER OF LEVELS — — — — 2 3 4 TIME STEPS 4000 8000 16000 32000 4000 4000 4000 TOTAL TIME (S) 148.6 3683 Et (%) 168 41. The excitation is imposed at 1.98 1.4 DIELECTRIC WAVEGUIDE YJUNCTION A dielectric waveguide Yjunction.85 — 41. The excitation is a modulated pulse 3 2 1 AMRFDTD FDTD 360x360 FDTD 2880x2880 Ez (V/m) 0 −1 −2 −3 1.7 .07 6.2 x 10 2.92 2313 39181 447497 394.0 2.14 is simulated [59].4 −4 FIGURE 6. 2 μm from the edge of port 2.
7 of the form: e −(y−y0 ) /W 2 2 −(t−t0 )2 /T 2 sin (2π f t) . for 40.4 compares the accuracy and computation time of AMRFDTD using a root mesh of 250 × 100 and 2–4 mesh levels.6 1.DYNAMICALLY ADAPTIVE MESH REFINEMENT IN FDTD 1 111 0.2 x 10 2. W = 1 μm.5 AMRFDTD FDTD 360x360 FDTD 2880x2880 1. and y0 corresponds to the center of the waveguide.4 1.3) where f = 200 THz.5 E z (V/m) 0 −0.8 Time (ns) 2 2. (6.5 −3 FIGURE 6.4 −4 −1 FIGURE 6. t0 = 3T.5 3 x 10 3.10: Electric ﬁeld obtained with the dynamic AMRFDTD at 1 μm from the edge of port 2 of the power splitter of Fig. T = 0.000 time steps.5 1 1.9: Electric ﬁeld at 1 μm from the edge of port 3 of the power splitter of Fig.7. 6.02 ps. 6. obtained with the dynamic AMRFDTD. indicating the absence of latetime instability . to a reference FDTD simulation using a 3 2 1 Ez (V/m) 0 −1 −2 −3 0 0. Table 6.5 2 Time (ns) 2.
at t = 300 t. one level2. 6. at t = 100 t.112 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS FIGURE 6. and one level3 mesh FIGURE 6.11: Electric ﬁeld magnitude distribution across the power splitter of Fig. one level2.7. 6.12: Electric ﬁeld magnitude distribution across the power splitter of Fig.7. There are three meshes in total: one level1. and one level3 mesh . There are three meshes in total: one level1.
σAMR = 2. 6.001. 2. P2.15.7. at t = 400 t. three level2. NAMR = 10.13: Electric ﬁeld magnitude distribution across the power splitter of Fig. Table 6. All AMRFDTD simulations use a reﬁnement factor of 2 for their successive mesh levels. and P3 denote ports 1. respectively .4 shows that AMRFDTD can achieve the accuracy of the reference FDTD method in a dramatically reduced computation time. and the reference FDTD simulation are shown and compared in Fig. and four level3 meshes (tracking transmitted and reﬂected wavefronts) 2000 × 800 mesh and several coarser FDTD schemes. Dimensions are given in units of μm. θg = 0. Field waveforms in time.8. θc = 0.14: Geometry of the dielectric waveguide Yjunction.DYNAMICALLY ADAPTIVE MESH REFINEMENT IN FDTD 113 FIGURE 6. with a 250 × 100 root mesh. extracted by a fourlevel AMRFDTD. P1. The AMR parameters are: θe = 0. and 3. 6. The dielectric constant of the waveguide is 9 and the dielectric constant of the surrounding medium is 1. demonstrating y 1 2 10 P1 2 waveguide 2 2 4 absorber 25 7 11 x P2 P3 FIGURE 6. There are eight meshes in total: one level1.
7 174. Moreover. A waveform extracted by a conventional FDTD implemented on the root mesh of the AMRFDTD domain alone is appended.5 78. 2 μm from the edge of port 2 . despite its reduced numerical cost.7 646. to indicate the accuracy improvement. 6. The level1 mesh is drawn in the ﬁgures and 2 1 Ez (V/m) 0 −1 −2 2 3 Time (ns) AMRFDTD FDTD 250x100 FDTD 2000x800 4 x 10 −4 FIGURE 6.20 show the evolution of the electric ﬁeld and the mesh tree.17–6. which uses a 125 × 50 root mesh and maximum level of 3. Fig. 6.14.4 0.8 — 174.4 77.4: Computation Time and TimeDomain Error Metric Et for the Dielectric Waveguide YJunction for Fields Recorded at Port 2 METHOD FDTD FDTD FDTD FDTD AMRFDTD AMRFDTD AMRFDTD MESH 250 × 100 500 × 200 1000 × 400 2000 × 800 250 × 100 250 × 100 250 × 100 NUMBER OF LEVELS — — — — 2 3 4 TIME STEPS 2500 5000 10000 20000 2500 2500 2500 TOTAL TIME (S) 15.114 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS TABLE 6. Figures 6.55 the excellent agreement of the AMRFDTD approach to the reference data.16 shows that the multilevel AMRFDTD is free from the late instability problems that characterize static subgridding techniques.1 2665 Et (%) 166.7 155. brought about by the mesh reﬁnement.0 1296 23418 184.15: Electric ﬁeld at the center of the dielectric waveguide Yjunction of Fig.
one level2. 6. indicating the absence of latetime instability FIGURE 6. There are three meshes in total: one level1. obtained with the dynamic AMRFDTD.14. one level2.5 −3 FIGURE 6. and one level3 mesh (enclosed in the black solid line) .5 2 2.18: Electric ﬁeld magnitude distribution across the Yjunction of Fig. and one level3 mesh FIGURE 6.5 1 1.000 time steps. 6. 6.5 4 x 10 4. There are three meshes in total: one level1. at t = 100 t.17: Electric ﬁeld magnitude distribution across the Yjunction of Fig. at t = 400 t.DYNAMICALLY ADAPTIVE MESH REFINEMENT IN FDTD 2 115 1 Ez (V/m) 0 −1 −2 0 0.14.16: Electric ﬁeld at the center of the dielectric waveguide Yjunction of Fig. 2 μm from the edge of port 2.5 Time (ns) 3 3.14. for 25.
two level2. 6. (6. is simulated [59]. Note that level2 and level3 meshes track the pulse into the junction FIGURE 6.4) .21. The excitation is a modulated pulse of the form: e −(y−y0 ) /W 2 2 −(t−t0 )2 /T 2 sin (2π f t) . at t = 800 t. previously modeled in [61] and shown in Fig. and one level3 mesh (enclosed in the black solid line). The edgetoedge distance between the ring and the waveguide is 0. The computational domain is 12 μm × 12 μm and it is truncated by a 1 μm thick matched absorber. one level2.24. There are four meshes in total: one level1. The width of the waveguide and the ring is 0. These ﬁgures indicate the wavefront tracking achieved by the dynamic AMRFDTD. The external radius of the ring is 2. 6. at t = 600 t.116 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS FIGURE 6.14.3 μm. The rectangular boxes embedded in the solid gray regions indicate the boundaries of level3 meshes.20: Electric ﬁeld magnitude distribution across the Yjunction of Fig.14. and their dielectric constant is 10.5 μm.19: Electric ﬁeld magnitude distribution across the Yjunction of Fig.5 DIELECTRIC RING RESONATOR A highly resonant THz structure of a dielectric ring resonator coupled to two dielectric waveguides. which are not drawn because they are too dense. 6. 6.232 μm. There are three meshes in total: one level1. and one level3 mesh (enclosed in the black solid line) the regions without grid lines are covered by level2 and level3 meshes.
P1.2 — 163.01. The excitation is imposed 1. W = 0. All AMRFDTD simulations use a reﬁnement factor of 2 for successive levels.8. σAMR = 2. TABLE 6.7 1.5: Computation Time and TimeDomain Error Metric Et for the Dielectric Ring Resonator for Fields Recorded at Port 2 METHOD FDTD FDTD FDTD FDTD AMRFDTD AMRFDTD AMRFDTD MESH 120 × 120 240 × 240 480 × 480 960 × 960 120 × 120 120 × 120 120 × 120 NUMBER OF LEVELS — — — — 2 3 4 TIME STEPS 20000 40000 80000 160000 20000 20000 20000 TOTAL TIME (S) 60.0 163. Dimensions are in units of μm.5 1080 4096 Et (%) 187.21: Geometry of the dielectric ring resonator coupled to two dielectric waveguides.1 μm from the end of the lower left waveguide port.6 36. and 4.56 . P2.3 x 0. NAMR = 10. t0 = 3T. θg = 0. and θe = 0. and the electric ﬁeld is recorded at 2 μm from the edge of each port.DYNAMICALLY ADAPTIVE MESH REFINEMENT IN FDTD 12 117 P3 1 0.3 y 2.71 703.232 P2 P1 FIGURE 6. T = 0. 2.7 μm.56%. and P4 denote ports 1.1 6904. P3. Table 6.5 compares accuracy and execution times of AMRFDTD and conventional FDTD for this largescale computational problem. 3.5 P4 12 0. a fourlevel scheme can reproduce the results of the reference FDTD simulation within a smaller execution time (by a factor of 33).4 133993 466. at a relative timedomain error of ﬁelds recorded at port 2 of 1. θc = 0. respectively where f = 200 THz.6 37.02 ps. and y0 corresponds to the midsection of the waveguide. Again.
22: Electric ﬁeld at 2 μm from port 2 of the ring resonator of Fig. It should be noted that the ﬁrst Gaussian pulse arrives at port 3 at about 0. 5 AMRFDTD FDTD 120x120 FDTD 960x960 Ez (V/m) 0 −5 1. the ﬁeld at port 3 is due to the disturbance caused by the excitation.15 x 10 3. which propagates through free space.4 2. Before that.22–6.2 ps.05 2.05 Time (ns) 3.6 2.1 3.8 2 2. the pulses circulating the ring resonator and coupled to the waveguides are tracked by the ﬁnest mesh and therefore their amplitude and group velocity are very accurately resolved even at very late stages of the simulation. However.05 AMRFDTD FDTD 120x120 FDTD 960x960 Ez (V/m) 0 −0. it is not tracked by the ﬁnest mesh and therefore it induces a relatively large error.4 1.95 3 3.6 1. 6. Since this disturbance is very small compared with the main pulse propagating along the waveguide.2 Time (ns) 2.9 2.2 1.118 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS The timedomain results of Figs.2 −3 FIGURE 6.21 .24 demonstrate the latetime stability and accuracy of the method.8 x 10 3 −4 0. 6.
1 AMRFDTD FDTD 120x120 FDTD 960x960 0.1 3.02 0.01 119 AMRFDTD FDTD 120x120 FDTD 960x960 Ez (V/m) 0.2 3.21 .01 −0.05 −0.23: Electric ﬁeld at 2 μm from port 2 of the ring resonator of Fig.05 3 3.005 −0.05 Ez (V/m) 0 −0.DYNAMICALLY ADAPTIVE MESH REFINEMENT IN FDTD 0.02 0 1 Time (ns) 2 x 10 −4 0.005 0 −0. 6.3 x 10 −3 Time (ns) FIGURE 6.1 2 3 4 Time (ns) 0.015 −0.015 0.05 x 10 −4 AMRFDTD FDTD 120x120 FDTD 960x960 Ez (V/m) 0 −0.
6 NUMERICAL ERROR ESTIMATION AND CONTROL There are ﬁve parameters controlling the accuracy and computation time of AMRFDTD. Since the structure is highly resonant.012 0.6.120 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS 5 Ez (V/m) 0 −5 0 0.004 0. obtained with the dynamic AMRFDTD.25. the S21 obtained by using AMRFDTD is shown in Fig.014 0. which agrees quite well with the results of [61]. which indicates a very good agreement between the two sets of results.000 time steps. The resonant frequencies of the present method and [61] are compared in Table 6.01 Time (ns) 0. a clear set of . 100. for 100.006 0. extracted by FDTD. The AMRFDTD uses a 120 × 120 root mesh and four levels.24: Electric ﬁeld at 2 μm from port 2 of the ring resonator of Fig. as described by the previous sections.000 time steps are executed. 6.002 0.25: S21 obtained with the fourlevel AMRFDTD Finally.008 0. indicating the absence of latetime instability 0 −2 −4 −6 −8 −10 185 Amplitude of S21 (dB) 190 195 200 Frequency (THz) 205 210 FIGURE 6.016 FIGURE 6. To use AMRFDTD as a CAD technique. 6.21. 6.
the error is deﬁned as: ref f m (tk ) − f m (tk ) 2 E= k.11 FDTD [61] 185.6: Resonant Frequencies (in THz) of the Dielectric Ring Resonator ORDER 25 26 27 28 29 AMRFDTD 186. The numerical error whose dependence on the AMRFDTD parameters is sought for is deﬁned as follows. this metric is much more effective in capturing generic error effects. such an approach is the only feasible.01 192.92 209.85 191. . Given a computational domain. for a 2D TE case with ﬁeld components (Ez. due to the arbitrary distribution of submeshes in a domain. the extraction of analytical error bounds seems to be impossible. The main objective of this section is to produce such guidelines. if ref a ﬁeld component recorded at the mth probe at time tk is denoted by f m (tk ) and f m (tk ) the corresponding reference ﬁeld value. due to the dynamic nature of mesh evolution in AMRFDTD. or the error in a frequency domain quantity (such as characteristic impedance or propagation constant).88 197. Hy ). given the nature of the problem at hand.m ref f m (tk ) k. that AMRFDTD is compared to. Then. The reference solution. the same methodology can be easily extended to 2D TM cases and 3D cases.m 2 (6. is provided by applying the FDTD method in a uniform mesh.03 198. than a norm that would be based on ﬁeld sampling at a port of a device. Therefore.94 guidelines for the choice of the ﬁve parameters is needed.08 210. Finally. an array of probes is considered.DYNAMICALLY ADAPTIVE MESH REFINEMENT IN FDTD 121 TABLE 6.90 203.06 204. that would allow a user to determine the AMRFDTD parameters for a given error tolerance [59]. This approach is analogous to a MonteCarlotype simulation and is aimed at rendering the obtained error less dependent on the device under test. Contrary to static subgridding methods. Hx . at a Yee cell resolution equal to the maximum achievable resolution of the AMRFDTD. This is chosen to be dense enough to guarantee convergence. but also mesh conﬁgurations.5) Note that different probes encounter different ﬁeld waveforms. where ﬁeld values are recorded at each time step. Furthermore.
7 3.8 −1.3 0.6 log10(Simulation time) (a) −1 −1.5 1 splitter waveguide Yjunction 1 2 2 3 4 3 3.2 0.8 −2 −2. and θc on the accuracy and simulation time is investigated.26 includes errorsimulation time curves.8 0.1 3 1 4 ε 10 2 3 4 log 3.5 0.3 log −1.5 0.2 splitter waveguide Yjunction 3. (b) σAMR .3 3.26: Error versus simulation time for variable (a) NAMR .4 3.9 3.7 3.7 0. These are the power splitter.5 −3 10 20 40 20 40 40 splitter waveguide Yjunction 3.8 4 10 20 log 10 ε 3 3.5 10 −2 −2.8 0. presented and simulated in the previous section.6 10 0.5 log10(Simulation time) (b) 3. σAMR .8 0.4 3. (c) θc . the effect of NAMR .5 −2 −2. deduced by changing one of these −1.8 3.5 ε 0.7 0.4 3.3 3. corrugated permittivity proﬁle waveguide and the Yjunction.4 −1. results obtained from simulating three dielectric waveguide ones are shown. Figure 6. First.6 3.2 3.1 3.7 0.3 3 3.122 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS While the AMRFDTD error performance on a number of structures was studied.6 log10(Simulation time) (c) FIGURE 6.5 3.2 3.
2 −3.6 1.26. All ﬁgures refer to fourlevel AMRFDTD simulations.5 10 g 0. The general pattern of these curves suggests that although increasing NAMR and σAMR or decreasing θc may improve accuracy (e.7 1.5 −2 −1. since marking a cell for reﬁnement depends on both thresholds.7 balances accuracy and efﬁciency well.5 −3 −2.. As expected. any change in the parameters merely increases the execution time.2 1.8 1. the error is mainly a function of θe and vice versa.3 1.4 1. 0) arctan θg C2 C4 (6.5 1. which is why their choice is rather straightforward.4 1.5 1.9 1.3 1.9 1 1.g. the error decreases as θe or θg decreases.5 log10θe −1 −0. as NAMR is increased the AMRFDTD tends to become equivalent to the reference FDTD technique).6 1.5 −2 −2. Figure 6.1 1.1 2.2 1. θg ) = C1 E(0. the choice of NAMR = 10. σAMR = 2.5 −4 2 −4. Based on this data E(θe .1 1.7 1. θg that are studied next.1 0.27 shows the dependence of the error E on θe and θg for all the cases studied. these parameters do not explicitly control the mesh reﬁnement procedure.2 1. 6. this improvement reaches a plateau beyond which. but twoand threelevel simulation results exhibit a similar behavior.27: Dependence of −lo g 10 E on θe and θg 1 1.2 1.9 1.5 1. θg ) arctan θe θe + C3 E(θe .9 0. This is not the case for θe . This conclusion is also supported by Fig.3 1 log θ −3 −3.8 1.1 1.5 1.6 2 0 . and θc = 0.6) .DYNAMICALLY ADAPTIVE MESH REFINEMENT IN FDTD 123 three parameters and ﬁxing the rest.1 FIGURE 6.5 −5 −4 2.9 1 1 1. θg ) can be approximated as: θg ˜ E(θe .4 1. −1 −1. In a number of versatile situations. Essentially. For very small values of θg .3 1.
5 10 g 0.124 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS −1 −1. C4 = 0.5 2. By simulating typical optical waveguide structures.1 −5 −4 −3.6 2 −2. As an example.02.7 −3. C4 = 0.6). 8 1. 6.05.2 −4.2 1.4θg .9 −4 2 1.3 1.01. θg ) = 0. 0) and E(0.7 1 1.28: Curveﬁtted dependence of −lo g 10 E on θe and θg based on (6. 1. C2 = 0. C2 . θg ). θg ) arctan θg θe C2 θe + C3 Ebe (θe .8) C4 (6.4 1. 0. θg = 0) arctan θg where C1 = C3 = 1.01. log θ −3 1.5 1.28 shows the curveﬁtted error obtained by (6.1 1 1. respectively.27.3 1. Therefore. two empirical error bound functions are obtained for the special cases θe = 0 and θg = 0. Fig. where C1 = C3 = 1. requiring only E(θe .1 0.6) ˜ where C1 .2 1 1.5 1.3 1. θg ) = C1 Ebg (θe = 0.5 −3 ˜ FIGURE 6.9 1 1.9 1.05. which depicts a very good agreement with Fig.37 .9 0. 6. C2 = 0. θg = 0) = 0. and C4 are ﬁnetuned to minimize the difference between E and E.02.6 2 1. Ebe (θe .9 1. 6.4 1.8 1. this approximation. we deﬁne a general error bound as Eb (θe . is useful as sufﬁciently accurate.5 1. C3 .29 and can be used to determine θe and θg given the desired accuracy.1 0.9 1 0.5 1. This general error bound is shown in Fig.9) .5 log θ 10 e −1 −0.4 1. 1.5 0 (6. Following (6.35 Ebg (θe = 0.15θe0.7) (6.6 1.6).5 −2 −1.5 −2 −2.8 1.3 1.
4 1.9 7 8 1. requiring that the error be less than 3%.41. 0.7.5 −4 −4 1. Then.3 1. similar to that of [62].24 and 4.5 1. The surrounding medium is air.2 1.82 x 0.0003 satisﬁed this criterion.5 1. Dimensions are given in μm. θg = 0.6) While the validity of these bounds can be conﬁrmed through all previously published examples of the dynamic AMRFDTD.7 −3.5 −3 −3. The dielectric constant of the waveguide and the center piece is 3. 0. is also discussed here. 5 −2.9 −0.8 0.1 −2 −1.30). 6. the case of a dielectric waveguide directional coupler (Fig.3 −1 0.5 0. letting NAMR = 10.8 0.2 1.9 1 1.3 1 1.2 1 0.35 P2 P3 FIGURE 6.5 log10θe FIGURE 6.001.DYNAMICALLY ADAPTIVE MESH REFINEMENT IN FDTD 125 −1 −1.29.4 1.5 0 . In this example. 0.6 1. The excitation was a sinewave y 0. θg and θe were chosen. respectively 1. from Fig.9 absorber 16 2. σAMR = 2 and θc = 0. θe = 0.5 −3 1. 6.5 P4 6 P1 waveguide 2.1 1.9 1 −2 1.9 1 0.29: Curveﬁtted dependence of error bound −lo g 10 Eb on θe and θg based on (6.30: Geometry of a directional coupler (not drawn to scale).1 10 g 0.8 log θ −2.25 0.
The excitation was imposed at 1. Fields and errors were recorded over 3200 time steps of the root mesh. AMRFDTD cannot be distinguished from the reference solution.1 50 100 150 Time (fs) (b) Port 4 FDTD 2560x960 AMRFDTD FDTD 320x120 200 250 FIGURE 6.5 −1 80 FDTD 2560x960 AMRFDTD FDTD 320x120 100 120 Time (fs) (a) Port 3 140 160 180 0.5 Ez (V/m) 0 −0. For AMRFDTD. The reference result was obtained by applying FDTD in a 2560 × 960 uniform mesh.7. The actual error was 0. An array of 14 × 8 probes was used to record the Ez ﬁeld throughout the computation domain. the size of the coarsest mesh was 320 × 120.2 μm from port 1. although it indicated that the deduced bound was rather conservative. Finally.05 −0. even 1 0. The time step for each mesh was determined by letting all mesh Courant numbers be equal to 0. 6.46%. four levels of mesh resolution were used.126 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS modulated Gaussian pulse with a center frequency of 230 THz and a pulse width of 20 fs.1 0. Fig. which satisﬁed the requirement.31: Electric ﬁeld at port 3 and 4 of the directional coupler . Evidently.31 compares the ﬁelds obtained by AMRFDTD and FDTD in ports 3 and 4 of the coupler.05 Ez (V/m) 0 −0.
3 hours. Large speedup factors (ranging from 30 to 100) compared to the conventional FDTD were achieved. For that waveform. while all errors involved remained small. The AMRFDTD simulation lasts 2.DYNAMICALLY ADAPTIVE MESH REFINEMENT IN FDTD 127 for the weak ﬁeld at port 4. Guidelines for the a priori choice of these parameters were provided. 6. the coarse mesh FDTD result suffers from a signiﬁcant numerical dispersion induced error.7 CONCLUSION This chapter discussed the application of the multilevel. Furthermore. and its dependence on the parameters controlling the accuracy and performance of this technique was outlined. Numerical examples demonstrated the efﬁciency and latetime stability of the AMRFDTD technique. thus fulﬁlling an important condition for AMRFDTD to be considered as a CADoriented tool for electromagnetic and optical applications.8 hours. typically of the order of 1% or less. . the numerical error associated with the application of the AMRFDTD technique was studied. dynamically AMRFDTD technique to optical waveguide structure analysis. whereas the reference FDTD simulation 19.
.
Soc. 1968. F. Harrington. vol.1109/8. no.. Sommerfeld.491023 [12] W. 812–815. B. 6. Dec.. MA: Kluwer. Electromagn. [4] A. “A local mesh reﬁnement algorithm for the time domain ﬁnite difference method using Maxwell’s curl equations. Feb. 1996. vol.1109/22. New York: Macmillan. 438–440. Hoefer.” IEEE Trans. 1993. pp. 6. J. 1971.. pp. Rev. Prog. New York: J. Am.. Microwave Theory Tech. June 1990. Mittra and S. 1993. 1962.130985 [11] M. New York: Macmillan. 2. 302–307. Nakhla.. Wolff. Antennas Propagat.544542 [13] M.” IEEE Microwave Guided Wave Lett. vol. Yee. pp.” Proc. 1985. [3] R. 1996. Mar.” IEEE Trans.doi:10. “The transmissionline matrix method—Theory and applications. Microwave Theory Tech. Katehi. 555–561. Microwave Theory Tech. [8] E. R. no.” J. J. Antennas Propagat. pp. Dec. W. Analytical Techniques in the Theory of Guided Waves. Appl. Jin.. [5] J. B. Asymptotic Waveform Evaluation and Moment Matching for Interconnect Analysis. 1732–1737. [9] K. Schlager et al. vol. no.doi:10. MTT44.doi:10. Aidam and P. 52. “Geometrical theory of diffraction. . MTT33. New York: Academic Press. 12. 1964. Krumpholz and L. “Numerical solution of initial boundary value problems involving Maxwell’s equations in isotropic media.” IEEE Trans. vol. W. S. 10. Keller. Hoefer. Kim and W. AP14. Oct. vol. L. Annu. Partial Differential Equations in Physics. Mar.” IEEE Trans. Lee.. 1966.273296 [10] I. MTT38.. Werthen and I. R. 882–893. S. [2] R. pp. 41. pp. 1994. The Finite Element Method in Electromagnetics. S. pp. “Relative accuracy of several ﬁnitedifference timedomain methods in two and three dimensions. pp. [7] W. no. [6] K. P. Russer. Apr. 116–130. “A novel wavelet based time domain approach.1109/75. Comput. Chiprout and M. Wiley & Sons. Norwell. 4. “Application of biorthogonal B—spline wavelets to Telegrapher’s equations.1109/22..129 Bibliography [1] J.” IEEE Trans. no. Opt. 983–990. Field Computation by Moment Methods. 1998. “MRTD: New time domain schemes based on multiresolution analysis.doi:10. vol.
1998.doi:10. vol. E. 1563–1573. “A hybrid FDTDwaveletgalerkin technique for the numerical analysis of ﬁeld singularities inside waveguides..1109/20.1109/36.” Proc. “Application of MRTD to printed transmission lines. T. 1563–1566.” IEEE MTTS Int. Magn. Kantartzis. vol.doi:10. vol. 10. and K. vol.doi:10. 1301–1304. [23] F. 1999. [19] E. Z.” IEEE Trans. B.130 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS [14] M. and D. 573–576. “A 2D TLM and Haar MRTD real time hybrid connection algorithm. pp. July 2000. M. [18] M. Antennas Propagat. Tentzeris et al. and N. B. New York: Academic Press. 2692–2698. Dec. Robertson.1109/8. 1397–1400. Oct. Kim. [20] E. V. no. Hoefer. 1998.. Carin. Dig.. pp. 1568–1578.and timeadaptive gridding using MRTD technique. pp.. 1997. 337–340. R.” IEEE MTTS Int.964093 [17] E. 1997. A Wavelet Tour of Signal Processing. 1. G. Kosmanis. Microwave Symp. I. 1013–1020. Wolff and M. Microwave Symp. Dig. Prog. P. Microwave Theory Tech..899673 [16] X.” IEEE MTTS Int. IEEE APS Symp. vol. pp. . vol. no. 11. Fujii and W. pp. Tsiboukis. and L. Rev. S. 6.” IEEE MTTS Int.. 37. vol. 1996. P.” IEEE MTTS Int. 2000. 36.H. [15] S. “Multiresolution timedomain analysis of planewave scattering from general threedimensional surface and subsurface dielectric targets. Nov.. Walter. Roselli. Ju. Microwave Symp. Microwave Symp. pp.. 48. Huang. vol. “A threedimensional Haar waveletbased multiresolution analysis similar to the 3D FDTD method—derivation and application. D.” IEEE Trans. Comput.. “Adaptive transient solution of nonuniform multiconductor transmission lines using wavelets. [24] I. Katehi. 3. Krumpholz. Katehi. M. pp. L. Ritter. Dig. “MRTD applied to complex airdielectric boundary structures. “A generalized 3D subgrid technique for the FDTD method. B. “Nonlinear circuit characterization using a multiresolution time domain technique (MRTD).803416 [26] T. 49. Mallat. “An FDTD/MoM hybrid technique for modeling complex antennas in the presence of heterogeneous grounds. 1999. 3.” IEEE MTTS Int.. [28] H. pp. 2000. “The modeling of lumped elements using the Haar wavelet multiresolution time domain technique. J. and L. pp. Fujii et al. no.” IEEE MTTS Int. 2000. 2463–2475. 1999. Arndt and J. 1997. Annu. Dig.. Goverdhanam et al. pp. vol... [22] S. G. Electromagn. Microwave Symp. “An algorithm for realizing Yee’s FDTD method in the wavelet domain. pp. L. M. “Applications of a multiresolution based FDTD multigrid. Zhu and L.” IEEE Trans. R. 4.877589 [27] M. Microwave Symp.1109/8. Demarest. M. Tentzeris. Plumb. vol...” Proc. [21] K. Geosci. and L. pp.. 333–336. pp. Dig. P.doi:10. “Space. Dig. Nov. Katehi. Bae. 1463–1466. 1999. 1. Microwave Symp. vol. MTT46. 902–906. 2001. Appl. Remote Sensing. 3. Tentzeris.” IEEE Trans. Tentzeris. [25] R. Dig. GrivetTalocia.” IEEE Trans. no. Antennas Propagat. 4.
Dec. Feb. B. FL: CRC Press.” IEEE Trans. M. no. pp. “An efﬁcient analysis of planar microwave circuits using a DWTbased Haar MRTD scheme. 2261–2270. Philadelphia. B. 243–246.doi:10. Battle. 6. Boca Raton. D..402284 [36] J. 110.1109/8. 774–784. “Zur Theorie der Orthogonalen Funktionensysteme. no. R. Computational Electrodynamics: The FiniteDifference TimeDomain Method. Phys.1007/BF01456326 [34] G. [32] G. 2.. Daubechies. no. Finite Difference Schemes and Partial Differential Equations.” IEEE Trans.. 50. 1987.1109/22. G. Microwave Theory Tech. vol. Strikwerda. vol. Wiart. Antennas Propagat. vol. pp.” Math. 2001. Goverdhanam et al.. pp.2003. “Fundamental gridding related dispersion effects in MRTD schemes. Real Analysis. Ann. vol. 12...doi:10.1109/22.doi:10.809620 [30] I. 1992. Antennas Propagat. 51. Dec.doi:10.. 793.. 12.2002. 2000. “An efﬁcient numerical interface between Haar MRTD and FDTD: Formulation and applications. Citerne.982230 [41] K. and S. Sarris and L. Dec. 69. E. 3rd edn.. vol. no. Huber. 331–371.. June 2002. and P. Krumpholz. P. Russer. Katehi. 2248–2257. pp.1017657 [38] C. Carat. 1999. 49. 1989. vol. pp. pp. “Application of Haarwaveletbased multiresolution timedomain schemes to electromagnetic scattering problems. pp.” IEEE Trans. no.” Proc. D. Modeling Numer.” IEEE Trans. Aug. and J. 50.BIBLIOGRAPHY 131 [29] C.971607 [39] A. Gedney. [37] T. Englewood Cliffs.1007/BF01205550 [35] C. 4.doi:10. MA: Artech House. [40] E. “A ﬁeld theoretical comparison of FDTD and TLM. pp.546249 [43] J. Part I : Lemarie functions. 1146–1156.1109/22. PA: Society for Industrial and Applied Mathematics. Royden.” Math. “A perfectly matched layer formulation for Haar wavelet based MRTD. 12. Norwood. 1995. Apr. 1360–1369. pp. Ten Lectures on Wavelets.doi:10.1109/TMTT. 2003.898973 .. Sarris and L.” IEEE Trans. 48. 2002. Bacry. doi:10. [42] S. 7. Microwave Theory Tech. Papanicolaou. “A wavelet based spacetime adaptive numerical method for partial differential equations. Haar. Microwave Conf. p. vol. Microwave Theory Tech. no. no. 1988. Taﬂove. 1992. L. 44. Microwave Theory Tech. Carin.” IEEE Trans. “Multiresolution timedomain (MRTD) adaptive schemes using arbitrary resolutions of wavelets. 601–615.1109/TAP.” IEEE Trans. Anal. 43. 1996. 1910. vol.. “A block spin construction of ondelettes. Mallat. Gillard.1109/22. Microwave Theory Tech. Tentzeris et al. D. NJ: Prentice Hall. vol. Katehi.” Comm. “An anisotropic perfectly matched layerabsorbing medium for the truncation of FDTD lattices.doi:10. C. 1935–1950. 1995. Eur. vol. 26.. Math. M. P. [31] H. pp. 501–516. Dogaru and L.doi:10. [33] A.
” IEEE Trans. R.. Feb. Phys.” Proc. 1985. 484–512. 46.. pp. Comput. Syst. Okoniewski. Heyman. 1414–1417. no. vol..132 MESH REFINEMENT FOR TIMEDOMAIN NUMERICAL ELECTROMAGNETICS [44] M. Sheen. 1997. pp. Man. G. B. B. W.. no.558657 [45] M. Sullivan. Liu.doi:10. W. Ali. Electromagn. and R. [55] M. and J.1109/TMTT. M.95580 [47] H.” Appl. Schuster. “Application of moving window FDTD to prediction of path loss over irregular terrain. M. 1997. no. doi:10. Oliger. 1991. Microwave Theory Tech. [51] R.” IEEE Trans. vol.1109/22. Kastner./Oct. 45. [50] R. Sept.1109/8. Fidel. 3.1016/00219991(84)900731 [53] Y. M. pp.” IEEE Trans.. vol.55775 [49] R. Mar. doi:10. Mar. “An algorithm for point clustering and grid generation.” Proc. vol. 45.. Berger and I. and E. “Hybrid rayFDTD moving frame approach to pulse propagation. “Three dimensional subgridding algorithm for FDTD. “Adaptive mesh reﬁnement for hyperbolic partial differential equation. 1994. June 2003. 1278–1286. no. IEEE APS Symp. vol. Kuzyk. A. 2006. 1006–1009. Wu. pp. 377–382. Berger and J. vol. July 2000. Ziolkowski. 689–703. 1997. 2. M. Winful. Abouzahra. 1984. “Threedimensional optical pulse simulation using the FDTD method. Sarris. Katehi. IEEE APS Symp. pp. pp. 385–393. Mar. [52] M. vol. “Absorbing boundary conditions in the context of the hybrid rayFDTD moving window solution. Lett. Stuchly.doi:10. 422–429. July 1990. pp. and Linda P. 849–857. Okoniewska. Microwave Theory Tech. Ziolkowski. 38. 2. 54.. pp. vol.” IEEE Trans.1109/21. Fidel. 7. J. Liu and C. A. vol. 5. 21.” IEEE Trans. Kong.862660 [54] G. D.1109/22. Heyman.” J..848495 [46] H. “Efﬁcient modeling of microwave integrated circuit geometries via a dynamically adaptive mesh reﬁnement (AMR)FDTD technique. and R. Microwave Theory Tech. 48..563337 [48] D. 527–529. S.” Proc. Krumpholz.1109/22.. IEEE APS Symp. vol. “Nonlinear time domain modeling by multiresolution time domain (MRTD). Winful. 3.2005. doi:10.. Mur. Kastner. E. 53. no.doi:10. 2. no. vol. and J. Phys. pp. “Application of the threedimensional ﬁnitedifference timedomain method to the analysis of planar microstrip circuits.. pp.120081 .” IEEE Trans. 1981. vol. G. Computat. 3.1063/1. Antennas Propagat. 610–613.doi:10. pp. Rigoutsos. 1127–1133. 7. D. pp.” IEEE Trans. M. Cybern. Luebbers. Microwave Theory Tech.doi:10. EMC23. B. E. and K. “Absorbing boundary conditions for the ﬁnitedifference approximation of the timedomain electromagnetic ﬁeld equations.. “Pulse compression in optical ﬁber ﬁlters.
Chu and S. 11. 186. Rev. 24. D. 2. Computat. [58] S. and I. Microwave Theory Tech.BIBLIOGRAPHY 133 [56] A. pp. 1997. Annu. pp. “What determines the speed of timediscrete algorithms. “FDTD microcavity simulations: Design and experimental realization of waveguide coupled singlemode ring and whisperinggallerymode disk resonators. Comput. “Fast timedomain simulation of optical waveguide structures with a multilevel dynamically adaptive mesh reﬁnement FDTD approach. 1989.1109/JLT. T.doi:10. Zivanovic. Ho. Liu and C. 2.doi:10.” IEEE MTTS Int. J.41625 . W. no. vol. vol. M. T.” IEEE/OSA J. Phys. vol.1109/22. Taﬂove. “Interpolating wavelet collocation method of time dependent Maxwells equations: Characterization of electrically large optical waveguide discontinuities. T.. R. Prog.doi:10. 39. pp. R. Appl. So. K.” IEEE/OSA J. 8. “A ﬁnitedifference timedomain method for the design and analysis of guidedwave optical structures.. Hoefer. [57] T. Mar. Hoefer. Fujii and W. 594–601. pp. Raﬁzadeh. and K. and P. 2033–2038. vol. Russer. V. Mangold.” Proc. “A novel adaptivity for em time domain methods: Scale adaptive time steps (SATS). vol. 666–689. Sarris. C.” J.. no. S. Electromagn. pp.1109/50.” IEEE Trans. Y. Rennings. no. 2. K. vol. 2154–2165. Itoh. May 2001. Rebel. 2003. Hagness. P. Microwave Symp.. S. no.1109/50. 757–760. 15. vol. 2000. 12.876899 [60] M. Nov. pp. 471–479. Qian.2006.1016/S00219991(03)000913 [61] S. P. Dec. J. pp. K. Waldow. 7. P. Apr. Chaudhuri. 3235–3247. Dig. no. J..doi:10. Mei. 3. D. 1991. Lightwave Tech. Lightwave Tech.75289 [59] Y. Toni.. Lightwave Tech. Aug.641537 [62] S.. Wolff. and A. S.doi:10. 2006. Yee.” IEEE/OSA J. “A subgridding method for the timedomain ﬁnitedifference method to solve Maxwells equations.
.
University of Toronto. Greece. and a M. . respectively. in 1997. Canada. in Electrical Engineering.D. University of Toronto.Sc. He received a student paper award in the 2001 International Microwave Symposium for his work on a hybrid FDTD/MRTD numerical scheme. In November 2002. a Canada Foundation for Innovation New Opportunities Fund Award in 2004 and an award for excellence in undergraduate teaching from the Department of Electrical and Computer Engineering. wavepropagation in complex media and metamaterials and electromagnetic compatibility/interference (EMI/EMC) problems. as well as applications of timedomain analysis to wireless channel modeling. meshadaptive techniques.Sc. with emphasis in highorder. ON. He is currently involved with basic research in novel numerical techniques. and a M. 1998 and 2002. in 2002. Rogers Sr. where he is currently an Assistant Professor. in Applied Mathematics from the University of Michigan. Department of Electrical and Computer Engineering (ECE). including Hellenic Fellowship Foundation (1993–1997) and Technical Chamber of Greece (1994–1997) awards for academic excellence and an NTUA 1997 class bronze medal. he joined the Edward S. Prof. Sarris received a Ph. Ann Arbor. His research interests are in the area of computational electromagnetics. Toronto.135 Author Biography Costas D. He also received a Diploma in Electrical and Computer Engineering (with distinction) from the National Technical University of Athens (NTUA). Sarris has received a number of scholarship distinctions.