Advanced Formulations and Applications of Finite

Difference Time Domain Analysis

by

Ravi Chandra Bollimuntha

M.Tech., Banaras Hindu University, Varanasi, 2012

A thesis submitted to the

Faculty of the Graduate School of the

University of Colorado in partial fulfillment

of the requirements for the degree of

Doctor of Philosophy

Department of Electrical, Computer & Energy Engineering

2018
 This thesis entitled:
Advanced Formulations and Applications of Finite Difference Time Domain Analysis
written by Ravi Chandra Bollimuntha
has been approved for the Department of Electrical, Computer & Energy Engineering

Prof. Melinda Piket-May

Prof. Mohammed Hadi

Date

The final copy of this thesis has been examined by the signatories, and we find that both the
content and the form meet acceptable presentation standards of scholarly work in the above
mentioned discipline.
Bollimuntha, Ravi Chandra (Ph.D., Electrical Engineering)

Advanced Formulations and Applications of Finite Difference Time Domain Analysis

Thesis directed by Prof. Melinda Piket-May

This dissertation deals with advanced formulations and applications of finite difference time

domain (FDTD) method. This is composed of four themes. The first deals with the development

of a plane wave excitation formulation in FV24, a finite-volume based higher-order FDTD variant.

With its excellent phase error performance even for coarse grid, FV24 can be applied to electrically

large problems. The plane wave excitation method based on Total field/Scattered field formulation

and Discrete planewave technique is demonstrated and validated for FV24. In the latter part of

this work different near to farfield transformation approaches possible in FDTD are compared for

accuracy.

The second part of the research deals with application of FDTD to glass weave-induced skew

(GWS) problem. The GWS on a differential pair can cause increased bit error rates affecting the

robustness of the digital system, and increased radiated emissions causing compliance failures. The

frequency dependence of glass and resin materials properties are modeled using auxiliary differential

equation formulation for FDTD, to estimate GWS on a differential pair. The skew numbers are

benchmarked with the available commercial solvers. Also, the use of graphical processing units

(GPUs) to accelerate the skew simulations is demonstrated.

The third part of the research deals with the derivation of numerical dispersion relation

(NDR) for spherical FDTD, and the sensitivity study of the associated numerical wave number.

Elementary functions native to spherical coordinates are used in the derivation of the numerical

dispersion. Given the non-uniform nature of the spherical FDTD grid, the NDR and the cor-

responding numerical wave number are shown to be position dependent. The latter part of this

research includes a study to derive the stability criterion for spherical FDTD and challenges involved

therein.
 iv

The final part of the research studies the effectiveness of different absorbing boundary con-

dition formulations for spherical FDTD in absorbing the waves. It is shown that the split-field

formulation of perfectly matched layer (PML) is not as effective as stretched-coordinate formula-

tion. This work includes derivation of continuous-space PML reflection coefficient, update equations

for implementation of stretched-coordinate PML in spherical FDTD and analysis of reflection error

for different PML parameters.

 Dedication

In dedication to my mother Dhanalakshmi, my teachers and my wife Suganya Manoharan

 vi

Acknowledgements

Firstly, I would like to thank my advisers Prof. Melinda Piket-May and Prof. Mohammed

Hadi and Prof. Atef Elsherbeni for their guidance and advice throughout my journey as doctoral

student. Prof. Piket-May has been of constant support at every milestone- from choosing the

courses, preparing for the preliminary, comprehensive and final defense examinations. Many thanks

to her patience and feedback while getting me ready for the conference presentations.

Special thanks to Prof. Mohammed Hadi for helping me shape the dissertation right from the

beginning. This dissertation would not have been possible with out his constant guidance, feedback,

and availability for research reviews, and are greatly appreciated. His deep understanding of FDTD

schemes, dedication to research, and attention to detail are motivational. Thanks to Prof. Atef

Elsherbeni for his deep insights and feedback on the new formulations in the dissertation, and for

his time for the research reviews despite his busy schedule. The FDTD codes from his text books

have been tremendously helpful in quick implementation of the new formulations in this work.

I would like to thank Prof. Edward Kuester for being part of dissertation committee and for

offering two wonderful courses: waveguides and transmission lines, and electromagnetic boundary

problems. These courses helped me in not only understanding the theoretical basis, but are also

instrumental in removing mathematical road blocks in my research - thanks to the rigorous and

involved math in the courses. His clarifications on my research problems and feedback during the

comprehensive exam are appreciated. His dedication to teaching is inspiring and commendable.

Thanks to Prof. Dejan Filipovic for guidance and support on securing teaching assistantship

while he was the graduate director of the department. I am thankful for his kindness in lending
 vii

compute power and software licenses for my research. Thanks to Dr. Mohamed Elmansouri for

agreeing to be on the committee in short notice. Working in the same lab, I observed him over four

years, his dedication to research and guiding other students in the lab are inspiring.

Thanks to Dr. Eric Bogatin’s for the financial support and his high speed digital design course

inspired me to think of application of FDTD method to practical problems such as glass weave

induced skew dealt with in the thesis. His style of teaching, engineering and simulation philosophy

influenced the way I approached my research. Thanks to Prof. Robert Marshall for interesting

discussions on spherical perfectly matched layer and for sharing his equations and computer code

on its implementation.

Thanks to Alpesh Bhobe, Soumya De, Mike Sapozhnikov and Amendra Koul from Cisco Sys-

tems Inc., and Kevin Zhu, David Fernandez at Ansys Inc. for discussions about my research during

my internships. Thanks to research group members Alec Weiss, Ryan Smith, Sanjay Dmello, Vinit

Vyas, Dharmateja Paladugu, Neeti Sonth, and fellow PhD students Tim Wang Lee and Fadi Deek

for participating in numerous discussions about each other’s research and their feedback. Thanks

to fellow graduate students Saurabh Sanghai, Prathap Valaleprasannakumar, Ehab Etellisi, Elie

Tianang, Aman Samaiyar, Adbulaziz Haddab, and Prathamesh Pednekar for creating conducive

atmosphere in the lab for research, their insight, advice and support.

I would like to thanks Adam Sadoff, Susan Callihan, Laramie Rose, Pamela Aguila, Kelly

Payton, Christine Ralston in Dept. of ECEE for their administrative support and helping me

navigate the graduate school, TA/RA appointment, and payroll paperwork.

Finally, I would like to thank my wife Suganya Manoharan for bearing with me patiently

on numerous days while I conduct research and write thesis. Thanks to her helping hand in

typesetting equations and being the first proofreader of the thesis draft. Her moral support and

constant encouragement helped me in finishing my thesis on time.

 Contents

Chapter

1 Introduction: Review of Finite Difference Time Domain Analysis 1

1.1 Conventional FDTD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.2 Numerical Dispersion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

1.3 Field Sources in FDTD Grids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

1.4 Higher-Order FDTD: The FV24 Method . . . . . . . . . . . . . . . . . . . . . . . . . 5

1.5 Grid Truncation: Perfectly Matched Layer and Convolutional Perfectly Matched Layer 6

1.6 Organization of Rest of the Chapters . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2 Planewave Injection in Higher-Order FDTD Grid 9

2.1 Planewave Excitation in Standard FDTD . . . . . . . . . . . . . . . . . . . . . . . . 10

2.1.1 Total Field/Scattered Field Formulation . . . . . . . . . . . . . . . . . . . . . 10

2.1.2 The Discrete Planewave Technique . . . . . . . . . . . . . . . . . . . . . . . . 11

2.1.3 Auxiliary One-Dimensional Grid . . . . . . . . . . . . . . . . . . . . . . . . . 11

2.1.4 Properties of 1D Grid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

2.1.5 TF/SF Corrections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

2.2 Planewave Excitation in FV24 Grid . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

2.2.1 FV24 Update Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

2.2.2 One-Dimensional Update Equations for FV24 . . . . . . . . . . . . . . . . . . 21

2.2.3 TF/SF Corrections for FV24 . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

 ix

2.2.4 Waveform Excitation on 1D grid . . . . . . . . . . . . . . . . . . . . . . . . . 25

2.3 Validations of the Injection Technique . . . . . . . . . . . . . . . . . . . . . . . . . . 29

2.4 Comparison of Near-to-Farfield Transformation Techniques for FDTD . . . . . . . . 30

2.4.1 Arithmetic and Geometric Averaging . . . . . . . . . . . . . . . . . . . . . . . 34

2.4.2 The Separate and Mixed Surface Approach . . . . . . . . . . . . . . . . . . . 35

2.5 Frequency Domain Near-to-Farfield Transformation in FDTD . . . . . . . . . . . . . 37

2.6 Error Comparison with a No Scatterer . . . . . . . . . . . . . . . . . . . . . . . . . 38

2.6.1 Resolution Sweep . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

2.6.2 Incident Angle Sweep . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

2.6.3 Equivalent Surface Size Sweep . . . . . . . . . . . . . . . . . . . . . . . . . . 41

2.6.4 Bistatic RCS Comparison for a Dielectric Cube . . . . . . . . . . . . . . . . 42

3 Analysis of Glass Weave Induced Skew in Differential Pairs 45

3.1 Introduction: Glass Weave Skew in High Speed PCBs . . . . . . . . . . . . . . . . . 45

3.1.1 FDTD Modeling of Glass-Resin Composites: Motivation . . . . . . . . . . . . 49

3.1.2 Implementation of Glass Weave Structure in FDTD Grid . . . . . . . . . . . 51

3.2 Frequency Dependent Losses: Dispersive Medium . . . . . . . . . . . . . . . . . . . . 51

3.2.1 Adapting Djordjevic-Sarkar Model to FDTD . . . . . . . . . . . . . . . . . . 53

3.2.2 Implementation of Multi-Pole Debye Model in FDTD . . . . . . . . . . . . . 56

3.3 Glass Weave Skew Analysis using FDTD . . . . . . . . . . . . . . . . . . . . . . . . . 58

3.4 Comparison with HFSS Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

3.4.1 Acceleration of FDTD in MATLAB . . . . . . . . . . . . . . . . . . . . . . . 61

4 Spherical FDTD Dispersion Relation and Stability Analysis 62

4.1 Introduction to Spherical FDTD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

4.1.1 Spherical FDTD Update Equations . . . . . . . . . . . . . . . . . . . . . . . . 63

4.2 Numerical Dispersion Relation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

4.3 Derivation of Spherical FDTD Numerical Dispersion Relation . . . . . . . . . . . . 69

 x

4.3.1 Position Dependence of Numerical Dispersion Relation . . . . . . . . . . . . . 72

4.3.2 Convergence Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

4.4 Numerical Wave Number Sensitivity Analysis . . . . . . . . . . . . . . . . . . . . . . 73

4.4.1 Sensitivity to Mesh Resolution . . . . . . . . . . . . . . . . . . . . . . . . . . 73

4.4.2 Sensitivity to Mode Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

4.4.3 Sensitivity to Elevation Angle . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

4.5 Stability Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

4.5.1 Existing Stability Analysis for Spherical FDTD . . . . . . . . . . . . . . . . . 78

4.5.2 Stability Criterion based on Position Dependent Numerical Dispersion Rela-

tion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

4.5.3 Challenges in Validating Stability Criterion . . . . . . . . . . . . . . . . . . . 82

5 Spherical FDTD PML Analysis 83

5.1 Perfectly Matched Layer in Spherical FDTD . . . . . . . . . . . . . . . . . . . . . . . 83

5.2 Split-Field PML Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84

5.3 Lossy Absorbing Shell as a Boundary Condition . . . . . . . . . . . . . . . . . . . . . 86

5.4 Stretched-Coordinate PML Formulation . . . . . . . . . . . . . . . . . . . . . . . . . 88

5.4.1 Electric Field Stretched-Coordinate PML Equation . . . . . . . . . . . . . . 88

5.4.2 Update Sequence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92

5.4.3 Magnetic Field Stretched-Coordinate PML Equation . . . . . . . . . . . . . 94

5.5 PML Reflection Analysis in Continuous Medium . . . . . . . . . . . . . . . . . . . . 96

5.6 PML Performance Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98

5.6.1 Geometric Grading of PML Parameters . . . . . . . . . . . . . . . . . . . . . 98

5.6.2 Simulation Space and Parameters . . . . . . . . . . . . . . . . . . . . . . . . . 99

5.6.3 Performance Analysis and Comparison . . . . . . . . . . . . . . . . . . . . . . 101

5.7 Why Split-Field PML doesn’t work in Spherical FDTD? . . . . . . . . . . . . . . . . 104

 xi

6 Conclusion 107

6.1 Original Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111

Bibliography 113
 xii

Tables

Table

2.1 Indexes for field location of S22 Ex update equation . . . . . . . . . . . . . . . . . . 15

2.2 Indexes for field location of S22 Hx update equation . . . . . . . . . . . . . . . . . . 16

2.3 Indexes for some field location of FV24 Ex update equation . . . . . . . . . . . . . . 21

2.4 Indexes for some field location of FV24 Hx update equation . . . . . . . . . . . . . . 22

3.1 Loss-less Material Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

3.2 Material properties of glass and resin at 1 GHz . . . . . . . . . . . . . . . . . . . . . 54

3.3 FDTD Simulation Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

4.1 Legendre Polynomials Pnm (cos θ) for some lower order n and m = 0 . . . . . . . . . . 75

5.1 Simulation parameters for PML/ABC implementations in spherical FDTD . . . . . . 100

5.2 Legendre Polynomials Pnm (cos θ) for some lower order n and m = 0 . . . . . . . . . . 104
 Figures

Figure

1.1 Standard FDTD unit cell showing electric and magnetic field component locations [1]. 2

1.2 Computational voxel for FV24 algorithm vis-a-vis conventional FDTD unit cell. . . . 6

2.1 TF/SF formulation showing scatterer, total and scattered field regions, and interface

surface between them [2]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2.2 Auxiliary 1D grid, used in Discrete Planewave technique, in relation to the TF/SF

formulation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

2.3 The projections of E and H field locations are offset by ∆r/2 on the 1D grid. . . . . 17

2.4 A 2D view of the projections of E and H locations showing the uniformity of the

1D grid. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

2.5 Left: A set of Ex , Ey and Ez components having same projection on the 1D grid.

Only a part of FDTD grid with E component locations is shown. Right: A set of

Hx , Hy and Hz components having same projection on the 1D grid. Only a part of

FDTD grid with H component locations is shown. . . . . . . . . . . . . . . . . . . . 18

2.6 Left: A 3D view showing projections of all E and H locations. Right: Another 3D

view that shows several E projections line up (similarly for H) forming a uniformly

1D grid. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

2.7 Computational voxel for FV24 algorithm vis-a-vis conventional FDTD unit cell. . . . 20
 xiv

2.8 Left: Total Field Ex (on Y planes) for which the whole H-plane (containing nine Hz

components) fall in scattered field region. Right: Total Field Ex (that belong to X

plane nodes or fall on Z planes of the TF/SF surface) for which only the three of

the nine components on H-plane fall in scattered field region. Also included are the

special cases at the corners. Respective incident H fields need to be added to these

H fields that fall in scattered field region. . . . . . . . . . . . . . . . . . . . . . . . . 24

2.9 Left: 3D view of the Y planes, with blue and red being part of SF and TF respectively.

Right: 2D view showing Scattered Field Ex (on Y plane) for which the whole H-

plane or part of it fall in total field region. Respective incident H fields need to be

subtracted from these H fields. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

2.10 Left: Scattered Field Ex (on Z planes) for which part of the H-plane fall in total

field region. Respective incident H fields need to be subtracted from these H fields.

Right: Scattered Field Ex (on X planes) for which part of the H-plane fall in total

field region. Respective incident Hz fields need to be subtracted from these H fields.

Also included are some special cases at the corners. . . . . . . . . . . . . . . . . . . . 26

2.11 Scattered Field Ex (diagonally outside the TF along the Y axis) for which part of the

H-plane fall in total field region. Respective incident H fields need to be subtracted

from these H fields. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

2.12 Top: Ex on 1D grid after 300 time-steps Bottom: Hz on 1D grid after 300 time-steps 31

2.13 Top: Incident field Ex in TF on a Z plane at the center of TF after 300 time-steps.

Bottom: Incident field Hz in TF on a Z plane at the center of TF after 300 time-steps. 32

2.14 Top: Leakage of Ex into SF on a Z plane at the center of TF after 300 time-steps.

Bottom: Leakage of Hz into SF on a Z plane at the center of TF after 300 time-steps. 33

2.15 Schematic of the FDTD problem space, showing the dielectric scatterer, TF/SF

regions and the Equivalent surface. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

2.16 The tangential electric and magnetic fields on two separate surfaces, red and blue. . 34
 xv

2.17 Left: Average of two Electric field components highlighted red. Right: Average of

four Magnetic field components highlighted red. . . . . . . . . . . . . . . . . . . . . . 35

2.18 Left: Magnetic field-produced electric currents are placed on Se . Right: Electric

field-produced magnetic currents are placed on Sh . . . . . . . . . . . . . . . . . . . 36

2.19 Maximum farfield error in dB for different grid resolutions. . . . . . . . . . . . . . . 39

2.20 Average farfield error in dB for different grid resolutions. . . . . . . . . . . . . . . . . 40

2.21 Maximum farfield error in dB for different plane wave incident angles. . . . . . . . . 40

2.22 Average farfield error in dB for different plane wave incident angles. . . . . . . . . . 41

2.23 Maximum farfield error in dB for various equivalent surface sizes. . . . . . . . . . . . 41

2.24 Average farfield error in dB for various equivalent surface sizes. . . . . . . . . . . . . 42

2.25 The θ component of bistatic RCS in different farfield planes. FDTD resolution is 40

cells/wavelength at 1 GHz. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

3.1 Stack-up of a fourteen layer PCB [3], as an example. . . . . . . . . . . . . . . . . . . 46

3.2 Different glass weave styles and their style numbers. . . . . . . . . . . . . . . . . . . 46

3.3 A CAD model showing woven glass fiber and resin composite material. Also shown

are glass-rich and resin-rich areas and a microstrip trace . . . . . . . . . . . . . . . 47

3.4 A model showing two traces of a microstrip differential pair, one that falls entirely

on glass fiber bundle and one that glass in between glass fiber bundles. . . . . . . . . 48

3.5 Voltage waveforms show slight glass-weave induced timing skew on a differential a

pair. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

3.6 Glass fiber face modeled as an ellipse in x − y plane as shown above and is swept

along the z-axis modulating the y-coordinate of the center of the ellipse as shown

below. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

3.7 Various materials assigned to the cells in the FDTD grid modeling a PCB . . . . . . 52

3.8 Multi-pole Debye model curve fitted to Djordjevic-Sarkar Modem for Glass and Resin

materials. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
 xvi

3.9 Process flow showing sequence of steps involved in implementing dispersive FDTD. . 58

3.10 Pulse propagation on the differential pair obtained from FDTD simulation, and the

timing skew between them. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

3.11 Top view of PCB Models for HFSS (left) and FDTD (right). . . . . . . . . . . . . . 60

3.12 Conversion of S-parameter data to time domain. . . . . . . . . . . . . . . . . . . . . 60

3.13 Pulse propagation on the differential pair obtained from applying inverse Fourier

transform to HFSS simulation results, and the timing skew between them. . . . . . . 60

4.1 Spherical FDTD unit cell showing electric and magnetic field components’ locations. 63

4.2 Spherical FDTD unit cell showing positional offsets of electric and magnetic field

components. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

4.3 Edge lengths and face areas in spherical FDTD grid . . . . . . . . . . . . . . . . . . 67

4.4 A section of spherical FDTD grid showing discretization in r, θ, and φ directions. . . 67

4.5 Special cells in spherical FDTD at poles and origin, also shown is a standard six-faced

spherical FDTD unit cell. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

4.6 Numerical dispersion relation derivation process in general and steps involved in the

derivation for Cartesian FDTD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

4.7 Real and imaginary parts of wave number vs. normalized distance from origin at

different resolutions (θ = ∆θ, n = 3, and m = 0). . . . . . . . . . . . . . . . . . . . 74

4.8 Real and imaginary parts of wave number vs. normalized distance from origin for

different mode numbers (θ = ∆θ, R= 40, and m = 0). . . . . . . . . . . . . . . . . . 75

4.9 Real and imaginary parts of wave number vs. normalized distance from origin for

different mode numbers m. Here θ = ∆θ, R= 40, and n = 4 . . . . . . . . . . . . . 76

4.10 Real and imaginary parts of wave number vs. normalized distance from origin at

different elevation angles (R = 40, n = 3, and m = 1). . . . . . . . . . . . . . . . . . 77

4.11 Problem space that excludes regions around origin to observe stability criterion prac-

tically for spherical FDTD. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80

 xvii

4.12 Practically observed stability criterion for different PEC sphere radii and for different

∆θ. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

4.13 Theoretical stability factor τ at various PEC sphere radii for different modes n, m =

0. Also shown is the observed τ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

5.1 PML boundary condition surrounding problem space in spherical FDTD. . . . . . . 85

5.2 Spherical FDTD unit cell showing the field locations and their offsets. . . . . . . . . 93

5.3 The position of PML region and PEC wall relative to origin. . . . . . . . . . . . . . 97

5.4 Variation of conductivity σr in the PML region for different orders nσ . . . . . . . . 99

5.5 Problem Space for Spherical FDTD with PML . . . . . . . . . . . . . . . . . . . . . 100

5.6 Problem space for reference simulation in spherical FDTD . . . . . . . . . . . . . . 101

5.7 Comparison of Refection Error for Split-Field PML and Absorbing Shell ABC for

nσ = 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102

5.8 Comparison of Reflection Error for stretched-coordinate PML for various PML poly-

nomial orders nσ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

5.9 Reflection Error for different mode numbers n and m = 0 in Continuous and FDTD

PML cases. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

 Chapter 1

Introduction: Review of Finite Difference Time Domain Analysis

This chapter introduces the finite difference time domain method and various tools associated

with it that are used in later chapters. The finite difference time domain (FDTD) method intro-

duced by Yee in [4] has been applied to a variety of electromagentic wave (EM) problems. Numerous

tools such as perfectly matched layer (PML) [5–8], auxiliary differential equation method [9, 10],

plane wave excitation techniques [11–13], near to farfiled transformation [1, 14, 15], to name a few,

have been developed to apply to various kinds of problems.

1.1 Conventional FDTD

The finite difference time domain method is an explicit formulation, i.e., there is no need for

matrix inversions to solve a set of linear equations. In conventional FDTD, time domain Maxwell’s

curl equations are discretized in both space and time using central difference scheme, producing

second-order accurate update equations for electric and magnetic field components.

The FDTD unit cell in figure 1.1 [1] shows the positions of electric and magnetic field com-

ponents. The update equations for Ex and Hx components are given in equations 1.1,1.2 [1, 12].

The update equations for Ey,z are obtained from the update equation 1.1 for Ex by replacing x

with y, y with z, and z with x.

The update equations for Hy,z are obtained in a similar fashion from the update equation
 2

Figure 1.1: Standard FDTD unit cell showing electric and magnetic field component locations [1].

1.2 for Hx . The complete FDTD update equation set can be found in [1, 12].

1 1
Exn+1 (i + , j, k) = Exn (i + , j, k)
2 2 
∆t n+ 1 1 1 n+ 12 1 1
+ Hz 2 (i + , j + , k) − Hz (i + , j − , k)
∆y 2 2 2 2
 
∆t n+ 1 1 1 1
n+ 2 1 1
− Hy 2 (i + , j, k + ) − Hy (i + , j, k − ) (1.1)
∆z 2 2 2 2

n+ 21 1 1 n− 1 1 1
Hx (i, j + , k + ) = Hx 2 (i, j + , k + )
2 2  2 2 
∆t n 1 n 1
+ Ey (i, j + , k + 1) − Ey (i, j + , k)
µ∆z 2 2
 
∆t 1 1
− Ezn (i, j + 1, k + ) − Ezn (i, j, k + ) (1.2)
µ∆y 2 2

1.2 Numerical Dispersion

Because of the disctretization in space and time, the FDTD method doesn’t produce the

same dispersion relation as that of free space, except for a special case in one dimension [1]. It

implies that different frequencies will propagate with different velocities in the FDTD grid. The

numerical dispersion relation for FDTD method is obtained by substituting plane wave solution in
 3

the update equations and reducing the resultant equations [1]. It is shown in equation 1.3.
!2 !2 !2
sin ω∆t/2 2
 
sin β̃x ∆x/2 sin β̃y ∆y/2 sin β̃z ∆z/2
µ = + + , (1.3)
∆t/2 ∆x/2 ∆y/2 ∆z/2

where, ∆x, ∆y, ∆z and ∆t are spatial and temporal discretization steps, respectively. ω is the

angular frequency, and β̃ is the numerical wave number. This relation also implies that the phase

velocity of numerical waves in FDTD grid is direction-dependent (anisotropic) [12].

The numerical dispersion relation can be solved for numerical wave number β̃. At low grid

resolutions the numerical wave number in FDTD can be complex [12, 16] even for free space. The

frequency spectrum of a wave being propagated in the FDTD grid can be wide containing low as well

as high frequency components. For a given grid resolution, this might mean that the high frequency

components are poorly resolved. This results in attenuation and faster-than-light (superluminal)

propagation of these high frequency components [12].

Chapter 4 of this work deals with the derivation of numerical dispersion relation and analysis

of numerical wave number for spherical FDTD (FDTD for spherical coordinates).

1.3 Field Sources in FDTD Grids

Sources are important in FDTD because they initiate the waves in the grid. Depending on

the application, there are different kinds of sources we might want to implement in the FDTD grid,

such as point-wise field sources, voltage and current sources, plane waves etc.

Pointwise field sources are the simplest to implement and are treated in [12, 17]. There are

two flavors of field sources: a hard source and a soft source. The hard source is implemented

by hard coding the field value at a point in the grid. For example, Exn |(i,j,k) = f (n∆t) fixes

the value of electric field component Ex to a particular value at every time step (at every in-

stant n∆t) based on the choice of the source function f (t). As such, this component’s value at

that location is not influenced by other field components surrounding it. As a consequence of

this, hard source will reflect the wave coming towards it similar to a perfect electric conductor

(PEC). On the other hand soft source doesn’t reflect and is transparent. It is implemented as
 4

Exn |(i,j,k) = usual update equation + f (n∆t), and the value of Exn |(i,j,k) is influenced by other

field components surrounding it because of the usual update equation involved. Comparing the

soft source implementation to Maxwell’s curl equation with a current source, ∇ × H =  ∂E

∂t + J, it

is observed that the soft source is not actually a field source, but a current source.

The voltage and current sources can be implemented in the FDTD grid to excite a wave on

a microstrip transmission line, for example. They can be lumped or distributed over a plane or

a volume, and are dealt with in [1, 18, 19]. For voltage source implementation, the coefficients of

the update equations are modified to include the internal resistance of the voltage source. And

the voltage source term is added to the update equation as shown in equation 7b in [18]. Current

sources are implemented in a similar fashion.

Plane wave sources, as the name implies, excite a plane wave in the FDTD grid and can be

used to irradiate a scatterer to study its radar cross section, for example. Orignially implemented by

hard sourcing fields in the entire grid as a inital value problem by Yee [4], this plane wave excitation

has serious limitations [12]. Plane wave sourcing based on total field/scattered field formulation

or pure scttered field formulation, which overcome these limitations, are discussed in [1, 11–13, 20].

Of these two, the total field/scattered field (TF/SF) formulation provides an excellent dynamic

range for total fields as the scattered field formulation suffers form a phenomena called subtraction

noise [12].

There are various techniques to calculate the incident fields required to implement TF/SF

formulation. Incident field array (IFA) technique [12] is based on auxiliary one dimensional grid

and needs interpolation to calculate incident fields in the main grid. This increases field leakage

into scattered field region and results in errors in scattering behavior [12]. One other technique

to calculate incident fields is the auxiliary field propagator (AFP) technique [20]. While this

method helps reduce spurious leakage into scattered field region, it needs expensive Fourier and

inverse Fourier transforms to calculate incident fields at the TF/SF interface [12]. Moreover, it

produces noncausal scattered field behaviour [12]. While this technique in its original form is

not based on auxiliary 1D grid, its optimized version O-AFP [21] is based on auxiliary 1D grid.
 5

Discrete planewave (DPW) technique [22] based on 1D-MAP [13] methodology provides efficiency

in calculation of incident fields and more than 300 dB of isolation between the total and scattered

fields, thus drastically reducing the spurious leakage to machine precision levels.

Chapter 2 deals with implementation of the Discrete planewave technique for FV24, a finite-

volumes based higher-order FDTD [23].

1.4 Higher-Order FDTD: The FV24 Method

Conventional or standard FDTD is second-order accurate method in both time and space.

Many higher-order schemes are available in the literature for FDTD [24–27] either to control phase

errors caused by numerical dispersion, or for better higher-order accuracy, or both [23]. Modified

FDTD (2,4) or M24 [28] was introduced to model two-dimensional electrically large problems with

high phase accuracy. Its three dimensional counterpart is the finite-volumes based higher-order

FDTD called FV24 [23]. They both are second order in time and fourth order in space algorithms.

The computational voxel or stencil for FV24 algorithm is shown in figure 1.2 [23]. While the

stencil shown in the figure is for updating Ex , the other five stencils are similar. In addition to

the four field components conventional FDTD considers, FV24 includes nine field components on

each of the four parallel faces of a cube of size 3h × 3h × 3h as shown in figure 1.2. The cube is

centered at the location of the field component that is being updated (Ex in this case). The nine

new field components are separated into three categorizes: b for axial nodes, c for surface axial

nodes, and d for surface diagonal nodes [23]. The contributions from these three categories are

weighted differently in the update equations to account for their different radial distances. And

these weights are optimized to minimize the phase errors [23].

Chapter 2 of this work deals with the plane wave excitation method for FV24.
 6

FV24
Voxel

+
h

b – axial H nodes 3h
c – surface axial H nodes
d – surface diagonal H nodes

Figure 1.2: Computational voxel for FV24 algorithm vis-a-vis conventional FDTD unit cell.

Source: M. F. Hadi, “A finite volumes-based 3-D low dispersion FDTD algorithm,”

1.5 on Antennas
IEEE Transactions Gridand Truncation: Perfectly
Propagation, vol. 55, no. 8, pp.Matched
2287–2293, Layer and Convolutional Perfectly 1
Aug. 2007.
Matched Layer

Effective truncation of FDTD grid without spurious reflections to simulate free space is of

importance in open-region problems such as calculation of antenna radiation pattern or radar cross

section (RCS) of a sctterer etc. In genreal, grid truncation can be classified in to two classes:

radiation boundary conditions (RBCs) and absorbing boundary conditions (ABCs). RBCs are

based on analytical operators and many of them were introduced in the EM literature to emulate

infinite free space. For a review of these methods, one can refer to [12]. They estimate the field value

on the boundary using the known field values near the boundary based on differential operator (for

example Sommerfeld radiation operator). While some advanced methods in this class can provide

reflection errors down to -100 dB [12], they are suitable only to certain problems. The other methods

in this class might not as effective in grid truncation as they suffer from spurious reflections.

The other class is the Absorbing boundary conditions (ABCs). An absorbing medium sur-

rounds the main problem space in the directions where free space is required in a similar fashion

to the absorbing material of an anechoic chamber, although physical taperings are often absent in

numerical implementation. Early attempts as in [29] used a lossy, impedance-matched medium to

 7

absorb waves and truncate the FDTD grid. However, it has limited application as the absorption

is not effective for non-normal incident angle cases [12]. A more comprehensive ABC called the

perfectly matched layer (PML), based on exploiting additional possibilities created by splitting the

field components in Maxwell’s equation, was intoduced by Berenger [5]. It is shown to be very

effective in absorbing the waves in FDTD grid and other numerical techniques as well.

In continuous space, PML is reflection less, however the PML in FDTD is of finite thickness

and is discretized, i.e., it is implemented as a multi-layered medium. This often results in spurious

reflections. So, the PML parameters are geometrically graded to avoid reflections [12]. Later,

PML has been re-interpreted as an uniaxial medium (UPML) by Gedney [6] and in stretched-

coordinate form independently by Chew-Weedon [30] and Rappaport [31]. Complex frequency

shifted PML (CFS-PML) [32, 33] was introduced to effectively absorb the evanescent modes that

have long time interactions with PML. Its efficient implementation based on stretched-coordinate

formulation called the convolutional PML (CPML) was introduced in [8]. Today, CPML stands as

the effective way among ABCs to truncate FDTD.

Chapter 5 of this work deals with the implementation of PML for spherical coordinate

FDTD.

1.6 Organization of Rest of the Chapters

This thesis is composed of four themes discussed in each of the next four chapters.

Chapter 2: The first theme deals with the development of a plane wave excitation formu-

lation in FV24, a finite-volume based higher-order FDTD variant. With its excellent phase error

performance even for coarse grid, FV24 can be applied to electrically large problems. The plane

wave excitation method based on total field/scattered field formulation and Discrete Planewave

technique is demonstrated and validated for FV24. In the latter part of this chapter different Near

to Farfield transformation approaches possible in FDTD are compared for accuracy.

Chapter 3: The second part of the research deals with application of FDTD to an interest-

ing problem encountered in printed circuit boards. The glass weave-induced skew in a differential
 8

transmission line can cause a variety of problems as in increased bit error rates affecting the ro-

bustness of the digital system, and increased radiated emissions causing compliance failures etc. In

this chapter, the auxiliary differential equation formulation for FDTD, that helps model frequency

dependence of material properties of real materials, is used to estimate glass weave-induced skew in

a differential pair. The skew numbers are benchmarked with the available commercial solvers. Also,

the use of graphical processing units (GPUs) to accelerate the skew simulations is demonstrated.

Chapter 4: The third part of the research deals with the derivation of numerical dispersion

relation for spherical FDTD, and the sensitivity study of the associated numerical wave number.

Elementary functions native to spherical coordinates are used in the derivation of the numerical

dispersion relation in contrast to plane waves used in the literature. Given the non-uniform nature

of the spherical FDTD grid, the dispersion relation and the corresponding numerical wave number

are expected and are shown in this work to be position dependent. The latter part of this chapter

includes a study to derive the stability criterion for spherical FDTD and challenges involved therein.

Chapter 5: The final part of the research demonstrates the effectiveness of different absorbing

boundary condition formulations for spherical FDTD in absorbing the waves without spurious

reflections. It is shown that the split-field formulation of perfectly matched layer (PML) is not

as effective in truncating the spherical FDTD grid as its stretched-coordinate formulation. This

chapter includes derivation of continuous-space PML reflection coefficient, update equations for

implementation of stretched-coordinate PML in spherical FDTD and analysis of reflection error for

different PML parameters.

 Chapter 2

Planewave Injection in Higher-Order FDTD Grid

Many electromagnetic wave (EM) applications involve a plane wave impinging on an object

that can be either a receiving antenna or scatterers such as aircraft and ships in radar applications.

Initiating a plane wave in a FDTD grid in order to simulate these real-life situations is of interest to

the applied EM community, for example to estimate radar cross section (RCS) in radar applications

or plot the radiation pattern of an antenna.

Different techniques exist for injecting this plane wave in a traditional FDTD grid, however

the error-free and efficient technique has been the Discrete Planewave formulation [13, 22].

In this chapter, we will show how this technique, which uses the total field-scattered field

formulation to implement plane wave incidence, is implemented for a 3D finite volumes-based,

extended-stencil FDTD (FV24) algorithm [23]. Its 2D verion is M24 discussed in [28]. The FV24

method is based on the second-order in time and fourth-order in space finite-difference scheme

whereas the conventional FDTD is obtained from second-order finite differences in both time and

space. This method is designed to counter the effect of relatively large phase errors in propagated

waves when a coarse grid is used [23] in modeling electrically large problems spanning hundreds of

wavelengths. This work has also been published and available in [2]

Other tools available for FV24 are the convolutional perfectly matched layer (CPML) ab-

sorbing boundary conditions [34], conformal perfect electric conductor (PEC) modeling [35], and

graphical processing units (GPU) code optimization [36].

Using the Discrete Planewave formulation, we can gain computational efficiency when the
 10

three-dimensional FV24 grid is converted, exploiting the plane wave properties, into an auxiliary

one-dimensional grid along the direction of plane wave propagation. The field leakage errors are

eliminated through the dispersion-match between the one and three dimensional grids. As a conse-

quence, the leakage errors into the scattered-field region are -300 dB below the incident field values

and are observed to be independent of angle of incidence. This is a limitation set by the machine

precision itself, not by the accuracy of the method.

Time-stepping along the aforementioned one-dimensional grid and the three-dimensional

FV24 grid are run in parallel. The consistency corrections, needed for the update equations at

the total field/scattered field boundary, are made using the field values from the one-dimensional

grid. In contrast with the consistency corrections needed for plane wave incidence in traditional

FDTD grid, more corrections are needed in FV24 grid due to its extended stencil. The compli-

cated mapping between the corresponding field locations on the one-dimensional grid and the main

three-dimensional grid will be demonstrated.

In the later part of the chapter, different interpolation schemes used in the process of trans-

forming nearfields to farfields (near to farfield transformation) are compared.

2.1 Planewave Excitation in Standard FDTD

Planewave excitation technique in conventional FDTD based on an auxiliary one dimensinoal

(1D) grid and total field/scattered field (TF/SF) formulaiton, as proposed by Tan and Potter [13]

[22] is reviewed first, and later will be extended to FV24 algorithm.

2.1.1 Total Field/Scattered Field Formulation

TF/SF formulation [12] isolates the problem space into a total field and scattered field regions

as shown in figure 2.1. The scattering object is placed inside the total field region and the virtual

surface between the total field and scattered field regions is used to place fictitious electric and

magnetic surface currents. These currents excite the required plane wave within the total field

region, while allowing only scattered waves from the scatterer to pass through it. It is one of the
 11

many flavors of the equivalence principle.

Equivalent Currents - Jump Conditions

Js & Ms

Js

Ms
Total Field Region

Scattered Field Region

Total Field = Incident Field + Scattered Field

Figure 2.1: TF/SF formulation showing scatterer, total and scattered field regions, and interface
surface between them [2].

2.1.2 The Discrete Planewave Technique

The Discete Planewave method [13, 22], which uses a 1D auxiliary grid that inherently prop-

agates a plane wave with an identical numerical behavior as that of the main three dimensional

grid, is the most effective TF/SF variant to date. The idea of auxiliary 1D in relation to TF/SF

interface is shown in figure 2.2. The plane wave fields propagated on the 1D grid are mapped back

to the main 3D grid and are used as fictitous surface currents on the TF/SF interface. This results

in virtually no leakage of non-scattered fields outside the TF/SF interface -i.e., in scattered field

region [2].

2.1.3 Auxiliary One-Dimensional Grid

The central theme behind implementing an error-free plane wave incidence using TF/SF

formulation is the perfect numerical dispersion-matched 1D auxiliary propagator. In other words,

the dispersion matching between the auxiliary 1D grid and the main FDTD grid makes it possible

to reduce the spurious field leakage into the scattered field region to the order 10−15 or -300 dB

below the field levels in total field region. The idea is to convert the same update equations as
 12

Figure 2.2: Auxiliary 1D grid, used in Discrete Planewave technique, in relation to the TF/SF
formulation.

that of the main 3D grid into 1D form for propagating the plane wave. This 1D implementation,

as it becomes evident later in this section, is much more efficient than the 3D counterpart. For a

3D problem space, the FDTD update equations (consider equation 2.2, for example) operate on

three dimensional E and H arrays. However, when we deal with a uniform plane wave, the same

update equations can be modified so that they now involve operations on E and H that are just

one dimensional arrays. This section will explain how this conversion (transformation or mapping)

from 3D to 1D is done.

Every location, where either E or H field component is located in the main 3D FDTD grid,

is projected onto an imaginary line in the direction of plane wave propagation. It is convenient to

visualize this imaginary line as passing through the lower left corner of the TF region. The unit

vector along this direction, in terms of the azimuth and elevation angles, is P = cos φ sin θax +

sin φ sin θay + cos θaz or P = px ax + py ay + pz az .

 13

As an example to demonstrate the projection mechanism, consider X = ((i+ 12 )∆x, j∆y, k∆z),

the actual location of Ex associated with the node (i∆x, j∆y, k∆z). The projection of this loca-
px ∆x py ∆y
tion along the imaginary line is P · X = px ∆x(i + 21 ) + py ∆yj + pz ∆zk. If mx = my =
pz ∆z
mz = ∆r (the rational angle condition [37]), for some odd integers mx , my and mz , then
mx
P · X = (mx i + my j + mz k + 2 )∆r. It means that this location’s projection is d = P · X

units from origin (the lower left corner of the TF region) along the imaginary line. Observe that

the choice of mx , my and mz will decide the direction of plane wave propagation through the

relations in equation 2.1, and theoretically infinite such directions are possible.

my mz
φ = tan−1 , θ = cos−1 q (2.1)
mx m2x + m2y + m2z

The collection of all the projections, of every Ex,y,z and Hx,y,z component location of the 3D

main grid, forms the whole 1D grid. The projection’s distance from origin of 1D grid in terms of
mx
∆r is (mx i + my j + mz k + 2 ), for example in the above case. This coefficient of ∆r is generally

a real number, however, it can be converted (as explained later) into an integer that can act as an

index for the 1D electric and magnetic field arrays in computer memory.

Individual 1D grids can be formed by taking the projections of Ex locations alone (also

applicable for Ey,z and Hx,y,z ) as shown in [22]. However, the term 1D grid refers to the whole

1D grid, formed from projections of all Hx,y,z and Ex,y,z locations on the 3D grid, in the rest of

the paper unless pointed otherwise as individual. Although, we maintain six different 1D arrays

for each of the six electric and magnetic field components that store the field values on respective

individual 1D grids.

According to the property of the uniform plane wave that all the locations on a wave front

have same phases and magnitudes, the projection of every point of the wave front onto the imaginary

line is same (as the wave front is perpendicular to the direction of propagation). In other words,

the Ex locations that have the same projection will have same field values at every instant. The

same applies to Ey,z or Hx,y,z . This property makes propagating a plane wave on 1D grid very

attractive because of high memory and time efficiency [22] since we don’t need to store and perform
 14

same update operations on exactly same field values (duplicates) repeatedly.

The conventional FDTD (standard second order in space and time or S22) update equation for

Ex in a loss-less homogeneous and isotropic medium, showing the exact field locations (half integer

offsets), is given by equation 2.2. This equation operates on 3D arrays and can be converted into

an update equation containing only 1D arrays (the 1D form) using the projections of the exact field

locations in the 3D update equations.

1 1
Exn+1 (i + , j, k) = Exn (i + , j, k)
2 2 
∆t n+ 1 1 1 n+ 12 1 1
+ Hz 2 (i + , j + , k) − Hz (i + , j − , k)
∆y 2 2 2 2
 
∆t n+ 1 1 1 n+ 1 1 1
− Hy 2 (i + , j, k + ) − Hy 2 (i + , j, k − ) (2.2)
∆z 2 2 2 2

However, as seen earlier, the projections are generally not integer multiples of ∆r making it

difficult to use them directly as an index for 1D arrays in computer memory. If mx , my and mz
1
are all odd integers, this problem can be rectified by simply adding 2 to the non-integer coefficient

of ∆r of the projection, i.e., the index on 1D grid of Ex at location ((i + 21 )∆x, j∆y, k∆z) on the
mx +1
3D grid is (mx i + my j + mz k + 2 ), for example.
1
However, care must be taken to avoid adding 2 if a coefficient turns out to be integer by

itself because of two fractions adding up. This happens for H components. For example, consider

Hz at (i + 12 , j ± 21 , k) or Hy at (i + 21 , j, k ± 12 ) in equation 2.2, whose coefficients of projection

my
distances are d
∆r = I = (mx i + my j + mz k + m2x ± 2 ) and d
∆r = I = (mx i + my j + mz k + m2x ± m2z )

respectively. These are already integers and can be used as index directly, i.e. there is no need to
1
add corrective 2 to coefficients of H component projections to extract a valid index i.

However, the above seemingly simple index extraction process for obtaining update equations

in 1D form becomes complicated in practice, as it turns out that the only reference algorithm has,

to choose right-hand-side Hy,z components on the 1D grid, is the index of Ex , for example. So,

we need to express the positions of Hy,z components relative to this index (of Ex ). For example,
my
the projection of Hz at (i + 21 , j + 21 , k) is ( 2 )∆r units away from the projection of Ex on 1D

grid. This relative distance, which is not an integer coefficient of ∆r, is first converted into relative
 15

my +1 d mx +1
index 2 . Adding this to the corrected coefficient ∆r = (mx i + my j + mz k + 2 ) of Ex

would have given the absolute index for Hz . However, it doesn’t for the subtlety that we have

over-estimated the index by correcting twice –the two halves add up and pick H at an unintended

position. This can be rectified by adding −1 to compensate, so the relative index for this Hz is
mx +1 my +1 mx +1 my −1
i = (mx i + my j + mz k + 2 + 2 − 1) or i = (mx i + my j + mz k + 2 + 2 ).

Observe that we indeed would arrive at the same value directly from the projection of the

location (i+ 12 , j+ 12 , k) of this Hz , and can be used as consistency check (we would know this location

beforehand by picturing the Yee grid in mind or by looking at the update equation, however, the

algorithm can only deduce this by taking the location of Ex as reference and hence the rotund
my
procedure). The projection of Hz at (i + 12 , j − 21 , k) is also ( 2 )∆r away from the projection of Hx

on 1D grid, however, on the opposite side on 1D grid. An integer index is obtained by subtracting
my +1 mx +1 my +1
2 from the reference index of Ex , i.e., I = ((mx i + my j + mz k + 2 − 2 ). In this case

as the two halves cancel rather than adding up, there is no need to compensate. Similar procedure

can be followed to obtain Hy locations.

Table 2.1: Indexes for field location of S22 Ex update equation

Field Location Index

1 mx +1 mx +1
Ex (i + 2 , j, k) mx i + my j + my k + 2 =K+ 2 = ref
1 1 mx +1 my −1 my −1
Hz (i + 2 , j + 2 , k) K+ 2 + 2 = ref + 2
1 1 mx +1 my +1 my +1
Hz (i + 2 , j − 2 , k) K+ 2 − 2 = ref − 2
1 1 mx +1 mz −1 mz −1
Hy (i + 2 , j, k + 2 ) K+ 2 + 2 = ref + 2
1 1 mx +1 mz +1 mz +1
Hy (i + 2 , j, k − 2 ) K+ 2 − 2 = ref − 2

The same approach is followed in order to convert the update equation 2.3 for Hx , showing

exact locations of the field components involved, into 1D from. The projection of Hx at (i, j + 12 , k +
1 d my mz
2) does not need any correction since ∆r = (mx i + my j + mz k + 2 + 2 ) is already an integer

and used as index directly. Again, as the only reference that the algorithm has to choose Ey,z

components on the 1D grid is the index of Hx , we need to express the positions of Ey,z components

relative to this index. For example, the projection of Ey at (i, j + 12 , k + 1) is ( m2z )∆r away from
 16

the projection of Hx on 1D grid. This relative distance, which is not an integer coefficient of ∆r, is
mz +1 my
converted into relative index 2 . Adding this to the coefficient d
∆r = (mx i+my j +mz k+ 2 + m2z )
my mz mz +1
of Hx ’s projection, we get the index for this Ey as i = (mx i + my j + mz k + 2 + 2 + 2 ).

The projection of Ey at (i, j + 12 , k) is also ( m2z )∆r away from the projection of Hx on 1D grid,
mz +1
however, on the opposite side on 1D grid. Subtracting 2 from the index of Hx would result in a
my
integer (mx i + my j + mz k + 2 + m2z − mz2+1 ). However, it would become evident after consistency

check that we under-estimated the index and so a compensatory 1 is added. Finally, the correct
my
integer index for this Ey , i.e. I = (mx i + my j + mz k + 2 + m2z − mz2−1 ). The summary of indexes

for all field components in S22 update equations for Ex and Hx are listed in tables 2.1 and 2.2,

respectively.

n+ 12 1 1 n− 1 1 1
Hx (i, j + , k + ) = Hx 2 (i, j + , k + )
2 2  2 2 
∆t 1 1
+ Ey (i, j + , k + 1) − Eyn (i, j + , k)
n
µ∆y 2 2
 
∆t n 1 n 1
− Ez (i, j + 1, k + ) − Ez (i, j, k + ) (2.3)
µ∆z 2 2

Table 2.2: Indexes for field location of S22 Hx update equation

Field Location Index

my mz my mz
Hx (i, j + 21 , k + 12 ) mx i + my j + my k + 2 + 2 =K + 2 + 2 = ref
my mz mz +1
Ey (i, j + 1
2, k + 1) K+ 2 + 2 + 2 = ref + mz2+1
my mz −1
Ey (i, j + 1
2 , k) K+ 2 + mz
2 − 2 = ref − mz2−1
1 my mz my +1 m +1
Ez (i, j + 1, k + 2) K+ 2 + 2 + 2 = ref + y2
my mz my −1 m −1
Ez (i, j, k + 21 ) K+ 2 + 2 − 2 = ref − y2

2.1.4 Properties of 1D Grid

If all the space is meshed (i.e. not confining ourselves to the truncated region of interest) and

projections of all possible 3D locations of E and H components are taken on the imaginary line

along the direction of plane wave propagation, and are sorted (in ascending order) after eliminating

the duplicates, we will have a uniformly spaced discrete set of points, the so called 1D grid, along the
 17

1 3 5 7 9
Δ𝑟 Δ𝑟 Δ𝑟 Δ𝑟 Δ𝑟
2 2 2 2 2

0Δ𝑟 1Δ𝑟 2Δ𝑟 3Δ𝑟 4Δ𝑟 5Δ𝑟

Ex,y,z Projections Hx,y,z Projections

Figure 2.3: The projections of E and H field locations are offset by ∆r/2 on the 1D grid.

imaginary line. Because of the inherent spatial offset (staggered in space) of E and H components

in the FDTD grid (it is assumed that the E components are edge centered while H components
mx,y,z ∆r
are face centered), the projections of E components are d = (mx i + my j + mz k)∆r + 2 (non-
(my +mz )∆r
integer multiple of ∆r) and of H components will be either d = (mx i + my j + mz k)∆r + 2 ,
(mx +mz )∆r (mx +my )∆r
or d = (mx i + my j + mz k)∆r + 2 or d = (mx i + my j + mz k)∆r + 2 (integer

multiple of ∆r).

Let the projection of H component associated with some node (i1 , j1 , k1 ) be d = n∆r for

some integer n, and for a given combination of (mx , my , mz ). It is interesting to observe that
∆r
there is another node (i2 , j2 , k2 ) whose associated E component’s projection is d = n∆r + 2 .
∆r
The consequence of the observation is that the 1D grid is uniform and has a spacing of 2 with

alternate E and H component projections. This property of the 1D grid is shown in figure 2.3.

Figure 2.4 offers a 2D view of projections of E and H locations and shows the uniformity of the

1D grid.

Also, the electric field components Ex , Ey , or Ez can have same projection. For example,

consider mx = 5, my = 3 and mz = 1. This combination makes the the projections of (i+ 21 , j, k+2),

(i, j + 32 , k), and (i, j + 1, k + 32 ), the 3D locations of Ex , Ey and Ez components respectively, co-

located on the 1D grid as shown in figure 2.5. Similar observation can be made for H components

as well. Figure 2.6 offers a 3D view of projections of E and H locations. It also shows a 3D view

that displays how several E projections, displayed by solid lines, line up and fall at a set of locations

on the imaginary line. Similarly, several H projections, displayed by dotted lines, line up and fall
 18

Ey , Hx Hz Ey , Hx Hz

∆r/2
Ex , Hy
Ez

Ex,y,z Projections
Ey , Hx
Hz
H x,y,z Projections
Ey , Hx Hz
∆r/2

Ez

Ex , Hy Ez Ex , Hy

Figure 2.4: A 2D view of the projections of E and H locations showing the uniformity of the 1D
grid.

at another set of location on the imaginary line. These sets of projections form the 1D grid and

the E and H projections on it alternate.

Hz
H
E x
x
E
z

Hy

Ey

Figure 2.5: Left: A set of Ex , Ey and Ez components having same projection on the 1D grid.
Only a part of FDTD grid with E component locations is shown. Right: A set of Hx , Hy and Hz
components having same projection on the 1D grid. Only a part of FDTD grid with H component
locations is shown.

2.1.5 TF/SF Corrections

In FDTD grid, the surface currents are introduced at the TF/SF interface by means of

consistency corrections to the field update equations [12], and are detailed in section 2.2.3.
 19

Z
Y X

Figure 2.6: Left: A 3D view showing projections of all E and H locations. Right: Another 3D
view that shows several E projections line up (similarly for H) forming a uniformly 1D grid.

2.2 Planewave Excitation in FV24 Grid

In this section, we will demonstrate the plane wave excitation method discussed above for

FV24 algorithm.

2.2.1 FV24 Update Equations

The computational voxel (stencil) for FV24 algorithm is shown in figure 2.7 [23]. While the

stencil shown in figure 2.7 is for updating Ex , the other five stencils are similar. In addition to the

four field components conventional FDTD considers, FV24 includes nine field components on each

of the four parallel faces of cube of size 3h × 3h × 3h centered at the location of the field component

that is being updated (Ex in this case.)

The FV24 update equation for Ex is as follows:

1 1
Exn+1 (i + , j, k) = Exn (i + , j, k)
2  2 
n+ 1 1 1 n+ 1 1 1
+ cb1x Hz 2 (i + ,j + , k) − Hz 2 (i + ,j − , k)
2 2 2 2
 
n+ 1 1 3 n+ 1 1 3
+ cb2x Hz 2 (i + ,j + , k) − Hz 2 (i + ,j − , k)
2 2 2 2

n+ 1 3 3 n+ 1 1 3
+ cb3x Hz 2 (i + ,j + , k) + Hz 2 (i − ,j + , k)
2 2 2 2
 20

FV24
Voxel

+
h

b – axial H nodes 3h
c – surface axial H nodes
d – surface diagonal H nodes

Figure 2.7: Computational voxel for FV24 algorithm vis-a-vis conventional FDTD unit cell.

Source: M. F. Hadi, “A finite volumes-based 3-D low dispersion FDTD algorithm,”


n+ 1 1 3 n+ 1 1 3
IEEE Transactions on Antennas and Propagation, Hz55,2no.
+vol. + , k + 1) + Hz 2 (i + , j + , k − 1)
8, pp., j2287–2293,
(i +
2 2 2 2 1
Aug. 2007. 
n+ 21 3 3 n+ 12 1 3
− cb3x Hz (i + , j − , k) + Hz (i − , j − , k)
2 2 2 2

n+ 21 1 3 n+ 21 1 3
+ Hz (i + , j − , k + 1) + Hz (i + , j − , k − 1)
2 2 2 2

n+ 1 3 3 n+ 1 1 3
+ cb4x Hz 2 (i + , j + , k + 1) + Hz 2 (i − , j + , k + 1)
2 2 2 2

n+ 1 3 3 n+ 1 1 3
+ Hz 2 (i + , j + , k − 1) + Hz 2 (i − , j + , k − 1)
2 2 2 2

n+ 1 3 3 n+ 1 1 3
− cb4x Hz 2 (i + , j − , k + 1) + Hz 2 (i − , j − , k + 1)
2 2 2 2

n+ 21 3 3 n+ 21 1 3
+ Hz (i + , j − , k − 1) + Hz (i − , j − , k − 1)
2 2 2 2
 
n+ 1 1 1 n+ 1 1 1
− cb1x Hy 2 (i + , j, k + ) − Hz 2 (i + , j, k − )
2 2 2 2
 
1
n+ 2 1 3 1
n+ 2 1 3
− cb2x Hy (i + , j, k + ) − Hy (i + , j, k − )
2 2 2 2

n+ 1 1 3 n+ 1 1 3
− cb3x Hy 2 (i + , j + 1, k + ) + Hy 2 (i + , j − 1, k + )
2 2 2 2

n+ 1 3 3 n+ 1 1 3
+ Hy 2 (i + , j, k + ) + Hy 2 (i − , j, k + )
2 2 2 2

n+ 1 1 3 n+ 1 1 3
+ cb3x Hy 2 (i + , j + 1, k − ) + Hy 2 (i + , j − 1, k − )
2 2 2 2
 21

n+ 1 3 3 n+ 1 1 3
+ Hy 2 (i + , j, k − ) + Hy 2 (i − , j, k − )
2 2 2 2

n+ 1 3 3 n+ 1 1 3
− cb4x Hy 2 (i + , j + 1, k + ) + Hy 2 (i − , j + 1, k + )
2 2 2 2

n+ 1 3 3 n+ 1 1 3
+ Hy 2 (i + , j − 1, k + ) + Hy 2 (i − , j − 1, k + )
2 2 2 2

n+ 1 3 3 n+ 1 1 3
+ cb4x Hy 2 (i + , j + 1, k − ) + Hy 2 (i − , j + 1, k − )
2 2 2 2

n+ 12 3 3 n+ 12 1 3
+ Hy (i + , j − 1, k − ) + Hy (i − , j − 1, k − ) . (2.4)
2 2 2 2

2.2.2 One-Dimensional Update Equations for FV24

Similar to the conventional FDTD, integral based FV24 FDTD update equation can be

reduced to 1D form. Consider the FV24 update equation 2.4 for Ex , for which the indexes of some

field locations are listed in table 2.3. The indexes for rest of the field locations can be constructed

based on those listed in the table. Similarly, for the FV24 update equation of Hx , the indexes of

some field locations are listed in table 2.4.

Table 2.3: Indexes for some field location of FV24 Ex update equation

Field Location Index

mx +1 mx +1
Ex (i + 21 , j, k) mx i + my j + my k + 2 =K+ 2 = ref
1 3 mx +1 3my −1 3my −1
Hz (i + 2, j + 2 , k) K+ 2 + 2 = ref + 2
1 3 mx +1 3my +1 3my +1
Hz (i + 2, j − 2 , k) K+ 2 − 2 = ref − 2
3 3 mx +1 3my −1 3my −1
Hz (i + 2, j + 2, k + 1) K+ 2 + mx + 2 + mz = ref + mx + 2 + mz
1 3 mx +1 3my +1 3m +1
Hz (i − 2, j − 2, k − 1) K+ 2 − mx − 2 − mz = K + ref − mx − 2y − mz

2.2.3 TF/SF Corrections for FV24

The six E and H update equations converted to 1D form (that manifest the 1D propagator)

are run along with the time stepping of the main FV24 grid. That is, at each time step, we update E

and H at every location on the 1D grid first and then on the main grid later. The TF/SF corrections

are required for the consistency of main grid’s update equations at the Huygens surface which acts
 22
Table 2.4: Indexes for some field location of FV24 Hx update equation

Field Location Index

my mz my mz
Hx (i, j + 21 , k + 12 ) mx i + my j + my k + 2 + 2 = K + 2 + 2 = ref
my mz
Ey (i, j + 23 , k + 2) K+ 2 + 2 + my + 3m2z +1 = ref + my + 3m2z +1
my
Ey (i, j − 21 , k − 1) K+ 2 + mz
2 − my − 3m2z −1 = ref − my − 3m2z −1
my mz
Ey (i + 1, j + 32 , k + 2) K+ 2 + 2 + mx + my + 3m2z +1 = ref + mx + my + 3mz +1
2
my
Ey (i − 1, j − 12 , k − 2) K+ 2 + mz
2 − mx − my − 3m2z −1 = ref − mx − my − 3mz −1
2

as a boundary between the total field and the scattered field regions (the Huygens surface itself

is assumed to be part of total field region and contains some E and H field locations as was done

in [12]).

Consider the main grid’s update equation for an Ex falling on the boundary (the Huygens

surface that belongs to total field). This will have some of the right-hand side Hy and Hz falling

in total-field region and some in scattered field region making the update equation for this Ex

inconsistent. The H fields falling in the scattered field region need consistency corrections, i.e., we

need to convert them to total fields by adding the incident H fields at those respective locations as

shown in equation 2.5. The incident H fields are present on the 1D grid, more specifically, in the

three 1D arrays for Hx , Hy and Hz which have been computed already. Based on the location of

the main grid’s H field that needs correction, we construct the index by calculating its projection

on the 1D grid, then pick the H field value on the 1D grid at that index and add it to correct the

main grid’s H field.

n+ 1 n+ 12 n+ 12 n+ 12 n+ 12
Exn+1 |tot = Exn |tot ± {Hz,y 2 |tot } ± {Hz |scat + Hz |inc } ± {Hy |scat + Hy |inc } (2.5)

In contrast with the consistency corrections needed for plane wave incidence in S22 grid, we

need far more corrections here in FV24 grid. Consistency corrections are needed even for some Ex

fields falling in the scattered field region (the case not present for S22). For these locations, we

will have some of the right-hand side Hy and Hz falling in total-field region and some in scattered

field region. This will make the update equation for Ex inconsistent. The total H fields are to

be converted to scattered fields by subtracting the incident H fields in order to make the update
 23

equation consistent as show in equation 2.6.

n+ 1 n+ 12 n+ 12 n+ 12 n+ 12
Exn+1 |scat = Exn |scat ± {Hz,y 2 |scat } ± {Hz |tot − Hz |inc } ± {Hy |tot − Hy |inc } (2.6)

In the rest of the section we highlight the corrections to be made to the Hz fields of the Ex

update equation at locations where it becomes inconsistent. Similar correction methodology can be

applied for Hy components of Ex update equation and further to rest of the five update equations.

The six faces of TF/SF (Huygens) surface are i = i0 , i = i1 , j = j0 , j = j1 , k = k0 and k = k1

and they correspond to the terminating YZ, XZ and XY planes (or simply X, Y and Z planes)

respectively.

First, consider the corrections to the Ex that fall on/inside the total field (TF) region. The

Ex locations that are on the two Y planes, i.e. j = j0 and j = j1 , need similar corrections. This

is because, for each one of these Ex locations, one of the two H-planes that contains the nine Hz

components falls completely in the scattered field (SF) region. To be more precise, for each of the

Ex on j = j0 , the H-plane that is at j = j0 − 23 is completely in SF as shown in figure 2.8:left, where

3
as for the Ex on j = j1 , the H-plane that is at j = j1 + 2 is completely in SF. So, the respective

incident Hz fields, picked from the 1D grid’s Hz (1D array) after forming the index for each of the

nine main grid’s Hz locations, are added to the right hand side of the Ex update equation.

Also, for the Ex that fall inside the TF on j = j0 + 1 and j = j1 − 1, one of the H-planes falls

completely in SF region, also shown in figure 2.8:left, and similar corrections as the above case are

needed.

The Ex locations that are on the Z planes k = k0 , k1 and that belong to the nodes on X

planes i = i0 , i = i1 of the TF/SF cubic-boundary, there are Hz fields falling in both TF and SF as

shown in figure 2.8:right. Again, those falling in SF need consistency corrections, i.e. the incident

Hz fields at these locations available on the 1D grid need to be added the RHS of the main grid’s

Ex update equation.

Second, consider the corrections to the Ex that fall outside the total field (TF) region. For

the Ex locations that are part of the Y plane j = j0 − 1(which is in SF) shown in figure 2.9:Left,
 24

݆ ൌ ݆଴ +1

݆ ൌ ݆଴
݇ ൌ ݇ଵ

݅ ൌ ݅଴ ݅ ൌ ݅ଵ

‫ܧ‬௫
͵݄
ʹ ‫ܧ‬
௫

͵݄
‫ܪ‬௭
ʹ
‫ܪ‬௭

݇ ൌ ݇଴

Figure 2.8: Left: Total Field Ex (on Y planes) for which the whole H-plane (containing nine Hz
components) fall in scattered field region. Right: Total Field Ex (that belong to X plane nodes or
fall on Z planes of the TF/SF surface) for which only the three of the nine components on H-plane
fall in scattered field region. Also included are the special cases at the corners. Respective incident
H fields need to be added to these H fields that fall in scattered field region.

some or all of the Hz fall in the total field (TF) region and are highlighted in figure 2.9:Right. So,

the Ex update equation at each of these SF locations are inconsistent and the incident Hz fields

on the 1D grid at the indexes corresponding to these highlighted main grid’s locations need to be

subtracted from the RHS of the main grid’s Ex update equation. similar corrections are needed for

the Ex falling on Y plane j = j1 + 1.

The corrections needed for the update equation at Ex locations that either fall on Z planes

k = k0 − 1, k = k1 + 1 or that belong to the nodes on X planes i = i0 − 1, i = i1 + 1 are shown in

figure 2.10. Almost all the locations on the above Z planes (or all E − x locations that belong to

the the above X planes) have similar corrections with exceptions only at corners.

Lastly, the Ex that are located diagonally outside the TF and along the Y-axis, i.e. those

falling on the lines formed by the intersection of plane-pairs k = k0 − 1 & i = i0 − 1, k = k0 − 1 &

i = i1 , k = k1 + 1 & i = i0 − 1 and k = k1 + 1 & i = i1 as shown in figure 2.11, may have a lone

 25

݆ ൌ ݆଴
݇ ൌ ݇ଵ
݆ ൌ ݆଴ െ ͳ

݅ ൌ ݅଴ ݅ ൌ ݅ଵ

݇ ൌ ݇଴

Figure 2.9: Left: 3D view of the Y planes, with blue and red being part of SF and TF respectively.
Right: 2D view showing Scattered Field Ex (on Y plane) for which the whole H-plane or part of it
fall in total field region. Respective incident H fields need to be subtracted from these H fields.

Hz falling inside TF. This inconsistent update equation can be made consistent by subtracting the

incident Hz field picked from the 1D grid.

2.2.4 Waveform Excitation on 1D grid

In order to initiate a waveform on the 1D grid, the first few nodes of it are hard wired at every

time step using either analytical expressions for E and H (analytic source method), or Optimized

AFP (O-AFP) [21]. For both the methods, the dispersion relation given by equation 2.7 is solved

using a root-finding technique to find the numerical wavenumber k̃. The procedure to populate

fields at initial nodes on the 1D grid is similar in both the methods. Given a time series f (n) at a

reference point on 1D grid, use the numerical wavenumber k̃ to delay f (n) appropriately and find

the incident field time series fE,H (n) at those few initial nodes.
 2
h ω∆t
sin2 ( ) = p̃2x + p̃2y + p̃2z =
c∆t 2
 26

݇ ൌ ݇ଵ
݇ ൌ ݇ଵ

݅ ൌ ݅଴ ݅ ൌ ݅ଵ

݅ ൌ ݅଴ ݅ ൌ ݅ଵ

݇ ൌ ݇଴
݇ ൌ ݇଴

Figure 2.10: Left: Scattered Field Ex (on Z planes) for which part of the H-plane fall in total field
region. Respective incident H fields need to be subtracted from these H fields. Right: Scattered
Field Ex (on X planes) for which part of the H-plane fall in total field region. Respective incident
Hz fields need to be subtracted from these H fields. Also included are some special cases at the
corners.

݇ ൌ ݇ଵ

݅ ൌ ݅଴ ݅ ൌ ݅ଵ

݇ ൌ ݇଴

Figure 2.11: Scattered Field Ex (diagonally outside the TF along the Y axis) for which part of the
H-plane fall in total field region. Respective incident H fields need to be subtracted from these H
fields.

   2
1 Kc
Ka sin (k̃x h/2) + sin (k̃x h/2) · Kb + (cos (k̃y h) + cos (k̃z h)) + Kd cos (k̃y h) cos (k̃z h)
3 2
 27
   2
1 Kc
+ Ka sin (k̃y h/2) + sin (k̃y h/2) · Kb + (cos (k̃x h) + cos (k̃z h)) + Kd cos (k̃x h) cos (k̃z h)
3 2
   2
1 Kc
+ Ka sin (k̃z h/2) + sin (k̃z h/2) · Kb + (cos (k̃x h) + cos (k̃y h)) + Kd cos (k̃x h) cos (k̃y h)
3 2
(2.7)

The first method is similar to the incident field array (IFA) method [38], however, the neces-

sity for interpolation doesn’t arise because every location on the 3D grid has an equivalent point on

the 1D grid. Having solved for the numerical k̃ at the center frequency of the pulse, the approximate

time series (time waveform) at any further nodes can be found by applying spatial phase delay to

the waveform f (n) at reference location as shown in equation 2.8. η is the impedance of the homo-
n∆t−n0 2
geneous medium. Considering a modulated Gaussian pulse, f (n)|ref = eω0 (n∆t−n0 ) e[− nσ
]
, in

a dispersive medium (such as an FV24 grid), the envelop travels with the group velocity, where as

the carrier propagates with its phase velocity assuming the approximation k̃(ω) = k̃0 + k̃00 · (ω − ω0 )

is valid. Here, ω0 /k̃0 is the phase velocity of the carrier and 1/k̃00 is the group velocity (at carrier

frequency ω0 ) of the envelope.

1
fE (n)|ir −1/2 = Re f (n)|ref · e−j k̃0 (ir − 2 )∆r ,


1
Re f (n − 1/2)|ref · e−j k̃0 ir ∆r

fH (n − 1/2)|ir = (2.8)
η

The implementation given in equation 2.8 is overly simplistic as it delays only the carrier and

neglects the group velocity. It considers the wavenumber at every frequency in the pulse to be same

and equal to the wave number of the carrier. This causes distortion of the pulse at the point where

the transition happens on the 1D grid from hard-wired part to the 1D update equation. One would

need to delay the envelope considering the group velocity to reduce this distortion. Moreover, it

is observed that the numerical wave number in the FV24 grid closely follows the free-space wave

number for a sufficiently band limited pulse, so we can approximate the group velocity as equal to

the phase velocity of the carrier. This formulation is given in equation 2.9.

1
− n∆t−n0 −k̃0 (ir − 12 )∆r 2

fE (n)|ir −1/2 = cos ω0 [n∆t − n0 ] − k̃0 [ir − ]∆r · e nσ ,
2
(n−1/2)∆t−n0 −k̃0 ir ∆r 2
 

−
fH (n − 1/2)|ir = cos ω0 [(n − 1/2)∆t − n0 ] − k̃0 ir ∆r · e nσ (2.9)
 28

Ideally, the numerical wave number should be solved for at every frequency in the pulse,

multiply e−j k̃(ω)ir ∆r with the Fourier transform of the time series at the reference point, then

apply the inverse Fourier transform to get time series at locations that fall ir ∆r units beyond the

reference point on the 1D grid. This method was initially applied to the the 3D grid and called

the Analytic Field Propagator (AFP) [39], and later adapted to the 1D grid by Tan and Potter as

Optimized AFP (O-AFP). As we need fields only at few initial points using this method, a single

phase velocity (of the most energetic frequency component, the carrier) is used. The formulation

is given in equation 2.10. For a discussion on choice of number of points in FFT and inverse FFT,

one can refer to [21, 39].

1
fE (n)|ir −1/2 = Re F F T −1 F F T [Re{f (n)|ref }] · e−j k̃0 (ir − 2 )∆r ,


1
Re F F T −1 F F T [Re{f (n − 1/2)|ref )} · e−j k̃0 ir ∆r


fH (n − 1/2)|ir = (2.10)
η

For both the methods, once we have the time series for electric and magnetic fields, respective

components (Ex,y,z and Hx,y,z ) can be expressed using the polarization projections in equations

2.12. Here, ψ is the angle the vector E makes with ẑ × P̃ and it is assumed to be known [12]. Those

parameters with ˜ represent numerical counterparts of continuous world. The FV24 numerical

angels can be calculated using equations 2.11 as given in [21], where p̃2x,y,z are obtained from the

right hand side of equation 2.7 after solving for k̃.

p̃y p̃2x + p̃2y

tan φ̃ = , tan2 θ̃ = (2.11)
p̃x p̃2z

Ex |nir −1/2 = fE (n)|ir −1/2 · (cos ψ sin φ̃ − sin ψ cos θ̃ cos φ̃)

Ey |nir −1/2 = fE (n)|ir −1/2 · (− cos ψ cos φ̃ − sin ψ cos θ̃ sin φ̃)

Ez |nir −1/2 = fE (n)|ir −1/2 · (sin ψ sin θ̃)

n−1/2
Hx |ir = fH (n)|ir · (sin ψ sin φ̃ + cos ψ cos θ̃ cos φ̃)
n−1/2
Hy |ir = fH (n)|ir · (− sin ψ cos φ̃ + cos ψ cos θ̃ sin φ̃)
n−1/2
Hz |ir = fH (n)|ir · (− cos ψ sin θ̃) (2.12)
 29

The number of initial nodes that are hard wired for each component is selected based on

avoiding the negative indexes that may arise in the 1D update equations. Also, a common number

of hard wired nodes for all field components will not work because the first Ex,y,z nodes that should

be updated on the 1D grid are not co-located. To see this subtlety, consider that the corner of

the TF region to be (x, y, z) and its projection on the 1D grid to be m0 ∆r. At any time step, the

location of the first Ex (of the TF region) on the 3D grid is (x + h2 , y, z) whose projection and index
my
on the 1D grid are (m0 + m2x )∆r and m0 + mx2+1 ) respectively. For Ey , they are (m0 + 2 )∆r and
my +1 mz +1 my +mz my +mz mx +my
m0 + 2 . Similarly index is m0 + 2 for Ez , m0 + 2 , m0 + 2 , and m0 + 2

for Hx,y,z respectively. Observing that they are offset by different amounts from m0 , the field

component updates on the 1D grid should start in the same staggered manner. Otherwise, the

time and space offsets inherent to any FDTD scheme are not taken care of and wrong field values

from right locations are picked up on the right hand side of the update equations by the algorithm.
my +1
Therefore, the hard wiring is done up to index m0 + mx2+1 − 1, m0 + 2 − 1, m0 + mz2+1 − 1 and
my +mz mx +mz mx +my
m0 + 2 − 1, m0 + 2 − 1, and m0 + 2 − 1 for Ex,y,x and Hx,y,z respectively.

2.3 Validations of the Injection Technique

The above TF/SF method for exciting a plane wave in an FV24 grid is implemented with no

scatterers inside the TF region for the sake of demonstration. As we would expect to see no signal

in the SF region in the absence of scatterers, the lattice truncation using PML or CPML is not

necessary to implement, however, their inclusion is straight forward. The wave form initiated on
n∆t−n0 2
the 1D and so on 3D grid is a modulated Gaussian pulse f (n)|ref = eω0 (n∆t−n0 ) e[− nσ
]
where

n0 and nσ (despite n in their name, n0 and nσ have units of time) are used to vary the bandwidth

and the time shift of the Gaussian pulse. n0 should be selected such that the pulse rises gradually

from zero. If n0 is not sufficiently large, the time series Re(f (n)) will start abruptly in time and

also in space (on the 1D grid near the transition from hard sourcing to take over by 1D update

equations). This results in distortion of the waveform on 1D grid because of the high frequency

components associated with an abrupt change.

 30

The choice of nσ decides the width of the pulse in time and frequency domain, the larger

the value the broader is the pulse in time domain and narrower in frequency domain. Smaller

nσ will increase the highest frequency component in the pulse which may distort the pulse as it

propagates on the 1D grid (manifests itself on the 3D grid as well) because of insufficient grid

resolution and resultant FV24 numerical dispersion. However, pulse distortion because of too small

n0 and/or nσ will not effect the field leakage in to the SF region as long as the TF/SF corrections
√
2 2.3
are done properly. For a distortion free pulse propagation, we chose nσ = π∆f and n0 = 4.5nσ as

recommended in [1]. Here ∆f is the range of frequencies of the Gaussian pulse spectrum around

carrier frequency with a amplitude of 10% of the maximum or more.

The simulations are carried out with a carrier frequency of f0 = 2 GHz, uniform grid resolution
h 3
of 20 cells per wavelength at 1 GHz, and ∆t = √
c 3 |3−4Kb −2Kc −4Kd |
(the upper limit provided

in [23]). The TF region size is 40X40X40, and the SF region extends 10 cells from the TF region in

every direction. The integers (mx , my , mz ) = (9, 3, 13) and the polarization angle ψ = π/2. With

hard sourcing similar to IFA (first of the two methods described above), the wave forms for Ex and

Hz after propagating for 300 time steps on the 1D grid are shown in figure 2.12. The incident fields

Ex and Hz in TF region at a Z plane cross-section are shown in figure 2.13.

Also, the leakage field for Ex and Hz at the same Z plane are shown in figure 2.14 after 300

time steps. As expected, the leakage for all the field components is -300 dB below the respective

incident field strengths.

2.4 Comparison of Near-to-Farfield Transformation Techniques for FDTD

Equivalence theorems in EM come very handy to efficiently calculate/project farfields from

a radiating source or a scatterer. Extending the problem space to include the farfield region is

not often computationally feasible due to time and memory costs involved. Equivalence theorems

stipulate that farfield radiation or scattering profile (of a source or scatterer) can be evaluated from

the nearfields. The nearfields in the form of fictitious electric and magnetic surface currents (Js

and Ms ) are chosen on an equivalent surface enclosing the source or the scatterer. The surface
 31
E x on 1D grid
0.8

0.6

0.4

0.2
V/m

-0.2

-0.4

-0.6

-0.8
0 500 1000 1500 2000 2500
i r node number along 1D grid

Hz on 1D grid
×10 -3
1

0.8

0.6

0.4

0.2
A/m

-0.2

-0.4

-0.6

-0.8

-1
0 500 1000 1500 2000 2500
i r node number along 1D grid

Figure 2.12: Top: Ex on 1D grid after 300 time-steps Bottom: Hz on 1D grid after 300 time-steps

currents on the equivalent surface are used to calculate the farfield vector potentials A and F [40].

These vector potentials are then used to obtain either radiation pattern in case of sources or radar

cross section (RCS) in case of scatterers. The surface currents, (Js and Ms ), are in turn calculated

from nearfield tangential magnetic and electric fields (to the equivalent surface), respectively. This

process is referred to as near-to-farfield transformation.

Near-to-Farfield transformation in FDTD has been applied to satisfactorily predict the farfield

radiation/scattering profiles. One peculiar thing about constructing equivalent surface in FDTD

is that, it is not possible for any single closed surface to house both the tangential electric (E)

and magnetic (H) nearfields (that are used to calculate surface currents Js and Ms ). The reason

being, the E and H fields are not co-located in the FDTD grid (staggered in space). This makes it
 32
Incident Field in TF-Ex

0.6

0.4

0.2

V/m
0

-0.2

-0.4

-0.6
0 60
10 50
20 40
30 30
40 20
50 10
node number along X axis 60 0 node number along Y axis

Incident Field in TF-H z

-3
×10
1

0.5
A/m

-0.5

-1
0 60
10 50
20 40
30 30
40 20
50 10
node number along X axis 60 0 node number along Y axis

Figure 2.13: Top: Incident field Ex in TF on a Z plane at the center of TF after 300 time-steps.
Bottom: Incident field Hz in TF on a Z plane at the center of TF after 300 time-steps.

necessary to interpolate fields from neighboring Yee cells in order to bring E and H (and thereby

Js and Ms ) onto the same surface.

Different interpolation schemes available in the literature are arithmetic averaging [1], geo-

metric mean [41] and the mixed-surface approach [42]. In this work, we compare FDTD farfield

profiles of a empty scattering region, for which farfields should be ideally zero. The farfields are

obtained from surface currents calculated using different interpolation techniques mentioned above,

and also a separate surface approach dealt later in this chapter.

The equivalent surface, on which near-to-farfield transformation is performed, is placed in

the scattered field region as shown in figure 2.15. The scattered fields on this equivalent surface

are used to calculate surface currents, which are used in the surface integrals to calculate vector

potentials.

However, as mentioned earlier, in FDTD a single equivalent surface will not house both the
 33
Leakage into SF-E x

×10 -15

1.5

0.5

V/m
0

-0.5

-1

-1.5
70
60
50 70
60
40 50
30 40
20 30
20
node number along Y axis 10
10 node number along X axis
0 0

Leakage into SF-Hz

×10 -18

1
A/m

-1

-2

-3

-4
70
60
50 70
60
40 50
30 40
20 30
20
10
10
0 0
node number along Y axis
node number along X axis

Figure 2.14: Top: Leakage of Ex into SF on a Z plane at the center of TF after 300 time-steps.
Bottom: Leakage of Hz into SF on a Z plane at the center of TF after 300 time-steps.

Figure 2.15: Schematic of the FDTD problem space, showing the dielectric scatterer, TF/SF
regions and the Equivalent surface.

tangential magnetic and electric fields (or surface currents), because of the staggered field locations

in FDTD grid, as shown in figure 2.16.

 34

Figure 2.16: The tangential electric and magnetic fields on two separate surfaces, red and blue.

2.4.1 Arithmetic and Geometric Averaging

One of the techniques to bring the surface currents (Js and Ms ) on to the same surface is to

interpolate electric and magnetic fields using arithmetic average. Average of four H fields and two

E fields (time-domain fields) brings currents onto the same surface and to the same location, as

demonstrated in [1]. This is shown in figure 2.17. Discrete Fourier Transform (DFT) is applied on

the time-domain average to obtain frequency-domain current components, at the desired frequen-

cies. These frequency-domain currents are obtained at discrete locations, covering all the six faces

of the equivalent surface.

The geometric-mean interpolation is performed on the same fields as used by arithmetic

averaging, shown in figure 2.17. However, a different sequence of steps is followed. First, the

DFT is applied on the four time-domain H fields and the two time-domain E fields at the desired

frequencies, then the geometric mean of the complex-valued frequency-domain currents are obtained

at all the discrete locations of the equivalent surface. Directly applying geometric mean on time-

domain fields would force us take the square root and fourth root of negative real values. First

applying DFT and then taking the geometric mean would help avoid this. While applying geometric

mean on complex values, care should be taken to convert (wrap) the angle of complex values from
 35

Figure 2.17: Left: Average of two Electric field components highlighted red. Right: Average of
four Magnetic field components highlighted red.

[−π, π] scale to [0, 2π] scale. This allows the complex geometric mean to bisect the angle between

the two/four complex operands.

2.4.2 The Separate and Mixed Surface Approach

Another technique to overcome staggered-nature of planes containing tangential currents in

FDTD, for performing near-to-farfield transformation, is mixed-surface approach introduced in [42].

Similarities exist between mixed-surface approach and TF/SF formulation. For example, the fic-

titious surface currents introduced at the TF/SF interface in the TF/SF formulation, in the form

of consistency corrections to the FDTD update equations, where they excite equivalent planewave

fields inside the total field region, required no interpolation. This mixed-surface approach is coun-

terpart (dual) to the TF/SF formulation, in the sense that it also does not use field interpolation

to launch equivalent farfields outside the equivalent surface.

During the consistency corrections to the electric field in TF/SF formulation, for example,

we add or subtract incident magnetic field to the right hand side of the update equation. This is

analogous to placing a magnetic field-generated electric surface current at the electric field loca-

tion, i.e., shifting the location of electric surface current from the location of its associated magnetic

field. Similarly, the location of magnetic surface current is changed from the location of its asso-
 36

ciated electric field. This mixing of field and current locations, when applied to near-to-farfield

transformation, is referred to as the Mixed-Surface approach [42].

The implementations of the mixed-surface approach is shown in equations 2.13 and figure

2.18. Se and Sh represent two surfaces on which the tangential electric and magnetic fields are

present, respectively, in FDTD grid. When evaluating the surface integral for N in equations

2.13 using the magnetic fields on Sh , the surface electric currents caused by the magnetic fields

(Js = n̂ × H) are assumed to be placed on Se . Therefore, the distance (between the reference point,

i.e. the center of the volume enclosed by equivalent surface, and the current location on Se ) re0 is

used in the exponential inside the integral.

Similarly, when evaluating the surface integral for L in equations 2.13 using the electric fields

on Se , the surface magnetic currents caused by the electric fields (Ms = −n̂ × E) are assumed to

be placed on Sh . Thus, the distance rh0 is used in the exponential term inside integral.

‹ ‹
jkr̂·r0e 0 0
N= n̂ × H|Sh e dS = Js |Se ejkr̂·re dS 0
Sh Sh
‹ ‹ (2.13)
jkr̂·r0h 0 jkr̂·r0h 0
L=− n̂ × E|Se e dS = Ms |Sh e dS
Se Se

Figure 2.18: Left: Magnetic field-produced electric currents are placed on Se . Right: Electric
field-produced magnetic currents are placed on Sh

Finally, a separate-surface approach, which does not involve any mixing of field and current
 37

location and similar to mixed-surface approach in all other aspects, is implemented.

2.5 Frequency Domain Near-to-Farfield Transformation in FDTD

The frequency domain near-to-farfield projection is defined in terms of vector potential func-

tions A and F in equations 2.14. They are functions of frequency and farfield position unit vector r̂

along r as shown in figure 2.15. The unit vector also represents farfield angles θ and φ. The closed

surface integral is over the equivalent surface. The frequency domain currents Js and Ms (complex-

valued) are functions of position on the equivalent surface (represented by primed position vector

r0 as shown in figure 2.15) and the frequency.

‹
µ0 e−jkR 0 µ0 e−jkR
A(r̂, ω) = Js (r0 , ω)ejkr̂·r dS 0 = N,
4πR 4πR
S
‹ (2.14)
0e−jkR 0 jkr̂·r0 0 0 e−jkR
F(r̂, ω) = Ms (r , ω)e dS = L
4πR 4πR
S

The auxiliary vectors N and L in equations 2.14, that represent only the surface integrals, are then

used to calculate the θ and φ components of farfield Electric field, given by equations 2.15.

e−jkR
Eθ (θ, φ, ω) = − (Lφ + η0 Nθ ),
4πR (2.15)
e−jkR
Eφ (θ, φ, ω) = (Lθ − η0 Nφ ).
4πR
where k and η0 are the free-space wave-number and the intrinsic impedance, respectively. The

surface integration in equations 2.14 are carried as discrete summations (Riemann Sum). And,
1 2
Pinc is the power density of the incident planewave, given by Pinc = 2η (Eθ + Eφ2 ) · |F (ω)|2 , where

F (ω) is the Fourier transform of time-series f (t) which is defined in equations 2.18, sampled at the

desired frequency ω.

The θ and φ components of the auxiliary vectors N and L are defined by equations 2.16 as
 38

given in [1].
‹
0
Jx cos(θ)sin(φ) + Jy cos(θ)sin(φ) − Jz sin(θ) ejkr cos(ψ) dS 0 ,


Nθ =
S
‹
0
− Jx sin(φ) + Jy cos(φ) ejkr cos(ψ) dS 0 ,


Nφ =
S
‹ (2.16)
0
Mx cos(θ)sin(φ) + My cos(θ)sin(φ) − Mz sin(θ) ejkr cos(ψ) dS 0 ,


Lθ =
S
‹
0
− Mx sin(φ) + My cos(φ) ejkr cos(ψ) dS 0


Lφ =
S

2.6 Error Comparison with a No Scatterer

As a way to benchmark the accuracy of the above interpolation schemes, an empty region

(no-scatterer) is illuminated by a plane wave using Discrete Planewave technique with TF/SF

formulation described earlier in the chapter. Then, the farfields are obtained from the nearfields

which are interpolated using different schemes dealt with in the previous section, and compared.

As one might expect, the no-scatterer case should ideally result in zero farfield. Any farfield electric

field in equation 2.15 observed is an error (the noise floor). Errors can be because of:

• Numerical Dispersion

• Js and Ms calculation

• Surface integration

• Discretization in space and time

• Truncation (Finite word length)

Although this work doesn’t attempt to separate the above, the error in farfield caused by the

miscalculation (interpolation) of Js and Ms is the main differentiator. This is because, all the other

error contributors itemized above are identical for all the interpolation techniques. Therefore, it is

expected that the interpolation scheme that gives minimum farfield performs better than the rest

in terms of accuracy.
 39

The above interpolations schemes are compared for three different simulation parameters:

(1) Resolution of the FDTD grid,

(2) Incident angle of the plane wave,

(3) Size of the equivalent surface.

2.6.1 Resolution Sweep

Figure 2.19 shows how different interpolation schemes compare for different resolution of the

grid. The vertical axis is the maximum of Eθ · ∆t or Eφ · ∆t in dB, observed in all the farfield

principal planes (XY, XZ, and YZ planes). Here ∆t is the FDTD discretization time step. The

horizontal axis is the grid resolution in terms of cells per wavelength. The frequency at which the

farfields are calculated is 2 GHz.

-210

-220

-230

Arthmetic Mean
dB

-240
Mixed Surface
Separate Surface
-250 Geometric Mean

-260

-270
10 15 20 25 30 35 40
Resolution, Cells per Wavelength

Figure 2.19: Maximum farfield error in dB for different grid resolutions.

On the other hand, figure 2.20 shows average of Eθ · ∆t and Eφ · ∆t observed in all the farfield

principal planes.

The incident angle for the plane wave is φinc = 35.50 , θinc = 380 and the size of equiva-

lent surface is (1λ)3 . As expected, the farfield error decreases as resolution increases because the

dicretization errors in FDTD get minimized as the grid gets finer and finer.
 40
-220

-230

-240

-250 Arthmetic Mean

dB
Mixed Surface
-260 Separate Surface
Geometric Mean
-270

-280

-290
10 15 20 25 30 35 40
Resolution, Cells per Wavelength

Figure 2.20: Average farfield error in dB for different grid resolutions.

2.6.2 Incident Angle Sweep

Figures 2.21, 2.22 shows how maximum and average errors compare for different interpolation

schemes at different incident angles of the plane wave. The vertical axis is same as described before

and the horizontal axis specifies the θinc direction of the plane wave while the φinc = 23.20 . The size

of the equivalent surface is (5λ)3 and the grid resolution is 10 cells per wavelength. The frequency

at which the farfields are calculated is 2 GHz.

-180
Arthmetic Mean
-190 Mixed Surface
Separate Surface
Geometric Mean
-200
dB

-210

-220

-230

-240
0 10 20 30 40 50 60 70 80 90
Incident angle ( inc
), degrees

Figure 2.21: Maximum farfield error in dB for different plane wave incident angles.

The trend suggests that as incident plane wave direction gets close to the axes directions

(θinc = 0, 900 ), the farfield errors increase. This might be because of FDTD numerical dispersion

errors becoming worse for waves propagating along the axes directions as demonstrated in [12].
 41
-210

-220

-230
Arthmetic Mean

dB
Mixed Surface
Separate Surface
-240 Geometric Mean

-250

-260
0 10 20 30 40 50 60 70 80 90
Incident angle ( inc
), degrees

Figure 2.22: Average farfield error in dB for different plane wave incident angles.

2.6.3 Equivalent Surface Size Sweep

Figures 2.23, 2.24 shows how maximum and average errors compare for different interpolation

schemes for different equivalent surface sizes. The vertical axis is same as described before and the

horizontal axis specifies the size of equivalent surface as multiple of λ in 6(nλ)2 . The incident angle

for the plane wave is φinc = 35.50 , θinc = 380 and the grid resolution is 10 cells per wavelength. The

frequency at which the farfields are calculated is 2 GHz. As expected, the farfield errors increase

-180

-200

-220
dB

-240

Arthmetic Mean
Mixed Surface
-260 Separate Surface
Geometric Mean

-280

1 2 3 4 5 6 7 8 9 10
Size of Equiv. Surface, n in 6(n ) 2

Figure 2.23: Maximum farfield error in dB for various equivalent surface sizes.

as size of equivalent surface increases. This is because, the phase errors in FDTD accumulate more

as the wave travels in the grid longer and longer.

The comparison suggests that the mixed surface and arithmetic mean interpolation schemes

perform consistently better when compared to the geometric mean interpolation and separate

surface approach.
 42
-180
Arthmetic Mean
Mixed Surface
-200 Separate Surface
Geometric Mean

-220

dB
-240

-260

-280

1 2 3 4 5 6 7 8 9 10
Size of Equiv. Surface, n in 6(n ) 2

Figure 2.24: Average farfield error in dB for various equivalent surface sizes.

2.6.4 Bistatic RCS Comparison for a Dielectric Cube

To further validate the performance of different FDTD interpolation schemes, the bistatic

RCS of a dielctric cube (r = 5, µr = 1) is calculated at a single frequency of 1 GHz. This FDTD

farfield scattering profile is compared with those obtained from integral equation solvers available

in FEKO and HFSS.

The auxiliary vectors N and L in equations 2.14, that represent only the surface integrals,

are used to calculate the θ and φ components of RCS, given by equations 2.17.

k2
RCS θ (θ, φ, ω) = |Lφ + η0 Nθ |2 ,
8πη0 Pinc
(2.17)
k2
RCS φ (θ, φ, ω) = |Lθ − η0 Nφ |2 ,
8πη0 Pinc
The polarization of the planewave is defined by Eφ and Eθ , same as how planewave polariza-

tion is defined in HFSS and is similar to FEKO’s definition. In FEKO, the planewave polarization

is defined, slightly different, as the angle (η) Electric field vector makes with −θ̂ (unit vector at

planewave arriving angles). As an example of an identical setup in the three solvers, choosing the

angle η as 1800 in FEKO and Eθ = 1 & Eφ = 0 in FDTD and HFSS for the planewave will make

it theta-polarized. The polarization coefficients of the electric and magnetic fields in the X, Y and
 43

Z directions are given by equations 2.18, as given in [1].

Einc,x = [Eθ cos(θ̃inc )cos(φ̃inc ) − Eφ sin(φ̃inc )]f (t),

Einc,y = [Eθ cos(θ̃inc )sin(φ̃inc ) + Eφ cos(φ̃inc )]f (t),

Einc,z = −Eθ sin(θ̃inc )f (t),

−1 (2.18)
Hinc,x = [Eφ cos(θ̃inc )cos(φ̃inc ) + Eθ sin(φ̃inc )]f (t),
η0
−1
Hinc,y = [Eφ cos(θ̃inc )sin(φ̃inc ) − Eθ cos(φ̃inc )]f (t),
η0
1
Hinc,z = Eφ sin(θ̃inc )f (t)
η0

The dielectric cube of size λ/2 on each side (at 1 GHz) is illuminated with a plane wave

incident at θinc = 38.00 and φinc = 35.50 . This incident angle is rendered by the choice of integers

(mx ,my ,mz ) as (7,5,11) in the perfect TF/SF formulation described in the initial sections of the

chapter. These angles in FDTD indicate the direction of planewave propagation (direction of

propagation vector). Contrary to this, HFSS and FEKO would require the direction the planewave

comes from (opposite to the direction of propagation vector). Consequently, the planewave arrival

angles, θarrival = 180 − θinc and φarrival = 180 + φinc , are used in these solvers. The time-profile

of the theta-polarized planewave is a modulated Gaussian pulse, with frequency spectrum centered

around 1 GHz.

The RCS results from the four interpolation schemes of interest –arithmetic averaging, geo-

metric mean, mixed-surface, separate-surface approach– are compared with RCS profiles obtained

from FEKO-MoM and HFSS-IE in figures 2.25.

These figures show clear advantage that the mixed surface and arithmetic averaging schemes

have over the other two, geometric mean and separate surface approach. Here, the FDTD grid

resolution used is 40 cells/wavelength at 1 GHz. Here, only RCSθ comparison is shown on the

three farfield cuts (principal planes) as the RCSφ results do not deviate from each other to a

perceivable degree.
 44
Radar Cross Section, Farfield θ=90 (XY Plane)
-5

-10

-15

-20

RCSθ, dBm2
-25

-30
FEKO-MoM
-35 HFSS-IE
Arithmetic Averaging
Geometric Mean
-40 Mixed-Surface
Separate-Surface
-45

-50
-150 -100 -50 0 50 100 150
Farfield φ, degrees
(a)

Radar Cross Section, Farfield φ=0 (XZ Plane)

0

FEKO-MoM
-5 HFSS-IE
Arithmetic Averaging
Geometric Mean
-10 Mixed-Surface
RCSθ, dBm2

Separate-Surface

-15

-20

-25

-30
-150 -100 -50 0 50 100 150
Farfield θ, degrees
(b)

Radar Cross Section, Farfield φ=90 (YZ Plane)

-5

-10

-15

-20
RCSθ, dBm2

-25

-30 FEKO-MoM
HFSS-IE
Arithmetic Averaging
-35 Geometric Mean
Mixed-Surface
-40 Separate-Surface

-45
-150 -100 -50 0 50 100 150
Farfield θ, degrees
(c)

Figure 2.25: The θ component of bistatic RCS in different farfield planes. FDTD resolution is 40
cells/wavelength at 1 GHz.
 Chapter 3

Analysis of Glass Weave Induced Skew in Differential Pairs

This chapter deals with the timing skew between the two signals on a differential pair because

of glass fiber weave inside the PCB composite. As the interconnects on a high-speed printed circuit

boards carry more bits per second, skew becomes more crucial to predict and control. Accurate

simulation models to estimate the glass-weave skew and thereby predict the bit error rate, and the

differential-to-common signal conversion are very helpful in the robust design of high-speed boards.

Finite difference time domain (FDTD) technique with its inherent capability to produce

transient (time-domain) and wide-band (frequency-domain) responses might better suit this skew

study. The challenges in FDTD are to implement frequency-dependent dielectric losses (dispersive

complex permittivity) with high degree of accuracy.

This work demonstrates that FDTD, with the help of the techniques available in the literature,

can successfully estimate the signal skew in high-speed PCBs. Skew can be extracted directly from

the FDTD time-domain responses without the need for impulse responses constructed by inverse

Fourier transforming S-parameters. Also, the FDTD simulation data in the form of S-parameters

are bench-marked with frequency-domain finite element solver in HFSS. It also demonstrates that

the acceleration of FDTD using the GPU in MATALB leads to faster FDTD simulations.

3.1 Introduction: Glass Weave Skew in High Speed PCBs

Modern printed circuit boards (PCB) contain many dielectric (core and prepeg) and metalic

layers stacked on top of each other as shown in figure 3.1. The dielectirc layer is a composite
 46

material having woven glass fiber and resin (epoxy). The resin binds the glass fiber and hence

aptly called binder. The binder material is also referred to as the matrix [43]. Core is a cured

dielectric available to PCB manufacturers with copper cladding on both sides, where as prepreg is

a uncured dielectric and acts as a glue when all the layers are pressed together to form a PCB [3].

Figure 3.1: Stack-up of a fourteen layer PCB [3], as an example.

Glass fiber weave is available in many different styles as shown in figure 3.1 [44], and appro-

priate selection of glass style is important in mitigating glass weave induced skew (discussed later)

and for a robust current high-speed PCB designs.

ϳϲϳϴ Ϯϭϭϲ Ϯϭϭϯ

ϮϭϮϱ ϭϬϴϬ ϭϬϲ

Figure 3.2: Different glass weave styles and their style numbers.

Figure 3.3 shows a CAD model of the woven glass fiber and resin composite material. As
 47

shown in the figure, dielectric layer is an in-homogeneous material of glass and resin with glass-

rich and resin-rich areas. Moreover, these two materials, glass and resin, have inherently different

electrical properties. Comparatively, glass is a high permittivity, low loss material while resin is a

low permittivity, high loss material. This results in the microscopic variation of material properties.

Therefore, the electrical properties of copper traces (microstrip or striplines for example) etched

on such a material vary with position depending whether they see more glass or more resin. For

example the microstrip trace shown in figure 3.3 falls entirely on glass fiber and so sees more glass.

That means the propagation delay of two conductive traces might vary even though they have same

physical properties and are parallel to each other.

Figure 3.3: A CAD model showing woven glass fiber and resin composite material. Also shown are
glass-rich and resin-rich areas and a microstrip trace

In todays high speed PCBs, the differential signaling is widely used because of its immu-

nity to cross talk, ground bounce, simultaneous switching noise, and less radiated emissions when

compared to single-ended signaling [45], [46]. In differential signaling scheme on PCBs, a pair of

conductive traces (Differential Pair) carry balanced signals that are out of phase (180 degrees phase

difference). The two signals in the differential signaling should ideally be out of phase by 180 de-

grees at transmitter side and receiver side, canceling each other when added. However, because of

asymmetries in the drivers, terminations, and in the channel (interconnects), there is certain imbal-

ance between the two signals on the differential pair. The imbalance in time delay or propagation
 48

delay between the two traces is often referred to as skew.

When the skew is imparted because of one trace ’seeing’ more glass (or resin) than the other

trace, i.e., because of difference in dielectric properties of glass fiber and resin, it is referred to as

Glass Weave skew (GWS). The signal on the trace that sees more glass will propagate slower than

the signal on the trace that sees more resin. This effect is called Fiber Weave Effect (FWE). As

shown in figure 3.4, Trace 1 in the differential pair falls on glass fiber, i.e., in the glass-rich region,

where as the Trace 2 falls in between glass fibers, i.e., in the resin-rich region. The effective dielectric

constant seen by the former is more than that of what is seen by the latter. This results in signal

on it traveling slower than the signal on the other trace, giving rise to glass weave induced skew. In

today’s high bandwidth networking devices (switches and routing platforms), where digital signals

pass through traces tens of inches long, the skew can accumulate to several pico-seconds.

Figure 3.4: A model showing two traces of a microstrip differential pair, one that falls entirely on
glass fiber bundle and one that glass in between glass fiber bundles.

The skew results in a non-zero average of the two signals at the receiver end, i.e., results in

a common signal on the differential pair. This common signal, sometimes referred to as common-

mode signal, can severely effect the integrity of the signal at the receiver. For example, it causes

resonances in the SDD21 which is the differential insertion loss, leading to closure of eye diagrams

and increasing bit error rates. It might also cause unintended radiation resulting in electromagnetic
 49

compatibility (EMC) issues.

3.1.1 FDTD Modeling of Glass-Resin Composites: Motivation

The estimation of skew can be helpful to the system designers to see if the worst case skew is

within tolerable levels, even before they start thinking about mitigation techniques. Full wave 3D

field solvers can estimate the skew for ideal PCB board designs. Although, the real PCB boards are

bound to have some manufacturing defects/variations, a reliable estimate of skew is expected from

the 3D field solvers to aid system designers. Commercial frequency domain solvers such as HFSS,

can be used to obtain S-parameters (scattering parameters) over a range of frequencies the board

operates in. These S-parameters are in turn used to construct time domain pulse (with non-zero

rise time) response of individual traces in the differential pair. Finally, comparing the two pulses

gives an estimate of skew. Inverse Fourier transform is invariably involved in this frequency domain

to time-domain conversion.

The conversion of frequency domain S-parameters to get time-domain skew is expected to

have certain errors. The choice of number of frequency points, the highest frequency for which data

is available, and the input pulse rise-time during the inverse Fourier transform operation might

lead to miscalculation of skew. Moreover, the need to solve for S-parameters accurately over a

large number of discrete frequencies to get better resolution in time and thereby obtain accurate

estimate of skew is the main disadvantage with frequency-domain solvers.

This disadvantage and the uncertainty in estimation of skew can be overcome by using fi-

nite difference time domain (FDTD) method. With its inherent ability to produce time domain

responses, FDTD can estimate the glass weave skew with minimal post processing of EM fields.

While the finite element solvers have the ability to model the inhomogeneities of the PCB more

accurately because of the tetrahedral meshing, the simplest FDTD method uses rigid rectangular

cuboid grid. This should have minimal impact as long as a finer mesh is used in FDTD. Using

finer meshes increases the computational cost, a trade-off for accuracy. Given that the FDTD is

easily and massively parallizable, and with the cheap compute power available, the accuracy can
 50

be maintained.

The differential trace pair shown above in figure 3.4, with a loss-less dielectric and conductor

material properties as in table 3.1, is simulated using FDTD. The update equations, sourcing,

probing and PML boundary implementation are as provided in [1]. The waveform at the source

Table 3.1: Loss-less Material Properties

Material Dielectric Constant (r )

Resin Dielectric 3.2
Glass Dielectric 5.6

side of two traces is a Gaussian pulse having frequency components in excess of 40 GHz which is

the typical bandwidth of today’s high-speed digital signals. Two probes (Probe 1 and 2) are placed

on receiver side of Traces 1 and 2 in figure 3.4. Figure 3.5 shows the voltage waveforms sensed by

probes 1 and 2. Implementing the loss less or flat-loss dielectric materials using in-house FDTD

codes is straight-forward as given in [1].

Waveforms on Trace 1 and Trace 2, Showing Skew

0.5
Probe 1
Probe 2

0.4

0.3
Volts

0.2

0.1

-0.1
4 5 6 7 8 9 10
time,sec 10-11

Figure 3.5: Voltage waveforms show slight glass-weave induced timing skew on a differential a pair.
 51

3.1.2 Implementation of Glass Weave Structure in FDTD Grid

In the FDTD material grid, a glass weave fiber face is implemented mathematically as an

ellipse in x − y plane as given in equation 3.1 below:

 2  2
x − x0 y − y0
+ = 1, (3.1)
a b

where a and b are major and minor axes radii. This ellipse is centered around (x0 , y0 ).

The center of this ellipse can be swept along a sinusoidal path along the PCB lengthwise

direction (along z axis, for example) as shown in equation 3.2below:

 2  2
x − x0 y − sin(wg · z)
+ = 1, (3.2)
a b

where wg is wiggle factor that denotes how frequent are the undulations along the z direction.

It can be noticed that the x-coordinate of the ellipse’s center is kept constant while the y-

coordinate of the ellipse’s center depends on the position along the z-axis. The above procedure

is depicted in figure 3.6. This models a glass fiber bundle in the PCB’s dielectric layer. This

fiber bundle is replicated multiple times along the woof as well as the warp directions as shown

in the figure to form a weave-like material structure. The FDTD cells whose centers fall inside

the ellipse are assigned dielectric constant of the glass. Comparatively, microstrip trace and resin

material implementation is easy in Cartesian FDTD because of their rectangular or square shape

that conforms well with Cartesian coordinate system. Figure 3.7 shows a two-dimensional cross

section of FDTD grid that models a PCB, including the glass weave structure, resin, and copper

traces. The ground plane is implemented as zero-thickness PEC boundary condition and is not

shown.

3.2 Frequency Dependent Losses: Dispersive Medium

Loss-less materials are used for the simulation in the previous section. Similarly, flat-loss (σ

is constant across frequencies) materials can be implemented simply by including the conductivity

σ in the FDTD update equations. However, real-world materials are lossy, i.e., they have complex
 52

Figure 3.6: Glass fiber face modeled as an ellipse in x − y plane as shown above and is swept along
the z-axis modulating the y-coordinate of the center of the ellipse as shown below.

Figure 3.7: Various materials assigned to the cells in the FDTD grid modeling a PCB

dielectric constant and these losses are frequency dependent. These materials are causal i.e., the

real and imaginary parts of complex permittivity satisfy Kramers-Kronig relationship.

This frequency dependence is often modeled using Djordjevic-Sarkar model as in equation

3.4. Implementing the frequency dependent losses of a PCB material for time domain FDTD

simulation is not as straight forward as it is for frequency domain simulations. The logarithmic
 53

terms in Djordjevic-Sarkar model are even more complex to transform to time domain. On the other

hand, Debye (given in equation 3.3) and Lorentz models [10], which have ω frequency dependence,

can be transformed to time domain and implemented relatively easy using Auxiliary Differential

Equation (ADE) method.

X Ak (s − ∞ ) σDC
r = ∞ + + (3.3)
1 + ωτk ω0
k
 
(s − ∞ ) ωB + ω σDC
r = ∞ +   ln + , (3.4)
ln ω2 ωA + ω ω0
ωA

where:

r : frequency dependent relative permittivity,

s : relative permittivity at DC,

∞ : relative permittivity in the optical frequency range,

Ak : a dimension less Debye model coefficient,

ω : the angular frequency,

τk : inverse of relaxation frequency,

σDC : the conductivity at DC,

0 : free space permittivity,

ωA , ωB : lower and upper angular transition frequencies, respectively.

3.2.1 Adapting Djordjevic-Sarkar Model to FDTD

The Djordjevic-Sarkar model is suitable for FR4-type materials used in PCBs [47]. The

advantage of this model is that the material properties can be calculated from minimum measured

data points. For example, r and tan δ in table 3.2 for glass/resin dielectrics measured at 1 GHz

(Courtesy: Cisco Systems, Inc.) are used in conjunction with equation 3.4 to extrapolate complex

permittivity over wide frequency ranges. This method is used in frequency domain solvers such as

finite element method in HFSS for broadband simulation of PCB models.

As mentioned earlier, the logarithmic terms in this model are a bottle neck for its imple-

mentation in FDTD. A workaround is to first generate complex permittivity values from available
 54
Table 3.2: Material properties of glass and resin at 1 GHz

Material Dielectric Constant (r ) Loss tangent (tan δ)

Resin 3.65 0.02
Glass 6 0.0058

measured data by using Djordjevic-Sarkar model over a wide frequency range, then fit a multi-pole

Debye model to this data. In the process, we obtain the multi-pole Debye coefficients Ak in equation

3.3.

The procedure followed in HFSS [48] to generate broadband material properties from a single

set of measurement using Djordjevic-Sarkar model is given here for completeness. First, let us

assume that the real part of dielectric constant 1 and dissipation factor tan δ1 are available at

frequency f1 . Then, the frequency dependent real and imaginary parts of relative complex dielectric

constant can be generated by following the set of equations below.

∆ 1 tan δ1 − σωDC
K =  =  1 0 (3.5)
ωB −1 ωB
ln ωA tan ω1
 2 !
K ωB
∞ = 1 − ln +1 (3.6)
2 ω1
∆ = 10 tan δ1 ∞ (3.7)

DC = ∆ + ∞ (3.8)
ωB
ωA = ∆ (3.9)
e( K )  2
ωB + ω 2

K
Re [r (f )] = ∞ + ln 2 + ω2 (3.10)
2 ωA
       
−1 −1 ω −1 ω
Im [r (f )] = · σDC + K0 ω tan − tan , (3.11)
ω0 ωA ωB

where,

ω : angular frequency,

∞ : real part of relative permittivity at optical frequencies,

DC : real part of relative permittivity at DC,

1012
ωB = 2π Hz

The relaxation frequencies (1/τk ) in equation 3.3 are chosen over a large bandwidth (20 kHz
 55

to 200 GHz) and curve fitting is done using non-linear regression. MATLAB’s nlinfit function is

used in the current study. Figure 3.8 shows a multi-pole Debye model curve fitted to Djordjevic-

Sarkar model. The real and imaginary parts of relative permittivity are compared for glass and

resin materials between the two models. Eight poles at relaxation frequencies (1/τk ) equally spaced

from 20 KHz to 200 GHz (one each in the decades from 10 kHz to 1 THz) are used for curve-fitting.

Real part( r )
4.2
Multi-Pole Debye
4 Djordjevic-Sarkar

3.8

3.6

3.4
103 104 105 106 107 108 109 1010 1011
Frequency, Hz
Imag part( r )

0.08

0.06

0.04 Multi-Pole Debye

Djordjevic-Sarkar
0.02

0
103 104 105 106 107 108 109 1010 1011
Frequency, Hz

Glass Material Property Models

6.3
Multi-Pole Debye
6.2 Djordjevic-Sarkar
Re( r)

6.1

5.9
103 104 105 106 107 108 109 1010 1011
Frequency, Hz

0.05

0.04

0.03
Im( r)

0.02
Multi-Pole Debye
0.01 Djordjevic-Sarkar

0
103 104 105 106 107 108 109 1010 1011
Frequency, Hz

Figure 3.8: Multi-pole Debye model curve fitted to Djordjevic-Sarkar Modem for Glass and Resin
materials.
 56

3.2.2 Implementation of Multi-Pole Debye Model in FDTD

Once the multi-pole Debye coefficients for glass and resin are obtained, we use ADE method

in [10] to include material dispersion in FDTD. The polarization current (or the displacement

current) in dielectric medium modeled by P -pole Debye model, in equation 3.3, is given by equation

3.12:
P P
X X Ak (s − ∞ )
Jpol = J¯k = ω0 Ē. (3.12)
1 + ωτk
k=1 k=1

The frequency domain polarization current because of k th pole is given by equation 3.13:

Ak (s − ∞ ) ζk
J¯k = ω0 Ē = ω Ē (3.13)
1 + ωτk 1 + ωτk

Consequently, the Maxwell’s curl equation for magnetic field is modified as in equation 3.14,

and the curl equation for electric field remains unmodified as in equation 3.15. Finally, the polariza-

tion current of k th pole in equation 3.13 can be transformed to time domain Auxiliary Differential

Equation form as in equation 3.16.

P
∂ X
∇ × H̄ = 0 ∞ Ē + σDC Ē + J¯k (3.14)
∂t
k=1
∂
∇ × Ē = −µ0 µr H̄ (3.15)
∂t
∂ ¯k
J ∂
J¯k + τk = ζk Ē (3.16)
∂t ∂t

The equations 3.14-3.15 in combination with auxiliary equation 3.16 form the basis for im-

plementing material dispersion in FDTD using multi-pole Debye model. The first two of these

equations are discretized in time and space as in conventional FDTD, and the third equation only

in time, to obtain a set of update equations for Ē, H̄, J¯ respectively. As shown in [10], the update

sequence in the FDTD time marching loop is summarized as follows. At every step of the iteration:

(1) update magnetic fields using previous magnetic fields (from a full time step ago) and electric

fields from previous iteration (from half a time step ago)

(2) store the electric fields from previous iteration. These are needed to update polarization

currents.
 57

(3) update electric fields using previous electric fields (from a full time step ago), current

magnetic fields (from half a time step ago), and sum of previous polarization currents

(from a full time step ago)

(4) update polarization currents for each pole using previous polarization currents (from a full

time step ago), current electric fields (at the same time step), and stored previous electric

fields (from a full time step ago).

The pseudo code of the above steps in shown in equations 3.18-3.19, where n represents the

nth time step. The update coefficients are shown in simplified notation for illustration.

H n+0.5 = k1 · H n−0.5 + g · E n (3.17)

temp = E n

E n+1 = k2 · E n + r · H n+0.5 + ρ · Jkn (3.18)

Jkn+1 = k3 · Jkn + σ · (E n+1 − E n ), (3.19)

where g, r, σ, ρ and k1,2,3 are appropriate FDTD update coefficients. As shown above, unlike

conventional FDTD, this implementation requires that the electric fields, at least at the cells where

there is dispersive material if not all, be stored in computer memory before they are updated. The

sequence for updating electric, magnetic fields and polarization currents is shown in the 3.9.

As mentioned earlier, the multi-pole Debye model obtained by curve fitting it to Djordjevic-

Sarkar model as shown in section 3.2.1 is implemented using the discretized form of equations

3.14-3.16. In [10], the discretized update equations given for the above equations are for a multi-

pole Debye model. These update equations are adapted in this work. This involves calculating

polarization currents in equation 3.16 for all the multiple poles we have, one by one and summing

them to be used in electric field update equation (discretized form of equation 3.14).
 58

Increment time

Update
Update
polarization
magnetic fields
currents

Store electric
Update electric fields from
fields previous
iteration

Figure 3.9: Process flow showing sequence of steps involved in implementing dispersive FDTD.

3.3 Glass Weave Skew Analysis using FDTD

The above procedure to include material dispersion in FDTD is applied to analyze the skew

induced in PCBs because of glass weave. The glass weave model is implemented as described in

section 3.1.2. A PCB model of dimensions 150x100x4.72 mils (milli inches) with a differential pair

of microstrip traces is simulated. The traces are 150 mils long, 5.31 mils wide and 2.065 mils thick

and are separated edge to edge by 23.89 mils between them. The traces are placed so that one

of them falls on the glass fiber bundle and the other fall in between the bundles. The simulation

parameters are given in the following table 3.3.

Table 3.3: FDTD Simulation Parameters
Parameter Value
No of Cells per Wavelength 1020 at 40 GHz (free-space)
Total Cells including air gaps and CPML 12 million
Spatial Step: ∆x 7.5 µm
Time Step: ∆t 12.9 fsec
Time Steps 7700
Grid Truncation CPML (8 layers)
Input Pulse Gaussian Pulse (10% bandwidth of 40 GHz)
 59

Figure 3.10 shows the output pulses on the two traces after traveling 150 mils from FDTD

that include material dispersion from glass and resin. As expected, the signal on trace that falls on

glass travels slower than that of the trace that falls on resin. The timing skew for this configuration

is about 1 psec for 150 mils.

Figure 3.10: Pulse propagation on the differential pair obtained from FDTD simulation, and the
timing skew between them.

3.4 Comparison with HFSS Simulations

A HFSS model is also built and solved with the finite element solver to benchmark the skew

numbers from FDTD. Figure 3.11 shows the models built for FDTD and HFSS.

As mentioned earlier, the broadband frequency domain (0 to 40 GHz) S-parameters (insertion

loss for the two traces) are obtained from HFSS and converted to time domain following the

procedure shown in figure 3.12. The inverse Fourier transform process mentioned in [49]is used

here. Glass weave-induced skew studies using HFSS simulation process can be found in [50, 51].

Figure 3.13 shows the time domain pulses obtained from inverse Fourier transform process.

The timing skew between the two pulses is 0.97 psec which is a close match to 1 psec obtained from

FDTD.
 60

Figure 3.11: Top view of PCB Models for HFSS (left) and FDTD (right).

Obtain S-Par Multiply by

ࡴሺࢌሻ
ܵଶଵ ሺ݂ሻ Input Pulse
Spectrum

Negative
Up to 40 GHz Frequencies: Inverse Fourier
∗
‫ܪ‬ሺെ݂ሻ ൌ ܵ
 ݂ Transform

Figure 3.12: Conversion of S-parameter data to time domain.

Figure 3.13: Pulse propagation on the differential pair obtained from applying inverse Fourier
transform to HFSS simulation results, and the timing skew between them.

The discrete sweep in HFSS from 0 to 40 GHz with a 20 MHz frequency step took around
 61

57.5 hours on a single core (limited by licenses available) of Intel i7-6700 HQ processor running

at 2.6 GHz. It consumed about 4.6 GB of RAM. While the FDTD simulation in MATLAB took

around 7.97 hours consuming 2.5 GB on the same machine.

3.4.1 Acceleration of FDTD in MATLAB

MATLAB Parallel Computing Toolbox can be leveraged easily to exploit the CUDA-enabled

graphical processing unit (GPU) available on a machine. Reasonable speedup over running on a

CPU is possible with minimal code changes. The gpuArrays are used for storing FDTD fields

and update-coefficients. This facilitates running the FDTD update equation on GPU. Vector-

ized element-wise operations in CPML and Debye update equations can be implemented with

arrayfun()and bsxfun(), which can speed-up the operations further [52]. Significant speedups

are possible with CUDA or OpenCL programming which might require specialized knowledge.

The same simulation problem in the above section run on GPU from MATLAB using single

precision took around 1.49 hours, a 5x speed-up when compared to CPU implementation, and

consumed 1.25 GB of GPU RAM. The GPU on the machine is a NVIDIA GeForce GTX 960M

running at 1176 MHz with 640 CUDA Core and graphical RAM of 4 GB running at 5010 MHz.

The skew is 1 psec, same as before. Up to 30x speedup are reported for scattering problems with

plane wave and dipole antenna excitation in MATALB [52].

 Chapter 4

Spherical FDTD Dispersion Relation and Stability Analysis

Spherical FDTD is interesting in the context of modeling conical and spherical structures,

and earth-ionosphere system [53], when compared to Cartesian FDTD. The numerical dispersion

relationship of an FDTD algorithm predicts the effect of spatial and temporal discretizations on

the numerical wave number. It is useful in predicting phase errors in FDTD wave solutions and can

be used to choose the optimal grid parameters for a required phase-error tolarence. The numerical

dispersion relation for Cartesian FDTD is obtained by substituting plane waves harmonics (elem-

ntary functions native ot Cartesian coordinates) into the discretized Maxwell’s equations. On the

other hand, for Cylindrical FDTD, it is obtained by substituting cylindrical wave harmonics into

the discretized Maxwell’s equations [54]. In this chapter, numerical dispersion relation for spherical

finite-difference time-domain method (FDTD) based on spherical wave functions (elementary func-

tions native to spherical coordinates) is derived and the sensitivity of numerical wave number with

respect to spherical grid parameters is studied. Some sections in the chapter are adapted from our

work in [55], [56]. Also, based on the numerical dispersion derived, an attempt is made to derive

the stability criterion for spherical FDTD.

4.1 Introduction to Spherical FDTD

Before delving into details of numerical dispersion relation and its derivation, let’s see how the

spherical FDTD grid and the corresponding update equations look. Figure 4.1 shows a standard

six-faced spherical FDTD unit cell and the locations of electric and magnetic field components
 63

Er,θ,φ and Hr,θ,φ in the grid. The spherical grid is similar to Cartesian FDTD grid shown in figure

1.1 in that the electric fields are along the edges and magnetic fields are normal to the face centers

of the unit cell. One major difference though is the curvature of spherical unit cell edges in θ and

φ directions. As a consequence of this, the size of the unit cell is not constant across the grid, it

rather depends on how far the cell is away from the origin.

‫ܪ‬ఏ
‫ܪ‬థ
‫ܧ‬௥
‫ܪ‬௥
‫ܧ‬థ
‫ܧ‬ఏ

Figure 4.1: Spherical FDTD unit cell showing electric and magnetic field components’ locations.

4.1.1 Spherical FDTD Update Equations

In this section, the update equations for spherical FDTD are revisited. The spherical FDTD

update equations are avialble in different forms in the literature [53], [57], [58], and the ones listed

below are adapted mainly form [53]. The discretized update equations in [53], [58] are derived

based on Maxwell’s equations is integral form rather than the differential form and are slightly

more accurate than the ones in [57].

Let i, j, k represent the indexes in r, θ, φ directions respectively. Figure 4.2 shows a unit cell

cornered at (i, j, k) and positional offsets of the field components with respect to (i, j, k), similar

to Cartesian FDTD. The free-space update equations for Er,θ,φ and Hr,θ,φ can be derived from
 64

source-free modified Ampere’s law and Faraday’s law in integral form, respectively. Considering

the latter given in equation 4.2, we observe there are surface and closed-contour integrals.
ˆ ˛
d
0 E · dS = H · dl (4.1)
dt S
ˆ ˛C
d
−µ0 H · dS = E · dl (4.2)
dt S C

Referring to unit cell in figure 4.2 at (i, j, k), the Faraday’s law for Hr involves surface integration

‫
ܧ‬ቚ
,, .
‫
ܧ‬ቚ
.,,
‫ ܪ‬ቚ
.,, .

‫
ܪ‬ቚ
.,.,

‫
ܪ‬ቚ
,., .

‫ ܧ‬ቚ
,.,

Figure 4.2: Spherical FDTD unit cell showing positional offsets of electric and magnetic field
components.

over bottom surface (to which Hr r̂ is perpendicular to), and closed-contour integral on the edges

enclosing that (bottom) surface. The surface integral for Hr , centered at (i, j + 12 , k + 12 ), can be

written as in equation 4.3 and simplified as in equation 4.7 following the procedure below.

ˆ ˆ
k∆φ+∆φ ˆ
j∆θ+∆θ

Hr dS = Hr (rdθ)(r sin θdφ) (4.3)

S
k∆φ− ∆φ
2
j∆θ−∆θ

ˆ
k∆φ+∆φ ˆ
j∆θ+∆θ

= Hr r2 sin θ dθdφ (4.4)

k∆φ−∆φ j∆θ−∆θ

r2 ∆φHr |(i,j+ 1 ,k+ 1 ) cos(j∆θ)

 
= − cos((j + 1)∆θ) (4.5)
2 2
 65
   
2 ∆θ ∆θ
= 2r ∆φHr |(i,j+ 1 ,k+ 1 ) sin j∆θ + sin (4.6)
2 2 2 2
= Ar |(i,j+ 1 ) Hr |(i,j+ 1 ,k+ 1 ) , (4.7)
2 2 2

where Ar |(i,j+ 1 ) = 2r2 ∆φ sin (j + 21 )∆θ sin ∆θ

2 represents the surface area which is a function
2

of radius and elevation angle only (indexes i and j only).

And the corresponding contour integral on the right hand side of equation 4.2 (for Hr term

on the left-hand side) is as shown in equation in equation 4.8:

˛
E · dl = Eθ |(i,j+ 1 ,k) lθ |(i) − Eθ |(i,j+ 1 ,k+1) lθ |(i) +
2 2
C

Eφ |(i,j+1,k+ 1 ) lφ |(i,j+1) − Eφ |(i,j,k+ 1 ) lφ |(i,j) , (4.8)

2 2

where edge length lθ |(i) = r∆θ is a function of radius (index i only) and edge length lφ |(i,j) =

r sin(j∆θ)∆φ is a function of radius and elevation angle (indexes i and j only). It is noted here

that r = i∆r.

Similary, the integrals are simplified for other electric and magnetic field components and the

spherical FDTD update equations in final form as given in [59] are as follows:

∆t
Er |n+1
(i+ 1 ,j,k)
= Er |n(i+ 1 ,j,k) + ×
2 2 εr |(i+ 1 ,j,k) Ar |(i+ 1 ,j)
2 2
" 
1 1
n+ 2 n+ 2
Hθ |(i+ 1 ,j,k− 1 ) − Hθ |(i+ 1 ,j,k+ 1 ) lθ |(i+ 1 ) +
2 2 2 2 2
#
n+ 12 n+ 21
Hφ |(i+ 1 ,j+ 1 ,k) lφ |(i+ 1 ,j+ 1 ) − Hφ |(i+ 1 ,j− 1 ,k) lφ |(i+ 1 ,j− 1 ) (4.9)
2 2 2 2 2 2 2 2

∆t
Eθ |n+1
(i,j+ 1 ,k)
= Eθ |n(i,j+ 1 ,k) + ×
2 2 εθ |(i,j+ 1 ,k) Aθ |(i,j+ 1 )
2 2
" 
n+ 12 n+ 12
Hr |(i,j+ 1 ,k+ 1 ) − Hr |(i,j+ 1 ,k− 1 ) lr −
2 2 2 2
#
n+ 12 n+ 21
Hφ |(i+ 1 ,j+ 1 ,k) lφ |(i+ 1 ,j+ 1 ) + Hφ |(i− 1 ,j+ 1 ,k) lφ |(i− 1 ,j+ 1 ) (4.10)
2 2 2 2 2 2 2 2

∆t
Eφ |n+1
(i,j,k+ 1 )
= Eφ |n+1
(i,j,k+ 1 )
+ ×
2 2 εφ |(i,j,k+ 1 ) Aφ |(i)
2
" 
n+ 12 n+ 12
Hr |(i,j− 1 ,k+ 1 ) − Hr |(i,j+ 1 ,k+ 1 ) lr +
2 2 2 2
 66
#
n+ 1 n+ 1
Hθ |(i+21 ,j,k+ 1 ) lθ |(i+ 1 ) − Hθ |(i−21 ,j,k+ 1 ) lθ |(i− 1 ) (4.11)
2 2 2 2 2 2

n+ 1 n− 1 ∆t/µr |(i,j+ 1 ,k+ 1 )

2 2
Hr |(i,j+
2
1
,k+ 1 )
= Hr |(i,j+
2
1
,k+ 1 )
− ×
2 2 2 2 Ar |(i,j+ 1 )
2
"
 
Eθ |n(i,j+ 1 ,k) − Eθ |n(i,j+ 1 ,k+1) lθ |(i) +
2 2
#
Eφ |n(i,j+1,k+ 1 ) lφ |(i,j+1) − Eφ |n(i,j,k+ 1 ) lφ |(i,j) (4.12)
2 2

n+ 1 n− 1 ∆t/µθ |(i+ 1 ,j,k+ 1 )

2 2
Hθ |(i+21 ,j,k+ 1 ) = Hθ |(i+21 ,j,k+ 1 ) − ×
2 2 2 2 Aθ |(i+ 1 ,j)
2
"
 
Er |n(i+ 1 ,j,k+1) − Er |n(i+ 1 ,j,k) lr −
2 2
#
Eφ |n(i+1,j,k+ 1 ) lφ |(i+1,j) + Eφ |n(i,j,k+ 1 ) lφ |(i,j) (4.13)
2 2

n+ 1 n− 1 ∆t/µφ |(i+ 1 ,j+ 1 ,k)

2 2
Hφ |(i+21 ,j+ 1 ,k) = Hφ |(i+21 ,j+ 1 ,k) − ×
2 2 2 2 Aφ |(i+ 1 )
2
"
 
Er |n(i+ 1 ,j,k) − Er |n(i+ 1 ,j+1,k) lr +
2 2
#
Eθ |(i+1n ,j+ 1 ,k) lθ |(i+1) − Eθ |n(i,j+ 1 ,k) lθ |(i) (4.14)
2 2

The expressions for the edge lengths and face areas that appear in the above equations, and

shown on the spherical FDTD grid in figure 4.1.1, are given by:

lr = ∆r (4.15)

lθ |i = (i · ∆r)∆θ (4.16)

lφ |i,j = (i · ∆r)∆φ sin(j · ∆θ) (4.17)

Ar |i,j = 2(i · ∆r)2 ∆φ sin(∆θ/2) sin(j · ∆θ) (4.18)

Aθ |i,j = (i · ∆r)∆r∆φ sin(j · ∆θ) (4.19)

Aφ |i = (i · ∆r)∆r∆θ. (4.20)

It should be noted that in the above equations for edge lengths and face areas, we can have

half integer i, j, k depending on field component’s location in the grid.

 67

̂

݈௥
‫ܣ‬஘ 
݈ம
‫ܣ‬ம

݈஘ ‫ܣ‬௥



Figure 4.3: Edge lengths and face areas in spherical FDTD grid

Figure 4.4 shows how the spherical coordinate is discretized in radial (r), elevation (θ), and

azimuthal (φ) directions and just one ”φ-section” is shown for illustration. It also shows how cell

sizes vary, i.e., decrease as we move towards, and increase as we move away from the origin. This

position-dependence of spherical FDTD cells needs to be carefully accounted for in the numerical

dispersion relation and stability criterion derivations.

෡
ી

࢘ො

෡
ࣘ

Figure 4.4: A section of spherical FDTD grid showing discretization in r, θ, and φ directions.
 68

Moreover, because the spherical grid cells converge at singularities (the origin r = 0 and poles

θ = 0, 1800 ), as shown in figure 4.4, there are some special cells in the spherical FDTD whose shape

varies from that of the standard cell shown in figure 4.1. These special cells are shown in figure

4.5. It shows four-faced cell at r = 0, θ = 0, five-faced cells at θ = 0 and a five-faced cell possible

at r = 0. Similarly, four or five-faced cells are also possible at θ = 1800 .

6 Faced Cell
5 Faced Cell ࢘ො
࢘ො
North Pole

࢘ො

4 Faced Cell
5 Faced Cell

Figure 4.5: Special cells in spherical FDTD at poles and origin, also shown is a standard six-faced
spherical FDTD unit cell.

4.2 Numerical Dispersion Relation

This section introduces what numerical dispersion relation is and how it looks for Cartesian

FDTD. This section is adapted from our work in [55]. Wave propagation within finite-difference

time-domain (FDTD) meshes deviate from continuous space predictions due to the discrete nature

of the underlying update equations. For any given FDTD variant, a numerical dispersion relation

can be derived by populating its update equations with the corresponding set of elementary wave

functions. The resulting relation, which approaches the continuous space at the limit (ω 2 µ =

β 2 in a lossless medium) faithfully predicts numerical wavenumber solutions. Armed with such

knowledge, an optimal selection of mesh and update equations parameters can be ascertained,

while controlling the simulation errors. For the standard Cartesian FDTD, the numerical dispersion
 69

relation is obtained by substituting plane wave solutions in the discretized Maxwell’s equations [1]

to produce:

µDt2 = Dx2 + Dy2 + Dz2 , (4.21)

sin β̃ζ ∆ζ/2

where Dt =  sin∆t/2
ω∆t/2
, Dζ =  ∆ζ/2 , ζ → x, y, z, where ∆x, ∆y, ∆z and ∆t are spatial and

temporal discretization steps respectively, and β̃ is the numerical wave number. This relation can

be solved for β̃ for any given set of frequency and discrete step values.

4.3 Derivation of Spherical FDTD Numerical Dispersion Relation

As mentioned earlier, numerical dispersion relation derivation process generally consists of

following steps and these are pictorially shown in figure 4.3.

• Express field components in terms of elementary functions native to the coordinate system,

• Substitute the field components in discretized Maxwells curl equations,

• Reduce the set of equations from the above step to obtain numerical dispersion relation.

The figure also shows steps involved for the Cartesian FDTD numerical dispersion relation

derivation, as an example, in which a plane wave represented by e(ωt−β·r) is substituted in dis-

cretized Maxwell’s equations, and reduced to finally obtain the dispersion relation.

The dispersion relation in spherical FDTD for T Mr , transverse magnetic to radial direction,

is derived here. The analysis with T Er mode also results in the same dispersion relation as shown

in [56]. The elementary functions for spherical coordinates are the solutions to the Helmholtz

equation [60] given in equation 4.22. These elementary functions are given in the equation 4.23
(2)
[60] and it consists of spherical Hankel functions (used by Schelkunoff ) βrhn (βr), associated

Legendre functions Pnm (cosθ) and exponential functions e(−mφ) to represent r, θ, and φ dependency,
(2)
respectively. The function βrhn (βr) represent an outward traveling wave analogous to e(−βz)

traveling along the positive z direction.

∂2A ∂2A
 
1 ∂ ∂A 1
2
+ 2 sinθ + 2 2 + k2 A = 0 (4.22)
∂r r sinθ ∂θ ∂θ r sin θ ∂φ2
 70

Fields in Discretized
Obtain
terms of Maxwell’s
Dispersion
Elementary Curl
Relation
Functions Equations

௫௡ାଵ , , 

௫௡ , , 

 ௡ାଵ 1 ௡ାଵ 1
݁
⋅

௭
 ଶ ,  ,   ௭ ଶ ,   , 
2 2
௧ଶ  ௫ଶ  ௬ଶ  ௭ଶ

 ௡ାଵ 1 ௡ାଵ 1
 ௬ ଶ , ,   ௬ ଶ , ,  

 2 2

Figure 4.6: Numerical dispersion relation derivation process in general and steps involved in the
derivation for Cartesian FDTD

−(mφ−ωt)
A = βrh(2) m
n (βr)Pn (cosθ)e (4.23)

The spherical field components’ equations are given in terms of these elementary functions

as in equations 4.24-4.29 [60] [61].

 2 
1 ∂ 2
Er = +β A
ω ∂r2
n(n + 1) m
= Pn (cos θ)e−j(mφ−ωt) (4.24)
βr
1 ∂2A
Eθ =
ωr ∂r∂θ
!
(2)
hn (βr) 0 (2) 0
= Eθ sin θ + hn (βr) Pnm (cos θ)e−j(mφ−ωt) (4.25)
βr
1 ∂2A
Eφ =
ωr sin θ ∂r∂φ
!
(2)
1 hn (βr) 0 (2)
= Eφ + hn (βr) Pnm (cos θ)e−j(mφ−ωt) (4.26)
sin θ βr
Hr = 0 (4.27)
1 ∂A
Hθ =
r sin θ ∂φ
1 (2)
= Hθ h (βr)Pnm (cos θ)e−j(mφ−ωt) (4.28)
sin θ n
1 ∂A
Hφ = −
r ∂θ
 71
0 −j(mφ−ωt)
= Hφ sin θ h(2) m
n (βr)Pn (cos θ)e (4.29)

These field equations in terms of elementary functions are then substituted in spherical update

equations 4.9-4.14. An intermediate step after the substitution is shown in equation 4.30, as a

example.

eω∆t/2 − e−ω∆t/2 Hθ
 β̃h(2) (β̃r)Pnm (cos θ) =
∆t sin θ n
e−m∆φ/2 − em∆φ/2 n(n + 1)
− Er β̃ βr)Pnm (cos θ)
r sin θ∆φ βr
N (β̃r+ ) − N (β̃r− ) Eφ m
+ P (cos θ) (4.30)
r∆r sin θ n

The resultant set of five (not six because Hr = 0 for T Mr modes) equations can be concisely

represented in the form of a matrix equation as given below:

   
E E
−Dt 0 0 −Dφ +Dθ   Er 
   
 0

−Dt 0 0 E
−Dr   Eθ 
  
   
   
−Dt +DrE  ×  Eφ  = 0. (4.31)
 0 0 0   

   
   
−DH +DrH −µDt
0 0    Hθ 
 
 φ
   
+DθH −DrH 0 0 −µDt Hφ

The expressions for the above terms are given in equations 4.34-4.40. For a non-trivial field

solution to matrix equation 4.31, the determinant should be made zero as shown in equation 4.32.
 
 E E

−Dt 0 0 −Dφ +Dθ 

 
 0 −Dt 0 0 −DrE 
 
 
 
 0 0 −D +D E 0 =0 (4.32)
 t r 
 
 
−DH 0 +DrH −µDt 0 
 φ
 
+DθH −DrH 0 0 −µDt 
 

Making the determinant zero finally produces the numerical dispersion relatio 4.33 for spher-

ical FDTD, which is the same dispersion relationship as that for T Er modes in [56].

µDt2 = DrE DrH + DθE DθH + DφE DφH , (4.33)

 72

where

sin ω∆t/2
Dt =  (4.34)
∆t/2
(2) (2)
β̃(r + ∆r)hn [β̃(r + ∆r)] − β̃rhn (β̃r)
DrE =  0 (2)
 , (4.35)
(2)
∆r hn (β̃r+ ) + β̃r+ hn (β̃r+ )
N+ − N−
DrH = (2)
, (4.36)
β̃r∆rhn (β̃r)
(2) 0 0
βrhn (βr) ∆θ/2 sin2 (θ + ∆θ)Pnm [cos(θ + ∆θ)] − sin2 θPnm (cos θ)
DθE = · , (4.37)
n(n + 1) sin ∆θ/2 r sin θ+ ∆θPnm (cos θ+ )
n(n + 1) Pnm (cos θ+ ) − Pnm (cos θ− )
DθH = , (4.38)
(2)
βrhn (βr) r sin θ∆θPn0 m (cos θ)
sin m∆φ/2 ∆θ/2 1
DφE = − , (4.39)
∆φ/2 sin ∆θ/2 r2 sin2 θ+
sin m∆φ/2
DφH = − , (4.40)
∆φ/2

and
0
N ± = h(2) ± ± (2) ±
n (β̃r ) + β̃r hn (β̃r ), (4.41)

with r± = r ± ∆r/2, θ± = θ ± ∆θ/2.

4.3.1 Position Dependence of Numerical Dispersion Relation

This numerical dispersion relation is position-dependent, i.e., a function of absolute position

along r and θ directions, an anticipated dependence since mesh cell sizes in spherical FDTD vary

with radius and elevation angle, (r, θ) as shown in figure 4.4.

4.3.2 Convergence Analysis

The numerical dispersion relation in 4.33 can be tested for convergence in the farfield where

r → ∞, or in the continuum limit where ∆r, ∆θ, ∆φ → 0. In the continuum limit, it converges

to that of free-space dispersion relation ω 2 µ = β 2 , and in the farfield it converges to that of in

equation 4.42 [56]. If we replace ∆r with ∆x in this equation, it is essentially same as the dispersion
 73

relation for one-dimensional FDTD in [62].

sin ω∆t/2 sin β̃∆r/2

= , (4.42)
∆t/2 ∆r/2

where, β̃ is the numerical wave number. This makes total sense because, at larger and larger radii,

the spherical FDTD grid appears more and more as a one-dimensional FDTD grid.

4.4 Numerical Wave Number Sensitivity Analysis

In this section, the sensitivity of numerical wave number to grid parameters is studied. This

analysis will help us understand the errors in wave propagation within spherical FDTD and select

better simulation parameters with appropriate error control. This section is adapted from our

work [55]. Equation 4.33 is solved for the numerical wavenumber β̃ using fsolve function in

MATLAB. The chosen time step ∆t is stable expression from [53];

h q i−1
∆t ≤ c lr−2 + lθ−2 + lφ−2 . (4.43)

where c is the speed of light in vacuum and lr , lθ , lφ are minimum non-zero edge lengths in the grid.

The resulting numerical wave number is typically complex-valued, especially near the origin,

which is a departure from the continuous free-space wave number (β0 ). This behavior is further

studied for different radii, mesh resolutions, elevation angles, and mode numbers.

As the cell size decreases when we approach singularities, origin or poles, the wave number

is expected to behave eccentric. And as we move away from the origin towards the farfield, the

spherical wave propagation approaches that of one dimensional plane wave. Then, we expect the

wave number to approach that of one dimensional FDTD mesh which when sufficiently resolved

produces free space wave number.

4.4.1 Sensitivity to Mesh Resolution

The numerical dispersion relation (4.33) is solved for different mesh resolutions ∆r = λ0 /R
π
where R = 10, 20, 40 free-space cells per wavelength at a frequency of 1 GHz, ∆θ = 20 , and
 74

2π
∆φ = 40 . The time step ∆t is chosen as maximum value of (4.43), where lr = ∆r, lθ = ∆r∆θ and

lφ = ∆r∆φ sin(∆θ). The elevation angle θ is chosen as ∆θ.

The normalized real and imaginary parts of the numerical wave number with respect to

normalized distance from the origin are shown in figure 4.7 for different mesh resolutions. As we

move away from the origin, the Re(β̃/β0 ) converges to one signifying the numerical wave number is

same as the free space wave number. However, as we move towards the origin, Re(β̃/β0 ) increases

beyond unity. This signifies that the wave travels slower than speed of light in free space. The

Im(β̃/β0 ) quantity is totally a spurious error and causes either solution attenuation or growth near

the origin, depending on its sign. As we approach origin, the wave is initially attenuated and then

amplified in close proximity to the origin. Also, as one would expect, figure 4.7 shows that the

higher mesh resolutions simulate wave propagation better near the origin.

15
R=10
10 R=20
R=40

0
0 0.5 1 1.5 2

1
R=10
0.5 R=20
R=40

-0.5
0 0.5 1 1.5 2

Figure 4.7: Real and imaginary parts of wave number vs. normalized distance from origin at
different resolutions (θ = ∆θ, n = 3, and m = 0).
 75

4.4.2 Sensitivity to Mode Numbers

Next, the mode numbers n or m are varied while keeping the other integer constant. Solutions

for higher n modes get affected more as we move towards the origin for a fixed m as shown in figure

4.8. The increased sensitivity with increasing n can be attributed to the fact that higher order

Legendre functions vary rapidly with θ, and become sparsely sampled for a fixed θ resolution. For

example, when m = 0, Pnm (cos θ) for n = 1, 2, 3 are listed in the table 4.1.

Table 4.1: Legendre Polynomials Pnm (cos θ) for some lower order n and m = 0

n Pn0 (cos θ)
1 cos θ
1 2
2 2 (3 cos θ − 1)
1 2
3 2 cos θ(5 cos θ − 3)

15
n=1
10 n=2
n=3

0
0 0.5 1 1.5 2

1
n=1
0.5 n=2
n=3

-0.5
0 0.5 1 1.5 2

Figure 4.8: Real and imaginary parts of wave number vs. normalized distance from origin for
different mode numbers (θ = ∆θ, R= 40, and m = 0).

As shown in figure 4.9, it is also observed that the effect of changing m for a fixed n is minimal

on the behavior of numerical wave number.

 76
15
m=0
10 m=1
m=2
m=3
5

0
0 0.5 1 1.5 2

1
m=0
m=1
0.5 m=2
m=3
0

-0.5
0 0.5 1 1.5 2

Figure 4.9: Real and imaginary parts of wave number vs. normalized distance from origin for
different mode numbers m. Here θ = ∆θ, R= 40, and n = 4

4.4.3 Sensitivity to Elevation Angle

Figure 4.10 shows wave number sensitivity to different elevation angles, which is relatively

minimal compared to radial distance. This is surprising result, as we expect the numerical wave

number to show some dependence, if not strong, on θ because the cell sizes reduce considerably

moving towards the poles. This might be because, the edge length lθ = r∆θ of a cell doesn’t depend

on elevation angle (θ) and the FDTD is constrained by this edge length.

4.5 Stability Analysis

Having derived the position-dependent numerical dispersion relation and seen its sensitivity

in the previous sections, we will attempt to study the stability of spherical FDTD scheme. The

stability criterion sets the maximum time step ∆tmax that we can use in spherical FDTD beyond

which the fields will grow unbounded (instability). The stability study is interesting because of the

non-uniform grid in spherical FDTD; grid cells are small close to the origin and poles, and grow
 77

15
=
10 =3
=6
=9
5

0
0 0.5 1 1.5 2

1
=
0.5 =3
=6
=9
0

-0.5
0 0.5 1 1.5 2

Figure 4.10: Real and imaginary parts of wave number vs. normalized distance from origin at
different elevation angles (R = 40, n = 3, and m = 1).

larger as we move away from them. Given that, we expect that grid cells close to the origin and

poles set the stability condition or criterion for spherical FDTD.

As shown for Cartesian FDTD case in [12], the stability criterion is a by product of numerical

dispersion relation. The derivation of stability criterion involve the following steps [63]:

• Obtain numerical dispersion relation

• Obtain temporal and spatial mode growth,

• Apply the condition that maximum spatial mode growth <= maximum temporal mode

growth. This results in stability criterion (a bound on ∆t)

Following the above steps for Cartesian FDTD produce:

√ 2  2  2  2
sin β̃y ∆y/2
• Dispersion relation : µ sin∆t/2
ω∆t/2
= sin β̃x ∆x/2
∆x/2 + ∆y/2 + sin β̃z ∆z/2
∆z/2

√ 2
• Maximum temporal mode growth µ ∆t
 78
r  2
2 2 2 2 2



• Maximum spatial mode growth ∆x + ∆y + ∆z

• Stability criterion:
√
µ
∆t <= r  2 = ∆tmax . (4.44)
1

2 1 1 2


∆x + ∆y + ∆z

4.5.1 Existing Stability Analysis for Spherical FDTD

Before attempting to obtain stability criterion for spherical FDTD based on the above pro-

cedure, existing stability conditions available in the literature are reviewed.

1. One such condition is available in [53]. The stability condition is given as below:

1
∆t ≤ q (4.45)
c · lr−2 + lθ−2 + lφ−2

where, lr = ∆r, lθ = rmin ∆θ, lφ = rmin sin∆θ∆φ are the minimum edge lengths of Yee cells in the

problem space.

This is an approximate stability criterion directly adapted from Cartesian FDTD grid’s given

in equation 4.44. And there is no dispersion relation associated with it.

2. The next one is given in [64]. The numerical dispersion relation leading to this stability criterion

is obtained by substituting plane wave solutions in spherical FDTD update equations and the

criterion is given below:

c1 ∆l
c∆t ≤ π

(4.46)
N sin 2 4N

where,

• c1 is a constant (no expression given),

πR π

R·sin(∆θ)·∆φ
• ∆l = N sin 2N = 2 the “smallest distance between cells”,

• R is the radius of Earth,

• the stability condition is valid only when ∆r  ∆l ,

• when ∆r  ∆l , “the stability limit is determined mainly by ∆r ”

 79

• 2N, N are the number of points in longitude and latitude directions.

3. The third one is available in [65].

1
∆t ≤ q (4.47)
1 4 4
c· (∆r)2
+ (∆r∆θ)2
+ (∆r∆φsin∆θ)2

This stability condition does not depend on absolute radial position (r) in the grid. Again,

the Numerical Dispersion relation is obtained by substituting plane wave solutions in Spherical

FDTD update equations.

4.5.2 Stability Criterion based on Position Dependent Numerical Dispersion Re-

lation

Having seen the stability conditions available in the literature, an attempt is made to obtain

the stability criterion for spherical FDTD from position dependent numerical dispersion relation

derived earlier in the chapter. The dispersion relation given in equation 4.33 is repeated here for

convenience:

µDt2 = DrE DrH + DθE DθH + DφE DφH , (4.48)

The condition for stability is that the maximum spatial mode growth should be less than are

equal to the maximum temporal mode growth, i.e.,

 E H
+ DθE DθH max + DφE DφH max ≤ µDt2 max
     
Dr Dr  (4.49)
max

 2
2
 
where µDt2 max = µ ∆tmax . Therefore, the stability criterion can be written as:
√ r
∆r µ ∆r    
∆tmax = ;τ = |DrE DrH |max + DθE DθH max + DφE DφH  (4.50)
 
τ 2 max

For a given mode (n, m), we calculate τ as given in above expression and obtain theoretical

∆tmax . One can notice from figure 4.4 that as we move away from the origin, the cell sizes

increase and we expect the stability criterion to loosen. The stability criterion can also be observed

practically by setting a simple free space simulation problem, and by using the update equations

4.9-4.14. We can vary time step size and observe the threshold ∆t (∆tmax ) below which the fields
 80

are bounded and above which the fields grow unbounded. We can exclude grid cells near origin by

making them part of PEC (all the fields are zero inside it and tangential electric fields are zero on

the surface). Figure 4.5.2 shows such a problem space which contains PEC sphere at its center.

The radius of this sphere is varied and the stability is observed practically.

W^ƉŚĞƌĞ

Figure 4.11: Problem space that excludes regions around origin to observe stability criterion prac-
tically for spherical FDTD.

Figure 4.5.2 show how stability criterion varies as the PEC size is increased. As expected, the

stability criterion becomes less stringent (τ decreases and so ∆tmax increases by virtue of equation

4.50). Also, the figure show how stability criterion changes when ∆θ changes, i.e., for different

discretizations in elevation angle. When the grid is made coarser (∆θ2 ), we can observe that the

stability becomes less stringent compared to finer grid case (∆θ1 ).

As mentioned earlier, the elementary functions in spherical co-ordinates are composed of

(1,2)
Lengendre functions (Pnm (cos θ)) and spherical Hankel functions (hn (βr)). Higher order n and m

causes these functions to fluctuate more, and so the need for higher grid resolution for these higher-

order functions to get sampled enough. So, we expect the higher-order modes to have stringent

stability criterion than the lower-order modes. This is shown in figure 4.5.2 where theoretical τ is

plotted for different central PEC sphere radius at different n and fixed m = 0. From the figure,

we can see that higher-order modes (higher n) have stringent stability criterion (larger τ ) than the

lower-order modes. Also shown is the observed τ in addition to the theoretically obtained τ .
 81
KďƐĞƌǀƌĞĚdĂƵ;EdŚĞƚĂсϭϳͿ KďƐĞƌǀĞĚdĂƵ;EƚŚĞƚĂсϭϯͿ
Δ
Δ

ϯ
Δଵ
Ϯ͘ϱ
KďƐĞƌǀĞĚʏ sĂůƵĞƐ
Δଶ

IJ Ϯ
W W
ϭ͘ϱ
 
Δଵ  Δଶ 
ϭ 17  1 13  1
ϭ Ϯ ϯ ϰ ϱ ϲ ϳ ϴ ϵ ϭϬ
Center PEC Sphere Radius /ǻr

Figure 4.12: Practically observed stability criterion for different PEC sphere radii and for different
∆θ.

Surprisingly, the observed τ does not show correspondence to any one particular mode. Even

though, the excitation pattern is changed to excite a particular mode (n, m) in the grid, the stability

criterion didn’t change, hence a single curve for observed τ . As the smallest cell size increases (i.e.,

as PEC sphere radius increases), the observed stability factor τ decreases slower than the modes’

theoretical stability factors. It appeared as if a combination of modes is getting excited in the grid

and and the mode mixture is controlling the stability.

^ƚĂďŝůŝƚLJƌŝƚĞƌŝŽŶ;EdŚĞƚĂсϭϳͿ
ϯ KďƐĞƌǀĞĚ
Ŷсϭ
Ϯ͘ϱ ŶсϮ
Ŷсϯ
ʏ Ϯ Ŷсϰ
Ŷсϱ
ϭ͘ϱ
Ŷсϲ
Ŷсϳ
ϭ
ϭ Ϯ ϯ ϰ ϱ ϲ ϳ ϴ ϵ ϭϬ Ŷсϴ
W^ƉŚĞƌĞZĂĚŝƵƐͬǻr

Figure 4.13: Theoretical stability factor τ at various PEC sphere radii for different modes n, m = 0.
Also shown is the observed τ .
 82

4.5.3 Challenges in Validating Stability Criterion

Exploring the reasons why the observed τ didn’t match the theoretical τ might lead one to

the following questions.

• How to excite a single mode (n,m) to verify the stability criterion?

• How to find the highest-order mode being excited that might be deciding the stability

criterion?

• Is stability criterion same for all modes with same n?

• Is it possible that the algorithm renders different mode number (n) for Legendre and

Spherical Hankel Functions?

• What is the effect of radial node number (p)?

To address the first question of the above, we need spherical FDTD update equations for

a single mode (n, m) as was done for BOR FDTD (2-dimentiosal cylindrical FDTD) in [54]. In

the literature for spherical FDTD, the mode specific update equations are discussed and derived

in [66]. An attempt was made to obtain the stability criterion as well in that work. Nonetheless, the

theoretical and observed stability criterion fail to obey and finally an estimate based on heuristic fit

(empirical fit) is obtained for a stability factor B (analogous to τ in this work), and it was shown

that the observed stability factor lies close to this value. So, the perfect stability criterion for

spherical FDTD remains still elusive. Also, it was shown that the observed and estimated stability

factors both scale linearly with n and show little dependence on m.

For obtaining the maximum value under the square root in equation 4.50, the numerical wave

number must be used which in turn is obtained from solving the numerical dispersion relation.

However, in the work mentioned above, the free space wave number is used instead. One can

enhance the stability estimate (of aforementioned work) by obtaining the numerical dispersion on

similar lines to this present work and using the numerical wave number to obtain the stability

factor.
 Chapter 5

Spherical FDTD PML Analysis

This chapter deals with the Perfectly Matched Layer (PML) implementation and its ab-

sorption analysis for spherical FDTD. As mentioned in introductory chapter, PML is an effective

Absorbing Boundary Condition (ABC) to truncate FDTD grids and was introduced by Berenger

in 1994 [5] [12]. Berenger introduced the PML in split-field form for Cartesian FDTD, where each

of the six field components are split into two orthogonal components [12]. Later, PML has been

re-interpreted as anisotropic uniaxial medium (UPML) by Gedney [6] and in stretched-coordinate

form independently by Chew-Weedon [30] and Rappaport [31].

5.1 Perfectly Matched Layer in Spherical FDTD

The PML has been extended to spherical FDTD by Teixeira [67, 68] in 1997 using the

stretched-coordinate form of PML. However, complete discretized equations for stretched-coordinate

PML implementation in spherical FDTD remained unavailable in the literature until recently. Re-

cent work by Bao and Teixeria tried to fill this gap. The authors analyzed the performance of PML

in absorbing the incident waves in spherical FDTD [69,70], while the discretized PML equations are

available in [70]. However, these discretized equations contain typographical errors making it hard

to make use of them to implement PML in spherical FDTD. Moreover, the above works were based

on backward difference in PML region. Here in this work, the discretized PML equations (based

on stretched-coordinate formulation for spherical coordinates) are based on central-differencing as

in the non-PML region.

 84

In addition to the above, the absorption capabilities of the stretched-coordinate PML are an-

alyzed and compared with split-field PML formulation. Also, a closed-form expression for reflection

coefficient is derived for continous space PML in spherical coordinates that acts as a benchmark for

the comparing effectiveness of different formulations. Moreover, the efficacy of a lossy absorbing

shell, where a lossy medium is used to absorb waves similar to implentation in [29], to truncate

spherical FDTD grid is analyzed. It will be shown clearly why split-field implementation of PML

is not as effective in truncating spherical FDTD and is not as efficient absorber as it is in Cartesian

FDTD. It is also shown that split-field PML is only as efficient as lossy absorbing shell in spherical

FDTD.

5.2 Split-Field PML Formulation

The split-field PML in Cartesian FDTD involves splitting the field component perpendicular

to normal of PML surface into orthogonal portions: one that propagates along the normal to the

PML surface and one that propagates perpendicular to the normal. The portion that propagates

in the direction of normal to the PML surface is applied a loss factor in the form of electrical

conductivity σ (σe ) or magnetic conductivity σ ∗ (σm ) and gets absorbed by the PML.

The PML boundary condition in spherical FDTD has the following properties: (a) it sur-

rounds the problem space as shown in figure 5.1 (in the form of a shell around an hazelnut or similar

to thick rind surrounding red/pink flesh of a watermelon), and (b) the spherical FDTD PML region

is made of layers (because of radial discretization in spherical FDTD) and normals to these layers

are parallel to radial direction. Therefore, we have a scenario similar to split-field PML implemen-

tation in Cartesian coordinates. Similar to Cartesian case, the field components perpendicular to

the normal of the PML surface are split into two orthogonal portions. This means that only the

components other than the radial component, i.e Xθ,φ are split into Xθr,θφ and Xφr,φθ , in split-field

PML implementation for spherical FDTD. Here, X represents both electric and magnetic fields.

The portions Xθr and Xφr are applied the PML artificial loss factors σr and σr∗ in spherical FDTD,
σr σr∗
where they satisfy the condition  = µ .
 85

Problem Space

PML

Figure 5.1: PML boundary condition surrounding problem space in spherical FDTD.

This way of field components’ split is evident from the equation sets 5.1-5.6 and 5.7-5.14.

The first set is the Maxwell’s equations in differential form in spherical coordinates.
 
∂Er 1 ∂(sin θHφ ) ∂Hθ
 = − (5.1)
∂t r sin θ ∂θ ∂φ
 
∂Eθ 1 1 ∂Hr ∂(rHφ )
 = − (5.2)
∂t r sin θ ∂φ ∂r
 
∂Eφ 1 ∂(rHθ ) ∂Hr
 = − (5.3)
∂t r ∂r ∂θ
 
∂Hr 1 ∂(sin θEφ ) ∂Eθ
−µ = − (5.4)
∂t r sin θ ∂θ ∂φ
 
∂Hθ 1 1 ∂Er ∂(rEφ )
−µ = − (5.5)
∂t r sin θ ∂φ ∂r
 
∂Hφ 1 ∂(rEθ ) ∂Er
−µ = − (5.6)
∂t r ∂r ∂θ

The second set below is the split-field formulation for PML in spherical FDTD. The fields Eθ,φ

and Hθ,φ are split and only Eθr , Eφr , Hθr , Hφr (the portions that propagate in radial direction) are

applied the loss factors σr or σr∗ as shown in equations 5.8, 5.10, 5.12, and 5.14. We can also observe
∂
that these equations have ∂r on the right hand side, signifying the field component portion on the

left hand side of these equations propagates along radial (r) direction. These split-field equations

and their discretized versions are available in the chapter [71].

∂Eθφ 1 ∂Hr
 = (5.7)
∂t r sin θ ∂φ
∂Eθr 1 ∂(rHφ )
 + σr Eθr = − (5.8)
∂t r ∂r
 86
∂Eφθ 1 ∂Hr
 = − (5.9)
∂t r ∂θ
∂Eφr 1 ∂(rHθ )
 + σr Eφr = (5.10)
∂t r ∂r
∂Hθφ 1 ∂Er
−µ = (5.11)
∂t r sin θ ∂φ
∂Hθr 1 ∂(rEφ )
−µ − σr∗ Hθr = − (5.12)
∂t r ∂r
∂Hφθ 1 ∂Er
−µ = − (5.13)
∂t r ∂θ
∂Hφr 1 ∂(rEθ )
−µ − σr∗ Hφr = (5.14)
∂t r ∂r

5.3 Lossy Absorbing Shell as a Boundary Condition

Absorbing shell is a poor man’s absorbing boundary condition, where the inner region is

surrounded by a lossy medium. This lossy medium absorbs the incident waves, reflecting only

a part of the incident wave. In Cartesian coordinates, this technique [29] perfectly absorbs the

incident waves only for normal incidence [12]. Therefore, it is interesting to study its behaviour in

spherical coordinates. In this section, the spherical FDTD equations given in [53] are modified to
σ∗
include the loss factors σe and σ ∗ that satisfy the reflection-less condition σe
0 = µ0 . These modified

update equations are listed in equations 5.15 to 5.20. The performance of this absorbing shell

boundary condition is compared with PML formulations later in the chapter.

∆tσe
!
1 − 20 ∆t
Er |n+1
i+ 12 ,j,k
= ∆tσe
Er |ni+ 1 ,j,k +  ×
1 + 20 2
εr |i+ 1 ,j,k Ar |i+ 1 ,j 1 + ∆tσ e
2 2 20
"
n+ 1 n+ 1
(Hθ |i+ 12,j,k− 1 − Hθ |i+ 12,j,k+ 1 )lθ |i+ 1 +
2 2 2 2 2
#
n+ 21 l1 1 n+ 12
Hφ |i+ 1 ,j+ 1 k |i + , j + − Hφ |i+ 1 ,j− 1 ,k lφ |i+ 1 ,j− 1 (5.15)
2 2 2 2 2 2 2 2
!
1 − ∆tσ20
e
∆t
Eθ |n+1
i,j+ 1 ,k
= ∆tσe
Eθ |ni,j+ 1 ,k +  ×
2 1 + 20 2
εθ |i,j+ 1 ,k Aθ |i,j+ 1 1 + ∆tσ e
2 2 20
"
n+ 1 n+ 1
(Hr |i,j+2 1 ,k+ 1 − Hr |i,j+2 1 ,k− 1 )lr −
2 2 2 2
#
n+ 12 n+ 12
Hφ |i+ 1 ,j+ 1 ,k lφ |i+ 1 ,j+ 1 + Hφ |i− 1 ,j+ 1 ,k lφ |i− 1 ,j+ 1 (5.16)
2 2 2 2 2 2 2 2
 87
∆tσe
!
1− 20 ∆t
Eφ |n+1
i,j,k+ 1
= ∆tσe
Eφ |n+1
i,j,k+ 1
+  ×
2 1+ 20
2
εφ |i,j,k+ 1 Aφ |i 1 + ∆tσe
2 20
"
n+ 1 n+ 1
(Hr |i,j−2 1 ,k+ 1 − Hr |i,j+2 1 ,k+ 1 ) +
2 2 2 2
#
n+ 12 n+ 12
Hθ |i+ 1 ,j,k+ 1 lθ |i+ 1 − Hθ |i− 1 ,j,k+ 1 lθ |i− 1 (5.17)
2 2 2 2 2 2

∆tσ ∗
!
n+ 21 1− 2µ0 n− 1 ∆t/µr |i , j + 12 , k + 12
Hr |i,j+ 1 ,k+ 1 = ∆tσ ∗
Hr |i,j+2 1 ,k+ 1 − 
∗
 ×
2 2 1+ 2µ0
2 2
Ar |i,j+ 1 1 + ∆tσ 2µ0
2
"
(Eθ |ni,j+ 1 ,k − Eθ |ni,j+ 1 ,k+1 )lθ |i +
2 2
#
Eφ |ni,j+1,k+ 1 lφ |i,j+1 − Eφ |ni,j,k+ 1 lφ |i,j (5.18)
2 2

∆tσ ∗
!
n+ 12 1− 2µ0 n− 1 ∆t/µθ |i+ 1 ,j,k+ 1
Hθ |i+ 1 ,j,k+ 1 = ∆tσ ∗
Hθ |i+ 12,j,k+ 1 −  2 2
×
1+ ∆tσ ∗
2 2
2µ0
2 2
Aθ |i+ 1 ,j 1 + 2µ0
2
"
(Er |ni+ 1 ,j,k+1 − Er |ni+ 1 ,j,k )lr −
2 2
#
Eφ |ni+1,j,k+ 1 lφ |i+1,j + Eφ |ni,j,k+ 1 lφ |i,j (5.19)
2 2

∆tσ ∗
!
n+ 21 1− 2µ0 n− 1 ∆t/µφ |i+ 1 ,j+ 1 ,k
Hφ |i+ 1 ,j+ 1 ,k = ∆tσ ∗
Hφ |i+ 12,j+ 1 ,k −  2 2 ×
∗
2 2 1+ 2µ0
2 2
Aφ |i+ 1 1 + ∆tσ
2 2µ0
"
(Er |ni+ 1 ,j,k − Er |ni+ 1 ,j+1,k )lr +
2 2
#
Eθ |i+1n ,j+ 1 ,k lθ |i+1 − Eθ |ni,j+ 1 ,k lθ |i , (5.20)
2 2

where,

lr = ∆r (5.21)

lθ |i = r|i ∆θ (5.22)

lφ |i,j = r|i ∆φ sin θ|j (5.23)

Ar |i,j = 2r2 |i ∆φ sin(∆θ/2) sin θ|j (5.24)

Aθ |i,j = r|i ∆r∆φ sin θ|j (5.25)

Aφ |i = r|i ∆r∆θ. (5.26)

 88

5.4 Stretched-Coordinate PML Formulation

This section covers the theory behind implementation of stretched-coordinate PML in spher-

ical FDTD. The electric and magnetic field update equations in PML region are derived. The no-

tations for the variables and derivation process of the PML update equation are adapted from [69].

5.4.1 Electric Field Stretched-Coordinate PML Equation

Frequency domain Ampere’s law in complex coordinate space, stretched only along radial

direction, in spherical coordinate system are given by (similar to Faraday’s equation in [68]):
 
1 ∂ ∂
ωEr (ω) = (sin θHφ (ω)) − Hθ (ω) (5.27)
r̄ sin θ ∂θ ∂φ
 
1 1 ∂ ∂
ωEθ (ω) = (Hr (ω)) − (r̄Hφ (ω)) (5.28)
r̄ sin θ ∂φ ∂ r̄
 
1 ∂ ∂
ωEφ (ω) = (r̄Hθ (ω)) − Hr (ω) , (5.29)
r̄ ∂ r̄ ∂θ

where,
ˆ r
r̄(r, ω) = sr (r0 , ω) dr0 , (5.30)
0
σr (r)
sr (r, ω) = ar (r) + . (5.31)
ω

Here, σr is the electrical conductivity in radial direction analogous to σx,y,z for Cartesian

FDTD PML. sr (r, ω) is the stretched-coordinate variable for radial direction, and r̄(r, ω) is the

complex radius variable. Given these, complex radius variable in equation 5.30 can be rewritten

as:
ˆ r
σr (r0 )

0 ∆r (r)
r̄(r, ω) = ar (r ) + dr0 = br (r) + , (5.32)
0 ω ω
´r
where ∆r (r) = 0 σr (r0 ) dr0 , not to be confused with radial discretization ∆r.

Also, it is observed from equation 5.30 that the relationship between partial derivatives in

complex and real space is given by:

∂ ∂ ∂r 1 ∂
= · = . (5.33)
∂ r̄ ∂r ∂ r̄ sr ∂r
 89

Applying the above manipulations to the equations (5.27-5.29), and after some rearrange-

ment, we get:
 
1 ∂ ∂
ωr̄(ω)Er (ω) = (sin θHφ (ω)) − Hθ (ω) , (5.34)
sin θ ∂θ ∂φ
1 ∂ ∂
ωsr (ω)Ēθ (ω) = (sr (ω) · Hr (ω)) − H̄φ (ω), (5.35)
sin θ ∂φ ∂r
∂ ∂
ωsr (ω)Ēφ (ω) = H̄θ (ω) − (sr (ω) · Hr (ω)) . (5.36)
∂r ∂θ

In the above equations, new field terms in stretched-coordinate PML region have been used

and they are defined as:

H̄θ,φ (ω) = r̄(ω)Hθ,φ (ω), (5.37)

Ēθ,φ (ω) = r̄(ω)Eθ,φ (ω). (5.38)

Till now all the equations are given in frequency domain. However, they need to be trans-

formed to time domain for FDTD implementation. In time domain, the above equations translate

to:
   
∂ ∆r (r) 1 ∂ ∂
 br (r) + Er (t) = (sin θHφ (t)) − (Hθ (t)) (5.39)
∂t  sin θ ∂θ ∂φ
 
∂ σr (r) 1 ∂ ∂
 ar (r) + Ēθ (t) = (sr (t) ∗ Hr (t)) − (H̄φ (t)) (5.40)
∂t  sin θ ∂φ ∂r
 
∂ σr (r) ∂ ∂
 ar (r) + Ēφ (t) = (H̄θ (t)) − (sr (t) ∗ Hr (t)) (5.41)
∂t  ∂r ∂θ
Hθ,φ (t) = r̄−1 (t) ∗ H̄θ,φ (t), (5.42)

Eθ,φ (t) = r̄−1 (t) ∗ Ēθ,φ (t). (5.43)

The frequency domain terms sr (r, ω) and r̄(ω), given by equations 5.31 and 5.32 respectively,

too are converted to time domain as follows:

 
−1 σr (r) σr (r)
sr (r, t) = F ar (r) + = ar (r)δ(t) + u(t), (5.44)
ω 
   
1 1 ω
r̄−1 (t) = F −1  ∆r (r)
= F −1  
br (r) + ω br (r) ω + ∆ r (r)
br (r)
 
1 d −
∆r (r)
t
= e br (r) u(t) . (5.45)
br (r) dt
 90

br (r)
Defining ∆r (r) = τ0 (r), we can write equation 5.45 as follows:
 
−1 1 − τ t(r) 1 − τ t(r)
r̄ (t) = e 0 δ(t) − e 0 u(t) (5.46)
br (r) τ0 (r)
 
1 1 − τ t(r)
= δ(t) − e 0 u(t) . (5.47)
br (r) τ0 (r)

Following this conversion to time domain, the convolution operation in the equations (5.39-

5.43) is expanded in time domain integral form as follows:

ˆ t
σr (r)
sr (t) ∗ Hr (t) = ar (r)Hr (t) + Hr (t − t0 ) dt0 , (5.48)
 0

r̄−1 (t) ∗ Ēθ (t) = Eθ (t)

 ˆ t 0

1 1 0 − τ t(r) 0
= Ēθ (t) − Ēθ (t − t )e 0 dt (5.49)
br (r) τ0 (r) 0

The integral in the above equation is implemented in discrete time domain for FDTD using

piece-wise constant summation as:

ˆ t 0 n=m ˆ (n+1)∆t 0
!
1 0 − τ t(r) 0 1 X − τ t(r) 0
Ēθ (t − t )e 0 dt ≈ e 0 dt Ēθ (m∆t − n∆t)
τ0 (r) 0 τ0 (r) n∆t
n=0
n=m
− τ ∆t − τn∆t
X
= (1 − e 0(r) ) e (r)
0 Ēθ (m∆t − n∆t) (5.50)
n=0

The summation in the above equation can be implemented recursively as follows:

n=m
− τ ∆t − τn∆t − τ ∆t
X
(1 − e (r)
0 ) e 0(r) Ēθ (m∆t − n∆t) = (1 − e 0 (r) )Q(n∆t), (5.51)
n=0

where,
− τ ∆t
Q(n∆t) = Ēθ (n∆t) + e 0 (r) Q((n − 1)∆t), Q(0∆t) = 0. (5.52)

This recursive summation is efficient for implementation in FDTD schemes as it involves

reusing previous summation in the current time step.

Finally, the update equations for FDTD implementation of the time domain stretched-
 91

coordinate PML equations (5.39-5.43) are given by:

(
∆t∆r (i+ )

n+1 1 +
Er (i, j, k) = + br (i ) − Ern (i, j, k) +
br (i+ ) + ∆t∆2r (i ) 2
" n+ 1 n+ 1
Hφ 2 (i, j, k) sin(j + ∆θ) − Hφ 2 (i, j − 1, k) sin(j − ∆θ)
Cerh(i, j, k) −
∆θ
n+ 1 n+ 1
#)
Hθ 2 (i, j, k) − Hθ 2 (i, j, k − 1)
, (5.53)
∆φ
∆t∆θ
2
Cerh(i, j, k) = ;
sin(j∆θ)(i, j, k) sin( ∆θ
2 )

( 
1 ∆tσr (i)
Ēθn+1 (i, j, k) = ∆tσr (i)
ar (i) − Ēθn (i, j, k) +
ar (i) + 2
2
"  
1 n+ 21 n+ 12
Ceth(i, j, k) ar (i) H r (i, j, k) − Hr (i, j, k − 1) +
∆φ sin(j + ∆θ)
 !
σr (i) n+ 21 n+ 12
Hrs (i, j, k) − Hrs (i, j, k − 1) −

n+ 1 n+ 1 #)
H̄φ 2 (i, j, k) − H̄φ 2 (i − 1, j, k)
, (5.54)
∆r
∆t
Ceth(i, j, k) = ;
(i, j, k)

( 
1 ∆tσr (i)
Ēφn+1 (i, j, k) = ∆tσr (i)
ar (i) − Ēφn (i, j, k) +
ar (i) + 2
2
n+ 12 n+ 12
"
H̄θ (i, j, k) − H̄θ (i − 1, j, k)
Ceph(i, j, k) −
∆r
 
1 n+ 12 n+ 12
ar (i) Hr (i, j, k) − Hr (i, j − 1, k) +
∆θ
  !#)
σr (i) n+ 12 n+ 12
Hrs (i, j, k) − Hrs (i, j − 1, k) , (5.55)

∆t
Ceph(i, j, k) = ;
(i, j, k)
 92

n+1 1 h n+1 − ∆t
i
Eθ,φ (i, j, k) = Ēθ,φ (i, j, k) − (1 − e τ0 (i) )Qn+1
θ,φ (i, j, k) , (5.56)
br (i)
− τ∆t
Qn+1 n+1
θ,φ (i, j, k) = Ēθ,φ (i, j, k) + e
0(i) Qnθ,φ (i, j, k), (5.57)

Q0θ,φ (i, j, k) = 0; (5.58)

and

n+ 12 n− 12 n+ 21
Hrs (i, j, k) = Hrs (i, j, k) + ∆tHr (i, j, k). (5.59)

At each time step, we calculate the discrete summation in equation 5.59 from magnetic

fields before the electric field updates in equations (5.54-5.56). This summation corresponds to the

integral term in equation 5.48.

In the above equations i, j, k and ∆r, ∆θ, ∆φ correspond to grid cell indexes and discretization

steps in radial, elevation, and azimuthal directions, respectively. FDTD time step is given ∆t. And

i± = i ± 0.5 (similarly j ± = j ± 0.5 and k ± = k ± 0.5) depending on concerned field’s location

in the Yee cell for spherical FDTD as shown in figure 5.2. The discretization is based on central

differencing in both time and space.

5.4.2 Update Sequence

Since there are new field terms in the PML region (Eθ,φ , Ēθ,φ , Hθ,φ , H̄θ,φ , the question arises

as to which field needs to updated first. Moreover, because of the presence of additional field terms

in stretched-coordinate formulation, field updates at the PML and non-PML interface needs to be

handled carefully. Assuming the interface is at r = r0 , the following update sequence for the above

update equations addresses these issues.

(1) Update Er :

• Update Er in inner non-PML region.

• Update Er in PML region using equation 5.53.

 93

‫
ܧ‬ቚ
,, .
‫
ܧ‬ቚ
.,,
‫ ܪ‬ቚ
.,, .

‫
ܪ‬ቚ
.,.,

‫
ܪ‬ቚ
,., .

‫ ܧ‬ቚ
,.,

Figure 5.2: Spherical FDTD unit cell showing the field locations and their offsets.

(2) Update Eθ :

• Update Eθ in inner non-PML region.

• Calculate Hrs in equation 5.59 using magnetic fields.

• Calculate H̄φ near PML interface (yet, that still belongs to inner region, i.e., r =
∆r
r0 − 2 ) using H̄φ = rHφ . This is needed for Ēθ update.

• Update Ēθ in PML region using equation 5.54.

• Calculate Qθ in PML region using equation 5.57.

• Calculate Eθ in PML region using equation 5.56.

(3) Update Eφ :

• Calculate H̄θ near PML interface (yet, that still belongs to inner region, i.e., r =
∆r
r0 − 2 ) using H̄θ = rHθ . This is needed for Ēφ update.

• Update Ēφ in PML region using equation 5.55.

• Calculate Qφ in PML region using equation 5.57.

 94

• Calculate Eφ in PML region using equation 5.56.

As mentioned earlier, at the interface between the inner region and PML, the update to

tangential Electric fields in equations 5.54 and 5.55, that can be considered part of PML, require

magnetic fields in the inner non-PML region. To put these non-PML magnetic fields in the format

of equation 5.37 as required by update equations 5.54 and 5.55, we multiply them with radius, i.e,

H̄θ,φ (ω) = rHθ,φ (ω). It is to be noted that the radius r is not a function of frequency ω, as these

these magnetic fields are located in inner non-PML region that is not stretched.

5.4.3 Magnetic Field Stretched-Coordinate PML Equation

In a similar fashion to the above derivation of electric field update equations in spherical

PML region, magnetic field update equation can be derived. Frequency domain Faraday’s law in

complex coordinate space, stretched only along radial direction, in spherical coordinate system are

given by:
 
1 ∂ ∂
ωµHr (ω) = − (sin θEφ (ω)) − Eθ (ω) (5.60)
r̄ sin θ ∂θ ∂φ
 
1 1 ∂ ∂
ωµHθ (ω) = − (Er (ω)) − (r̄Eφ (ω)) (5.61)
r̄ sin θ ∂φ ∂r
 
1 ∂ ∂
ωµHφ (ω) = − (r̄Eθ (ω)) − Er (ω) , (5.62)
r̄ ∂ r̄ ∂θ

In time domain, the above equations translate to:

   
∂ ∆r (r) 1 ∂ ∂
µ br (r) + Hr (t) = − (sin θEφ (t)) − (Eθ (t) (5.63)
∂t  sin θ ∂θ ∂φ
 
∂ σr (r) 1 ∂ ∂
µ ar (r) + H̄θ (t) = − (sr (t) ∗ Er (t)) + (Ēφ (t)) (5.64)
∂t  sin θ ∂φ ∂r
 
∂ σr (r) ∂ ∂
µ ar (r) + H̄φ (t) = − (Ēθ (t)) + (sr (t) ∗ Er (t)) (5.65)
∂t  ∂r ∂θ
Eθ,φ (t) = r̄−1 (t) ∗ Ēθ,φ (t), (5.66)

Hθ,φ (t) = r̄−1 (t) ∗ H̄θ,φ (t). (5.67)

Similar to Ampere’s law, the Faraday’s law’s update equations for FDTD implementation of
 95

the time domain stretched-coordinate PML equations (5.63-5.67) are given by:
( 
n+ 21 1 ∆t∆r (i) n− 12
Hr (i, j, k) = b r (i) − Hr (i, j, k) +
br (i) + ∆t∆ r (i)
2
2
"
Eφn (i, j + 1, k) sin((j + 1)∆θ) − Eφn (i, j, k) sin(j∆θ)
Chre(i, j, k) − +
∆θ
#)
Eθn (i, j, k + 1) − Eθn (i, j, k)
, (5.68)
∆φ
∆t∆θ
2
Chre(i, j, k) = ;
sin(j + ∆θ)µ(i, j, k) sin( ∆θ
2 )

(
∆tσr (i+ )

n+ 1 1 n− 12
H̄θ 2 (i, j, k) = ∆tσr (i+ )
ar (i+ ) − H̄θ (i, j, k) +
ar (i+ ) + 2
2
"
1
Chte(i, j, k) − ar (i+ ) [Ern (i, j, k + 1) − Ern (i, j, k)] +
∆φ sin(j∆θ)
!
σr (i+ ) n n
[Ers (i, j, k + 1) − Ers (i, j, k)] +

#)
Ēφn (i + 1, j, k) − Ēφn (i, j, k)
, (5.69)
∆r
∆t
Chte(i, j, k) = ;
µ(i, j, k)

(
∆tσr (i+ )

n+ 1 1 + n− 12
H̄φ 2 (i, j, k) = ∆tσr (i+ )
ar (i ) − H̄φ (i, j, k) +
ar (i+ ) + 2
2
"
Ēθn (i + 1, j, k) − Ēθn (i, j, k)
Chpe(i, j, k) − +
∆r

1
ar (i+ ) [Ern (i, j + 1, k) − Ern (i, j, k)] +
∆θ
!#)
σr (i+ ) n n
[Ers (i, j + 1, k) − Ers (i, j, k)] , (5.70)

∆t
Chpe(i, j, k) = ;
µ(i, j, k)
 96

 
n+ 1 1 n+ 21 − τ∆t n+ 12
Hθ,φ 2 (i, j, k) = H̄θ,φ (i, j, k) − (1 − e 0 )Rθ,φ (i, j, k) ,
(i) (5.71)
br (i)
n+ 1 n+ 1 − τ∆t n
Rθ,φ 2 (i, j, k) = H̄θ,φ 2 (i, j, k) + e 0(i) Rθ,φ (i, j, k),
1
2
Rθ,φ (i, j, k) = 0;

and

n n−1
Ers (i, j, k) = Ers (i, j, k) + ∆tErn (i, j, k). (5.72)

At each time step, we calculate the discrete summation in equation 5.72 from electric fields before

the magnetic field updates in equations (5.69-5.71).

5.5 PML Reflection Analysis in Continuous Medium

In this section, the PML reflection coefficient is derived in continuous space. This will later act

as benchmark for comparing the absorption peromance of different PML and ABC formulations

given above. Ideally, if the PML medium extends to infinity in continuous space, the reflection

coefficient is zero. However, for practical PML implementation in methods such as FDTD, PML

thickness is finite and a perfect electric conductor (PEC) wall terminates the grid as shown in figure

5.3. As shown in the figure, the PML region stretches from radius r0 to r1 . The incident wave gets

attenuated while passing through the PML, reaches the PEC wall and gets reflected. This reflected

wave travels towards the inner region. Although it gets attenuated in the process agian, it is of

finite amplitude. Therefore, a non-zero reflected wave will appear in the inner region because of the

PEC wall backing PML. The reflected wave, and by extension, the associated reflection coefficient

can be mathematically derived. A similar analysis for continuous-space cylindrical PML is given

in [72].

The incident wave for T Er case (transverse electric to radial direction) in the inner region in

terms of spherical elementary functions dealt in chapter 4 and [60] can be written as:

0 −(mφ)
Eφ+ = −β sin θ h(2) m
n (βr)Pn (cos θ)e , (5.73)
 97

PML
Region
Origin ‫ݎ‬ଵ

‫ݎ‬଴

PEC
Wall

Figure 5.3: The position of PML region and PEC wall relative to origin.

(2)
where, hn (βr) represents an outward traveling wave.

Because of the stretching along the radial direction as given in equation 5.30, the incident

wave in the PML region transforms to:

 ˆ r 
0
+
Eφ,P ML = −β sin θh(2)
n β(r0 + s(r, ω) dr ) Pnm (cos θ)e−(mφ) ,
0
(5.74)
r0

where r0 is the inner radius of the PML.

The reflected wave from the PEC backing the PML at radius r1 is given by:
 ˆ r 
− 0
Eφ,P ML =k·β sin θh(1)
n β(r0 + s(r, ω) dr ) Pnm (cos θ)e−(mφ) .
0
(5.75)
r0

(1)
Here, hn (βr) represents an outward traveling wave. The dimension-less constant k can be

found from the tangential boundary condition at the PEC as follows:

(2)
 ´r 
hn β(r0 + r01 s(r, ω) dr0 )
k=
(1)
 ´r , (5.76)
hn β(r0 + r01 s(r, ω) dr0 )

where, r1 is the outer radius of the PML (PEC wall).

Thus, the reflection coefficient is given by:

(2)
 ´ r1 0)

− h β(r + s(r, ω) dr (1)
Eφ,P M L |r=r0 n 0 r0 hn (βr0 )
Γ= = − ´r · . (5.77)
Eφ+ |r=r0
 
(1) (2)
hn β(r0 + r01 s(r, ω) dr0 ) hn (βr0 )
 98

This analytical reflection coefficient is used as benchmark for the PML/ABC formulations

discussed in earlier sections for spherical FDTD, in the following section where PML performance

is analyzed.

5.6 PML Performance Analysis

The formulations in the above sections (split-field PML, absorbing shell and stretched-

coordinate based PML) are used to truncate a free space region in spherical FDTD, and the

reflection errors are analyzed for different PML/ABC parameters. In this section, the absorption

performance of these formulations are compared. The continuous-space reflection coefficient for

spherical PML in equation 5.77 derived in the previous section is used as benchmark for compari-

son.

5.6.1 Geometric Grading of PML Parameters

For implementation of PML in an FDTD scheme (decretized space), the PML parameters

such as σ and a in equation 5.31 are changed gradually from one FDTD PML cell to the next,

instead of an abrupt change. This minimizes the numerical reflections [12]. These parameters’

values are changed typically from 0 (at the interface between inner region and PML region) to a

maximum value just before the PEC backing. For example, these maximum values are given in

the equations 5.78 as σmax and amax . However in continous space, even an abrupt change of the

PML parameters will work (not cause any spurious refelection) as long as the impedence match
σ∗
condition ( σ = µ ) is satisfied. In the following equations, this gradual change is implemented

using polynomial grading as in [12].

 nσ
r − r0
σr (r) = σmax (5.78)
N ∆r
r − r0 na


ar (r) = 1 + (amax − 1) (5.79)
N ∆r
N ∆r r − r0 nσ +1
 
∆r (r) = σmax (5.80)
nσ + 1 N ∆r
 99
 na +1
N ∆r r − r0
br (r) = r + (amax − 1) (5.81)
na + 1 N ∆r

Here nσ and na are the orders of grading for PML parameters σr and ar , respectively, and N

is the number of PML layers. σmax and amax are the maximum values of PML parameters σ and a

and they will reach their respective maximum values towards the end of the PML region (near the

PEC wall). For example, the variation of conductivity σr in PML region for different orders nσ is

shown in figure 5.4.

0.9
n =1
0.8 n =2

0.7 n =3
n =4
0.6 n =5
r

0.5

0.4

0.3

0.2

0.1

0
51 52 53 54 55 56 57 58 59 60 61 62
Cell Index in PML

Figure 5.4: Variation of conductivity σr in the PML region for different orders nσ

5.6.2 Simulation Space and Parameters

This section discusses the simulation space and parameters chosen to study the PML perfor-

mance. The problem space in figure 5.5 shows the positions of source and observation, and sizes

of inner non-PML and outer PML regions simulation parameters are listed in table 5.1. A much

larger simulation as shown in figure 5.6 is run to get the reference fields that are not contaminated

by any reflection from grid truncation.

The normalized reflection error, defined as the normalized maximum difference between the

fields from reference simulation (of problem space shown in figure 5.6) and from the test problem

at hand (in figure 5.5), can be calculated according to the equation 5.82. This normalized error in
 100
Table 5.1: Simulation parameters for PML/ABC implementations in spherical FDTD

Variable Value
Resolution 40 cells/wavelength at f = 1 GHz (ω = 2πf )
Number of grid cells in r, θ, φ (Nr , Nθ , Nφ ) = (61, 13, 17). Nr includes PML layers
directions
Source Location 2nd cell (Solid spherical shell close to origin shown in
figure 5.5)
Observation Location 31st cell
Number of PML Layers (N ) 10
amax 0

discretized PML implementation is equivalent of reflection coefficient in continuous-space PML.

  !
Eφ,test − Eφ,ref 
max
N ormalized Ref lection Error, dB = 20 log  
Eφ,ref  (5.82)
max

KďƐĞƌǀĂƚŝŽŶ
ϭ͘Ϯϱʄ
ϰϱϬ Ϭ͘ϳϱʄ

^ŽƵƌĐĞ

WD>
Ϭ͘Ϯϱʄ

Figure 5.5: Problem Space for Spherical FDTD with PML

Two similar yet subtly different phenomena contribute to the normalized error in discrete

PML implementation:

(1) reflections from PEC wall backing the PML region and,

(2) reflections form layers in discretized PML because of finite change in PML parameters,

given in equations 5.78-5.81, from layer to layer.

 101

KďƐĞƌǀĂƚŝŽŶ

ϭϮ͘ϱʄ
ϰϱϬ
Ϭ͘ϳϱʄ

^ŽƵƌĐĞ

Figure 5.6: Problem space for reference simulation in spherical FDTD

The PML parameters such as the polynomial orders nσ , na and σmax can be tweaked to

minimize the reflection error caused by above phenomena as is done in Cartesian FDTD PML

implementation.

5.6.3 Performance Analysis and Comparison

To begin the PML performance analysis, the split-field PML formulation for spherical FDTD

given in section 5.2 and the Absorbing Shell boundary condition for spherical FDTD given in section

5.3 are implemented numerically. And, the normalized reflection error given in equation 5.82 is

compared for these two formulations in figure 5.7. The figure shows how the normalized reflection

error varies with maximum conductivity σmax of equation 5.78 at different elevations angles θ in

the grid. Since the excitation and the problem space are spherically symmetric, it is expected that

the PML performance doesn’t depend on elevation angle θ or azimuth angle φ as can be observed

in the figure.

However, surprisingly, the comparison shows that the performance of about -30 dB reflection

for split-field formulation in spherical FDTD is identical to that of absorbing shell boundary con-
 102
-10

Split-Field PML, =1050

Normalized Difference, dB Split-Field PML, =450
-15
Absorbing Shell, =1050
Absorbing Shell, =450

-20

-25

-30
10-1 100 101 102
max
, S/m

Figure 5.7: Comparison of Refection Error for Split-Field PML and Absorbing Shell ABC for
nσ = 2

dition. This is in contrast to less-than -50dB of reflection errors easily attainable with split-field

PML formulation in Cartesian FDTD. The reason why split-field PML performance is not at par

with its Cartesian coordinate counterpart is analyzed in later sections of the chapter.

Next, the stretched-coordinate PML formulation for spherical FDTD given in section 5.4

is implemented numerically for the same problem space and relevant simulation parameters as

above. Figure 5.8 shows how normalized reflection error varies with respect to σmax for stretched-

coordinate PML formulation. The minimum reflection error possible with stretched-coordinate

PML for nσ = 2 is about -80 dB when compared to -30 dB from split-field PML formulation. The

figure also shows how the reflection error varies as nσ is changed.

Also, similar to Cartesian coordinate PML case, there is an optimum σmax for every nσ

at which the reflection error is minimum. Below this optimum value, the reflection error quickly

increases with decreasing σmax . Above this optimum value, the reflection error increases albeit

slowly. This is because, for the former case where σmax is low, the incident wave is not attenuated

enough in PML region by the time it reaches PEC backing, where it gets reflected. While for the

latter case where σmax is high, the reflections caused by the descretized PML parameters dominate
 103
0

Normalized Difference, dB
-20

-40

-60

-80

-100
10-1 100
max
, S/m

Figure 5.8: Comparison of Reflection Error for stretched-coordinate PML for various PML poly-
nomial orders nσ

the reflection from the PEC wall. Moreover, it is shown in figure 5.9 that in continuous PML, i.e.,

in the absence of descretization, the reflection error continues to decrease with increasing σmax .

This is because, having traveled through highly lossy medium, there is only so much incident wave

amplitude left to get reflected by the PEC backing.

20
Normalized Reflection Error, dB

0
-20
-40
-60
-80
-100
-120
-140
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
max
, S/m

Figure 5.9: Reflection Error for different mode numbers n and m = 0 in Continuous and FDTD
PML cases.
 104

Figure 5.9 compares the reflection error from stretched-coordinate PML for spherical FDTD

and the continuous-space reflection coefficient given by equation 5.77 for various mode excitations

(with variable n and m = 0). As mentioned before in table 5.1, the source is a spherical shell

of radius 2∆r, where ∆r is the radial discretization step. The sourcing is done by adding a time

series to a field component (soft sourcing) as shown in equation 5.83, that acts as a current source.

And as shown in the equation, the soft source has a θ-dependent profile to excite a certain n mode

(n = 1 case is showed in the equation as an example).

  
1
Hr |q(2,j,:) = Hr |n(2,j,:) + cos j− ∆θ · f (q∆t), (5.83)
2
1
f (q) = [10 − 15 cos(ωq∆t) + 6 cos(2ωq∆t) − cos(3ωq∆t)] . (5.84)
32

Here j is index for elevation angle θ, and ∆θ is discretized elevation angle step. And, f (q∆t)

is the time series, here shown as Blackman-Harris waveform in equation 5.84, as an example, where

q represents the qth time step and ∆t is dicretization time step of the FDTD simulation. This
2π
waveform is a periodic waveform containing pulses of period ω . Only one such pulse is introduced

to let it propagate outward and observe it after getting reflected by PML, as in a pulsed-radar. As
2π
such, the sourcing is time-gated, i.e. introduced only in the time limits 0 < n∆t < ω .

Associated Legendre functions for different n are given in table 5.2, and mode n can be

excited in the FDTD grid by introducing respective θ profile in equation 5.83.

Table 5.2: Legendre Polynomials Pnm (cos θ) for some lower order n and m = 0

n Pn0 (cos θ)
1 cos θ
1 2
2 2 (3 cos θ − 1)
1 2
3 2 cos θ(5 cos θ − 3)

5.7 Why Split-Field PML doesn’t work in Spherical FDTD?

In Cartesian FDTD, the split-field PML formulation will lead to stretched-coordinate PML

formulation and vice-versa as shown in the equations below. Starting with scalar Maxwell’s differ-
 105

ential equation for Ey and its split-field notation, we arrive at the stretched-coordinate notation

for Ey .

∂Hx ∂Hz
(ω + σ) Ey = − (5.85)
∂z ∂x
∂Hx 1 ∂Hx
ωEyz + σz Eyz = <=> ωEyz = (5.86)
∂z Sz ∂z
∂Hz 1 ∂Hz
ωEyx + σx Eyx = − <=> ωEyx = − (5.87)
∂x Sx ∂x

Adding the later parts of the above two equations we arrive at the following equation:

1 ∂Hx 1 ∂Hz
ωEy = − . (5.88)
Sz ∂z Sx ∂x

Starting with Maxwell’s equations in complex stretched-coordinate space and then modifying

it to include stretching variable as shown below, we arrive at equation 5.91 which is the same as

equation 5.88 .

∂Hx ∂Hz
ωEy = − (5.89)
∂ z̃ ∂ x̃
∂ 1 ∂ ∂ 1 ∂
= , = (5.90)
∂ x̃ Sx ∂x ∂ z̃ Sz ∂z
1 ∂Hx 1 ∂Hz
ωEy = − (5.91)
Sz ∂z Sx ∂x

where,
 
σx,z
Sx,z = 1+ ω ,
x
x̃ = ∫ Sx (x0 ) dx0 , and
0
y
ỹ = ∫ Sy (y 0 ) dy 0 .
0

Therefore, split-field and stretched-coordinate PML formulation in Cartesian coordinates are

identical although their interpretation and mathematical representations are different.

On the other hand for the PML in spherical FDTD, however, the split-field and stretched-

coordinate formulations do not lead to one another as shown below. Starting with split-field

equations 5.7 and 5.8 for Eθ , and using the stretched-coordinates variable sr , we get the field
 106

equation for Eθ in stretched-coordinates as shown in equation 5.98 following the process below.

1 ∂Hr
ωEθφ = (5.92)
r sin θ ∂φ
 
σr 1 ∂ (rHφ )
ω 1 + Eθr = − (5.93)
ω r ∂r
 
σr (r)
sr = 1+ (5.94)
ω
1 1 ∂ (rHφ )
ωEθr = − (5.95)
sr r ∂r
1 ∂Hr
ωEθφ = (5.96)
r sin θ ∂φ
1 ∂Hr 1 1 ∂ (rHφ )
ωEθ = − (5.97)
r sin θ ∂φ sr r ∂r
1 ∂ (sr · Hr ) ∂ (rHφ )
ωsr rEθ = − (5.98)
sin θ ∂φ ∂r

Whereas, if we start with Maxwell’s equation in complex stretched-coordinate space as in

equation 5.28 repeated here in equation 5.99 and modify it by introducing stretched-coordinate

variable Sr , we arrive at equation 5.102. This is essentially different from equation 5.98 that we

arrived at from the split-field formulation in the sense that radius r in equation 5.98 is real whereas

the r̄ in equation 5.102 is a complex (frequency dependent) stretched-coordinate radius.

 
1 1 ∂ ∂
ωEθ (ω) = (Hr (ω)) − (r̄Hφ (ω)) (5.99)
r̄ sin θ ∂φ ∂ r̄
∂ 1 ∂
= (5.100)
∂ r̃ Sr ∂r
 
1 1 ∂ ∂
ωEθ (ω) = (Hr (ω)) − (r̄Hφ (ω)) (5.101)
r̄ sin θ ∂φ Sr ∂r
1 ∂ (Sr · Hr ) ∂ (r̃Hφ )
ωSr r̃Eθ = − (5.102)
sin θ ∂φ ∂r

Thus, the split-field PML formulation for spherical FDTD doesn’t truly reflect the complex

stretched-coordinates as it does for the Cartesian case.

 Chapter 6

Conclusion

This thesis introduced some advanced formulations and applications of finite difference time

domain method while reviewing the existing methods.

The second chapter dealt with the development of a plane wave excitation formulation for

FV24, a finite-volume based higher-order FDTD variant. Now that the leakage-free and efficient

plane wave excitation method is available for FV24, and owing to its excellent phase error char-

acteristics, FV24 can be applied to electrically large problems such as a ship or an aircraft for

studying their scattering cross section, or for understanding propagation of mobile radio waves in

urban setting etc.

The plane wave excitation method based on total field/scattered field formulation and Dis-

crete planewave technique is reviewed for conventional FDTD before implementing it for FV24. In

the process, the concept of auxiliary 1D grid is elucidated clearly by showing its relation to the main

FDTD 3D grid. First, it is represented as an imaginary line along the direction of propagation of

plane wave. Then, the electric and magnetic field locations of the main grid are projected on to this

imaginary line forming the auxiliary 1D grid. It is shown that the 1D grid has co-located electric

fields separately from co-located magnetic fields. Multiple electric field locations (or magnetic field

locations) on the main 3D grid will have projection at the same location on the 1D grid. This

visual demonstration of the 1D grid and its properties, absent in the literature, will greatly help

in understanding the technique. Subsequently, the indexing for 1D update equations which help

propagate plane along the 1D auxiliary grid is demonstrated. It is pointed out why such a 1D grid
 108

will have the same dispersion relation as that of the main 3D grid.

Since FV24 computational stencil extends multiple Yee cells, the consistency corrections

needed to implement TF/SF formulation become complex and confusing, especially at the edges

and corners of the TF/SF interface. Hence the consistency corrections are dealt with systematically

and visually shown to aid easy understanding and replication. Initiation of the plane wave on the

auxiliary 1D grid by hard sourcing initial grid points is equally complex for FV24 given its extended

stencil. It is also well explained to avoid any pitfalls. Finally, the leakage free plane wave excitation

is demonstrated. The spurious leakage into the scattered field is below -300 dB that is same as the

noise floor set by the machine precision. It is made possible because of the numerical dispersion

match between the auxiliary 1D and the 3D grids.

In the later parts of this chapter, different near to farfield transformation approaches possible

in FDTD are compared for accuracy. The spatial interpolation of fields is required in FDTD to

calculate fictitious currents on the equivalent surface, that in turn are used to calculate farfields.

Different interpolation techniques such as the arithmetic averaging, geometric mean, Mixed-surface,

and separate-surface approaches are possible. The performance of these different interpolations in

estimating the farfields is compared. The empty scatterer case that ideally produces zero farfield

sets the baseline and is selected to compare the performance of above interpolation schemes. It is

shown that the arithmetic averaging and Mixed-surface interpolation schemes perform better than

the other two. It is also validated by calculating the radar cross-section of a canonical dielectric

cube using different FDTD interpolation techniques and comparing them with commercial solvers

such as FEKO and HFSS.

Therefore, we can combine the plane wave excitation method for FV24 dealt here and the

Mixed-surface or arithmetic averaging approach for performing near to farfield transformation, to

analyze radar cross section or radiation pattern of electrically large structures accurately.

The third chapter deals with application of FDTD to an interesting problem encountered in

printed circuit boards. The composition of composite dielectric materials used in PCBs, and the

root cause of the glass weave-induced skew are reviewed first. Then it is noted that the glass weave-
 109

induced skew in a differential transmission line can cause a variety of problems such as increased bit

error rates affecting the robustness of the digital system, and increased radiated emissions causing

compliance failures. This research applies auxiliary differential equation (ADE) formulation for

FDTD, that helps model frequency dependence of material properties of real materials, to estimate

glass weave-induced skew in a differential pair.

The implementation of ADE method is demonstrated using multi-pole Debye model. At

first, the broadband profile of real and imaginary parts of complex permittivity are obtained using

Djordjevic-Sarkar model using minimal input data points. Real and imaginary parts at just a single

frequency point often suffice to generate broadband profiles. Then the broadband Debye coefficients

are obtained by fitting multi-pole Debye model to Djordjevic-Sarkar model. The FDTD update

equations are subsequently modified to account for the multi-pole Debye model.

The geometric model of glass weave is implemented mathematically and the FDTD cells in the

grid are assigned different materials such as glass, resin, PEC (for microstrip traces), and free space.

This segregates the dispersive-material cells from the non-dispersive cells. The update equations at

the dispersive cells are modified to include the multi-pole Debye model using the ADE formulation.

The glass-weave induced skew is calculated for a sample differential pair and validated with the

results from HFSS. Also, the performance in terms of memory and time consumption are compared.

The advantages of using FDTD method for glass weave skew study are elucidated. Finally, the use of

graphical processing units (GPUs) to accelerate the skew simulations in MATALB is demonstrated.

The fourth chapter deals with the derivation of numerical dispersion relation for spherical

FDTD, sensitivity study of the associated numerical wave number, and the stability analysis of

spherical FDTD. At first, the structure of spherical grid, field locations, and update equations

are reviewed. The non-uniformity of spherical grid is visualized particularly near the origin. The

numerical dispersion relation for Cartesian FDTD is reviewed initially. Later, the derivation process

is generalized and applied to spherical FDTD.

The electric and magnetic fields are first given in terms of the elementary functions native

to spherical coordinate system. Then, they are substituted in spherical FDTD update equations.
 110

Finally. the resultant equations are reduced to obtain the numerical dispersion relation. The nu-

merical dispersion relationship thus obtained is position dependent demonstrating the non uniform

nature of the spherical FDTD grid.

This numerical dispersion relation is solved using MATLAB to obtain the complex numerical

wave number. It is observed that as we move close to the origin, the numerical wave number

deviates from that of the free space wave number. This signifies that the wave propagation in the

spherical FDTD grid is affected. To model the wave behavior at the origin with required accuracy,

the grid parameters such as resolution and time step need to be chosen with care. As we move

away from the origin, the numerical wave number converges to the free space wave number.

Later, the sensitivity of the numerical wave number to various grid parameters such as reso-

lution, elevation angle and mode numbers is studied. In the later part of the chapter, an attempt is

made to derive the stability criterion for the spherical FDTD grid based on the position dependent

numerical dispersion relation derived earlier in the chapter. The stability criterion available in the

literature are reviewed in the current study and the challenges faced during the validation of the

stability criterion are documented.

The last chapter deals with the implementation of the perfectly matched layer in spheri-

cal FDTD. Firstly, the split-field formulation and the associated update equations are elucidated.

Secondly, the stretched-coordinates formulation and the associated update equations are derived.

Finally, an absorbing boundary condition based on lossy dielectric and the associated update equa-

tions are obtained. The effectiveness of the above three formulations in absorbing the waves in

spherical FDTD is compared. It is observed that the split-field formulation performs only as good

as the lossy dielectric- based absorbing boundary condition, while the stretched-coordinates for-

mulation performs better than the rest of the two. This is in contrast to the Cartesian coordinate

PML, where both the split-field and stretched-coordinates formulation perform identically.

The reflection coefficient of the continuous space PML backed by a PEC is derived and is

used as a benchmark for the above formulations. It is observed that the performance of stretched-

coordinates PML implemented for spherical FDTD compares well with that of the continuous space
 111

spherical PML until the numerical errors in spherical FDTD start to degrade the performance of

the spherical FDTD PML.

In the final section of the chapter, it is shown that the split-field formulation and the stretched-

coordinate formulation of the PML in spherical FDTD are not identical, whereas in Cartesian

coordinated system they are identical. This explains the reason, why the split-field PML formulation

in spherical FDTD is not as effective in absorbing the incident waves as in Cartesian coordinate

system.

6.1 Original Contributions

The original contribution of the dissertation work are:

• The plane wave excitation technique for FV24, a finite-volumes based higher order FDTD,

based on total field/scattered field (TF/SF) formulation and Discrete plane wave formu-

lation. Associated 1D grid update equations, TF/SF consistency corrections, wave initial-

ization technique on the 1D grid and mapping of 1D grid locations to the 3D grid location

and vice-versa for FV24.

• The visualization of auxiliary 1D grid, used in the above technique, as a set of projections

of electric and magnetic field locations on the main grid.

• Technique for comparing the accuracy of various field interpolations techniques available

in the literature that are used in the implementation of near to farfield transformation in

FDTD, and its implementation.

• Technique to obtain glass weave-induced skew using FDTD by combining the implemen-

tation of glass weave structure and dispersive material characterization using auxiliary

differential equation method for multi-pole Debye model. The multi-pole Debye coeffi-

cients needed for dielectric materials are obtained by curve-fitting it to Djordjevic-Sarkar

model.
 112

• Technique to compare glass weave-induced skew extracted from S-parameters obtained from

HFSS simulations, and that obtained from FDTD. The same Gaussian pulse that is used

as excitation in FDTD is used to window the S-parameters from HFSS and then the pulse

response is calculated from inverse Fourier transforming the windowed S-parameters.

• Visualization of spherical FDTD grid and various special cells near poles and origin.

• Derivation of numerical dispersion relation in spherical FDTD for T Mr modes.

• Method to observe stability criterion for spherical FDTD at different distances from origin

by making the cells close to the origin as part of PEC.

• Derivation of central difference update equations based on stretched-coordinate PML for-

mulation for spherical FDTD.

• Technique to excite various modes in spherical FDTD grid.

• Comparison of effectiveness of stretched-coordinate based PML, lossy dielectric-based ab-

sorbing shell, and split-field PML formulations for spherical FDTD.

• Derivation of continuous-space reflection coefficient of spherical PML as a benchmark for

comparing performance of various truncation techniques for spherical FDTD.

• Derivation that shows split-field and stretched-coordinate PML formulations in spherical

coordinates are not identical.

 Bibliography

[1] A. Z. Elsherbeni and V. Demir, The Finite-Difference Time-Domain Method for elec-
tromagnetics with MATLAB Simulations. Raleigh, NC: SciTech Publishing, 2009.

[2] M. J. Piket-May, R. C. Bollimuntha, M. F. Hadi, and A. Z. Elsherbeni, “Dispersion optimised

plane wave sources for scattering analysis with integral based high order finite difference time
domain methods,” IET Microwaves, Antennas Propag., vol. 10, no. 9, pp. 976–982,
2016. [Online]. Available: http://digital-library.theiet.org/content/journals/10.1049/iet-map.
2015.0755

[3] Altera Corporation, “PCB Stackup Design Considerations for Altera FPGAs,” Tech. Rep.,
2010. [Online]. Available: https://www.altera.com/content/dam/altera-www/global/en US/
pdfs/literature/an/an613.pdf

[4] K. S. Yee, “Numerical Solution of Initial Boundary Value Problems Involving Maxwell’s Equa-
tions in Isotropic Media,” IEEE Trans. Antennas Propag., vol. 14, no. 3, pp. 302–307, mar
1966.

[5] J.-P. Bérenger, “A Perfectly Matched Layer for the Absorption of Electromagnetic Waves,” J.
Comput. Phys., vol. 114, no. 2, pp. 185–200, 1994.

[6] S. D. Gedney, “An Anisotropic Perfectly Matched Layer-Absorbing Medium for the Truncation
of FDTD Lattices,” IEEE Trans. Antennas Propag., vol. 44, no. 12, pp. 1630–1639, dec
1996.

[7] W. C. Chew, J. M. Jin, and E. Michielssen, “Complex Coordinate System as a Generalized

Absorbing Boundary Condition,” in IEEE Antennas Propagat. Soc. Int. Symp., vol. 3,
Montréal, Canada, jul 1997, pp. 2060–2063.

[8] J. A. Roden and S. D. Gedney, “Convolutional PML(C-PML): an efficient FDTD implemen-

tation of the C-PML for arbitrary media,” Microw. Opt. Technol. Lett., vol. 27, no. 5, pp.
334–339, 2000.

[9] T. Kashiwa and I. Fukai, “A Treatment by the FD-TD Method of the Dispersive Characteristics
Associated with Electronic Polarization,” Microw. Opt. Technol. Lett., vol. 3, no. 6, pp.
203–205, 1990.

[10] A. Z. Elsherbeni and V. Demir, The Finite-Difference Time-Domain in Electromag-

netics, ser. Electromagnetic Waves. Institution of Engineering and Technology, 2015.
[Online]. Available: http://digital-library.theiet.org/content/books/ew/sbew514e
 114

[11] D. E. Merewether, R. Fisher, and F. W. Smith, “On Implementing a Numeric Huygen’s Source
Scheme in a Finite Difference Program to Illuminate Scattering Bodies,” IEEE Trans. Nucl.
Sci., vol. 27, no. 6, pp. 1829–1833, dec 1980.
[12] A. Taflove and S. Hagness, Computational Electrodynamics: The Finite-Difference
Time-Domain Method, 3 ed. Boston, MA: Artech House, 2005.
[13] T. Tan and M. Potter, “1-D multipoint auxiliary source propagator for the total-
field/scattered- field FDTD formulation,” IEEE Antennas Wirel. Propag. Lett., vol. 6,
pp. 144–148, 2007.
[14] K. Umashankar and A. Taflove, “A Novel Method to Analyze Electromagnetic Scattering of
Complex Objects,” IEEE Trans. Elec. Compat., vol. 24, no. 4, pp. 397–405, nov 1982.
[15] R. J. Luebbers, K. S. Kunz, M. Schneider, and F. Hunsberger, “A Finite-Difference Time-
Domain Near Zone to Far Zone Transformation,” IEEE Trans. Antennas Propag., vol. 39,
no. 4, pp. 429–433, apr 1991.
[16] J. B. Schneider and C. L. Wagner, “FDTD Dispersion Revisited: Faster-Than-Light Propaga-
tion,” IEEE Microw. Guid. Wave Lett., vol. 9, no. 2, pp. 54–56, feb 1999.
[17] M. F. Hadi and N. B. Almutairi, “Discrete Finite-Difference Time Domain Impulse Response
Filters for Transparent Field Source Implementations,” IET Microw. Antennas Propag.,
vol. 4, no. 3, pp. 381–389, mar 2010.
[18] M. Piket-May, A. Taflove, and J. Baron, “FD-TD Modeling of Digital Signal Propagation in
3-D Circuits with Passive and Active Loads,” IEEE Trans. Microw. Theory Tech., vol. 42,
no. 8, pp. 1514–1523, aug 1994.
[19] J. Mix, J. Dixon, Z. Popovic, and M. Piket-May, “Incorporating Non-Linear Lumped Elements
in FDTD: The Equivalent Source Method,” Int. J. Numer. Model. Electron. Networks,
Devices Fields, vol. 12, no. 1–2, pp. 157–170, 1999.
[20] J. B. Schneider, “Plane Waves in FDTD Simulations and a Nearly Perfect Total-
Field/Scattered-Field Boundary,” IEEE Trans. Antennas Propag., vol. 52, no. 12, pp.
3280–3287, dec 2004.
[21] T. Tan and M. Potter, “Optimized analytic field propagator (O-AFP) for plane wave injection
in FDTD simulations,” IEEE Trans. Antennas Propag., vol. 58, no. 3, pp. 824–831, 2010.
[22] ——, “FDTD Discrete Planewave (FDTD-DPW) Formulation for a Perfectly Matched Source
in TFSF Simulations,” IEEE Trans. Antennas Propag., vol. 58, no. 8, pp. 2641–2648, aug
2010.
[23] M. F. Hadi, “A Finite Volumes-Based 3-D Low Dispersion FDTD Algorithm,” IEEE Trans.
Antennas Propag., vol. 55, no. 8, pp. 2287–2293, aug 2007.
[24] J. Fang, “Time Domain Finite Difference Computation for Maxwell’s Equations,” Ph.D. dis-
sertation, University of California at Berkeley, Berkeley, CA, 1989.
[25] E. Turkel, “High-Order Methods,” in Adv. Comput. Electrodyn. The Finite-Difference
Time-Domain Method, A. Taflove, Ed. Boston, MA: Artech House, 1998, ch. 2, pp.
63–110.
 115

[26] A. Yefet and P. G. Petropoulos, “A Staggered Fourth-Order Accurate Explicit Finite Difference
Scheme for the Time-Domain Maxwell’s Equations,” J. Comput. Phys., vol. 168, no. 2, pp.
286–315, 2001.

[27] T. T. Zygiridis and T. D. Tsiboukis, “Development of Higher Order FDTD Schemes with
Controllable Dispersion Error,” IEEE Trans. Antennas Propag., vol. 53, no. 9, pp. 2952–
2960, sep 2005.

[28] M. F. Hadi and M. Piket-May, “A Modified FDTD (2,4) Scheme for Modeling Electrically
Large Structures with High-Phase Accuracy,” IEEE Trans. Antennas Propag., vol. 45,
no. 2, pp. 254–264, feb 1997.

[29] R. Holland and J. W. Williams, “Total-Field Versus Scattered-Field Finite-Difference Codes:

A Comparative Assessment,” IEEE Trans. Nucl. Sci., vol. NS-30, no. 6, pp. 4583–4588, dec
1983.

[30] W. C. Chew and W. H. Weedon, “A 3D Perfectly Matched Medium from Modified Maxwell’s
Equations with Stretched Coordinates,” Microw. Opt. Technol. Lett., vol. 7, no. 13, pp.
599–604, sep 1994.

[31] C. M. Rappaport, “Perfectly Matched Absorbing Boundary Conditions Based on Anisotropic

Lossy Mapping of Space,” IEEE Microw. Guid. Wave Lett., vol. 5, no. 3, pp. 90–92, mar
1995.

[32] M. Kuzuoglu and R. Mittra, “Frequency Dependence of the Constitutive Parameters of Causal
Perfectly Matched Anisotropic Absorbers,” IEEE Microw. Guid. Wave Lett., vol. 6, no. 12,
pp. 447–449, dec 1996.

[33] S. D. Gedney, “Perfectly Matched Layer Absorbing Boundary Conditions,” in Comput. Elec-
trodyn., 3rd ed., A. Taflove and S. C. Hagness, Eds. Boston, MA: Artech House, 2005, ch. 7,
pp. 273–328.

[34] M. F. Hadi, “Near-Field PML Optimization for Low and High Order FDTD Algorithms Using
Closed-Form Predictive Equations,” IEEE Trans. Antennas Propag., vol. 59, no. 8, pp.
2933–2942, aug 2011.

[35] M. A. Kourah, M. F. Hadi, and A. S. Al-Zayed, “Extending the Enlarged Cell and Uniformly
Stable Techniques to Modeling Curved Conductors in Two-Dimensional High-Order Finite-
Difference Time-Domain Algorithms,” Inte. J. Nume. Model. Electron. Networ. Devices
Fields, vol. 26, no. 3, pp. 238–250, 2013.

[36] M. F. Hadi, “Simplified Graphical Processor Acceleration of the Standard and High-Order
Finite-Difference Time-Domain Algorithms,” Electromagnetics, vol. 32, no. 7, pp. 401–410,
oct 2012.

[37] T. Tan and M. Potter, “On the nature of numerical plane waves in FDTD,” IEEE Antennas
Wirel. Propag. Lett., vol. 8, no. 1, pp. 505–508, 2009.

[38] A. Taflove, Ed., Advances in Computational Electrodynamics: The Finite-Difference

Time-Domain Method. Boston, MA: Artech House, 1998.
 116

[39] K. Abdijalilov and J. B. Schneider, “Analytic field propagation TFSF boundary for FDTD
problems involving planar interfaces: Lossy material and evanescent fields,” IEEE Antennas
Wirel. Propag. Lett., vol. 5, no. 1, pp. 454–458, 2006.

[40] C. A. Balanis, Advanced Engineering Electromagnetics. New York: John Wiley, 1989.

[41] D. J. Robinson and J. B. Schneider, “On the use of the geometric mean in FDTD near-to-
far-field transformations,” IEEE Trans. Antennas Propag., vol. 55, no. 11, pp. 3204–3211,
nov 2007.

[42] T. Martin, “An improved near- to far-zone transformation for the finite-difference time-domain
method,” IEEE Trans. Antennas Propag., vol. 46, no. 9, pp. 1263–1271, sep 1998.

[43] D. J. Vaughan, HANDBOOK OF COMPOSITES, S.T. Peters, Ed., 1977, vol. 32, no. 1.
[Online]. Available: http://linkinghub.elsevier.com/retrieve/pii/0010436170903162

[44] A. Morgan, “Developments in Glass Yarns and Fabric Constructions,” PCB Mag., 2014.

[45] D. Brooks, “Differential Signals The Differential Difference !” Print. Circuit Des. C. Publ.,
pp. 1–2, may 2001.

[46] L. W. Ritchey, “A Treatment of Differential Signaling and its Design Requirements,” SPEED-
ING EDGE, Tech. Rep. September, 2008.

[47] A. R. Djordjević, R. M. Biljić, V. D. Likar-Smiljanić, and T. K. Sarkar, “Wideband frequency-

domain characterization of FR-4 and time-domain causality,” IEEE Trans. Electromagn.
Compat., vol. 43, no. 4, pp. 662–667, 2001.

[48] Ansys, “HFSS Online Help,” 2016.

[49] E. Bogatin, B. Hargin, V. S. S. D. T. Paladugu, D. DeGroot, A. Koul, S. Baek, and

M. Sapozhnikov, “New characterization technique for glass-weave skew,” Signal Integr. J.,
no. March, 2017. [Online]. Available: www.signalintegrityjournal.com

[50] R. Bollimuntha, M. Piket-may, D. Paladugu, and E. Bogatin, “An Efficient Method for Glass
Weave Skew Simulations,” University of Colorado Boulder, Tech. Rep., 2017.

[51] C. Herrick, T. Buck, and R. Ding, “Bounding the Effect of Glass Weave through Simulation,”
in DesignCon, 2009.

[52] J. E. Diener and A. Z. Elsherbeni, “FDTD acceleration using MATLAB parallel computing
toolbox & GPU,” Appl. Comput. Electromagn. Soc. J., 2017.

[53] O. Franek, G. F. Pedersen, and J. B. Andersen, “Numerical Modeling of a Spherical Array

of Monopoles Using FDTD Method,” IEEE Trans. Antennas Propag., vol. 54, no. 7, pp.
1952–1963, jul 2006.

[54] M. F. Hadi, A. Z. Elsherbeni, M. J. Piket-May, and S. F. Mahmoud, “Radial Waves Based

Dispersion Analysis of the Body-of-Revolution FDTD Method,” IEEE Trans. Antennas
Propag., vol. 65, no. 2, pp. 721—-729, 2017.
 117

[55] R. Bollimuntha, M. Hadi, M. Piket-May, and A. Elsherbeni, “Numerical dispersion analysis

for spherical FDTD,” in 2018 Int. Appl. Comput. Electromagn. Soc. Symp. Denver,
ACES-Denver 2018, 2018.

[56] M. F. Hadi, R. C. Bollimuntha, A. Z. Elsherbeni, and M. Piket-May, “A spherical FDTD

numerical dispersion relation based on elemental spherical wave functions,” IEEE Antennas
Wirel. Propag. Lett., vol. 17, no. 5, pp. 784–788, 2018.

[57] R. Holland, “THREDS: A Finite-Difference Time-Domain EMP Code in 3D Spherical Coor-

dinates,” IEEE Trans. Nucl. Sci., vol. NS-30, no. 6, pp. 4592–4595, dec 1983.

[58] A. Soriano, E. A. Navarro, D. L. Paul, J. A. Portı́, J. A. Morente, and I. J. Craddock, “Finite

Difference Time Domain Simulation of the Earth-Ionosphere Resonant Cavity: Schumann
Resonances,” IEEE Trans. Antennas Propag., vol. 53, no. 4, pp. 1535–1541, apr 2005.

[59] O. Franek, G. F. Pedersen, and J. B. Andersen, “Numerical modeling of a spherical array

of monopoles using FDTD method,” IEEE Trans. Antennas Propag., vol. 54, no. 7, pp.
1952–1963, 2006.

[60] R. F. Harrington, Time-Harmonic Electromagnetic Fields. New York, NY: McGraw-

Hill, 1961.

[61] J. A. Stratton, Electromagnetic theory. McGraw-Hill Book Company, USA, 1941.

[62] A. Taflove, Computational Electrodynamics: The Finite-Difference Time-Domain

Method. Boston, MA: Artech House, 1995.

[63] A. Taflove and M. E. Brodwin, “Numerical Solution of Steady-State Electromagnetic

Scattering Problems Using the Time-Dependent Maxwell’s Equations,” IEEE Trans.
Microw. Theory Tech., vol. 23, no. 8, pp. 623–630, 1975. [Online]. Available:
http://ieeexplore.ieee.org/lpdocs/epic03/wrapper.htm?arnumber=1128640

[64] Y. Wang, H. Xia, and Q. Cao, “Stability and dispersion analysis of the finite-
difference time-domain algorithms in modelling the earthionosphere system,” IET
Microwaves, Antennas Propag., vol. 5, no. 4, p. 476, 2011. [Online]. Available:
http://digital-library.theiet.org/content/journals/10.1049/iet-map.2010.0184

[65] Y. W. Liu, Y. W. Chen, P. Zhang, and Z. X. Liu, “A spherical higher-order finite-difference

time-domain algorithm with perfectly matched layer,” Chinese Phys. B, vol. 23, no. 12,
2014.

[66] A. Mock, “Compact FDTD formulation for structures with spherical invariance,” IEEE
Trans. Antennas Propag., vol. 59, no. 3, pp. 987–993, 2011.

[67] F. L. Teixeira and W. C. Chew, “PML-FDTD in Cylindrical and Spherical Grids,” IEEE
Microw. Guid. Wave Lett., vol. 7, no. 9, pp. 285–287, sep 1997.

[68] ——, “Systematic Derivation of Anisotropic PML Absorbing Media in Cylindrical and Spher-
ical Coordinates,” IEEE Microw. Guid. Wave Lett., vol. 7, no. 11, pp. 371–373, nov
1997.
 118

[69] W. Bao and F. L. Teixeira, “Performance Analysis of Perfectly Matched Layers Applied to
Spherical FDTD Grids,” IEEE Trans. Antennas Propag., vol. 66, no. 2, pp. 1035–1039,
2018.

[70] W. Bao, “A Simulation and Optimization Study of Spherical Perfectly Matched Layers,” Ph.D.
dissertation, The Ohio State University, 2017.

[71] M. Hadi, A. Elsherbeni, R. Bollimuntha, and M. Piket-May, “FDTD in Cartesian and Spherical
Grids,” in Computational Photonic Sensors, M. F. O. Hameed and S. Obayya, Eds.
Cham, Switzerland: Springer Nature, 2019, ch. 7, pp. 153–175.

[72] J. Maloney, M. Kesler, and G. Smith, “Generalization of PML to Cylindrical Geometries,” in

Annu. Rev. 13th, Prog. Appl. Comput. Electromagn. Monterey, CA: NTIS, 1997,
pp. 900–908.