THE UNIVERSITY OF HULL

AN INVESTIGATION OF ALTERNATING-DIRECTION IMPLICIT

FINITE-DIFFERENCE TIME-DOMAIN (ADI-FDTD) METHOD IN

NUMERICAL ELECTROMAGNETICS

being a Thesis submitted for the degreeof Doctor of Philosophy

at the University of Hull

by
SayCheohOw M.Eng (Hons)
October2003
 ACKNOWLEDGEMENTS

I would like to thank Dr. Sunil Judahfor his assistancethroughoutthis researchwork, my parentsfor their
continuousmoral supportand Dr. David Stubbsfor his patience,support,encouragement,manyusefultechnical
discussionsandfor proof-readingthis thesis.
 CONTENTS

Acknowledgments ii
...................
Abstract vii
.....................
List of symbols Viii
....................

CHAPTER I- INTRODUCTION

1.1 Introduction 1

1.2 FDTD Method 2

..................
1.3 Finite-differenceapproximationto derivatives 3
..........
1.3.1 Explicit method 4
................
1.3.2 Implicit method 4
................
1.4 Alternating-direction implicit procedure 5
............

CHAPTER 2- FINITE-DIFFERENCE TIME-DOMAIN METHOD

2.1 Introduction 7

2.2 Maxwell's equationsin threedimensions 7

......... ..
2.3 FDTD algorithm 8
............... ..
2.4 Finite differenceexpressionof Maxwell's equationsin threedimensions 9
.. ..
2.5 Divergenceof FDTD algorithm 10
............ ..
2.6 Numericalstability of the three-dimensional
FDTD algorithm 11
.... ..
2.7 Numericaldispersionof the three-dimensional
FDTD algorithm 15
.... ..
2.8 Boundaryconditions 16
.............. ..
2.8.1 1" orderMur boundary condition .......... 17
..
2.9 Conductorboundaries 19
.............. ..
2.10 Dielectric boundaries 19
.............. ..
2.11 Excitation 20
................. ..
2.12 Simulationof a line-fed rectangularmicrostrippatch 21
...... ..
2.13 Extraction of voltage and current from the FDTD mesh .... 23
....

2.14 Extraction of Si, 24

............. ....
2.14.1 Two runs 24
............. ....
2.14.2 Singlerun 24
............. ....
2.15 Extraction of Zin 26
............. ....

2.16 Results 28
............... ....

2.17 Conclusion 30
.............. ....

CHAPTER 3- ALTERNATING-DIRECTION IMPLICIT

FINITE-DIFFERENCE TiME-DOMAIN METHOD

3.1 Introduction 31
............... ...
3.2 Three-dimensional
ADI-FDTD algorithm 31
........ ...
3.3 ADI-FDTD method ...
Physicalinterpretationof three-dimensional 35
...
3.4 Divergenceof ADI-FDTD algorithm 36
......... ...
3.5 Numericalstability 36
.............. ...
3.5.1 2-dimensional ADI-FDTD 36
.......... ...
3.5.2 ADI-FDTD
3-dimensional 39
.......... ...
3.6 Numericaldispersion 40
............. ...
3.6.1 2-dimensional
ADI-FDTD 40
.......... ...
3.7 Implementationof I" orderMur absorbingboundarycondition ...
42
...
3.7.1 Boundaryconditionwithinthetri-diagonal
matrix ..... 42
...
3.7.2 Boundaryconditionoutsidethetri-diagonal
matrix 44

3.8 Simulatedresults .............. 45

...
3.8.1 Three-dimensional cavity ........... 46
...
cavitywith inhomogeneous
3.8.2 Three-dimensional media 48
.... ...
3.8.3 Simulationof a line-fedrectangularmicrostrippatch 49
3.8.4 Three-dimensional line .....
cavitywith a transmission 50
...
3.9 Conclusion 51
............... ...
CHAPTER 4- MODIFIED ALTERNATING-DIRECTION IMPLICIT METHOD

4.1 Introduction 52
.............. ....
4.2 modifiedADI-FDTD algorithm .....
Three-dimensional 52
....
4.3 Weightingfactor in the modified ADI-FDTD algorithm 55
.... ....
4.4 Divergenceof modified ADI-FDTD algorithm 56
...... ....
4.5 Numericalstability 56
............. ....
4.5.1 2-dimensional modifiedADI-FDTD 56
...... ....
4.5.2 3-dimensional modifiedADI-FDTD 59
...... ....
4.6 Numericaldispersion 61
............ ....
4.6.1 2-dimensional
modifiedADI-FDTD 61
...... ....
4.7 Simulatedresults 63
............. ....
4.8 Relationshipbetweenattenuationandweightingfactorf 64
... ....
4.9 Conclusion 66
............... ....

CHAPTER 5- SIMULATING COPPER LAYER IN

ALTERNATING-DIRECTION IMPLICIT METHOD

5.1 Introduction 67
................. .
5.2 Three-dimensional
ADI-FDTD algorithmwith electricconductivityterm 67
.. .
5.3 Simulatedresults 71
................ .
5.3.1 Simulationof a line-fedrectangular
microstrippatch 71
5.3.1.1 Transientresponse 72
5.3.1.2 Frequency
response 74
5.3.1.3 Accuracyvs stabilityfactor 77
5.3.1.4 Run-timecomparison 77
........... .
5.3.1.5 Inputimpedance 78
............ .
5.3.2 Simulationof a line-fedrectangular microstrippatchwith threeparasiticpatches 81
5.3.2.1 Transient
response ............ 81
.
5.3.2.2 Frequencyresponse 83
5.3.2.3 Inputimpedance............. 85
.
5.4 Conclusion 86
................. .
CHAPTER 6- CONCLUSION AND FURTHER WORK

6.1 Overall Conclusion 87

........... ......
6.2 FurtherWork 88
............ ......
6.2.1 Cylindrical coordinatesystem 88
6.2.2 Microstripswith slotsandnotches 88
6.2.3 Gradedmesh 89

REFERENCES

(1] Papers go
....................
[2] Books 94
....................

APPENDICES

Al 2ndorder accuracyof centraldifferenceapproximation 95

........
BI Tri-diagonalmatrix equationsfor proceduresI and2 in ADI-FDTD method 96
.
B2 Tri-diagonalmatrix equationsfor proceduresI and2 in modified ADI-FDTD method 98

B3 Tri-diagonalmatrix equationsfor proceduresI and2 in ADI-FDTD methodwith

electricconductivityterm ............... 100

C1 Graphicalillustration of implicit/explicit ADI-FDTD method 102

......
 ABSTRACT

In this thesis,the alternating-directionimplicit method(ADI) is investigatedin conjunctionwith the finite-

difference time-domain method (FDTD) to allow crossing of the Courant-Friedrich-Levy(CFL) stability
criterionwhile maintainingstability in the FDTD algorithm.The main reasonfor this is to be ableto usea larger
numericaltime step than that governedby the CFL criterion. The desiredeffect is a significant reductionin
numericalrun-times. Although the ADI-FDTD method has been used in the literature, most analysisand
applicationhavebeenperformedon simplethree-dimensional
cavities.

This work makesoriginal contribution in two aspects.Firstly, a new modified aitcrnating-dircctionimplicit

method for a three-dimensionalFDTD algorithm has been successfullydevelopedand implementedin this
research.This new methodallows correctmodellingof a realistic physicalstructuresuchas a microstrippatch
with the ADI schemewithout causinginstabilityevenwhenthe CFL criterion is not observed.However,dueto
the inherentpropertyof this modified ADI-FDTD method,a decreasing'reflectioncoefficient is observedusing
this scheme.

is
The secondand more importantcontributionthis researchmakesin the field of numericalelectromagnetics
the developmentof a new method of simulating realistic complexstructuressuch as geometriescomprising
copper patch antennason a dielectric substrate.With this new method, for the first time, the ADl-FDTD
algorithmremainsstablewhile still in violation of the CFL criterion, evenwhen complexstructuresare being
modelled.

However,there is a trade-off betweenaccuracyand computationalspeedin ADI-FDTD and modified ADI-

FDTD methods.The larger the numericaltime step,the shorteris the simulation run-time but an increasein
numericaltime step causesa degradationin accuracyof numericalresults. Comparisonbetweenspeedand
accuracyis shownin this thesisand it hasto be mentionedherethat thesevaluesare very much dependenton
the structurebeingmodelled.
List of symbols viii

LIST OF SYMBOLS

k propagatingwavewavenumber
G scalarGreen'sfunction
8 dirac delta function
r radial distance
At time stepin numericalalgorithm
Ax numericalspacestepin x-direction
AY numericalspacestepin y-direction
Az numericalspacestepin z-direction
V wavevelocity
n nth time step
i ith spacestepin x-direction
i ith spacestepin y-direction
k kth spacestepin z-direction
i electricfield vector in volts per metre
h electricflux densityin coulombsper squaremetre
H magneticfield vectorin amperesper metre
B magneticflux densityvectorin wcbersper squaremetre
je electricconductioncurrentdensityin amperesper squaremetre

im magneticconductioncurrentdensityin volts per squaremetre

P magneticresistivity in ohmsper metre

a electricconductivityin siemensper metre
P magneticpermeabilityin Henrysper metre
C electricpermittivity in Faradsper metre
At time eigenvaluesdueto temporaldifferentiation
A, spaceeigenvaluesdueto spatialdifferentiation

complexnumber V--I

kx numericalwavenumberin x-direction

ky numericalwavenumberin y-direction

kz numericalwavenumberin z-direction
C speedof light in metreper second

,8 phaseconstantof propagatingwavein radiansper metre

CO angularfrequencyin radiansper second
r reflectioncoefficient
7 propagationconstantof propagatingwave
Introduction

CHAPTER 1
INTRODUCTION

1.1 Introduction

Frequencydomain analytical methodshave been extensively used [1.1],[2.1] to solve complicated

microstrip circuits. Generally, when using this approach,the thicknessof the microstrip substrateis

assumedto be muchthinnerthanthe shortestwavelengthof interest.As a result,thereis no field variation

throughoutthe thicknessof the substrateand this rendersit a two-dimensionalelectromagnetic problem.
In sucha casethe microstrippatchis modelledby applyinga magneticwall aroundit. The fringing fields,
however,haveto be accountedfor by usingempiricallyobtainedeffectivepatchdimensionsandeffective
permittivities.This is also known as the cavity model. Despitethesedisadvantages,
the cavity model is

simpleto implementand givesgreatphysicalinsightto the circuit operation.The scalarGreen'sfunction

is usedto analytically solvethis two-dimensionalelectromagneticproblemwith perfectwall boundaries.
In general,the scalarGreen's functionsolvesthe following scalarHelmholtzequation:

V2G+k2G= g(r-r,, ) (1.1)

wherek is the wavenumber in the medium andthe excitationis in the form of a dirac delta functionat

r=r, The Helmholtz equation(1.1) is solvedby first expanding the solution in termsof eigenfunctions
of the homogeneousHelmholtz equationfor the appropriatecoordinatesystemwith the applicationof
perfect boundaryconditions.By applying the methodof separablevariables,an exact solution to the
differential equationmay be found. The solution is generally in the form of a double seriesGreen's
function for a two-dimensionalproblem.This methodhas been widely used to analysevarious patch
circuits [2.1]. This doubleseriesGreen'sfunctionwas successfullyreducedto a single seriessummation
[1.1] by applying the reducedoperatormethodas describedin [2.2]. However,this type of solution is
restrictedto only modellingthin substrates.

To modelthick substratesand accountfor any fringing field effects,a full-wave electromagnetic

solution
is required. There are generally two categoriesof numerical methods for solving electromagnetic

scatteringproblems,namely, frequencydomain methodsand time domain methods.Frequencydomain

methods includethe finite-element method and the methodof moments[2.3] while the line
transmission
matrix (TLM) [1.2] - [1.3] andthe finite-differencetime-domain(FDTD) [1.4] are time domainmethods.
Full-wave frequencydomain methodshavebeenusedto model variousproblems,especiallythosewith
few selectedfrequencypoints of interest.This is because,in such methods,the data for the whole
frequencyrangeare calculatedone frequencyat a time. However,for widebandsolutions,time domain
methodsare generally preferableas a whole spectrumof frequencyresponsecan be obtainedfrom a
singlesimulationrun. By excitingthe time domainmodelwith a broad-bandGaussianpulse,for example,
Introduction

and then applying the Fourier transformon the time-domainresults,one can get the entire frequency
rangeof interest,all in a singlesimulationrun.

The FDTD methodhas beenextensivelyusedto solve two- and three-dimensionalscatteringproblems

[1.5] - [1.7]. The author has chosenthe FDTD method over other time-domainmethodsbecauseits
implementationis straightforward,directly derivedfrom Maxwell's equations.

1.2 FDTD Method

The finite-difference time-domain or commonly known as the FDTD method was first proposed by
K. S.Yee in 1966 [1.4]. The FDTD method is formulated by discretizing the differential form of
Maxwell's two curl equations over a finite volume and approximating the derivatives with centred
difference approximation to obtain a second order accuracy in time and space. Appropriate boundary

conditions are imposed on the source point, conductors and computational boundaries to model the real
structure. Indeed, FDTD is relatively simple, flexible and easy to implement. However, over the years,
FDTD applications have been restricted to solving electrically small structures. To obtain accurateresults
for large electrical structures, large amounts of CPU time and memory resources are required. These

expensive computer resourcescome from two modelling constraints.

1. The spatial step, Ah, must be at least 10 to 20 times smaller than the smallest

wavelength of interest for a negligible dispersion error and

2. The time step used in the algorithm must satisfy the Courant-Friedrich-Levy (CFL)

stability condition stated below (and derived in Chapter 2) :

At :5-I

whereAx, Ay andAz arethe spatialsteps,At the time stepandv the maximumwavevelocity in the media
beingmodelled.The implicationsof the abovetwo constraintsareconsiderable.

A physicalunderstandingof the CFL stability constraint(1.2) will be explainedbelow with the help of
Fig. 1.1which showsan elementalthree-dimensional building block in the FDTD mesh.Assumethat the
elementalblock is a cube,that is Ax = Ay = Az = Ah. In numericalFDTD, for example,whenmodellinga
wavespeed,v, the numericalwavetakes3 At to propagatediagonallyin the cube;that is, the wavetakes
3 time stepsto travel a distanceof vr3- Ah. The numericalwave speedis governedby the dielectric

constantusedin the simulation.Therefore,if a biggertime stepis usedto modelwavepropagatingat the

samespeed,v, the wave will seemto havetravelledfurther than it actually has.This gives rise to errant
simulationresults.SinceFDTD is a time-domainmethodwith eachtime-domainresultsfeedingbackto
the next time-domain algorithm, this error will accumulateand eventually grow as time progresses
resultingin an unstablesystem.
Introduction

J3-AIV'.

< Ah
Ah

Fig.1.1 : Elementalcubein FDTD mesh

Another modelling constraintis that not only must the spatial incrementalstep be small relativeto the
smallestwavelengthof interest,but, in order to model an electrically large structurewhich contains
discontinuitiesaccurately,the spatialstepmustalso be madefine nearthe discontinuities.The constraint
of equation(1.2) above meansthat the time step has to be small near the discontinuitiesin order to
maintainstability in the FDTD scheme.The smallerthe time step,the longer is the simulationrun-time
becausemorenumericaliterationsarerequiredto representa finite amountof real physicaltime. This can
leadto a prohibitively long simulationrun-time.

Oneway to model an electrically largestructurewith fine discontinuitieswithout incurring intensiveuse

of memorystorageis to sub-dividethe computationaldomain into regions,eachwith a different mesh
size;a finer meshsize is usedfor regionswith high irregularity and a larger meshsize for the rest of the
domain [1.8] - [1.12]. The finer mesh is obtainedby further meshingthe larger mesh.This processis
referredto assubgridding.In sucha situation,to maintainstability as definedby equation(1.2), eitherthe
time stepcorrespondingto the smallestmeshsizeis usedfor the whole computationaldomain,or the time
stepsare setseparatelyfor eachmeshregion.For a largeobject with a highly irregularstructure,the first
methodcan be computationallyexpensivein termsof simulationrun-time.Using the secondmethod,the
meshsizeshaveto be suchthat the time stepsare integermultiples of one another.Furthermore,space
and time interpolationsat the interfacesof the mesh may be required for accuratesimulation. Also,
numerical dispersion will vary throughout the different mesh sizes. All in all, this makes the
implementationof FDTD to meshgeometriesthat vary acrossa volume,a difficult and time consuming
taskboth in implementationandexecution.Despitethe savingin computerstorage,the overall simulation
run-timewill still be long dueto the run-timenecessary
over regionswith fine meshsize.

The key to modelling electrically large structureswith fine discontinuitieswithout incurring a huge
computationalburden in terms of simulationrun-time is if the CFL stability criterion can be violated
without causinginstability therebyallowing the useof biggertime stepsin the simulation.This is realized
whenthe alternating-directionimplicit (ADI) methodis appliedon the FDTD algorithm.

1.3 Finite-difference approximation to derivatives

Considera two-dimensionalparabolicequation(1.3) below:

au alu alu
=. ý +7
at
Introduction 4

1.3.1 Explicit method

One finite-difference approximation to (1.3) is

n+l un nn + un
ul,j .nnui,j uj+l,j -2 Ij + ui-l, j ul,j+l -2 un
Ij I,J-l
At AX2 Ay 2

where x= lAx, y= jAy and I= nAt . Equation (1.4a) can be written as :

uni.
t+ j,
2uj + u,»-1 un- 2un +un)
J,J+l J,j 1,1-1
u n+I -ý UI,n'j +At + (1.4b)
ij äx 2 AY 2
ll

and (1.4b) gives the unknownvaluesu at time step(n+I)At in termsof known valuesu at time stepnAt.
This is known as an explicit method.This explicit methodis simple but can be computationallyintensive
becausethe condition for its validity [2.4], shownbelow, limits the time step,At that canbe usedin order
to maintainstability in the system.

At :5
12+ 1
2( -T2)
Ax y

1.3.2 Implicit method

Anotherpossiblefinite-differenceapproximationto (1.3) is

n+l 2Un+l unJ. j - 2u"Ij

u ij - Uo I U"+'
1+1,j - I,j + Un+l
i-l. j + uin-l,j
+4
At 2 Ax2 &x 2

n+l
u J,j+l - 2un+l + n+l
ul'j-1 ui.nj+l -2 Un
j 1.1
+j
2 Ay 2 AY2

The unknown values u at time step (n+I)At are given in terms of the known values u at time step nAt and

also the unknown values u at time step (n+I)At. The unknown values u at time step (n+I)At are then

calculatedby solving (M-1)(N-1) simultaneousequationscomprisingthe known valuesu at time stepnAt

whereM is the numberof Ax spacestepsandN the numberof Ay spacesteps.For largevaluesof M and
N, the simultaneousequationswill be solved iteratively.This method is known as the Crank-Nicolson
implicit method. This implicit method is valid for all values of Ax, Ay and At, that is, there is no
constrainton the time step used. But it takes considerablymore computing power than the explicit
methodas the simultaneousequationsmay needto be solved iteratively and they involve finding the
inverseof the matricescontainingthe equations.
Introduction 5
1.4 Alternating-direction Implicit procedure
One crucial point about the FDTD method is that it is a fully explicit method of solving differential

equations. This means that the iterative field values are calculated from previously known values. As
mentioned above this method is attractive as it is simple to implement but it is limited to some extent by
the CFL stability criterion that limits the time step that can be used in the algorithm. Implicit methods, on
the other hand, have superior stability properties [2.4] where the time step is not limited to any size.
Unfortunately, an implicit method in two dimensions requires at each time step, the solution of large sets

of simultaneous equations, which is not always easy to accomplish directly. Morever, when applied in
conjunction with the three-dimensional FDTD algorithm, this method results in three three-dimensional
matrices which have to be solved simultaneously.

The most efficient methodto date that incorporatesthe implicit method is the one first proposedby
Peacemanand Rachfordin 1955 [1.13]. This methodrequiresthe line-by-line solution of small setsof
simultaneousequationsthat can be solvedby a direct,non-iterativemethod.This is calledan alternating-
directionimplicit (ADI) procedure.PeacemanandRachfordtestedthe ADI procedureby usingit to solve
the heat flow equationwith boundaryconditionsin two spacedimensionsand comparedthe solutions
with known formal solution.The two solutionsshowedgoodagreement.The ADI methodwasalsotested
by Peacemanand Rachfordon steady-state problemsin two dimensionsby solving Laplace'sequationin
a square.The stability of the ADI schemewas also discussedand analysedin their paper.Part of their
work in the ADI methodusedto find solutionsof an unsteady-state
heat-flow in a squareis described
below.
n+1 n n+l n+l
2ul, n+l nnn
Ul,j - Ui,j ui+l,j " j + Uj-j,J ui,j+l_, 2ui.j + ul,j-l.
At &2 AY2

In [1.13], the secondorder derivativeterm from (1.3), a2U1C6x2

is replacedby a secondorder difference

term evaluatedin terms of the unknownvaluesof u, that is implicit in the x-direction, while the other
derivative, a2Ula Y2 is replacedby a secondorderdifferenceterm evaluatedin termsof known valuesof

u. This results in sets of simultaneousequationsthat can be solved easily without iteration. If the
procedureis then repeatedfor a secondtime stepof equal size to the first time step and the difference
equationsare set implicit in they-direction,as shownin (1.6b),then Peaceman and Rachfordshowedthat
the overall procedurefor the two time stepswould be stablefor any size time step. This meansthat the
time stepusedis no longerrestrictedby the stability requirementof the system.

m+2 n+l n+l n+l

2ui, n+l n+2 n+2 n+2
Ui.j - Ui,j ui+l,j , + 2ui,
j Ui-i,j uj j+j - j + ul,j-l
+
At Ax2 Ay 2

Sinceits introduction,the ADI procedurehasbeenbroadlyusedto solvediffusion problems.This method

was first adaptedto solvewaveproblemsin FDTD meshby T.Namiki [ 1.14].With the applicationof the
alternating-directionimplicit procedureon the FDTD method,the well-known CFL stability criterion
statedin equation(1.2) can now be violated without causinginstability. The physicalunderstandingof
how violation of the CFL criterion will leadto instability of the FDTD systemwas discussedearlier in
Introduction 6

section1.2 and illustratedusing Fig I. I. This meansthat potentially bigger time stepscan be usedin the
simulationtherebyreducingthe overall simulationtime.

In this thesis,the ADI-FDTD methodis investigatedanda new modified ADI-FDTD methodis proposed
anddiscussed.The detailedimplementationof the ADI-FDTD methodwith respectto the structurebeing
modelled is also presentedin this thesis. Numerical simulations of a simple line-fed rectangular
microstrip patch are usedto verify both the ADI-FDTD and the newly proposedmodified ADI-FDTD
methods. A bigger patchwith threeparasiticpatchesis also simulatedto verify the applicationof ADI-
FDTD methodon an electrically large object. Wherepossible,the simulatedresultsare comparedwith
resultsfrom publishedliterature.

This thesisis organizedasfollows.

Chapter2: Discussesthe theory behindthe finite-differencetime-domainmethod,the implementation
procedures,and comparesthe simulatedresultsobtainedin this researchwork with thosein the published
literature[1.5]. A new, more efficient methodof extractingthe reflection coefficient from the simulated
datais alsopresented.

Chapter 3: Presentsthe theory of the alternating-directionimplicit method applied on the FDTD

algorithmand illustratesthe physicalinterpretationof the ADI-FDTD method.With the help of simulated
data,the problemsencounteredwhen a line-fed rectangularmicrostrip patch is simulated,are discussed.
This chapteralsoexplainsthe implementationof the absorbingboundaryconditionat the boundaryof the
computationaldomainwhich is critical for accurateandcorrectmodellingof the structure,

Chapter4: Proposesa new modified ADI-FDTD methodto surmountthe problem of modellingthe

microstrippatch in chapter3 using perfectelectricwall boundaryon the copperpatch.The limitation of
this newmethodis alsodiscussed.

Chapter5: Showsthat a different techniqueof implementinga boundarycondition helpseliminatethe

problemencounteredin chapter3. Simulatedresultsfor morecomplexstructuresarealsoshown.

Chapter 6: Concludes the researchundertaken by the author and suggestssome further work in this field

of research.
Finite-Difference Time-Domain Method 7

CHAPTER 2
FINITE-DIFFERENCETiME-DoMAIN METHOD

2.1 Introduction
The algorithm of finite-difference time-domain field analysis was first introduced by Kane Yee in 1966
[1.4] and has been widely used to solve electromagnetic scattering problems.

2.2 Maxwell's equations In three dimensions

In a regionof space,the time dependentMaxwell's equationsare given in the differential form by [2.51
Faraday'sLaw:

L9B vxz Jm
FT =.

Ampere'sLaw:

ýLD VX77. -is (2.2)

=
at
And the constituentrelationsare :
V-D =0 (2.3)

V-B =0 (2.4)
In linear,isotropicnon-dispersivematerials,

B= uH (2.5)

and D=cE (2.6)

In orderto accountfor the magneticlossin the system,the magneticcurrentdensityis given by :

J. =p H (2.7)

andsimilarly the electriccurrentdensityis :

Je = uE (2.8)

E is the electric field vector in volts per metre, D is the electric flux density vector in coulombsper

is field in ý
squaremetre, H the magnetic vector amperesper metre, is the magneticflux densityvector
in webersper squaremetre, J. is the electric conductioncurrent density in amperesper squaremetre,

J. is the equivalentmagneticconductioncurrentdensity in volts per squaremetre, p is an equivalent

magneticresistivity in ohmsper metreand a is the electricconductivityin siemensper metre.

Finite-Difference Time-Domain Method 8

Combining the assumptions of (2.5) to (2.8) and substituting into Maxwell's curl equations (2.1) and
(2.2), we obtain:

vxý P77 (2.9)

-
ju p
ý-E I
VXT - 62 (2.10)
at c Er

which governthe propagationof both electricandmagneticfields in any structure.

Writing out the vector components of the curl operator in (2.10) and (2.9) yields the following six coupled

equations equivalent to Maxwell's curl equations in a three-dimensional Cartesian coordinate system.

ýE, j(LHj_ My )
-aE, (2.11a)
at c cy oz

LEy aH
- aEy (2.11b)
at az ax

M= I(aHy aH' )
- aE. (2.11c)
at e ax ay

ýH, aEy
. at __L(aE, - +p (2.12a)
u az

ally ( aE, aEz

_L . +pH (2.12b)
at p az ax Y)

I (My aE, )
clH_,
_ +pH, (2.12c)
at u ax Cýy

This systemof six coupledpartial differentialequationsof (2.11) and (2.12) formsthe basisof the finite-
differencetime-domain(FDTD) numericalalgorithmfor electromagneticwaveinteractions.

2.3 FDTD algorithm

The FDTD algorithm solves for both electric and magnetic fields in time and space by solving the six

coupled Maxwell's curl equations (2.11) - (2.12). A physical model of the fields in a Cartesian grid is

in For Z
shown Fig. 2.1. programmingconsiderations,the numberingof the spatial location of and
H fields in Fig. 2.1 differ from that in the original Yee's cell.

2 WH
As illustratedin Fig. 2.1, the algorithm centresits and componentsin three-dimensionalspaceso

that every E componentis surroundedby four H componentsand vice versa. Every componentof
71 Z
cannow be obtainedby the loop integralof usingthe four surroundingE nodalvaluesaccordingto
Maxwell's curl equationof Z. A similar conditionholdsfor 77.
Finite-Difference Time-Domain Method 9

Ex(ij+l, k+l)
Iz
:: ýýEýý
y Ey(ij, k+l)
Ex(i, j, k+ 1) I j, k+ 1)
x
Hx( Ez(ij+ I,k) Ez(i+lj+l, k)
Hy(i- I j, k) E ij, k) Ez(i+ I j, k)
Hx(ij -I, k Hy Ex(ij+l, k)
Hz(ij, k)
Hz(i- I j, k), Ey(*, I,k) Ey(i+ I j, k)
Ex(ij, k)
Hz(i-lj-l, k), -
Hy(i-l. j, k-1) Hx(ij, >
k-1)
Hy(ij, k- 1)
Hx(ij-l, k-1)

Fig. 2.1 : Yee's staggeredcell

Z -H
We note that in the FDTD algorithm, the and nodes are disjointed by half a space step. In addition,
Z 77 fields 2
calculation of and are also disjointed by half a time step. This means that the and H_
fields are calculated at alternate half time steps. For this reason, this algorithm is called the leapfrog

method. The leapfrog time-stepping process is fully explicit; that is, the current field values are calculated

using previously stored field values. As a consequenceit is not necessary to solve sets of simultaneous
equations by involving matrix inversions.

It is worth noting that continuity of the tangential 2 -H is

and automatically maintainedacrossan
interfaceof dissimilar materialsif the interfaceis parallelto one of the grid coordinateaxes.Changein
materialsis specifiedusing the materialpermittivity and permeability.Theseare defined in the FDTD
equations.

2.4 Finite difference expression of Maxwell's equations In three dimensions

Assuminglosslessmaterials,discretizing(2.11) & (2.12) leadsto the approximationof Maxwell's curl

asfollows :
equationsin three-dimensions

n+112(ijk) n+112 I
H n+112(ijk) Hzn+112(ij-l, k) H OM-1)
- yy -H
En+l (ijk)
xx = E" (iik) + z-- (2.13a)
AY AZ

H n+112(ii, 1/2(ij n+I/2(I h+I/2 (i- 1,i. k) 1

At k) -Hn+ k-1) H ZZ ,j, k) -H
En+' (iik) = Ey (ijk) +- -x--- --AZ
---x (2.13b)
Ax
Finite-Difference Time-Domain Method 10

At H n+112
0j, A) -H n+112(i_lj, k) H n+1ý2(jjj) ,+1/2(ii- 1,k) I
E n+l (ii, k) =E" (ii, k) + -H (2.13c)
zzc -yyxx AX Ay

At Ez"(ij+l, k)-E, "(ijk) Eyn(ijk+l) - y(IJ,k)

Ey"
H n+112(ij, k) n-112(ijk) :y (2.14a)
x, =Hx ,- AY AZ
lu

Exnfljk+l)-Enfljk) I
At En(i+ljk) Enfljk)
H n+1/2(ijk) Hn-112 (,Jk) xzz - (2.14b)
y y AZ AV
JU

I En(i+ljk) En(ijk) Enq J+ 1,k) -En (IJ, k)

1/2 (ijk) At -
Hzn+112
(iik) = n
Hz- -- .yyx- --- x (2.14c)
Ax Ay
lu

where x= iAx, y= jAy, z= kAz and t= nAt .

This FDTD algorithm has secondorder accuracyin both spaceand time becausethe centraldifference
methodis appliedon both the spaceandtime derivatives[A. I].

2.5 Divergence of FDTD algorithm

While the FDTD algorithm solves for both electric and magnetic fields in time and spaceusing the
coupledMaxwell's curl equations,thereis no explicit enforcementof the Gauss'sLaw relationsfor both
the electricand magneticfields as statedin (2.3) and (2.4) for sourcefree regions.It is importantthat the
Gauss'sLaw is observedin the FDTD algorithm.

The time derivativeof the surfaceintegralsof the electricflux densityover all the surfacesof a free-space
Yee cell of Fig. 2.1 is given by [2.5] :

(ffD -dS [E.,(1,J, k) E.,(i 1,J, k) ] AyAz

- -

coa [Ey ]&vAz

+ (1,J, k) - Ey (1,j-1, k)
at

+ 'Oa [E, (i, j, k) - E, (i, j, k- ]

1) AxAy
at

Using the finite difference expressionsof the electric field (2.13a) - (2.13c), the electric field time
derivativesin (2.15) can be substitutedwith the magneticfield spatial finite differencesin eachof the
RHSterm in (2.15) producing:
Finite-Difference Time-Domain Method II

Ay AZ
(ff-D-dS= AyAz

Ay AZ

AZ AX
+ AXAZ

AZ Ax
A)- H-ý(Mjýk)
Fy0j,.,
Ax AY
+L AxAy

AX Ay

(2.16)
for all time steps.The RIIS termsof (2.16)canceleachotherout. Hence,

0 ffD- fffV
- d§ = -L -B dV

V-D

Therefore,the time derivative of the net electric flux leaving the surfacesof a cubic Yee cell is zero,
henceupholdingGauss'sLaw for the electricfield in charge-freespace.The samecan be shownfor time
derivativeof the net magneticflux leavingthe surfacesof a cubic Yee cell. This showsthat the FDTD
algorithm is divergence-frecin sourcefree regionsand implicitly enforcesGauss'sLaw for both electric
andmagneticfields in thoseregions.

2.6 Numerical stability of the three-dimensional FDTD algorithm

Numerical instability is an undesirablepossibility with explicit numericaldifferential equationsolvers

that can causethe computedresultsto spuriouslyincreasewithout limit as time-marchingprogresses. A
standardmethodto analysenumericalstability was presentedby von Neumannand Courant,Friedrich
andLevy [the CFL condition,2.4].

Electromagneticwaves propagatingin a finite-difference grid naturally results in the generationof

numericalwave modesor Fourier modes.In order to maintain stability in the finite-differencetime-
domainsystem,the spectrumof eigenvaluesfor thesemodesdue to the numericalspacedifferentiation
processmust be containedwithin the stablespectrumof eigenvaluesdeterminedby the numericaltime
differentiationprocess.The magnitudeof field growth at every time step,called growth factor, is limited
to a maximumvalueof unity. If the growth factor is greaterthan unity, the systemwill be unstable.

Without lossof generality,considera normalisedregionof spacewith p=1, e=1, cr = 0, p=0 and

Finite-Difference Time-Domain Method 12

c=I where c is the normalized wave velocity in a vacuum. We can re-write Maxwell's equations in a
more compact form [ 1.7] as:

av
jvxp = (2.17)
c1t
77+ j2
where

i= ý--l
and

Unlike compactderivationsin most literature,the following derivationsfor time and spaceeigenvalues

have beenexpandedand explainedin detail by the author where necessaryto easeunderstanding.To
of (2.17), considerthe following pair of eigenvalue
analysethe stability of the numericalrepresentation
problems:

11 V= AtP (2.18)
T'Inumerical

A
XP= AP (2.19)
numerical

where A, representsthe eigenvaluesdueto the numericaltime differentiationprocessand

As the eigenvaluesdueto the numericalspatialdifferentiationprocess.
,

First, considerthe time eigenvalues,from (2.18),

-n+1/2 -n-112
v-v -n
= AIV (2.20)
At

Now definea constantgrowth factorfor the numericalsolutionasa function of spacepoint I

-n+1/2 -n
Vi Vi
for all n time steps
Pi n-112
Vi

in order to maintainstability in the FDTD algorithm, I qj :51for all possiblespatial modesin the grid

andfor all points i. Substituting(2.21) into (2.20)yields

n- PnI
qjP qj -n
= AtV
At

pn [qj 0
or AlAt qj -1]
-

F(2
A, At At
qj 'ýAt (2.22)
=
2
Finite-Difference Time-Domain Method 13

AAt
We seethat I qj I if is purely imaginaryandhasa magnitudeof I which means
2
AtAt
< :5 +j
2
2. Ilmag(A,
or ,:. )l <2 (2.23)
At At

All possiblespatialmodesmusthaveeigenvaluesthat arewithin this stablerangeto ensurestabilityof the

algorithm.The spatialeigenvaluescanbe determinedby analysing(2.19) as follows.

yz
daZ
VInumerical X FX- (2.24)
aXInumerical 'Onumerical
numerical
V, VY VZ

At any time step,the instantaneousvaluesof the electric and magneticfields distributedin FDTD space
acrossthe grid can be Fourier-transformed
with respectto the grid coordinatesto provide a spectrumof
sinusoidalmodes,resultingin planewaveeigenmodesof the grid. Now let the following specifya typical
modeof this spatial frequencyspectrumhaving k, ky andk: as the x- , y- and z- componentsof its
numericalwavevectorrespectively.

(ý.
-i I &v + Ay + W,k Az)
V, e (2.25)

Then,
F(i+112, j, k) T(I-112, j, k)
6-V -
K
numerical
-te i(W, (1+112)&v+rjAy+rk&) j(W,(1-112)Av4, jAy+rkAz)
V
_ e
AX
j(ý,,'&4, J'y+rkAz)[ej(r, &v12) j
V. e C,
AX
-[j2sin(W, Ax/2)-
v
AX

(2.26)
Therefore,
dex (T,
j2 sin Ax/2
AX
(2.27a)
numerical

j2 sin Ay12
(2.27b)
AY
Oynumerical

aj2 sin Az/2

aFz
(2.27c)
numerical AZ
Finite-Difference Time-Domain Method 14
(T, (T. ) rk,
sin Ax/2 2 sin Ay12 sin Az/2
Let X=2 Y= and Z=2
AX AY AZ

Thencombining(2.19),(2.24)and(2.27)gives,

(! ]+ ! j[XVY )=A, P
j j[YV, -ZVYI - ýj[XV, -ZV., - YVj (2.28)

or
0z0 z y
'XI V'
xI=A, X1
- VVY VY
Y 0 V, V.
-x -
-As z-Y Vx
-Z -As x Vy =0 (2.29)
Y -X -As.
_V:.

Solving for eigenvalues A, gives

2= (X2
As . + y2 + Z2)

Substituting back X, Y and Z gives

ýkyAy12) ýk..
sin2 2 Azlj
sin
A2= + + (2.30)
sI -4ý
Ax2 AY2 A?

-2
As - 1:5 sin p --ýI, for all possiblek., ky andk: . we canboundthe rangeof A,:

Ilmag (A, )1
:921+I+
,NX2 Ay 26 =2

To satisfythe stability condition (2.23) for the arbitrarylattice spatialmode,all the eigenvaluesin (2.31)
mustlie within the rangespecifiedin (2.23) i.e.

2
-XI2 +7+ -A (2.32)
Z2 At

anddenormalizing(2.32)by the FDTD wavevelocity,where v= l1VIp---

At :5111 (2,33)
Vý-ý + -ýY-T -I
A?

This is generallyknown as the CFL (Courant-Friedrich-Levy)stability condition. In an inhomogeneous

region of space,it is difficult to determinea spectrumof A, equivalentto (2.31) for all possiblelattice
Finite-Difference Time-Domain Method 15

For
spatialmodes. absolute algorithmstability, (2.33) will be good enoughas it representsthe worst case
choice of time step, At. If At is selectedto be larger than the bound in (2.33), the FDTD numerical
algorithmwill definitely be unstable.This is true as long as the FDTD algorithmis completelyexplicit. It
is the aim of this work to investigatethe possibility of usingthe time-stepbeyondthe constraintof (2.33)
while maintainingstability of the system.Chapters3 and 4 will discusstwo different methodsemployed
to achievethis aim.

2.7 Numerical dispersion of the three-dimensional FDTD algorithm

Dispersionis defined as the variation of the propagatingwave's wavenumber k=2; r /A with angular
frequencyra = 27rf The analyticaldispersionof physicalwave propagationis an inherentpropertyof the
.
medium of propagationand is structuredependent.For example, in a microstrip patch circuit, the
effective permittivity changesas a function of frequency.This gives rise to analytical dispersion.The
k, is different from Twhich is the waveriumberof the
wavenumberof the continuousphysicalwave,
numericalsinusoidaltravelling waveof angularfrequencyco that is presentin the finite-differencegrid.
This difference between k, analytical wavenumber,and T, numerical wavenumber,gives rise to
numericalphaseand groupvelocitiesthat aredifferent from the exactvaluesobtainedfor physicalwaves.
This differencegives rise to numericaldispersionand consequentlyan errant simulationresult. Whilst
analyticaldispersionis an inherentcharacteristicof the microwavestructure,numericaldispersionis due
to discretizationof time and spatialstepsin the finite-differencealgorithm.The variationof the numerical
wavevelocity with wavepropagationangledueto numericaldispersionis shownin [2.5].

Numerical dispersioncan be found by analysing(2.17). Substitutingthe vector-field travelling-wave

expressionwith time dependence:
(ij, k) F j(ý,' &v+ry jAy+rs k Az-wn At) (2.33)
= -0e

into (2.17) results in :

(IX ) T n+l Pn(I,
AY) iz Az (IJ, k) J, k)
y Fn+1/2 -
-sin +-sin + sin x (1, J, k)
-2
2. _2 AZ 2
AX AY At

(2.34)

or

( Ty
T, Ax Ay ,R I 2jsin(o)At/2) P"e, j"A 1/2
jwAtl2
-2 sin -+ sin -+- sin xPe =
2) Ay 2 AZ 2 At

(2.35)

Comparing(2.35) with (2.28) and(2.30)anddenormalizingit to a non-unitywavevelocity v, givesus the

generalform of the numerical dispersionrelation for the full vector-field FDTD algorithm in three
dimensionsas :
Finite-Difference Time-Domain Method 16

TyAy )l 2 ( W.Az )]
)]2 1 (TAX 1
CoAt
- sin sin ý-2)] sin sin ý-2- (2.36)
VAIt2 Ax Ay 2

In contrastto the numericaldispersionrelation (2.36), the analyticaldispersionrelation for a planewave

in a continuouslosslessmediumis simply:

w kx 2+k2+k2 (2.37)

However, (2.36) will reduce to (2.37) in the limit as At, Ax, Ay and Az all go to zero, that is, if the FDTD

grid is madevery fine.

As the FDTD grid size increaseswith respectto the wavelengthof propagatingwaves,the deviationof
the numericalphasevelocity from the exactanalyticalphasevelocity increasesuntil the waveseventually
ceaseto This
propagate. numericallow-passfiltering effect is inherent in the FDTD grid. Consequently,
FDTD modelling of pulseswith high bandwidthwill result in progressivepulse distortion as the high
spatialfrequencycomponentswill propagatemore slowly than the low spatial frequency In
components.
addition, the very high spatial frequency componentswith wavelengthsless than 2 to 3 cells are
completelyrejected[2.5].

2.8 Boundary conditions

In a finite-differencescheme,the finite-differencemeshhasto be of finite extentdue to the limitation of

computerstoragecapacity. However, in many applications,the media to be modelled are of infinite

extent.Scatteringproblems,for example, lead to solutionsof fields in an unboundeddomain.imposing

boundaryconditions on the finite-differencemesh boundarymay give rise to reflectionsthat are not
representativeof the actualphysicalsituation.Consequently,absorbingboundaryconditionsare applied
on the mesh boundary, also known as the boundary,
computational in order to simulateinfinite or very
largegeometries.The algorithm on the truncationplaneshasto simulatepropagationof outgoingwaves
as if they were propagatingto infinity. This is by
accomplished enforcing an impedance
matchon the
computationalboundaryso that thereis no reflectionof outgoingwavesback into the domain.

Referringto Fig.2.1, if the finite-differencemeshterminateson the electric field cell, we can seethat all
componentsof the electric field on the boundaryare tangentialto the boundarywhile the componentsof
field it. While 77
the -field componentscan be calculatedfrom the respective
the magnetic are normalto
E -field componentsusing equations(2.14), the E -field componentscannotbe evaluatedin the same

way as this H that

would require -field components areoutside the mesh. For the structures in
considered
this thesis,the pulseson the microstriplineswill be normally incidenton the meshboundaries.Therefore,
a simple approximatecontinuousabsorbingboundarycondition,wherethe tangentialfields on the mesh
boundariesobeythe one-dimensionalwaveequationin the directionnormalto the meshwall, will suffice.
Finite-DifferenceTime-DomainMethod 17

Engquist and Majda derived a theory of one-way wave equationssuitable for absorbingboundary
condition in Cartesian FDTD grids [1.15]. It was further discussedby Mur in 1981 [1.16] and
implementedon a finite-differencescheme.

2.8.1 1storder Mur boundary condition

To deriveMur's first approximateabsorbingboundarycondition, considerTEM wave propagationon a
lossfreetransmissionline giving riseto voltageandcurrentas follows :

V, = Vlcospx + jI, Z,, sin,8x (2.38)

V'
Ix = I, Cos x+ sin (2.39)
zo

wherex is an arbitrary point on the transmissionline and I is the load end of the transmissionline. In
orderto haveno reflectionat the boundary,the line mustbe matchedat the boundary,therefore,
V,
= zo (2.40)
11

whereZ,,is the characteristicimpedanceof the line.

Then, (2.38) and (2.39) reduce to

VV COX (2.41)
x1
Ix I, eJflx (2.42)

which is a standardequationfor one-dimensionalpropagatingwave travelling towardsthe -x direction.

Assumingsinusoidaltime variation,we have,
j(, 8x +wt)
Vx = V,e (2.43)

Takingtime andspacederivativesof (2.43)gives

a V, a V,
T=jwV, and jflv.,
t

andequatingthe V,,termsresultsin
I
-LIDVI (2.44)
i jw at
16
or
V,, av"
-I =0 (2.45)
&v at

So, for wave propagatingin the -x direction, normal to the absorbingboundarywall, the Mur's first
approximateboundaryconditionis:
OEtangential I OEtangential

ax =0 (2.46)
v at
Finite-DifferenceTime-DomainMethod 18

while for wave propagating in the +x direction,

OEtangential aEtangential
+I
_0 (2.47)
&v at

is the E field tangential to the boundary wall and v is the velocity of the propagating
where Etangential

wave. (2.46) can be discretised so that the field components on the boundary walls are dependenton only
the field components on and just inside the walls.

Oneway to approximate(2.46) is to useforwarddifferencingfor both spaceandtime. This resultsin :

Ax (Ex'+'
EtEt -0 Et (2.48)
xi xo
vAt XO)
wherethe subscriptdenotesthe spacestepandthe superscript,the time step.Rearranging(2.48)givesus :

(, vAt VAt E
E t+l E ý ý,- -Z7 + (2.49)
X0 X0 6x

(2.49)gives I" orderaccuracyin the implementationof the Mur's I" approximateboundarycondition.

Another way of applying (2.46), in order to have second-orderaccuracy in the discretisedfinite-
differencemesh,is to impose(2.46)at half spaceandtime stepsas follows :

LEt,, II aEtangential It-f+1/2

tj,,,
ýgLehal
ý,, (2.50)
v at
xX x-x+1/2

)=1 I (Exll+/2
(E W/2 1
10+/2
EXII/2)
I -E 'o X ' (2.51)
AX x2 vAt

Sincethe valuesat the half grid pointsand half time stepsare not available,it is possibleto usea semi-
implicit approximation:
(E"+'
Em"'/2 4,1 + E." (2.52)
M 2 '" M)
! (E
and Em"+112$4, n+, +En (2.53)
2m M)

This is partially implicit becauseit usesan unknownvalueof E at time step(n+I)At.

Substituting(2.52) and(2.53) into (2.51) givesus

1[1 (Exi+I )] [ýI (Ex

+ Ext, - Et+' Exto i+I +E e+I E., (2.54)
- =1 xo - ,-E., 0
Ax 2 vAt 21
Finite-DifferenceTime-DomainMethod 19
Simplifying (2.54) gives

(+
Ex'+' -L
,(
vAt
+ -L
=
E.,
vAt
-L + (+'
Ex,
(AX
-L -1)ýE
vAt
f(-L
x, - -L)
äx vät
or
äx) EE1+
AX)
( vAt Ax )I (E,
- W Et (2.55)
X0 xi vAt + Ax XO)

the tangentialelectric field componentson the meshwall and Exi, the tangential
where Exorepresents
electric field componentsone node inside the mesh wall. Strictly, (2.55) is the absorbingboundary
conditionfor wavepropagatingin the -x direction,i.e. for useon the boundarywall at x=O.In this case,
E.,O,electricfield on the x=Owall is to the left of E.,j the electric field onenodeinsidethe x=Owall. For
,
a wave propagatingin the +x direction, although(2.47) showsa changein sign in the one-waywave
equation,(2.55) canstill be appliedasthe absorbingboundaryconditionat the boundarywall x=h simply
becausethe changein sign in equation(2.47) is equivalentto swappingthe electric field positionson the
boundarywall; in this case, E.,Ois the electric field on the wall which is to the right of Ej the electric
,
field onenodeinsidethe x=h wall. Similar expressionscanbe derivedfor otherabsorbingboundaries,i.e.
normalto y andz directions.

Even after applying the absorbing boundary condition, there is some reflection because true wave

propagation is not one-dimensional and also the wave velocity is not constant but a function of frequency.
Besides, the normal incidence assumption is not valid for the fringing fields, therefore the side walls

should be far enough away so that the effects are negligible on the walls.

2.9 Conductor boundaries

Conductingground plane and coppermetallizationlayer can be modelledas perfectelectric conductors

wherethe tangentialelectric fields are forcedto be zero.It is usualto assumethat theselayershavezero
thickness.In order to model the edgeof a conductor,tangential E fields are positionedexactly on the
edgeof the conductor.

2.10 Dielectric boundaries

The E and H fields in a dielectric region are calculatedusing equations(2.13) and (2.14) with the
dielectric constant,c, set to that of the dielectric insteadof unity. The field componentswhich lie on a

dielectric-air interfaceare the tangential E and the perpendicularH components.To calculate E, and
Ey at the dielectric-air interface, the averagevalue of c is used in (2.13) [1.6], i. e.

el + 62
(2.56)
2
where c, is the permittivity of the dielectric

ý 1, permittivity of air.
and 102
Finite-Difference Time-Domain Method 20

Generally,in order to accountfor fringing field effects at the edge of microstrip patch,the dielectric
constantto be usedis dependenton the relativepermittivity of the dielectricascomparedto that of air and
the relation betweenthe width and height of the dielectric below the microstrip patch being modelled.
Sincethe fringing field effectsare automaticallyconsideredin the FDTD simulation,equation(2.56) is
sufficefor useonly at the dielectric-airinterface.

(2,14) is still usedto calculatenormal H, as the value of p does not changeacrossthe dielectric-air
boundary.

2.11 Excitation
.
A Gaussianpulsehasbeenchosenasthe excitationpulsein all investigationsin this thesis.This is mainly
becausea Gaussianpulsehas a smoothwaveformin time and its Fourier transformis also Gaussianin
shapeand centredat zero frequency.This meansthat by adjustingthe width of the Gaussianpulse,the
frequencyresponsecan be obtainedfrom dc to the frequencyof interest.An ideal Gaussianpulsewill
havethe following expression:

(2.57)

andthe pulsewill be at its maximumat t=t,,. Fig. 2.2 showsa typical Gaussianpulse.

9(t)

T
I
e
..........

tý
Fig. 2.2 : Gaussianpulse

The choicesof T and t,, are subjectto two requirements.Firstly, the FDTD grid size, Ax, Ay andAz are

chosento be fine enoughto modelthe smallestdimensionof the structure.Also, in order to havea good
spatialor meshresolution,the grid size is set suchthat it is at least 1/20 of the shortestwavelengthof
interest. At is then calculatedfrom the CFL stability criterion as given by the bound in (2.33). The
Gaussianhalf-width is given by:

(' * "I
*
g(t) =
Finite-Difference Time-Domain Method 21

1tto')2
T) =

or I= to ± (2.58)

that is, T is the symmetric point from the centre t, point when g(t) drops to I/e of its maximum value.

The Gaussianhalf-width in time is then 2T. We know from Fourier Transform method that the
relationshipbetweenthe highestfrequency
of interest, andT in the Gaussian
pulseis given by :

(2.59)
2f..

Knowing the highestfrequencyof interest,T is calculatedfrom (2.59).This is to ensurethat the Gaussian

pulse is narrow enoughto have a wide spectrumin order to maintain a substantialvalue within the
frequencyrange of interest.At the sametime, the Gaussianpulse has to be wide enoughto contain
enough number of time steps for a good time resolution. More importantly, in order to minimize
numerical dispersionerror, we have found that, the Gaussian half-width, derived in (2.58), mustcontain
at least20 spacesteps in the direction of propagation. If the Gaussian pulse travels at a speed,
v, in the
directionof propagation,thenthe equivalentspatialhalf-width, TV,of the pulseis given by :

JY = 2Tv (2.60)
2Tv
Therefore,to have20 spacesteps, -ý-h- ;>-20

loAh
or (2.61)

where Ah is the spacestep in the directionof If

propagation. the half width of the pulse,calculatedfrom
(2.59) is not wide enoughto contain20 spacestepsthen,the spacestephasto be reduced.The time step
At will thenhaveto be re-calculatedto ensurethat the CFL stability criterion is still satisfied.

Secondly,t,, must be chosensuchthat the initial 'turn on' of the excitationwill be small and smoothto

avoid exciting high order In

modes. orderto havea smooth'turn on', t,, in all simulationsin this thesisis

setto threetimesthe valueof T.

2.12 Simulation of a line-fed rectangular microstrip patch

The finite-differencetime-domainequations(2.13) & (2.14) are usedwith the I" order Mur absorbing
boundaryconditionto simulatethe propagationof a broad-bandGaussianpulseon a line-fedrectangular
microstrippatchas shown in Fig. 2.3. This microstrippatchcircuit was chosenfrom a paper by
published
Abouzahraet al [1.5]. The finite-differencemeshparametersare chosento be the sameas in that paperto
allow direct comparisonof results.
Finite-DifferenceTime-DomainMethod 22
I' orderMur boundaryon all z
boundariesexceptgroundplane
Y
12.448m

D-'--6.795mm

e, = 2.2
1.945m T-
Perfect 2.334mm
electric wall

Fig. 2.3 : Line-fedrectangularmicrostrippatch

Mesh parameters :
Ax = 0.389 mm
Ay = 0.400 mm
Az = 0.265 mm
Thickness of substrate :3 Az
Air spaceabovesubstrate: 13Az
Rectangularmicrostrippatch: 32 Ax x 40Ay
Sourceplaneto edgeof rectangularpatch: 50 Ay
Monitoredreferenceplaneto edgeof rectangularpatch: 10Ay
Microstrip line width :6 Ax
Total meshdimensions: 60 x 100x 16 in i, ýand 1 directionsrespectively
Time stepAt = 0.441ps
Gaussianhalf-width T= 15ps
Time delay t,, = 3T
Af = 0.2 GHz

Sincethe substratethicknessis relatively small comparedto the wavelengthof interest,up to 20GIIz, we

canassumethat thereis no variationof electricfield in the vertical direction.Then,to pxcitethe dominant
mode,a Gaussianpulse in time is launchedinto the sourceplane, setting off the vertical electric field,
E: of the individual cell, Az, throughoutthe dielectricthicknessand acrossthe width of the feed line, 6
.
Ax.

In [1.6], an electric wall sourceis usedfor the remainingnodeson the sourceplane.An unwantedside
effect of this is that a sharpmagneticfield is induceddueto the high value of the spacederivativeof the
electric field. This results in the distortion of the pulse.To overcomethis problem,a magneticwall is
TI,
simulatedon the sourceplane as was done in [1.5]. Applying imagetheory, the tangential a node
inside the sourceplane is set to be the negativevalue of the tangential 77 node outsidethe source
,a
!
plane.Then the remaining field componentson the sourceplane may be calculatedfrom the finite
differenceequations.However,when wavesare reflectedback to the sourcefrom the microstrippatch,
the source plane has to be transparentto the waves.To simulate this, the I" order Mur absorbing
boundarycondition is switchedon oncethe excitationis completed.This meansthat the sourceplanehas
Finite-DifferenceTime-DomainMethod 23

to be a reasonabledistanceaway from the edgeof the microstrippatch suchthat the reflectedpulsedoes

not arrive backat the sourceplanewhile the sourceis still turnedon.

Initially, all fields in the computationaldomain are set to zero. As time-marchingprogresses,the

sequence of the algorithmimplementedis as follows :
verticalelectricfield is excitedwith Gaussianpulsebelow the strip
71fields
arecalculatedusing(2.14)
magneticwall is appliedon the sourceplaneon nodesotherthanthe sourcepoints
electricfield valuesarestoredfor later usein absorbingboundaryconditioncalculation
fields arecalculatedusing(2.13)

tangentialE fields aresetto zeroon the metallizedcopperpatch

2 fields the
tangential on computationalboundariesare calculatedusingthe Mur's first
approximateabsorbingboundarycondition(2.55)

The iteration proceedstill the responseis close to zero or until there are enoughdata to meet the
frequencyresolution.

2.13 Extraction of voltage and current from the FDTD mesh

In the FDTD simulation, excitation is in the form of the electric field and subsequently, the electric and

magnetic fields are calculated on the finite-difference mesh using equations (2.13) and (2.14). To extract
the voltage at the reference plane (electric field reference plane), firstly the vertical electric field

underneath the microstrip feed-line for each cell is multiplied by the Az to get the voltage for each cell;
this is repeated throughout the thickness of the dielectric and all these voltages added together give the
total voltage at the reference plane at a specific x location. In order to get a more accurate result, an
averagetotal voltage is obtained for the width of the strip.

To extractthe current,the magneticfield is integratedwith respectto the meshsizealongthe dottedpath

as shown in Fig.2.4. However,the magneticfield referenceplane is shifted by half Ay from the electric
field referenceplane. So, to get the current at the samereferenceplane as the voltage referenceplane
(electricfield referenceplane),the magneticfield is integratedtwice, first along the path of 11-reference
plane and then along the path Ay behind the H-referenceplane (see Fig.2.4). Assuming linearity, the
averageof the two integrationsgivesthe currentat the E-referenceplane.
Finite-DifferenceTime-DomainMethod 24

Microstrip feedline
. ... ......
.......... ........
.......................................

-ref plane
. .....
.
"J*--,.............. ................
-. 11-111-1 - -c- - -. <7z - <, - i

E-ref plane

11-refplane- Ay

E-field
11-rield
Fig.2.4 : Extractionof electricandmagneticfields

2.14 Extraction of Sil

2.14.1 Two runs

The microstrip patch is a one-port device and therefore its scattering matrix has only one element Si I

which is the reflection coefficient. The reflection coefficient is given by :

Ereflecteil
r= (2.62)
El,
cident

In order to obtain the reflection coefficient,the incident and reflectedwavesmust be known. In FDTD
simulation,however,the calculatedelectricfields arethe total electricf icids. Oneway to obtainthe fields
separatelyis to obtainthe incidentwaveform,Ej,,,jd,,,,, by simulatingonly the microstripfeed-linewhich
extendsright throughto the absorbingboundary.This incidentwaveformcan now be subtractedfrom the
total waveform,E,,,t,,,, obtainedwhen simulatingthe rectangularmicrostrip patch to yield the reflected
waveform.The reflectioncoefficientis thencalculatedusing :

Ereflected
Et,,,,,,-
r- Ej, - Ei,,
(2.63)
cldent Id,, t

This meansthat two runs of the FDTD routineare requiredin order to obtainthe reflectioncoefficientof
the circuit. This is inefficient andtime-consuming.

2.14.2 Single run

It is found in this researchthat by extractingfour parametersinsteadof one from the FDTD simulationof
the circuit, only a singlerun is necessary
to calculatethe reflectioncoefflcientof the circuit.
Finite-Difference Time-Domain Method 25

Consider the voltage and current at any point on a transmission line

V VIcoshyx + IIZ,, sinhvx (2.64)

L'
Ix I, coshv x + sinhv x (2.65)
Z.
V, ( Z, + Z,, tanhyx
then = Z" (2.66)
Ix ZO + Z, tanhyx

V (VIsinhvx + IIZ,, coshyx),v

also (2.67)

aix V1
11sinhy x + coshyxr (2.68)
& zo

thenat x=0, setasthe referenceplaneat the monitoringpoint, dividing (2.67) by (2.68)gives

ÖV. /ex ilz, Z 2

= (2.69)
all, /IDX V,/zýI Z,

andfrom (2.64) and(2.65),with x=0,

7- Zi (2.70)

Note that Z, is not the load impedancein the conventionalsensebut the load impedanceat the monitoring
as
point x=0 has been set as the plane
reference at the monitoring point. So, Z, is the input impedance
at
the monitoringpoint andthe reflectioncoefficientat the monitoringpoint is given by :

- Z,, (2.71)
r=Z,Z, + Z,

or
VFZ-II/Z.l -I
l +I
(2.72)
VFZ-II/Z.
Representing(2.72) in the form of (2.69)and(2.70)givesus

/1,
V1,
FX//'2'va-x
aVxal., (2.73)
ac
r kV.
aax/IX
/k
/ Lax
+
V,aix
It canbe seenfrom (2.73) that the reflectioncoefficientof the circuit can be calculatedfrom a singlerun
of the FDTD routine if four parameters,namely, V, , I., , aVI& andL91., /vX are extractedfrom the

FDTD simulation.

Ub,
;;: y
Finite-DifferenceTime-DomainMethod 26

The derivationaboveis donefor wavepropagationin the x-direction.It canbe appliedto any directionof
propagationas long as appropriateparametersare extracted.In all simulationsin this thesisthe direction
of propagation is in the y-direction while the excited tangential electric field is in the z-direction.
Therefore,the requiredparameters are V, , ly jaV, ley andaly Icy .

2.15 Extraction of Zin

The input impedance,Zin, of the microstrippatchantennaat the edgeof the patchcanbe calculatedfrom
extractedin section2.14 by transformingthe referenceplane from the monitoringpoint to the
the S11(co)
edgeof the antennaandapplyingthe equationbelow.

j2pl )
Slie
Zj" =ZI+ J2,81
(2.74)
'( I-Slie

whereP is the phaseconstanton the microstripand I is the length from the monitoringpoint to the edge
of the patchantennaandZ,, is the characteristicimpedanceof the microstrip line.

Pi, :y= 50AY

y= 40Ay
Pmon2:Y= 35Ay

Fig.2.5 : Planview of the microstrippatchantenna

Fig. 2.5 showsthe plan view of the simulatedmicrostrippatchantennawherethe monitoringpoint is at

Non, andthe input impedanceto be calculatedis at Pin-By applyingequations(2.70) and(2.73),we have
the input impedance,Zj,, n,,nj andreflectioncoefficient,Siljnonjat the monitoringpoint, Pmoni-
In orderto
_,
apply (2.74) to find Zin at Pin,we needto calculateP, the phaseconstanton the microstrip and Z,, the
characteristicimpedanceof the microstripline. Re-arranging(2.74),we get

I- S11 )
zo Zin-monl mod (2.75)
=
I+Sll_monl

which gives us the characteristicimpedanceof the microstrip line. In order to calculate0, the phase
constant, V, , ly j aVlq andalylc'y are monitored at anotherpoint Pinon2in addition to point P.. j.

Again, applying(2.73)to the dataextractedfrom point Pmon2p

we obtainthe reflectioncoefficient,SIl_mon2-
Finite-Difference Time-Domain Method 27

Let
Ereflected
S11 (2.76)
MOM - Eicident

then
Ereflected e -J, 6 11
Sll-mon2
Ei,, 11
J-8 (2.77)
ciden, e
Sll_mon, -J2.8 11
e

sincethe reflectedwave at is
Prnon2 delayedfrom PI by a phaselength PLI and the incidentwave at
Pn....2 is aheadof Pn I by the sameamountwhereLI is the lengthfrom PI to P2 which is setat My.
Let 01be the phaseof S,I at P ,, and02the phaseof S,I at Pmon2,
then (2.77) becomes

I Sl I I L,
dol Sl doýe-j2P (2.78)
1_mon2 =I I_monl

01 '02
then (2.79)
2L,

Substituting6 andZ,,backinto (2.74),we get the input impedanceof the patchantennaas

or
jo 1'02
L,
L2
I- Sll-monl I+ Sll e
Zj" Zi, monl (2.80)
= MOM I+Sll
- 01 . 02
moni 1 L2
-)ý. I-Sll-,,, e L,
O,,,

Sincethe monitoringpointschosenareat PmonlandPmon2,

L, - My andL2ýI OAy,then (2.80)reducesto

1- Sil N, (01 - 02)

+ SI
Z,. zm monl 1 ment eJ2
' = monl
- l+Sll-monl 1-Sil-mon, j2 (01 -0 2)
e
.(

Equation (2.81) is used to calculate Zin in one simulation.

Finite-DifferenceTime-DomainMethod 28

2.16 Results

10 12 14 16 18 20

I freq (GHz)
0- . W-W ---111-1-- 11111f- ---I - --
1%%I. --

-20-

-30-
Abouzahra's simulation
Abouzahra's measurement
-40
Conventional FDTD

-50

Fig. 2.6 : ComparisonbetweensimulatedandAbouzahra'spublishedresultsfor I S,II

M.D.Abouzahraet al [1.5] discussedthe application of the three-dimensionalFDTD methodto the

analysisof planarmicrostripcircuits,oneof which wasthe line-fed rectangularmicrostrippatchshownin
Fig.2.3. In order to validatethe FDTD program,the patch in Fig.2.3 is simulatedusing the samemesh
parametersas in the publishedpaper[1.5] to allow exactcomparison.Fig. 2.6 aboveshowsa comparison
betweenthe publisheddata and the data generated.The simulatedresult agreeswell with Abouzahra's
simulatedresults.However, at high frequencies,
both Abouzahra'ssimulatedresultsand the simulated
resultsgeneratedin this work shifted slightly towardsthe lower frequency.The discrepancybetweenthe
simulatedand measureddata may be due to an increasein numerical dispersionat high frequencies.
Furthermore,the results have been obtained in the time-domain and then convertedinto frequency
domainby applying Fourier Transformmethodon the time-domaindata.Consequently,a small error in
the form of truncationerror in the time-domainwill result in a more significant error in the frequency
domain. Besides,the experimentaldata here are assumedto be error-frce which may not be a sound
assumption.The discrepancybetweenthe simulateddataandAbouzahra'smeasurement may well be due
to measurementerror. The simulatedresult in Fig. 2.6 was obtainedusing a singlerun FDTD methodas
describedin 2.15.2
Finite-DifferenceTime-DomainMethod 29

02468 10 12 14 16 18 20
01 ---, .--- -- .
freq (GHz)

-5

-10
Co

-15

-20

-25

-30

Fig 2.7 : Comparison between two-run and single run simulated results

In orderto validatethe single run methodof obtainingthe reflection coefficient for the microstrippatch
circuit as derived in section2.14.2,two setsof data are generated,one using the single run methodand
the other using the double-runmethod.Fig. 2.7 shows a comparisonbetweenboth sets of reflection
coefficients.They show exact agreement.Sincethe single run method is more efficient, all subsequent
resultsfrom this point onwardshavebeengeneratedusingthe singlerun method.

80-

40-

30-

20-

10

0- freq (GHz)
7.1 72 73 74 ýA 76 77 7.8
-10-

-23-
- Abouzahra'sreal conv-fdtdreal
-3D-
---- Abouzahra'sImag conv-fdtdImag
-40-

-
_ED

Fig. 2.8 : ComparisonbetweensimulatedandAbouzahra'spublishedresultsfor Zi,,

Finite-DifferenceTime-DomainMethod 30

Fig.2.8 above shows a comparisonof real and imaginary parts of the patch antennainput impedance
betweenthe simulatedandAbouzahra'spublishedresults.In Abouzahra'spublishedpaper,the microstrip
is assumedto havea constantcharacteristicimpedance,Z. of 50 0 and an effectivepermittivity of 1.9 is

usedto calculatethe wavenumber,, P. In this researchwork, thesevalueshavebeencalculatedusingthe

simulationdataas discussedin section2.15.Due to an inherentdispersivecharacteristicof the microstrip,
the effective permittivity is no longer static but changesas a function of frequency.Sincea Gaussian
pulsewhich containsall the frequenciesof interestis usedas an excitation,the dispersivenatureof the
microstrip patch due to the inhomogeneous media is automaticallyincorporatedin the full-wave time-
domainsolutionprovidedby FDTD. The variationof the characteristicimpedancein the frequencyrange
of interestis also accountedfor in the simulateddata.It is thereforenot surprisingto seethe discrepancy
betweenthe calculatedinput impedanceandthat presentedin the publishedliterature[1.5].

2.17 Conclusion

The three-dimensionalfinite-differencetime-domainalgorithm for solving numerical electromagnetic

problems has been introduced in this chapter. The salient features and key considerations in the
'
implementationhavebeendiscussed.Followingthis, the FDTD algorithm in this work hasbeenshownto
produceresultsthat agreewith publishedresultsusing a new more efficient techniqueof extractingthe
reflection coefficient. This methodof data extractionhas been shown to give identical resultsto the
conventionalapproachand removesthe necessityof performingtwo separatesimulationruns,therefore
saving simulation run-time. Another important point is that the input impedancehas been calculated
directly from simulated data which automatically incorporate the dispersive characteristicof the
microstrip; this removesthe need for a-priori knowledge of the line characteristicimpedanceand
effectivepermittivity. Unlike in [1.5], thereis no needto assumethe characteristicimpedanceof the line
andusean effectivepermittivity to calculatethe wavenumber.
Alternating-DirectionImplicit Finite-DifferenceTime-DomainMethod 31

CHAPTER 3
ALTERNATING-DIRECTION IMPLICIT
FINITE-DIFFERENCETiME-DOMAIN METHOD

3.1 Introduction
The FDTD method has been widely used in solving a broad range of electromagnetic problems, The

accuracy of the simulation can be greatly improved with the use of finer spatial increments, especially

where there are discontinuities in the structure. This though leads to the requirement of having fine mesh

sizes in localized areas. In the past, to maintain stability, as defined by the Courant-Friedrich-Levy (CFL)

criterion in (2.33), the time step used would have to be small. This would lead to a prohibitively long

simulation run-time if the object was electrically large but had small localized discontinuities,

In this chapter,the author showsthat with the application of the alternating-directionimplicit (ADI)
methodon the FDTD, the CFL stability constraintis eliminated[1.14] and thereforea singletime step,
largerthan the one allowed by the CFL criterion, can be usedfor all meshsizesthroughoutthe model.
The time stepis no longergovernedby the stability but by the accuracyrequiredfor the simulation.This
is particularly useful, for example,when modelling a probe-fed circular patch where the probe is
extremely narrow compared to the diameter of the patch. In order to representthe effective input
impedanceof the probe-fedcircularpatchaccurately,a high FDTD meshdensityis appliedin the vicinity
of the probeandthe meshdensitydecreases graduallyaway from the probe.Anotherareawherethe ADI-
FDTD may be useful is in the modelling of a structurecomprisingnarrow slots or notches.Here, fine
meshesarerequiredaroundthe slotsandnotcheswith a consequentincreasein computationtime.

This chapterwill discussthe key featuresin implementationof ADI-FDTD with particularemphasison

absorptionboundaries.Note that all the finite-differencealgorithms from this chapteronwardswill be
expressed such that the electric and magneticterms are staggeredby half a spacestep as depictedin the
original Yee cell to enablethe derivationof numericaldispersionsfor the ADI methods.

3.2 Three-dimensional ADI-FDTD algorithm

The conventionalADI method has been widely used to solve many diffusion problems.As statedin
equations(1.6a)and (1.6b) in chapterone,the conventionalADI finite differenceequationsaresplit into
two proceduresfor a two-dimensionalADI scheme.Correspondingly,a three-dimensionalconventional
ADI method will require the finite difference equationsto be split into three procedures,each one
replacinga spatial derivative with an implicit differenceapproximation[2.4]. However,unlike in the
conventionalADI method, in this three-dimensionalADI-FDTD method,the formulation is split into
only two procedures,eachone replacingeachspatialderivativein the Maxwell's curl equationswith an
Alternating-Directionimplicit Finite-DifferenceTime-DomainMethod 32

implicit difference approximation. Procedure I is applied for advancement from nAt to (n+ 1/2),&t while

2 is for from
procedure used advancement (n+ 1/2)At
to (n+I)At.

Equations(3.1) - (3.2) showthe numericalformulation for procedureI of the ADl-FDTD method.The

electric and magneticfields are spatially staggeredas in the conventionalFDTD. Examiningequation
(3.1a), one notes that it has a form which is reminiscentof the conventionalFDTD as in (2.13a).
However,whilst all the H-field termsin the MIS of (2.13a)in the conventionalFDTD areexplicit, i.e. all
valuesareknown,the H-fields on the RHS of (3.1a)in the ADI-FDTD havetwo implicit termswhich are
yet to be calculated.The sameform runsthroughall subsequentequationsfrom (3.1a) to (3.2c),i.e. there
aretwo implicit termsin eachequation.

Procedure I
H "+112(1
+1/2j+1/2, k) ll, "+1/2(1+1/ 2J-1/2, k)
At AY
En+112(i + 112j, k) = E," (i + 1/2j, k) +
x 2c Hn(! +112jk+112) Hn(1+112jk-112)
yy -
AZ

(3.1a)

H"+1/2(/J+1/2, k+1/2) -H "+112(ij + 1/2, k-1/2)

xx
At AZ
(ij + 1/2,k) = Ey"'(ij + 1/2,k) +
En+112 ,
y 2c H"(1+1/2J+1/2, k) - Hn(1-112J+112,k)
AX
(3.1b)

lin+II2 (1+ 1/2j, k+ 1/2)_. Iln+II2 (1-1/2jk + 112)

y y
At är
1,rn+II2 (ij k +I/2)= En (ijk + 1/2) + 7-' '
3tZ H»(ij +I/2, k +1/2) - 11"(IJ-1/2,
k + 1/2)
AY
(3.lc)

E.»(IJ+I,k+I/2) - E'(Ijk +1/2)

ät AY----Z
, -in+ 1/2 (ij+112, k+112) =Hn (ij+I/2, k+I/2)
x x - Ep-,
E n+112 k)
(ij+112, k+I)_Eyn+112(ij+112,
y
AZ
(3.2a)

E"(1+112j, k+l) - E"(i+112jk)

AZ
n+1/2 (i+1/2jk+1/2)
H yy = Hn (i+112jk+112)
- -iP
Ezn+112(i+ljk + 1/ 2) - Ezn+112(IJ, k+I/ 2)

Ax

(3.2b)
 implicit Finite-DifferenceTime-DomainMethod 33
_Alternating-Direction

En(i+lj+112, k) - En(!J+112, k)
y
at ax -Y
H n+1/2(i+112j+112, k) =H n (i+1/2j+1/2, k) - ,
zz2, u E,"+1/2(i+ 112J+I, k) - E,"+1/2(1+1/2jk)
AY
(3.2c)

Becauseof the unknown implicit termson the RIIS, equations(3.1).cannotbe solveddirectly as in the
conventionalFDTD. To solve equations(3.1), the LHS electric fields haveto be expressedsuchthat the
RHS terms are all known values.In other words, the RIIS terms have to be in the form of previously
calculatedvalues. This can be accomplished by substituting equations (3.2) into (3.1) appropriately;
specifically,substituting(3.2c) into (3.1a)resultsin (3.3) below. In equation(3.3), the LIIS forms a tri-
diagonalmatrix of E., when E,, is scannedin the ý direction. The MIS of (3.3) now consistsof only

explicit, known terms.This tri-diagonalmatrix is a sparse

matrix that can be solvedefliciently [2.6].

ý, )2
U-cAy
(i+112j-l, k)-E.,,"+112(i+112jk)
E n+112 2+ + E,"+1/2(1+1/2j+l, k)
x it
.11
-2 Ay (Ay
NTYc [E"
E " (i +I/2j, k) , -L (I+lj + 1/2, k) - Ey"(ij + 1/2, A) - Ey"(W J-1/2, k) + En (ij-1 /2, k)
ý, Ax) yy
(p
L
EuAy [II"(i Ay [Hy
+ 112j + 1/2, k) - 11"(i + 1/2j-1/2, k)] + "(I + 1/2j, k+ 1/ 2) 11" (I + 1/2j, k-1/2)]
- y
AI) zz21

(3.3)

Ey and E, can be solved in a similar manner with the former resulting in a tri-diagonal matrix when

scannedin the direction of 1 and the latter in i in procedure 1. Once all the electric fields are computed,

the magneticfields canbe computeddirectly using(3.2a)- (3.2c).

Equations(3.4) - (3.5) showthe numericalformulationfor procedure2 of the ADI-FDTD method.Those

partial derivativesthat were replacedwith implicit approximationsin procedure I are expressedin
explicit approximationsin procedure2 and vice versa.The switching betweenthe implicit and explicit
expressions in the two half time stepsgivesriseto the namealternating-direction.

Procedure 2
H.""1120+112j+112, k) -H "+1/2(1+ 1/2j-1/2, k)
Ay
E"+' 0+ 1/2j, k) =E "+112(i+112jk) +
xx 2s Hn+l(i+112jk+112) H"+1(1+112jk-112)
- y
AZ
(3.4a)
Alternating-Direction Implicit Finite-Difference Time-Domain Method 34

(ij + 1/2, k + 1/2) 11n+112(ij + 1/2, k-1/2)

H n+112
x x
A
At AZ
E "+' (ij + 1/2, k) =E n+112(ij+112, k) + ,0
yy 2c H n+l (i+112j+112, k) -H"+' (i-1/2j+1/2, k)
,Zz
I Ax
(3.4b)

H "+112(i+ 1/2j, k+ 1/2) -H "+1/2(1-1/2jk+1/2)

yy
At
E.""(ij, k+112)=E.. "+'12 (ijk+112)+
A Hn+l ij+1/2, k+1/2)-lix"+'(/J-1/2, k+1/2)
Ay

(3.4c)

E"+' (ij+l, k+112)-E"+' (/Jk+1/2)

At Ay
H,n+l(ij+112, k+112) =H, x"+112(ij+1/2, k+1/2)- 2,u E n+1/2(ij +I/2, k+l) -E n+112(ij+ 1/ 2,A)
yy
Az

(3.5a)

E'+' (i+II2J. k+1)-E»+' (i+II2jk)

At .x
Hn+'
yy
(i+112jk+112) =H"+112 (i+112jk+112) - 2p . n+I/2 1/2
Er (i+Ijk+112)-E»+ (ijk+I/2)
-ZAx
(3.5b)

En+' (1+IJ+l12, k)-E'+' (ij+I/2, k)

-y -- -y
Hzn+I(i+112j+l12, k)=H»+112 (i+I/2j+1/2, k)- -iß
Z n+II2
Ex (1+112j+I, k) -Ex n+112(l+ 112jk)
AY

(3.5c)

Again, substitutingequation(3.5b) into (3.4a)andcollectingthe E, termson the left give rise to equation
(3.6) below where all the RHS termsare explicit terms, i.e. known values.Repeatingthe sameprocess
over equations(3.4) leadsto tri-diagonalmatricesfor E.,,,
Ey E,
and when the fields are in
scanned the i,
i and ý directionsrespectively.The full formulationof electric fields in both proceduresI and2 are in
Appendix B I.

1 1
ucAz
E; +'(i+112j, k-l)-E, "+'(i+112j, k) 2+ + En+'(i + 1/2jk + 1)
141
(AAi [E,
ucäz
E", 1/2(i+112ik)

(,
EI
cýI
At
+
x)
»"1/2(i+Ijk + 1/2) -
"
E, +1/2
(ijk + 1/2) - Ez»"/2 (i+Ijk-112) + E"+' /2 (ijk. 1/2)

uAz) [H;; 11/2(1 1/2jk [lln+I/2

(1 + 1/2j + 112,k) - lln+I/2 (i+I/2j-1/2, k)]
- U-At + + 1/2) - lly»"12 (i+I/2jk-1/2)

(3.6)
Alternating-Direction Implicit Finite-Difference Time-Domain Method 35

3.3 Physical Interpretation of three-dimensional ADI-FDTD method

x x
Hz
conv-fdtd
Hy
(n+1/2)At

Ex
time
adi-fdtd
I
procedure
I Iy/2

nAt (n+I/2)At
Ex

adi-fdtd
procedure2 I ly/2
(n+I)At

Fig. 3.1 : ComparisonbetweenconventionalFDTD andADl-FDTD

Fig. 3.1 above illustratesa physicalrepresentationof the ADI-FDTD formulation in comparisonto its
conventionalFDTD counterpart.The diagramshowsthat in the conventionalFDTD method,the electric
field at time step(n+I)At is calculatedusing the previouslycalculatedelectric field at time stepnAt and
the curl of the known (hence explicit) magnetic fields, fly and 11, at time step (n+ V2)At.In the ADI-
FDTD method, an intermediate electric field is calculated at time step (n+ 1/2)At.In procedure I of the
ADI-FDTD method, the electric field at time step (n+ V2)At is calculated using, again, the previously

calculated electric field at time step nAt and the curl of the magnetic fields. However, this time, half the

curl is performed on the known (explicit) value, i.e. Ify at time step nAt, and the other half of the curl is

performed on the unknown (implicit) value, i.e. H. at time step (n+ 1/2)At.This is immediately followed on
by procedure 2 of the ADI-FDTD method; now the known (explicit) value of II,, at time step (n+ Y2),
&t

and the unknown (implicit) value of Hy at time step (n+I)At are used in the curl formulation. The total
magnetic field over a full time step remains unchanged.Note that although the curl at each half time step
is separatedinto two different time instances,it is still performed at a same point in space.

A furthergraphicalillustrationcomparingimplicit, explicit andADI-FDTD methodswith referenceto the

Runga-Kuttamethodis shownin AppendixCl.
Alternating-DirectionImplicit Finite-DifferenceTime-DomainMethod 36

3.4 Divergence of ADl-FDTD algorithm

As in the conventionalFDTD algorithm,thereis no explicit enforcementof the Gauss'sLaw relationsfor

both the electricand magneticfields as statedin (2.3) and (2.4) for sourcefree regionsin the ADI-FDTD
algorithm. It is importantthat Gauss's Law is also observedin the ADI-FDTD algorithm. Although the
curl operationof the magnetic fields is performedover two half time steps,the total magneticfield over a
full time step remainsunchangedin the ADI-FDTD scheme.Therefore,the ADI-FDTD algorithmwill
still satisfyequation(2.16), i.e. the time derivativeof the net electric flux leavingthe surfacesof a cubic
Yee cell is zero,henceupholdingthe Gauss'sLaw for the electric field in charge-freespacein the ADI-
FDTD scheme.

3.5 Numerical stability

As in chapter two, numerical stability of the ADI-FDTD can be analysedusing the standardvon
Neumann and Courant, Friedrich and Levy (CFL) method. Assuming the spatial frequencyto be
kx ky andkz
9 as the x- , y- and z- componentsof its numerical wavevectorrespectively,the field

componentscanbe written as follows.

(3.7a)
E'(i+112, j, k) = E., expf-jrk,, (i+1/2)Ax ZyjAy ZkAz])
+ +
x

Ty(J+112)Ay T: kAz]) (3.7b)

E'(i, j+112, k) = En expf-jrkxiAx + +
yy

f-j rk, Ty T: ) (3.7c)

E' (i, j, k+I/ 2) = E" exp iAx + jAy + (k +I/ 2)Az]
z

Ty(J+112)Ay T, (k+1/2)Az]) (3.8a)

Hn(i, j+112, k+112) = Hnexpf-jrkxiAx + +
xx

rk, (3.8b)
Hn (i + 1/ 2, j, k+I/ 2) = Hn exp j (i + I/ 2)Ax + TyjAy + T, (k + I/ 2)Az]
yy
rkx Ty (3.8c)
HnQ+ 1/ 2, j+ 1/ 2, k) = H" exp j Q+ I/ 2)AX + (i +I/ 2)AY + T, kAZI)
z2

wheren, i, j, k, Ax, Ay andAz all havetheir usualmeaningsasdefinedearlier in chapterone.

3.5.1 2-dimensional ADI-FDTD

For the sakeof simplicity, we considerfirst the numericalstability of a 2-dimensionalTE waveconsisting
of the following fields :

Procedure I
(H n+112 1
ät (i+I/2, j+I/2) - H"+"2(i+I/2, j-1/2)
E"+' 12(i+112J) = E"(i+112, j) + 2s1 (3.9a)
AY

At H"(1+1/2, J+1/2) -Hn (/-1/2, J+1/2)

Eyn+112
(i, j+ 1/ 2) = Ey"O,i+ 1/ 2) ---z -- z (3.9b)
2c Ax
Alternating-Direction Implicit Finite-Difference Time-Domain Method 37

En(i+II2, j+II2) - Ey"(1,

J+112)
y
A' AX
Hn+'/2(i+I/2,
x9 j+1/2) = Hn(! +112, j+112) - 2p ý n+II2 n+I/2
Ex (1+112,j+1) - Ex (i+I/2, j)
AY

(3.9c)
Substituting(3.7a),(3.7b)and(3.8c)into (3.9) yieldsthe following :

(ý2
At
Ex *GFI = Ex - GFI*H 2j sin jyAy) (3.1Oa)
2cAy

At i
Ey *GFI = Ey + H., ý 2j sin(L2xAx) (3.1Ob)
CAX

(iL2 (L2 Ay)

At At
Hz *GFI = Hz + Ey ý- 2j sin x"AX) GFI* Ex2, 2j sin yi (3.1Oc)
,u Ax Ay

whereGFI is the growth factor in procedure1. Substituting(3.1Oa)and (3.1Ob)into (3.1Oc)gives:

(iL2--) (L2
At Ax At
H: 2j si, , 2j sin xAx)
x-
2cAx 2,u Ax
H, *GFI=H: +
GFI -I

At Ay TyAY
y At
GFI* GFI*H, 2jsin 2jsin -2)
2cAy 2p Ay
-4 GFI-I (3.11)

N-12 )] 2
k,, Ax At Ay
H., (GFI - 1)2 H, sin"ý2- )] GF12H sin
y (3.12)
c Ax c Ay 2

ý2- z Ay)
A, AX) Ay
Let mx sin xAx m=Y Y
y
Ax and -L' sin -2
Ay

and dividing (3.12) by H,

(I 2) (I 2)
GF12 + -Y, - 2GFI ++ -.!!,! =0 (3.13)
U-"C

Let a 1+ýHy" and cM 1+

then (3.13) becomes

aGF12 - 2GFI +c=0 (3.14)

Alternating-DirectionImplicit Finite-DifferenceTime-DomainMcthod 38

±41-ac
GFI =I
a
2p
andsince 0 :5 sin :51, then ac ;->I

GFI

Procedure 2
At H ,+112
(1+ 1/2, J+1/2) - , n+1/2(1+1/2, J-1/2)
E n+l(i + 1/ 2, j) =E n+112
(I+112, j) +z
x AY
(3.16a)

1
At Hn+'(1+112,J+112) - 11'+1(1-1/2, J+1/2)
E; "/2 j+I/2) (3.16b)
- 2c AX

E""1(1+1/2, J+1/2) - Ey"*'(I, J+II2)

Y
At
H'+'(i+112, j+112) = Hn+1/2(i+1/2, J+1/2) - 2p .
z E"+1/2(1+1/2,j+l) E"+"2(1+1/2, J)
xx
AY
(3.16c)
Again, substituting(3.7a),(3.7b)and(3.8c) into (3.16)will yield the following

HA 2jsin
(T
y
NY
AY (3.17a)
E" * GF2 = E,
AV
y2
(LXPI)
At x
E *GF2=Ey + GFVH, 2j
ý &x sin (3.17b)
y C,

(LX2 (Z2
i- AY
&t Ax At y
Hz *GF2 =H+ GF2*E 2jsin xi Ex 2jsin (3.17c)
2p Ax 2,u Ay

whereGF2 is the growth factor in 2.

procedure Applying the sametechniqueon 2,
procedure we get,

(I (I
GF12 + 2GFI + +0

±jvrac-I
GF2 =I (3.19)
c
Therefore,the total growth factor of proceduresI and2 combinedis given by :

GF =IGFII*IGF21
Jc
E
C (3.20)
Alternating-DirectionImplicit Finite-DifferenceTime-DomainMethod 39

Since the overall growth factor of the 2-dimensionalADI-FDTD is unity, the system is said to be
unconditionallystable[1.14]. However, if we note carefully, the growth factor (3.20) or the gain of the
ADI-FDTD system,is exactly unity; potentially any slight increasein the growth factor due to any
truncationerrors may causethe systemto go unstable.This 'exact' phenomenonis discussedin more
detail in [1.17].

3.5.2 3-dimensional ADI-FDTD

To analysethe numericalstability of a three dimensionalADI-FDTD, we apply von Neumannmethod
again on the three dimensionalADI-FDTD formulations (3.1) - (3.4). Denoting all the electric and
magneticfields asX matrix,procedureI canbe written in the form :

Xn+112 GFI * X"

= (3,21)
andprocedure2 as
X"' = GF2*X""12 (3.22)

The overall growth factor for the proposedschemeis then given by :

GF =I GFI I*I GF21 (3.23)

where

I Mx-My -J. Mz J-my

0 0
7y- ýY-c
Ny-p -e Ny-z
1 My -Mz j-Mz J.
: mx
0 o
Nz Nz.p e Nz-c Nz-c
Mx-Mz 1 -j-My J-Mx
0 iTx ýxc 0
GFI = Nxg -E Me (3.24)
j. Mz -j-My 1 Mx-Mz
0 0
Nzg Nzp Nz Nzg -c
-j. Mz j-Mx Mx-My 1
0 0
Nxp Nxg Nxp -c Nx
j-My -j. Mx My. Mz I
0 0 Ry-
Ny-p Ny-p Ny-p-e

and
1 Mx-M Mz j-My
0 0 "j.
TZ Nz-p -c Nz-E Nve
Mx-My 1 imz -j-Mx
0 0
Nx-g Nx Nz-e Nz.c
My. Mz I -j-My j-Mx
0 Ry 7iTy7e 0
Ny -g -c Ty
GF2 = -6 (3.25)
j.Mz -j-My I Mx-My
0 Ny -p NY Ny -p -E
0
Ny -g

-j-Mz ýmx 1 My. Mz

0 0 -Fz
Nz-p Nz-g Nz-p -c
i-my mx Mx. Mz I
-j. 0 0
*ýx- ,
RX
Nx-p Nx-p ve
Alternating-DirectionImplicit Finite-DifferenceTime-DomainMethod 40

where
! hAh) M 2h
sin( and Nh 1- 1+ h- X, Y, z (3.26)
h "2-L'
Ah 2 PC

It has been shown [1.18] that the eigenvaluesof GF all have magnitudesof unity. Again, with the
theoreticalgain of unity, the threedimensionalADI-FDTD is saidto be unconditionallystable.

3.6 Numerical dispersion

The numerical dispersion for the ADI-FDTD method can be found by substituting the vector-field
travelling-waveexpressionwith time dependenceshown below into the ADI-FDTD finite-diffcrcnce
equations.

(3.27a)
E'(i+112, j, k) = Eexp(jo)nAt-jrk, (/+1/2)Av + k-yjAy + k-,kAzl)
x

Ty(j+112)Ay TkAz]) (3.27b)

E"(i, j+112, k) =E exp(jo)nAl-jrk, /&v + +
yy

WkAz]) (3.27c)
Hn (i+112, j+112, k) =H expjjo)nAl-jrk,, (1+1/2)AX + k-y(J+112)AY +

3.6.1 2-dimensionaIAD1-FDTD
Again, for simplicity, we investigatethe numericaldispersionof a2-dimcnsionalTE wave [1.19].
Substituting(3.27) into (3.9) (procedure1) gives:

(7Ay) (i2 AY)

jw At/ 2 Al y
j jwAl/2 in
_I)n=jEx sin e (3.28a)

jwA112
( At )i) Lx
_ 1) En k7- XPI I/: - (3.28b)
y , Ax)sin(

)( (i2 Ay)
At At
( jwW/2 x E" y Jd#At/2Em
e _n 1) Hz sin sin ex (3.28c)
y(p -j Ay)

and into (3.16) (procedure2) gives:

(7Ay) (i2
A' Ay)
jw, &t jwAt12 J- jivAt/2
(e - e Exny =j sin e (3.29a)

(i2AX)
jmAI12 At ) ý!
(eimAl
-e )E' j((7c sin x. ejwA'Iln (3.29b)
y Ax)
) (1-2 ) (.LWyAy )
jwAl jwAI/2 At Ax) JwAl At jo)AI/2 En
(e -e )Hn j(r. - sin e En j(rfl sin e (3.29c)
u AX) y Ayj 2
Alternating-DirectionImplicit Finite-DifferenceTime-DomainMethod 41

Combining(3.28a)& (3.29a),(3.28b)& (3.29b)and(3.28c)& (3.29c)givesrise to the following:

A, Ay
y jw At 2 H, "
(ejo A' -I) E' ' 2i -Ay sin e (3.30a)
2

(2)
At ) kxAx (e jwAl
(ej' At I) Ey" ý-c -Ax) sin + I)H: n (3.30b)
-

( ) )x
) wx AX Ay
At At Y2
jwAl 2j Cp JwAl/2 E' (3.30c)
(ei"Al I) Hnf sin e +1) E" - sin e
- z ý7-,
K;) 2)y Ay)

(3.30)canbe simplified to :

(W At )' )Y AY
siný72 E" CA sin H" (3.31a)
= 2 Z
y

( kAx ) ý-O )H,

(W At
(± At )
sin 2-) E"Y = sin 2 coý 2 -1111 ' (3.31b)

(W At ) T, (COAt TYAY
At AX At Y)
y
sin 2 H.' sin CO
sN-) E" ýPAY sin 2 Eff (3.31c)
AX 2 Y

or

(W At LTYAY
At Y
sin 0 -AY)(sin 2

(W
N-)At (L (0i At E"
At AX X
0 sin J'L)
sin 2 cos(L27) E; 0
Y
[(_ At LZ
ýYAY ) ( LZ,, 4)
(a) At (wAt
(± )
Y At Ax
xL
sin 2 74AX sin 2 co sin 2-
iT A-y

(3.32)
Thusthe numericaldispersionrelation for a 2-dimensionalTE wave is given by makingthe determinant
of the matrix zero,i.e.

(WAt
(± (At)2 ýi ( ýL
2( WAt I 2(ix.
sin -)
2
sin -2e Ax
sin 2
COS2(wAt
2
('TyAy) 1 (3.33)
(co
(± At I( At)2 2 0
+ sin 2-) - PC rAy) sin -
2

or
2 (± At) 2(2 Ly Ay (7A1 )2 (w At
ly
2(TXAX 2(01L2 2y 2 (3.34)
sin Cos - +( T_) sin ' 22 sin -E-)
Ix 'y t

where c= IlV-pe
-
Altcrnating-DirectionImplicit Finitc-DiffercriceTime-DomainMethod 42

Comparing(3.34), the dispersionrelation for a 2-dimensionalADl-FDTD with the known numerical

dispersionrelationfor a 2-dimensionalFDTD :

T" AX 212a L- (71

2 2 YAY 2(0)Al (3.35)
sin +(71 sin = sin
2y(2) ')

(w At)
(-a2-
it can be seenthat there is a differenceof a factor of cos2 in one of the left-hand-sideterms,Le.

T,.
the term with numericalwavenumber The dispersionequation(3.34) was derivedfrom the 2D ADI-
FDTD equations of (3.9) & (3.16). In procedure 1, (3.9), the 11. field was calculated using explicit Ey and

implicit E,, while in procedure 2, (3.16) the 11, field was calculated from implicit Ey and explicit E,.

According to Peaceman and Rachford [1.13], the ADI method remains unconditionally stable so long as

the two procedures are repeated over the same time step, one after another. Indeed, there is nothing to

stop us writing out the 2D ADI-FDTD equations with the implicit and explicit terms interchanged. In
W) At
2ý2/Y
such a case, the factor cos will be imposed on the numerical wavenumber k- rather than k..

The additionalfactor in (3.34) meansthat the variation in the numericalphasevelocity due to numcrical
dispersionfor the ADI-FDTD schemechangesin a non-uniform manner,dependingon the directionof
(E2-)
wave propagation, as the time-step,At, is increased.
Consider(3.34), where the COS2(4'At factor is

imposedon k, For wavepropagatingin the direction of x, kym 0, then the numericalwave velocity is

from that of the standardFDTD due

scheme to the factor cos2("'). On the other hand,for
reduced 2

wave propagatingin the direction of y, k, = 0, then the numerical wave velocity in the ADI-FDTD
is
scheme the sameasthat of the FDTD
standard [
scheme 1.19].

3.7 Implementation of I't order Mur absorbing boundary condition

The secondapproximationof the I" order Mur absorbingboundarycondition (2.55) is applied to the
ADI-FDTD algorithm. The implementationof the absorbingboundarycondition Is shown here in two
ways. The first way is to implement the absorbing boundary condition simultaneouslywithin the tri-
diagonal matrix when the rest of the fields are calculatedand the secondway is to implementtile
absorbingboundarycondition recursively after the other internal fields are found by solving the tri-
diagonalmatrix. Although the first way may seemeasierto implement,it is indeedan incompleteway.
Despitethis, the authorfeels it is necessaryto describethis incompleteapproachand explain why it can
leadto incorrectresults.

3.7.1 Boundary condition within the tri-diagonal matrix

The first methodis to apply the boundarycondition as part of the tri-diagonal matrix so that all fidds
within the computationalboundaryare computedwhen the matrix is scannedthrough in eachdirection.
Alternating-Direction Implicit Finite-Difference Time-Domain Method 43

To illustratethe implementationof the absorbingboundarycondition with the tri-diagonalmatrix, the tri-

diagonalmatrix of (3.5) andthe I' orderMur absorbingboundaryequation(2.55) are repeatedbelow.

( ý--Uc Y I+
Ay
E; +1/2(i+112j-l, k)-E., "+1/20+112jk) 2+ý- E"' 1/2(1+112J+I, k)
xx

(EA 2 (-ýAy- [Ey

6A+Y)
(i+112jk) 'u (! +Ij+112, k) -EY"(IJ+112, k) -EY" (I+IJ-1/2, k) + Eyll(ij-112, k)
-En y
t
(.
NA
('UAY
LI) [11"(i+112j+II2, U,&y2 [Ily
k)-H, "(1+112j-112,k)I + -LAil)7 "(i+1/2jk+1/2)-Ilyn(1+1/2jk-1/2)I

(3.36)

(Nfp- 2
21: -VOiAY
Let b2+ -,
At

I" order Mur absorbing boundary equation is given by:

rýVät Axý(EI»'»'
En+I En + « En (3.37)
01 = ý7A7+ AX) - 0)

(vAt Ax)
Now, let Mur-x - ýv& -
+ Ax)

andre-arranging(3.37),we get
EO" - (Mur_x . EI""') - Ei' - (Mur_x Eo») (3.38)

Let all the termson the right handsideof (3.36) be known as 'rhs' and for the purposeof this Illustration,
we shall consideronly six spatialstepsin the y-direction.Now, incorporating(3.38) into the tri-diagonal
matrix (3.36) givesus a tri-diagonalmatrix of the form below.

0 0 0 0- En+112(1+112,0, k)" "Ein

1 - Mur x - (Mur_x. EO')"
1 b 1 0 0 0 En+"2(1+112,1, k) rhsat j-1
x
0 1 b 1 0 0 Exn412 (1+1/2,2, k) rhsat j=2
b 1 0 1 En+' 12(l+ 1/2,3, k)
(3.39)
0 0 1 rhsat j-3
x
0 0 0 1 b Exn+l12(1+ 1/2,4, k) rhsat j-4
LO 0 0 0 -Mur x 1 Ex»" 12(1+1/ 2,5, k)i L E4n - (Mur x. E; )j
- -5

The samecan be applied for Ey & E, in procedureI and then all the electric fields in procedure2.
However, upon close observationof the matrices,one can identify a problem with this methodof
implementation.In procedure1, only one out of the two normal incident fields at each boundaryis
implementedasa one-waywaveequation.In procedure1, E,,,seesabsorbingboundariesat the y-direction
boundaries,Eyat the z-directionboundariesand E, at the x-direction boundaries.Then in procedure2, Ex
is 'absorbed' at z-direction boundaries,Ey at x-direction boundariesand E, at y-dircction boundaries.
This is an incompleteimplementationas all normal incident fields should seethe appropriateabsorbing
boundariesfor both proceduresI and 2 at eachhalf time step.Therefore,althoughat first sight it may
seem simpler to implement the absorbingboundarycondition simultaneouslywithin the tri-diagonal
Alternating-Direction Implicit Finite-Difference Time-Domain Method 44

matrix, this methodcan in fact leadto incorrectresults.This problemcan be resolvedif the other half of
the normal incident fields that are not implementedas a one-waywave equationwithin the tri-diagonal
matrix arerecursivelycalculatedseparatelyfrom the tri-diagonalmatrix.

3.7.2 Boundary condition outside the tri-diagonal matrix

The secondmethod is to initially solve the tri-diagonal matrix for all the fields in the computational
domain except for those at the boundaries.The field values at the absorbing boundariesare then
calculatedrecursivelyusingthe field valuesfound from solving the tri-diagonal matrix. Theseboundary
field terms, however,haveto be taken into accountwithin the tri-diagonal matrix. For consistency,six
spatialstepsin the y-directionare used.This gives rise to only four internal ficids excludingthe fields at
the absorbingboundaries.First, we write the tri-diagonalmatrix for the four internalf iclds.

n+112(1+ 1/2,1,
b100 Ex k) rhs at j=1
n+1/2
1b10 Ex (l+ 1/2,2, k) rhsat j=2
(3.40)
01bI E"+' /2(1+1/2,3, k) rhs at j-3
x
01 b-. En+l /2 (1+1/2,4, k)j at j= 4_
-0 -rhs

Incidentally,(3.40) is the final tri-diagonalmatrix to be solvedif the computationalboundaryIs a perfect

electricwall boundary.

To includethe fields at the absorbingboundaries,we re-write (3.40) to give :

b100 Exn+112
(i+112,1, k) (rhsatj=l) E"+1/2(1+1/2,0, k)
x
Ib10E n+1/2(i+1/2,2, k) rhs at j=2
x (3.41)
01bI E"+' /2(1+1/2,3, k) Hisat j=3
x
01 b-,. En+l/2 (1+1/2,4, k), (rhsat j= 4) - E"+' /2 (1+ 1/ 2,5, k)
-0 x_ x

Now, re-arrangingthe I" orderMur absorbingboundarycondition,

Eo"" /2 = El" - (Mur_x Eo") - (Mur_x E,"*1/2) (3.42)

Substituting(3.42) into (3.41) andre-arranging(3.41) to havethe n+1/2 termson the Ll IS givesus

b+Mur x100" Em+l /2(1+1/2,1, k) (rhsat j= 1) - Et' + (Mur-x. Eo")

x
Ib10 Exn+112(j+ 1/ 2,2, k) rhsat j=2 (3.43)
01bI Exn+112
(i + 1/ 2,3, k) rhs at j-3
b+Mur-XJLExn+l /2(j+ 1/ 2,4. k)j
L001 L (rhs at j= 4) - E4" + (Mur-x. Esn)

The samecan be appliedfor Ey& E, in procedureI and then all the electric fields in procedure2. With
this methodof implementation,the fields at the absorbingboundarieshave to be calculatedseparately
from the tri-diagonal matrix. This can be donerecursivelyafter the internal fields of the computational
Alternating-DirectionImplicit Finite-DifferenceTime-DomainMethod 45

domain are obtained.In all simulationsusing the ADI-FDTD method, the I' order Mur absorbing
boundaryconditionis implementedusingthis technique.

3.8 Simulated results

Severaldifferent modelswere simulatedto showthe applicationof the ADI-FDTD method.Eventhough

the ADI-FDTD methodwill seeits time-savingbenefit most in a model wherethere is a variety of mesh
sizes,for the purposeof showingthe applicationof the ADI-FDTD method,a regular finite-difference
mesh has been usedin all the following This
models. meansthat the worst casehas beenmodelledwhen
the stability factor is increasedbeyondthe CFL constraint.Since in the ADI-FDTD methodthe time step
is no longer restrictedby the CFL criterion (2.33), the time step At , can be set to be greaterthan the
maximumallowed by the CFL criterionmodifiedbelow by including the stability factorterm.

stability factor
At (3.44)
111
v +7 -72 "r
y A. 2

wherea stability factor of 1.0impliesthe maximum At asallowed by the CFL constraint.

In all structuressimulated below, the conductor is treated as a perfect electric wall boundary.The
dielectric is modelledin the sameway as in the conventionalFDTD methoddescribedin chaptcr2 anda
Gaussiansourceis usedfor excitation.

Initially all fields are set to zero. As time-marching progresses,the sequenceof the implemented
algorithmis as follows :

Procedure1:
E fields are calculated by solving the tri-diagonal matrices, one of which Is (3.5)

tangential E fields are set to zero on the metallized patch

tangential E fields on the computational boundaries are calculated using the Mur's I st

orderabsorbingboundarycondition(2.55)
the vertical electricfield is excitedwith Gaussianpulsebelow the strip
H fields arecalculatedusing(3.2)
electric and magnetic field values are stored for later use in absorbing boundary
conditioncalculationandfor field calculationsin procedure2
the time stepis incrementedby half At

Procedure2:

E fields are calculatedby solvingthe tri-diagonalmatrices,oneof which is (3.6)

tangentialE fields aresetto zeroon the metallizedpatch

Alternating-DirectionImplicit Finite-DifferenceTime-DomainMethod 46
.
tangential E fields on the computationalboundariesare calculatedusing the Mur's I st
orderabsorbingboundarycondition(2.55)
H fieldsarecalculatedusing(3.4)
electric and magnetic field values are stored for later use in absorbingboundary
conditioncalculationandfor field calculationsin procedureI
the time stepis incrementedby half At

2
To model a cavity, the calculationof the tangential fields on the computationalboundariesis left out
becauseby solving the tri-diagonalmatrix in the form of (3.40), a perfect electric wall boundaryon the
computationaldomainis automaticallyassumed.

3.8.1 Three-dimensional cavity

Perfect
electricwa
boundary

Fig.3.2 : Three-dimensional
cavity

A simple three-dimensionalcavity, filled with air, and boundedby a perfect electric wall boundary,
shownabovein Fig.3.2 is usedto validatethe ADI-FDTD method.
Meshparameters:
Ax = 0.2mm AY = 0.2mm Az'- 0.2mm
Total meshdimensions: 60 x 60 x 60 in 1, ý andi directionsrespectively
Critical time stepAt = 0.3851626ps
Alternating-Directionimplicit Finite-DifferenceTime-DomainMethod 47

0.16-

0.14-

0.12-

0.1
I con%Lfdtd
0.08
adl-fdtd st I
w 0.06

0.04

0.02-

0 4-
200 250 300 350 400 450 500
time (ps)

Fig. 3.3 : ComparisonbctweenconvcntionalFDTD andADI-FDTD

rcsultswith stability factor I for a threc-dimensionalcavity

Fig. 3.3 shows the time-domain results for wave propagation in a three-dimensional cavity with perfect
electric wall boundary. The stability factor used here is 1.0; that is the time step used Is the critical time
step. It shows good agreement between the results using the conventional FDTD and the ADI-FDTD
methods. Fig-3-4 shows that the results from the ADI-FDTD are still stable with stability factors 2,5 and
10 although the effect of numerical dispersion begins to show when stability 5 and 10 are used. However,
Fig. 3.5 shows that using the conventional FDTD method, increasing the stability factor to 2.0, thus

violating the CFL stability criterion, immediately causesthe results to go unstable. In all casesabove, the

position of the monitoring point is not important as the main aim of these simulations is to show that

unlike the conventional FDTD method, with ADI-FDTD method, the results rcmain stable even when
CFL criterion is not observed. More detailed results showing the accuracy of ADI-FDTD results against
the stability factors used will be shown and discussed in chapter 5.

0.18-
0.16-
0.14-
0.12-
-adi-fdtd st I
E 0.1 *. adl-fdtd st 2
0.08 - adl-fdtd st. 5
0.06 -ý' ....... adl-fdtdat 10
0.04-
0.02
0
200 250 300 350 400 450 500
time (ps)

Fig. 3.4 : ADI-FDTD resultswith stability factors 1,2,5 and 10

for a three-dimensional
cavity
Alternating-Direction Implicit Finite-Difference Time-Domain Method 48

2.5E+306

2E+306

1.5E+306
'9
E 7ds,
ý,
1E+306 I-con%, Id f 2.0
LU

5E+305

0
DO

-5E+305
time (ps)

Fig. 3.5 : ConventionalFDTD resultswith stability factor 2.0

for a three-dimensional
cavity

3.8.2 Three-dimensional cavity with inhomogeneous media

Perfect
electricwe
boundary

cavity with inhomogcncousmedia

Fig.3.6 : Three-dimensional

In order to model inhomogeneous media,the three-dimensionalcavity is partitioned into two sections,

onewith its permittivity setto 64 andthe othersetto I as shownabovein Fig.3.6.
Alternating-DirectionImplicit Finite-DifferenceTime-DomainMethod 49

0.08

0.06

0.04

-? 0.02
E con*fdtd
adi-fdtd St 1
-0.02

-0.04

-0.06

-n np
time (ps)

Fig. 3.7: ComparisonbetweenconventionalFDTD andAD[-FDTD

resultswith stability factor I for a three-dimensional
cavity with
inhomogeneous
media

A comparisonbetweenthe resultsgeneratedusing the ADl-FDTD algorithm and thosegeneratedusing

the conventionalFDTD methodis shown in Fig. 3.7. Again, the results show very good agreement.
Fig.3.8 showsthat the resultsare still stablewhenusing stability factors2,5 and 10with the ADl-FDTD
algorithm.Again, the effectsof dispersion
numerical beginto appcarwhen the stability factor is Increased
beyond5

0.08

0.06

0.04

-? 0.02 -adi-fdtd St. I

E0 adi-fdtd st.2
N
adl-fdtd st 5
UJ
-0.02 ....... adl-fdtdSt. 10

-0.04

-0.06

-0.08
time (ps)

Fig. 3.8 : ADI-FDTD resultsfor stability factors 1.0,2.0,5.0and 10.0

3.8.3 Simulationof a line-fedrectangularmicrostriopatch

In orderto validatetheADI-FDTDscheme on a morecomplexstructure, a line-fedrectangular
microstrip
patchas shownin Fig.2.3 is modelled.As canbe seen,the resultsshownherein Fig. 3.9 eventually
became unstable. evenwhena stabilityfactorof 1.0is usedin the ADI-FDTDalgorithm.
This happens
ADI-FDTDhasbeenvalidatedin homogeneous
Sincethethree-dimensional andinhomogencous
media
Alternating-Direction Implicit Finite-Difference Time-Domain Method 50

different dielectric it
constants, was suspectedthat the introduction of the copper strip in the model
with
causedthe instability.

1.5

1
E
E 0.5
ýE
LU
0
00
-0.5

.1
time (ps)

Fig. 3.9 : Line-fedrectangularmicrostrippatchusing ADI-FDTD method

with stability factor 1.0

3.8.4 Three-dimensional cavity with a transmission line

To further examinethe stability problemencounteredhere, the three-dimensionalair-fillcd rectangular
cavity shownin Fig.3.2 is simulatedagain.This time, a transmissionline that extendsbetweentwo of the
boundaries
computational is introducedin the cavity, as shown in Fig. 3.10. The transmission line is
implementedasa perfectelectricwall boundarywhereall the tangentialelectric fields are setto zero.

t< y
x cavity with a transmissionline
Fig.3.10 : Three-dimensional

Fig. 3.11 showsthat the AD1-FDTD resultsare unstableeven when a stability factor of 1.0 is usedthus
confirming our initial The
guess. fact that this model has all the sameparametersas that which produced
the resultsin Fig. 3.2 exceptfor the inclusionof a transmissionline in the cavity is an Indicationthat the
transmissionline, i.e. a perfect electric wall boundarywithin the computationdomain, hasto be treated
with care.
Alternating-Direction Implicit Finite-Difference Time-Domain Method 51

0
00
-2

-4
E
-6
-
adi-itd
-8

. 10

-12

-14

-16
Ume(PS)

cavity with a transmissionline with stability factor 1.0

Fig. 3.11 : Three-dimensional

3.9 Conclusion
The two- and three-dimensional ADI-FDTD algorithms for solving numerical electromagnetic problems
have been introduced in this chapter. The important aspectsof this method have also been discussed.The
detailed technique of implementing the I" order Mur absorbing boundary condition in conjunction with

the ADI-FDTD schemehas been discussed.

Following this, the ADI-FDTD with the time steps2,5 and 10 times the maximumallowed by the CFL
stability criterion have been successfullyimplementedon an air-filled three-dimensionalrectangular
cavity and on a three-dimensional cavity
rectangular with inhomogeneous
media.Both setsof resultsarc
stable and agree well
reasonably with that producedusing the conventional FDTD method.This shows
that in the ADI-FDTD scheme,the time stepusedis no longerrestrictedby the CFL stability criterionbut
by the accuracyrequired in the model. This is a significant advancementin the field of numerical
as simulationrun-timecannow be reducedwithout causinginstability.
clectromagnetics

Nevertheless,thereare still someteethingproblemswith regardto implementingADI-FDTD on complex

structuresas seenwhen the ADI-FDTD method is applied on a microstrip patch, where the tangential
electric field on the copperpatch is forcedto be zero at eachtime step. Instability occursevenwhenthe
time stepis within the CFL constraint.This happenswhen a cavity with a transmissionline is simulated
with stability factor 1.0. This showsthat the implementation
of the transmissionline as a perfectelectric
wall boundary the
causes instability.

In the next two chapters,chapters4 and 5, we examinethis in greaterdetail and proposetwo approaches
of overcomingthis problem.
Modified Alternating-Direction Implicit Method 52

CHAPTER 4
MODIFIED ALTERNATING-DIRECTION IMPLICIT METHOD

4.1 Introduction

The introductionof the ADI-FDTD methodhashada greatimpact in numericalelectromagrictics.

For the
first time, simulationrun-timescan be speededup by using a bigger time step in the FDTD algorithm.
The time stepin the numericalalgorithmis no longer governedby the CFL stability criterion. In fact, the
uppermost limit of the time step used is restricted only by the Nyquist samplingtheory which statesthat
be
samplingmust carriedout at a frequency of at leasttwice the maximum frequencyof Interest
in order
to avoid aliasing. Therefore,the upper limit of the time step is an inverse of twice the maximum
frequencyof interest.

However,as explainedin chapter3, thereis a difficulty whenthe ADI-FDTD schemeis usedto modela
more realistic problem. Initially, as discussedin chapter 3, the ADI-FDTD methodwas applied to a
rectangularcavity and the simulatedresult agreedwith that obtainedby applying conventionalFDTD
methodeven when the CFL condition was violated in the ADI-FDTD algorithm. But when the AD[.
FDTD schemewasusedto modela complexgeometrysuchas a microstrip patchantenna,the resultwent
unstableeven when the time step used was within the constraint of tile CFL stability criterion. To
this
overcome problem, the authorproposesa new modified ADI-FDTD method (1.48].

In order to exploit the advantageousfeatureof the ADI methodwithout suffering from instability in a
three-dimensionalmodel, a factorf, where0 <f < 1, is introduced[1.20] in the ADI-FDTD routine.A
very important of
characteristic this modificationis that it is consistent
with physicalconsiderat
ions.This

will be illustratedsection4.3 later in this chapter.The modified ADI-FDTD allows us to violate the CFL
in
stability constraint a complex three-dimensional model without causing instability. Furthermore,no
gradedmeshis necessaryto maintainstability of the overall system.

In this chapter,the ADI-FDTD schemeis modifiedwherebya factorf is introducedas a direct weighting

factor on the implicit-explicit terms of the ADI-FDTD equations.This method is applied to the Yee's
staggeredcell to solve Maxwell's The
equations. growth factor of this method is derived for a three-
dimensionalmodified ADI-FDTD. This modified ADI-FDTD method is then usedto model a line-fcd
rectangularpatch and
antenna the resultsare discussed.

4.2 Three-dimensional modified ADI-FDTD algorithm

Equations (4.1) - (4.4) show the numerical formulation for the modified ADl-FDTD method. The electric
and magnetic fields are spatially staggeredas in the conventional FDTD. Without loss of generality, the
formulation is carried out for lossless media. The formulations are
split into two procedures, procedure I
is applied for advancementfrom nAt to (n+ Va)At while
procedure 2 Is used for advancementfrom
Modified Alternating-DirectionImplicit Method 53

to (n+I)At.
(n+ 1/2)At

Procedure I
(2-f) ,,. "+1/2(1+ 1/2j+ 1/2, A) - Ilz"'1/2fl+1/2j-1/2, k)
Ay
At'
E,n+l/2 0+112jk)= En (i+112ik) +
2c Hym(i+1/2jk+1/2) lly"(1+1/2jk-1/2)
-
AL.

(4.1a)

(2-f) 11n +112(ij + 1/2, k + 1/2) - Il, "+"2(lj+ 1/2. k - 1/2)

x
A' Az
k)+ .0
Eyn+112(ij+112, k)=Ey"(ij+112, 'I I I
2c lln(i+112J+II2, k). Iln(i-112J+II2, A)
(f)
Ax

(4.1b)

lin4l/2 (I+ 1/2j, k+ 1/2) _ ll;,, 1/2(1 _ 1/2j. k+1/2)

(2-f) y
Ax
En+112 (ii, k+ 1/2) E. "(ijk+112)+
2c I
(f) lln (Ij + 1/2. k + 1/2) Il, #'(/J-1/2. k + 1/2)
-
AY

(4.lc)

I E,n(Ij 1.k+ 1/2)

+ -k+I
At AY
H"+' 12(ij+1/2, k+ 1/2)= H,, (ij+112, k +1/2) TU 1/2 12 I
(2-f) Ey"+ (Ij+ 1/2,k + 1) - Ey"+' (Ij + 1/2.4-)
AZ
J,
(4.2a)

10
E,"(I+ 1/2j. k +1) - E,"(1+1/2j, A)

Hy"+'/2 (i+112jk+112)=Hy"(i+112jk+112)-A' 'IA:

2,u E,,"+1/2(1+ljk+1/2)
(2-f) . Em+112
(IJ, k+ 1/2)
Ax i
(4.2b)

Eyn(l+ I J+112, k) - Ey"(IJ+112, A)

(f
At Ax
H"+' /2 (+112! j+112, k)= Hn(1+112J+112, k) -lu,
(2-f) E,",+112(1+1/2j + 1.k) - En+1/2(1+112j. k)
AY
(4.2c)

Referringto the equations(4.1), for the electric field terms, in procedure1, the implicit terms havethe
weightingfactor of (2 -f) while the explicit terms havef. The same appliesfor the magnetic rieldsterms
in procedure1, the implicit termsare weightedby (2 -f) while the explicit terms byf. This is repeated
Modified Alternating-DirectionImplicit Method 54

for procedure2 in equations(4.3) and (4.4). As in the ADI-FDTD, the implicit and explicit terms are
switched between proceduresI and 2.

Procedure 2

(f Il. "+"2(1+1/2j+ 1/2, k) - Ii, "+1/2(1+1/2j- 1/2, k)

At AY
' (i+112jk)=Exýn+112(i + 1/2j, k) +y,
E,n+l I Hy'+'(i+112jk+112)-Ilxn+'(i+112jk-112)
(2-f')
Az

(4.3a)

1112(ij l'O
(f )I + 1/2, k + 1/2) 11 + 1/2, k -1/2)
At A--
Ey"+1(ij + 1/ 2,k) = Ey"+112(ij + 1/2, k) + ,
2c I/. "+' (i + 1/2j + 1/2, k) - ll, "+' (1-1/2j + 1/2, k)
(2-f)
Ax

(4.3b)

+ 1/2jk + 1/2) 11;+1/2

(1 I/ 2j. k + 1/2)
(f - -
At
E,"+' /2 k 1/2)
2c I/, "+' (Ij + 1/2. k + 1/2) - I/, "+' (IJ-1/2, k + 1/2)
(2-f)
Ay

(4.3

I. I
(2-f) E."+'(ij + I, k + 1/2) - E,"+'(Ij. k + 1/2)
Ay
H,n+l(ij + 1/2,k+ 1/2) = Hn+l12(IJ+1/2, k+1/2) - I,
2.u (Ij+ 1/2.k + 1) - Ey"+112
Ey"+112 (Ij+ 1/2.A)
(f
. A.-
(4.4a)

1
(2-f) E,,"+'(i + 1/2j, k+ 1) - El + I(I + I/ 2j. k)
,I,
A' A.-
(i + 1/2j, k+ 1/2) -
Hyn+l(i+ 1/2jk +1/2)= Hyn+112
2p I-
(f) E,"+'12(/+Ijk+1/2)-E, "+"2(ljk+1/2)
Ax
(4.4b)

I- I
(2-f) +Ij+ 1/2,A) - Ey"+'(Ij+ 112,k)
y
At Ax
H"+'(+1121J+112, k)=H'+"2(i+112J+II2, k) -
2p I
(f) Exn+112
(1+1/2j+l, k) - g, +112(1+1/2jk-)
Ay
(4.4c)
Modified Alternating-DirectionImplicit Method 55

Again, as in the ADI-FDTD methodequations(4.1) cannotbe solveddirectly due to the implicit terms
involvedon the RHS.To solveequations(4.1), the LHS electric fields haveto be expressedsuchthat the
RHS termsare all explicit. This is accomplishedby substitutingequations(4.2) into (4.1) appropriately;
specifically, substituting (4.2c) into (4.1a) results in (4.5) below, where the LIIS forms a tri-diagonal
matrix of E, when E, is scanned in the ý direction.The RI IS of (4.5) now consistsof only explicit terms.

(1VXAY 2
(72
1/2 Ifi2)]
E" +1/2 (i+112j-l, k)-E, + (i+112jk) 2+ + E,"'112(I+ 1/2j+ I, k)

(V'u )2 (T244
Ay I+( Ay JE.,
(i + 1/2j,, k) ' (i+ I J+112, k) Ey*Q+112, k) (I+ I k)
J-112, + Ey
-E"xt2 - -Ey"

(EAy) (TI (,
U4V2) fý
"(i+112J+II2, k). 11"0+112J-112,k)I +(7AA- F2Tf (1+1/2j. k+1/2)-//. *,(1+1/2jk-1/2)]
At f z)(
-f

(4.5)

Similarly, Ey and E, can be solved in this manner. Once all the electric fields are computed, the magnetic
fields can be computed directly using (4.2a) - (4.2c). The same approach can be applied for procedure 2.
Equation (4.6) shows the tri-diagonal matrix for E, in procedure 2. The full formulation of electric fields
in both procedures I and 2 are given in Appendix B2

( )2 (772

E,"*' (I + 1/2jk-1) ""'

E, fl+ 1/2jk) 2+ 'FUCA: + E,*+' (I+ 1/2j, H)
- At

(! 2+ (T
1/2(i+112jk) ucAz
PC [E,
-E, + -fi2 (i+ljk+ 1/2) - E,"*lll (ij. k+112) - E,"'lll 041jk. 1/2) + En#jk412)j
At V) f
(Tf )2)[H; (liýtAy 62,
+1/2(i+1/2jk+1/2). Ily'+1/2(1+1/2jk-1/2)I+ JAZ' ll,"*"2#+1/2j-1/2. A)j
2-ýfl
(4.6)

4.3 Weighting factor In the modified ADI-FDTD algorithm

Ex
Modified (2 -f)12
adi-fdtd
procedure I

Hy. f12
nAt ýII- 11.4)Lit
Ex fiz f12 E,,
Modified
adi-fdtd
procedure2
Hy.(2 -f )/2
(n+I/2)At /7"'
(n+I)At

of modifiedADl-FDTD algorithm
Fig 4.1 : Physicalrepresentation
Modified Alternating-DirectionImplicit Method 56

Fig. 4.1 above illustratesthe physicalrepresentationof the modified ADI-FDTD formulation. In the

modified ADI-FDTD method, an intermediate

electric field is calculatedat time step(n+ V3)At.
As in the
ADI-FDTD methodexplainedin chapter3, in procedureI of the ADI-FDTD method,the electricfield at
time step(n+ V2)Atis calculatedusingthe previouslycalculatedelectric field at time stepnAt andthe curl
of the magneticfields, part implicit andpart explicit. In the ADI-FDTD method, the weighting factorsfor

the magneticfields, implicit H,, and explicit fly are the same,i.e. Va for both of them. However,in the
modified ADI-FDTD the implicit term H, at time step (n+ V2)Atis weightedby (2 -f)/2 andthe explicit
term Hy at time stepnAt byf/2 where0 <f < 1. This is immediatelyfollowed on by procedure2; now tile
known(explicit) valueof H,.at time step(n+ I/i)At is weightedbyf/2 and the unknown(implicit) valueof
Hy at time step(n+I)At is weightedby (2 -f )/2. The total magneticfield over a full time stepremains
unchanged.

4.4 Divergence of modified AD[-FDTD algorithm

As in the ADI-FDTD algorithm,thereis no explicit enforcementof the Gauss'sLaw relationsfor bothtile

electric and magnetic fields in the modified ADI-FDTD algorithm. Although the curl operationof the
magneticfields is performedover two half time stepsand the magneticfields arc weighteddi(Tcrcntlyat
each half time step,the total magnetic field over a full time stepremains unchanged in the modified ADI-
FDTD scheme.Therefore,the modifiedADl-FDTD algorithmwill still result In zero divergencefor both
electric and magnetic flux i.e. the time derivative of the net magrictic/clectricflux leaving the surfacesof
a cubic Yee is
cell zero, thus upholdingthe Gauss'sLaw for the magnctic/clectricfield In charge-frce
spacein the modifiedADI-FDTD scheme.

4.5 Numerical stability

The numericalstability of the modified ADl-FDTD can be carried out in the sameway as in tile ADI-
T, Ty T,
FDTD. Assumethe spatial frequencyto be and as the x- , y- and z- componentsof Its

as
respectively
wavevector
numerical in (3.7)
equations & (3.8).

4.5.1 2-dimensional modified ADI-FDTD

For the sakeof simplicity, we shall considerfirst the numericalstability of a 2-dimensionalTE wave
consistingof the following fields :

Procedure I
At n+1/2(1+1/2, J+1/2) lln+112(1+1/2, J-1/2) 1ý
(2-f)[11 I--I -
E"+' /2 (1+112, j) = E,"Q+112, j) +
2c AY
(4.7a)

A,, H,"(i+1/2, J+1/2) ll, "(1-1/2, J+1/2)

E n+112
(Ij+112) = Ey"(!,J+112) - 2c
y Ar

(4.7b)
Modified Alternating-Directionimplicit Method 57

J+1/2) - Ey' J+1/2)

(f)[E"(1+1/2,
yy
At Ax
Hn+112(i+112,j+112)
zz2 = H"(i+112, J+112) - li
n+112
(2 f)[-E x (1+1/2, j+l) E"+"2(1+1/2,
- x J)
- Ay
(4.7c)

Substitutingthe field componentsin spectralforms, (3.7a), (3.7b) and (3.8c) into (4.7) will yield the
following:
iYAY
At
Ex *GFI=Ex GFI*(2-f)*H.. 2jsin (4.8a)
- 2cAy 2

TxAx
At
(4.8b)
E *GFI=Ey
Y +f*H, -2eAx 2jsin( 2)

(2LY
T AY
At At
* A
2jsin(IX-26x) GFI* (2-f)*E, 2jsin (4.8c)
HZ*GFI=Hz + Ey*f - 2,u Ay
2pAx

where GFI is the growth factor in procedure 1. Substituting (4.8a) and (4.8b) into (4.8c) gives :

)2pAx
T,AX At (Lx2-
At AX)
f 2H, 2j sin . 2jsin
2c Ax 2
Hz*GFI=I,
z GFI-I
Ly Ay E
At At yAy)
(2 _f)2 GFI * GFI * H, 2j sin 2jsin
2c Ay 2 2p Ay 2
+ GFI-I

(4.9)

.' -12 ( 2
("k-,,Ax II (I )[ Wý,
Ay
A,
H (GFI_1)2 = _f2 H.. sin(- -2)j (2 - f) 2 GJ'Fl2 11: 7c sin -2
: Ax Ay

(4.10)
(2 (E
At k. Ax At Ay
My y
let mx = AX
sin ') and
Ay
sin
2

anddividing (4.9) by H,
(I
kfv2
GFI 2 I+EY2 (2 -f)2 2GFI + f2 0
- +
uepc

(I+
"X2
Let a +E-y(2- f)2 and C. f2

then(4.11) becomes
Modified Altmating-Direction Implicit Method 58

aGF12 - 2GFI +c0 (4.12)

±Nrl-ac
GFI -I
a
ý-a-c-
andsinceac ý: I GFI -j (4.13)
a

Procedure 2
At H n+112(, + 1/2, J+1/2) /In+'/2(1+1/2, J-1/2) 11
- z
E,n+'(i+112, j) n+112(i+
Ex 1/2, j) -
= 2e Ay
(4.14a)

"+'(i+1/2, 11
At (2_f)[Ii, J+1/2) /I, "+'(/-1/2, J+1/2)
Ey"+'(1,i+I/ 2) = Ey'+'12(i, J+1/2) - A

(C I b)

I
Ey"*' 1+ 1/2, j+I/ 2) - E"+' (1,j+ 1/ 2)
(2 f)[
A' AX
H"+'(i+112, j+112) =H n+1/2(1+1/2, J+1/2)
2 E"+1/2(1+1/2, j+l) E n+1/2(1+1/2, J)
Ay 'I

(4.14c)

Again, substituting(3.7a),(3.7b)and(3.8c) into (4.14)will yield the following:

Lx Ay)
At
Ex * GF2 = Ex -f*H., 2j sin (4.15a)
2cAy 2

(LX2
At L
E *GF2=E + GF2*(2-f)*H, 2jsin xi (4.15b)
YY ý I)
cAx

(i (ý2
L-) AY
At At Y
H *GF2=H + GF2*(2-f)*Ey 2jsin j-2
XAX) f *Ex 2jsin (4.15c)
zz 2,q Ax 2p Ay

whereGF2 is the growth factor in procedure2. Applying the sametechniqueon procedure2, we get,

2+
GF2 (2 _f)2 2GF2 ++f20 (4.16)

(I+LY2
f)2 d- f2
Let +pC 12-j and

then(4.16)becomes
b GF2 2- 2GF2 +d=0 (4.17)
Modified Alternating-DirectionImplicit Method 59

11--bd
GF2 -I±
b

Jb-d- I
andsincebd >I GF2 (4.18)
-Ij b

Now,
lac-
j-, I jNf --I
-bd
GFI - and GF2 =Ii
a b
where
MY2 "fx2
a+ (2 _f)2 and cM+ f2

(I (1+442
Mx2
+pe (2 _f)2 and d- f2)
PC

Examining M, and M., we seethat

mx (At
At) 61-8) L i Ax 2 (ý2 Ax
2(x2
x2
sin sin
peT2) vx

Therefore,the total growth factorof proceduresI and2 combinedis given by :

GF GFI I GF21

Va Vb
I I (ý'
(TAýx
(L Ay
V, i 2 L2 2 y
I+ f I+ f 2
sin 2
sin -2
v2
x V2
y
(Ey )
Ll
(E AY
(2-f )2 V2 2
I+ -2 sin 1+ (2. f) 21
vx 2! 2 sin -
VY

(4.19)
As 0 <f< 1, then I< (2-f) <2

Therefore(4.19) or the overall growth factor of the modified ADI-FDTD algorithm will alwaysbe less
than unity. Consequently,the newly proposedtwo-dimensionalmodified ADI-FDTD methodis always
stablefor 0 <f < 1.

4.5.2 3-dimensional modified ADI-FDTD

To analysethe numerical stability of a three-dimensionalmodified ADI-FDTD, we apply tile von
Neumannmethodagainon the three-dimensional modifiedADI-FDTD formulations(4.1) - (4.4).

Followingthe samemethodas in chapter3 we get GFI andGF2 as shownbelow:

Modified Alternating-DirectionImplicit Method 60

GFI
I Mx-My.(2 - f) -f -j. Mz. f J.
My-(2- f)
0 0
Ny Ny-ý-e Ny-c Ny-c
1 My -Mz.(2- f)-f j. Mz.(2- f) Mx.
-J. f
0 0
Nz Nz-g-s Nz.e Nz-c
Mx.Mz.(2- f)-f 1 -i-my. f j. Mx. (2- f)
0 0
Nx-p-c Nx Nx-E Nx-c
j. Mz.(2- f) -j, My. f Mx. Mz. (2- f)-f
0 0
Nz-A Nz-g Nz.p g
-j. Mz-f j. Mx. (2- f) Mx. My. (2- f)-f 1
0 0
Nx-p Nx.g NxA -c
j. My.(2- f) -j. Mx.f My. Mz. (2- f)-f I
0 0 Ty
Ny-p Ny-p Ny-p-c
(4.20)

GF2 =
1 Mx.Mz.(2- f)-f -J. Mz. (2- f) J-My-f
0 0
Nz Nz. p -E Nz-c Nz, c
Mx.My. (2- f)-f 1 J-Mz.f J,Mx.(2- f)
0 0
Nx-p -s Nx Nx-E Nx-c
My Mz.(2- f)-f I -i-My-(2- f) J-Mx-f
0 0
Ny. g -E Ny Ny-E Ny-c
j. Mz-f -j. My. (2- f) I Mx. My. (2- i)-f
0 TY 0
Ny-p Ny-p Ny-p -g
Mz. (2- f) j-Mx-f I MY-Mz-(2- f)-f
-j. 0 0
Nz-g Nz-g Nz Nz. p -c
j-My -f -j. Mx-(2- f) Mx. Mz. (2- f)-f 1
0 0
Tx
Nx-p Nx. g Nx-p -E

(4.21)
where
At (khAh
ý, ) m Ih
Mh -Ah sin( and Nh =1+ h-x, y, z (4.22)
2 PC

In orderto solvethe overall growth factor for the three-dimensionalmodified ADI-FDTD, the following
assumptionis made.
Overall growth factor,GF = GFI *GF2 = GF2*GFI (4.23)

The assumptionis soundbecausethe proceduresarecommutative.

Solving(4.20) and(4.21)using(4.23),we get
(6f ) 2
GF (4.24)

modifiedADI-FDTD is alwaysstablefor 0 <f < 1.

Therefore,the newly proposedthree-dimensional
Modified Alternating-DirectionImplicit Method 61

4.6 Numerical dispersion

The numericaldispersionfor the modified ADI-FDTD methodcan be found by substitutingthe vector.

field travelling-wave expressionwith time dependence(3.27) into the modified ADI-FDTD finite
differenceequations
.

4.6.1 2-dimensional modified ADI-FDTD

Again, for simplicity, we investigatethe numericaldispersionof a 2-dimensionalTE wave [1.19].
Substituting(3.27) into (4.4) (procedure1) gives:

At AY)
jwAI12 y jwAI12 n
(e - I)E" = -j(ýA AY)sin eH (2-f) (4.25a)
x 2z

)sin (2)2
&, Ax
At
(ejwAI12- I)E n=j( ýc AD
H" f (4.25b)
y

) (L2 Ax) ) (i Ay)

At At
(e jwA112 J)H nj xL Enf_ jrý-, ) sin jwAl/2 En (2-f) (4.25c)
sin yy -2 e x
uAy

and into (3.14) (procedure 2) gives :

( At kyAy j
jwAt jivAt/2 )E" jwAl/2 H" f
(e -e = -j(-) sinýLT e (4.26a)
xc 5yAy

( At !
jwAt jwA112
J2x)
,6 jw AH, n (2 f)
(e )E" =e sin (4.26b)
-e y(x

) ) & Ay )c
At At
(ejwA' -e jWAt/2 )H' =( sin xAx) ejol& En (2-f)
-i( Cp -Ay)(2
sin jaiAt/2 E," f
Ax) 2 y

(4.26c)
Combining(4.25a)& (4.26a),(4.25b)& (4.26b)and(4.25c)& (4.26c)givesrise to the following -

LTy
At AAY
ý) jwAt/2
jwAI Hn
(e I) ,n =
Ex (4.27a)
- -2j(. -eAy) sin 2 e2

(iX2
At ) (e jwAI(2-f)+f)
(ei'Al - I)E" =j( ý7A-X)sin Hn (4.27b)
y

(, (iX2ýL) (i2
At )( At
(e jwAI I)H n=j x Ey"(ej'A'(2-f)+f) 2jý- jex
yAy) JvAt/2 E"
- sin - -
z u Ax) P ,,&y
Y)sin

(4.27c)

(4.27) can be simplified to

rw At Atý AY
" y
yAy) (4.28a)
siný- 2)E, , sin z
Ay 2H
Modified Alternating-DirectionImplicit Method 62

(w At ) (i Ax (co
At jk At
En =( f x
Cc-Ax) sin -2 Hn cosN + (4.28b)
sin Y

(co Att )zu (co At) Z),Ay

At Ax) jw At
nv
Hft X2 E" coSý + (I-f)e At/2 En
sin sin y sin x
2! 2 Ay) 2
(4.28c)

or

(co
±LAt
1 Afy
sin 2 0
en Ex
(w
("212
Att [Cos
wat jiwal/2
0 sin + (1-f)e E; 0
y
(oj At) lln
MY M., [cos(22
(co IAt,
+ (1-f)e sin
p
(4.29)

Thusthe numericaldispersionrelationfor a 2-dimensionalTE wave is givcn by making tile dcterminant

of the matrix zero,i.e.
22 ýý ýx)ýCOS ( 121
(co At)
) At (-,At)
At CoAt) jo)At/2
'(0. f2 -L
siný7 sin ý7L2L ý; sin 2 -f)e
2 pe 22)+
(i2
(T., Ay (4.30)
(w At
(±L2
+ sin CAy) sin
_L(At)2 0
pe

or
2+ )2 ýYAY 2
I)2 r lcos(ýL-At 2 At)
f2( Sin2 + (I-f)ej"ý12 sin sin2(02
AX) 22 Ay) 2 cAt

(4.31)

where c=.
Equation(4.31) revertsto the numericaldispersionrelationfor 2-dimensionalADI-FDTD (3.34)when
f=I as expected.Following the sameargumentas for (3.34),the implication of (4.31) is that numerical

wave velocity for wave propagatingin the direction of x, i. e. ky = 0, is scaledfrom that of tile standard

F (wAt 2,
FDTD scheme by the factor L2
Cos + (1-f)e which is larger (for 0 <f < 1) than the

(2)
factor CoS2(±0LAt in the ADI-FDTD method. On the other hand, for wave propagating in the direction

of y, 0, the numerical wave velocity in the modified ADI-FDTD scheme is the same as that of the

standardFDTD scheme.
Modified Altemating-DirectionImplicit Method 63
4.7 Simulated results

The newly proposedmodifiedADI-FDTD methodis usedto simulatethe line-fed rectangularmicrostrip

patch.The physicaldimensionof the patchsimulatedis shownin Fig. 2.3 in chapter2. The spacesteps
usedare Ax = 0.389mm,AY = 0.400mmandAz = 0.265mmand the total meshdimensionis 60 x 100x
16 in the x, y and z directionsrespectively.The patch is excitedwith a Gaussianpulse.A I" order Mur
absorbingboundary condition is applied on all the five surrounding walls. A perfect electric wall
boundaryis applied on the ground plane and the copper patch; this is done by forcing the tangential
electric fields on the copperpatchto be zero at all time steps. The dielectric constant,Cr,Is set to 2.2.
Fig.4.2 showsthe comparisonof resultsfrom the publishedliterature [1.5] with that from the proposed
modified ADI-FDTD with stability factor of 2.0, i.e. the time step is twice that allowed by the CFL
criterion. For this casef is set to 0.9. Fig 4.3 showsanothercomparisonof results,this time a stability
factorof 3.0 is usedandf is setto 0.8.

Isill
(dB)

05 10 15 20
0 frcq (GI lz)

-5
-10
-15
-20
-25
-30
-35
-40
Abouzahra'ssimulation
-45
Modificd ADI-FDTD
-W

Fig. 4.2: Modified ADI-FDTD with At =2* At critical andf - 0.9

From Fig. 4.2, it can be seenthat the magnitudeof reflection coefficient, S11,decreasesfrom that when
the conventionalFDTD methodis used.As explainedin chapter3, when ADI-FDTD is uscdto modela
conductor using perfect electric wall boundary, the simulation results grow exponentially as time
progressesuntil eventuallythe systembecomesunstable.Unlike in the ADI-FDTD method,in modifled
ADI-FDTD scheme,the implicit and explicit terms are not weighted equally as shown in (4.1)-(4.4).
Effectively,ADI-FDTD methodis in the form of a predictor-correctormethod,the explicit term beingthe
predictorterm and implicit term the correctorterm. The fact that ADI-FDTD result becomesunstablein
chapter3 implies that there is a gain in the system.By introducinga weighting factor greaterthan unity
on the implicit term in modified ADI-FDTD method,the correctorterm is weightedmoreheavilythanthe
predictorterm. This 'corrects' the resultsand maintainstability in the systembut the side effect of this is
a reductionin the magnitudeof S11asexpecteddueto the correctorterm beinggreaterthan unity.
Modified Alternating-DirectionImplicit Method 64

Isill
(dB)

5 10 15 20
1' frcq(Gliz)

-20-
-25-
a-

-35-
Abouzahra's measurement
-40 -----
Abouzahra's simulation
-45-
Modified ADI-FDTD

Fig. 4.3 : Modified ADI-FDTD with At =3* At critical andf - 0.8

4.8 Relationship between attenuation and weighting factor f

To understandthe relationshipbetweenthe decreasingof S11magnitudeand the weighting factor,f, used,

the simulation for the line-fed rectangularmicrostrip patch was run with severaldifferent weighting
factorsandthe SII plots areasshownbelow.

1.2
- cortvFDTD
----modADWDMf--O. 9
....... modADWDTDf--0.8
----- modAMFDTD t--0.7
, Z:: :
"**"..\
0.8 -

0.6

0.4

0.2
V 'k'f t;

0 5 10 Is 20
freq (GHz)

Fig. 4.4: ComparisonbetweenconventionalFDTD and modified ADI-FDTD resultswith

At =2 *At critical andf setat 0.9,0.8 & 0.7. At critical is usedwith conventionalFDTD.
Modified Alternating-DirectionImplicit Method 65

The resultsin Fig.4.4 showthat the attenuationincreasesasthe weightingfactorof the explicit term on
the modified ADI-FDTD,f, is reduced.

- corwFDM
----modADWDMsf=l
1.2
...... modADWDM gfm2
mod ADWDM sfwl.2
I

0.6

0.4

0.2

0
0 5 10 15 20
freq (GHz)

Fig. 4.5 : Modified ADl-FDTD resultswithf setat 0.9

Fig. 4.5 aboveshowsthe ISI II of the line-fed rectangularmicrctrip patch simulatedwith the modified
ADI-FDTD methodwith the weightingfactor,f, setto 0.9. Fig. 4.6 below showsthe resultsfor the same
patchwith the weighting factor setto 0.8. Changingthe stability factor and thereforethe time-stepusedin
the algorithmdoesnot changethe attenuationsignificantlyfor a particularweightingfactor.

1.2 convFDM
----mod ADWDM af-2
...... modADWDM SN3

0.8 -

0.6 -

0.4

0.2
.v*i

V
0 5 10 is 20
freq (GHz)

Fig. 4.6 : Modified ADI-FDTD resultswithf setat 0.8

Modified Alternating-DirectionImplicit Method 66

4.9 Conclusion

A new modified ADI-FDTD methodwith the introductionof a factorf hasbeentheoreticallyderivedand

numerically simulated.From the study of the numerical stability of the both
scheme, two- and three-
dimensionalmodified ADI-FDTD algorithmsare shown to be permanentlystableas long as 0 <f < 1.
modelof a patchantennaare stableeventhoughthe CFL criterion has
The resultsfor a three-dimensional
not beenobservedand they comparereasonably
well with the publishedresultsin termsof the resonant
frequencypoints. The copperpatch in the modified ADI-FDTD schemehas beenmodelledas a pcrrect
electric wall boundarywherethe tangentialelectric fields on the patchare set to zero at eachtime step.
The samestructure,when modelledusingthe ADI-FDTD where the copperpatch is modelledalso as a
perfect electric wall boundary,showed instability. Although the results from the modificd ADI-FDTD

show a decrease in its S11amplitude,this technique is easy to implement and it is useful as a quick
method to obtain accurate resonantfrequencies.

In Chapter5 we proposea new implementationmethodof the ADI-FDTD In order to ovcrcomethe

problemof instability without compromisingthe amplitudeof the response.
CHAPTER 5
SIMULATING COPPER LAYER IN
ALTERNATING-DIRECTION IMPLICIT METHOD

5.1 Introduction

The introductionof the ADI-FDTD techniquehas made it possibleto speedup simulationrun-time of

largeelectricalobjectsevenwhen thereare small discontinuitiesin the model without compromisingthe
stability of the system. In this researchwork, the ADI-FDTD method has been shown to work
numericallyon an air-filled three-dimensionalrectangularcavity and on a three-dimensionalrectangular
cavity with inhomogeneous media.To model a transmissionline as a perfect electric wall boundary,the
modified ADI-FDTD has to be implementedinstead of the ADI-FDTD to maintain stability. The
modified ADI-FDTD gives reasonablyaccurateresonantfrequencypoints althoughthe amplitudesare
attenuateddue to the inherentproperty of the algorithm. Although the newly proposedmodified ADI-
FDTD schemeis useful for quick numericalanalysisof microstrip circuits, for the ADI-FDTD schemeto
be genericallyuseful, it must be ableto modelmorecomplexstructuresincluding microstrippatcheswith
accuracy
reasonable both in amplitude and frequency points without going Indeed,
unstable. a methodof
modellinga copperpatch in conjunctionwith the ADI-FDTD technique without causingeither instability

or attenuationis desirable.

In this chapter,a copperpatchis modelledin the ADI-FDTD algorithm as a layer of materialwith a finite
electric conductivity. The finite-difference equationswhich include the electric conductivity term are
presentedand results for severalsimulationsare shownand comparedwith publishedresultsand results
obtainedfrom conventionalFDTD method.

5.2 Three-dimensional ADI-FDTD algorithm with electric conductivity term

Considerthe Maxwell's curl equationfor electric field shown in (2.12a) - (2.12c) which include the

electricconductivity term, to for

a, account electriccurrent loss in Equation
materials. (2.12a)is repeated
below for convenience.
aE, I (aH.. aH,
Tt -C(-2ý- -Fz- a E,

The centred-differencefinite-differenceapproximationof (5.1) is given by :

L' H. n+112(ij k) n+112(ij.,, k) H +112(ij k) -Hy n+1/2 (ijk-1)

En+l (iik (ii, k) + -H. y" (ii, k) E," +1 2aj, k)
AY Az -
c

(5.2)
Simulating Copper Layer In Alternating-Direction Implicit Method 68

Note that all the fields within the bracketon the right-handside are evaluatedat time stepn+ Y2At. Since
the electric at n+ V2At is not readily available,it is calculatedusing the semi-implicit approximation
below.
E" (i, j, k) En," (1, J, k)
+
En+112(1,i, k) -- -x (5.3)
,x

Substituting (5.3) into (5.1) gives the following :

c(ij, k) At
E'n +1 (ij, k) =Ex n (ij, k) 2c
c(ii, k) At

2c

At
IH
n+112(ijk) Hn+112(ij. l, k) lln+112 (ij, k) Il. +1/2 (IJk-1)
. y'
+c21.. y
a(ij, k) At AY A. -
+
2c

(5.4)

Applying the same technique on all the electric field calculations in the ADI-FDTD mcthod produces the
following equations. Note that the At is replaced with At/2 in both procedures I and 2 in the ADl-FDTD

equations. Also, since magnetic loss is not considered here, the magnetic field equations remain the same
as in the ADI-FDTD method discussedin chapter 3.

Procedure I
(I_ a(i+112jk) Atl"
4c
E,"' 12(i+112jk) = E,",(1+112jk)
I+ I" (i +I/ 2j, k) At
4c

At H, '+"2(1+1/2j+1/2, k)-il, "+1/2(i+1/2j-1/2, k)

2c Ay
+
(i +I/ 2jk) At. Ily (I + 1/2jk + 1/2) - Ily* (I + 1/2j. k-1/2)
4c .-A.
10 -

(5.5a)

a(ij+112, k) At."
4e
(ij+112, k) = Ey"Oj+112, k)
E n+112
y 'Ir Q+112, k) At
+
4c
10
r)I At In+ 1/2 oj +1/2, k + 1/2) - I-lit +"2(ij + 1/2.k 2)
-I/
2c Az
+
(ij+112. k) At, 11" (1+ 1/2j+1/2. k) -I/, " (1-1/2j+ 112.k)
+a
4s AX

(5.5b)
Simulating Copper Layer In Alternating-Direction Implicit Method 69

(Ia (ijk + 1/2) At

4e
E"" 12(ijk+112) = E"(ijk+112)
a(ijk+112) At
+
U
At jj; +1/2(j + 112jk + 1/2) - Ily "*'l20-1/2j. k+1/2)
Te
+ Air
(ij, k+ 1/2) At. Oj+ 1/2.k + 1/2) - /I, " (IJ-112.k + 1/2)
11,11
+,
4c AY

(5.5c)

II
E"(Ij+l, k+112) - E'(Ijk +1/2)
-1 2 --
At AY
H +1/2(ij+1/2, k+1/2) =Hxn (ij+112, k+112)-
En"12(ij+112, k+l)-Eyn*112(ij+112,A)
y
A:
(5.6a)

E"(1+112j, k+l) -E. ','(1+1/2J. A)

x
Hn+112(i+1/2jk+1/2)
yy = H" (i+1/2jk+1/2) - 2p En+112(i+ljk +1/2) - ER+112(ij.
k +1/2)
Ax
(5.6b)

En(i+IJ+112, k) - E"(IJ+112, k)
yy
At
Hzn+112(i+1/2j+1/2, k) =H, z" (1+1/2j+1/2, k)-
2p Exn+1/20+112j+l, k) _ Em1112(1
+ I/ 2J.A)
Ay

(5.6c)

Substitutingequations(5.6) into (5.5) appropriatelyresult in tri-diagonal matriccsof E,, E. and E, Tri-

diagonalfor E,,for procedureI is shownbelow,the restareshownin Appendix D3

(3E (I
E,"+'(i + 1/2jk - 1) - E' "l(l + 112jk) a(1+112jk)At)j "fl
2+ + + ER + 1/2j. k + 1)
4c

-2 ) (I
(...FpCAZ
afl+112jk) At.
(I + 112jk) -
At 4c
[Ez
n4j/2 "' 1/2 (ij. k. 112)
+(A! (i+ljk + 1/2) En+112(!Jk + 1/2) EM+1/2(I+ljk -1/2)+E,
ax) z 3
(ILIýj [Hn+112
('U'&Z' ) [11.
(i 1/2jk 1/2) Hy"+112
(I+ 1/2jk-1/2)] + n- 1/2(1+1/2j + 1/2.#j 1j,", 1/2(1+1/2j-1/2.
y + + - A, &y - iV]

(5.7)
Simulating Copper Layer In Alternating-Direction Implicit Mcthod 70

Procedure 2
a(i+112jk) ät 1
1-
12(i + 112jk) 4c
(i+II2ik) = En+'
ät
17
4c
At ll,,, +112(7
+ 1/2j + 112.k) - Il, *"12(1+1/2j k)
-I/2.
TC AY
ol (I + 1/2j, k) At U." +' (I + 1/2jk + 1/2) - /1; "' 0+ 1/2jk-1/2)
+
4e

arij+112, k) At
E n+1/2 (ij + 1/ 2, k)
U
En+'(ij+112, k) = y Q+112, k) At
+a
4c
At ll, "+1,12(7j+1/2. k + 1/2) - Il, "*1/2(lj+ 1/2.k -1 /2)
Te A:
(Ij+112,k)AI
17 Ilm+l (1+1/2j+1/2, A)-lit"' (1-1/2j+1/2.4)
1+
4c
(5.8b)

a(IJk+112)At
4c
E,""' (IJ,k+ 1/2) - E,"" /2 (ij, k+ 1/2)
(IJ,k + 1/2) At
4c
At (I + 1/2j. k + 1/2) - Il; "'2(1 -I/ 2jk + 1/2)
Ily"+112
Ts AT
(ijk+112)AI 11,0+1(Ij + 1/2.k + 1/2) - I/, "*' (IJ-1/2. k + 1/2)
4c Ay

(5.8c)

E,""(/J+I, *+1/2)-E: ""(/J. k+1/2)

At Ay
H n+l (ij+112, k+112) =H n+112(ij+1/2, k+1/2) - 2ju
x E*"1/2(/j +I/2, k+l) - E; 41/2(/J+1/2,k)
.v
ILI
(5.93)

I
E"" (1+1/2jk+l) -E, "" (1+1/2jh)
At -x
H"+' (i+112jk+112)
yy2, =H"+112(i+1/2ik+1/2)- ,
u En+112(1+ljk+112)-Etn'112(ij.k+112)
Ax
(5.9b)

E"+1(I+ lj+ II2, k)-E"" (IJ+1/2, A)

YY
At &V
H n+I (i+112j+112, k) =H n+1/2 (i+1/2j+1/2, k)- ' 0
2ju E,"+1/2(1+1/2J+I, k) - E"*1/2(1+ 1/2jk)
-x
AY
(5.9c)
Again, substitutingequations(5.9) into (5.8) appropriatelyresult in tri-diagonalmatriccsof E,, E. andE,
Tri-diagonalfor E,,for procedure2 is shownbelow,the restareshownin AppendixB3.
SimulatingCopperLayer In Alternating-DirectionImplicit Method 71

ý'u7fAz (i 112jk), &t )I

I +
E."+'(I + 112j,k- 1) - E,",+'(i+ 1/2j, k) 2+ I++E, "+'(I + 1/2j. k + 1)
4e

ýPCAZ 2 At
"+l'2(i+112j, a(1+112jk)
--E. k)
4e

) [E,
+ n+1/2(i+ljk + 1/2) En+1/2 (ijk + 1/2) E,n +112
(i+ljk-112) + E, Oj. k-112)
-
Ax
(PAZ [Hn+112 (JuA. 2 [11,
(I + 112jk + 1/2) - H; +1/2(i + 1/2jk. 1/2)] + n11120+ 1/2j + 112A) Il, "+"2 0+ 1/2j-1/2.4)]
-
At y AIAY)

6.3 Simulated results

5.3.1 Simulation of a line-fed rectangular microstrip patch

I" order Mur boundary on all

boundaries except ground plane

12.448m 16,00mm

0.795mm
Fr = 2.2
1.945m T-
a= 5.8 xI CýS/m 2.334mm

Fig. 5.1 : Line-fedrectangularmicrostrippatch

In chapters3 and 4, the microstrip is modelledas a perfectelectric wall boundarywherethe tangential

electricfields on the microstripareforcedto be zeroat eachhalf time step.This modelsthe copperon the
microstrip as an ideal conductorwith an infinite conductivity. In reality, the copper layer has a finite
conductivitytakenas 5.8 x 107S/rnwhich contributesto its finite electric loss In the form of conduction
currenton the copperlayer.

In order to validate the ADI-FDTD program with the added electric conductivity term, ilia rinitc.
differencetime-domainequations(5.5) - (5.6) and (5.8) - (5.9) are usedwith the I" order Mur boundary
conditionto simulatethe propagationof a broad-bandGaussianpulseon a line.fcd rectangularmicrostrip
patch as shownin Fig. 5.1. As in chapter3, the finite-differencemesh parametersare chosento be the
sameas in the publishedpaper[1.5] to allow direct comparisonof results.

Fig. 5.2 showsa comparisonof the time-domainresponsebetweenthe ADI-FDTD and the conventional
FDTD method when stability factor of I is used in the ADI-FDTD program. With the copper layer
modelledas a layer of materialwith finite copperconductivity,a, of 5.8 x1 07S/m, the resultsare stable
whenstability factorsof up to 8 areused in the ADI-FDTD program.
 Copper Layer In Alternating-Direction Implicit Method 72
_Simulating
5.3.1.1 Transient response

1.2
1
0.8
0.6

E 0.4 I con%&Idtd
7ý1

0.2 adl-ldidifI
ut-ol
N
w
0
00
-0.2

-0.4

-0.6
t1me (ps)

Fig.5.2 : ComparisonbetweenconventionalFDTD andADI-FDTD with stability factor I

1.4
1.2

1
0.8
_odl-fdtd It I
0.6 adl-fUldat. 2
.......
0.4 adl-fdtdat. 3
0.2 adl-fdtdat 4
0 adl-fdtdat. 5
00
-0.2
-0.4
-0.6
time (ps)

Fig.5.3 : ComparisonbetweenconventionalFDTD andADl-FDTD with stability factorsI to 5

in stepsof I

Fig. 5.3 showsthe comparisonbetweenFDTD and ADI-FDTD with stability factors I to 5 in stepsor 1.
The resultsare completelystable.As the stability factor is increased,the effect of numericaldispersion
beginsto appearin the resultsasthe Gaussianpulsebeginsto broaden.This effect is shownmoreclearly
are magnifiedfor a clearerview orthe transientresponse.It
in Figs. 5.5 and 5.6 wherethe time responses
can be seenfrom Fig.5.6 that the Gaussianpulse broadensas the time step used Is Increased.Fig.5.4
showsa goodagreement between the FDTD
conventional andADI-FDTD with stability factor 1.
Simulating Copper Layer In Alternating-Direction Implicit Method 73

1.2

1
0.8
0.6
E
E 0.4 con*fdtd
N 0.2 adl-Idtd tI
w
0

)0
-0.2
-0.4

-0.6
time (ps)

Fig.5.4 : Zoomedin comparisonbetweenconventionalFDTD and

ADI-FDTD with stabilityfactor I

1.2

0.8
0.6

E 0.4 dl-Idtd: f 121

0.2 di-fdtd f'
w
0

-0.2
-0.4

-0.6
time (ps)

Fig.5.5 : Slight broadeningof pulsein ADI-FDTD with stability factor2

1.4
1.2
1
0.8
-adl-fdtd at. I
0.6
E ....... adl-Idtdat. 2
Z; 0.4 adi-idtdat. 3
WN 0.2 adi-Idtdat 4
0 ------ adl-fdtdSt 51

-0.2

-0.4

-0.6
t1me(ps)

Fig. 5.6 : Effect of numericaldispersionin ADI-FDTD with stability

factors3,4 and5
Simulating Copper Layer In Alternating-Direction Implicit Method 74

5.3.1.2 Frequency response

0248 10 12 14 16 18 20
0 11V .......ýI... IIIII frcq(GI lz)

-10

-2D
"a

-40
Adi-fdtd sf I

-M

Fig. 5.7: Comparison of ADI-FDTD with stability I with Abouzahra [1.5]

10 12 14 16 18 20

-5

LXI -10
«a
con\;-fdtd
adl-fdtd st I
-15

-20

-25
freq (GHz)

Fig. 5.8 : Comparisonof conventionalFDTD with ADI-FDTD with stability factor I

Simulating Copper Layer In Alternating-Direction Implicit Method 75

0 68 10 12 14 16 18 20

convfdtd
-20 - adi-fdtdat 21
0
-25-

-30-

-35

-401
fro q (GHz)

Fig. 5.9: Comparisonof conventionalFDTD with AD[-FDTD with

stability factor2

02468 10 12 14 16 IS 20

-5

M, -10
convfdtd
adl-fUtdat. 31
-15

-20

-25
freq (GHz)

Fig. 5.10: Comparisonof conventionalFDTD with ADI-FDTD with

stability factor3

02468 10 12 14 16 18 20
0i.,

-5

9-10
convfdtF-
CO
t. 4]
adl-ldid &t
-15

-20

-25
fro q (GHz)

Fig. 5.11 : Comparisonof conventionalFDTD with ADI-FDTD with

stability factor4
 Simulating Copper Layer In Altemating-Direction Implicit Method 76

02468 10 12 14 16 18 20

.5

6-10
-convfdtd
....... adi-fdtdat 51
-15

-20

-25
freq (GHz)

Fig. 5.12: Comparisonof conventionalFDTD with ADl-FDTD with

stability factor5

2468 10 12 14 16 IS 20
0--

adl-ldid st. I
adl-fdtdst. 2
-20 - adl-fdtdit. 3
adl-ldid st 4
-25 -
adl-fdtdst 5
-30-

-35-

-40-
freq (GHz)

Fig.5.13 : Comparisonof ADI-FDTD with stability factorsI to 5

Fig.5.7 showsa good agreementbetweenAbouzahra[1.5] simulateddataandthe ADI-FDTD resultusing

stability factor 1. Fig. 5.8 shows very good agreementbetween the results generatedusing tile
conventionalFDTD and the ADI-FDTD methodwith stability factor 1. Fig.5.9 showsthat with stability
factor2 the frequencyresponseresultsstill agreereasonablywell with the conventionalFDTD resultsbut
for frequencyhigher than 14GHz,the responsebeginsto shift slightly towardsthe lower frequency.This
cffect of numericaldispersionincreaseswith the increaseof the stability factor and it is greaterat the high
frequencyrange.The broadeningof the pulsein time-domaincausesa compressionof the responseIn the
frequencydomain. Figs. 5.10,5.11 and 5.12 show the comparisonbetweenconventionalFDTD and
ADI-FDTD with stability factors3,4 and 5 respectively.Fig. 5.13 showsthe comparisonof AD[-FDTD
resultswith stability factorsI to 5. The Wect of numericaldispersioncan be seenclearly.
Simulating Copper Layer In Altemating-Direction Implicit Method 77

5.3.1.3 Accuracyvs stability factor

resonance resonance resonance

stabilityfactor at 7.4GHz %error at 12.OGHz. %error at 18.OGHz %error
1 7.4 0 12 0 17.9 0.56
2 7.4 0 12 0 17.8 1.11
31 7.4 0 11.8 1.67 2.22
4 7.35 0.67 11.8 1.6 17.5 2.78
5 7.2 1 2.7 1 11.6 1 3.33 17.3 1 3 89
_1
Table5.1 : Percentage 7.4GIIz, l2. OGIIzand 18.OGIlz usingADl-FDTD
errorsat resonances
methodwith stability factors1,2,3,4and5 ascomparedto the respectiveresonances
using tile
conventionalFDTD method.

5.3.1.4 Run-time comparison

Computati nal run-time in minutes

Stability factor conv-FDTD ADI-FDTID time steps
1 25.75 58.35 8500
2 29.283 4000
3 19.45 2667
4 14.683 2000
5 11.67 1600
Table 5.2 : Run-timecomparisonusingcomputerwith Athlon 1.2GI Iz proccssor

Table 5.2 above shows that for the microstrip patch circuit, a stability factor greater than 2.0 In ADI.
FDTD scheme is required to have any time-saving as far as computational run-time is concerned.
Referring to table 5.1, by using stability factor of 3.0 in the ADI-FDTD, the errors are 1.67% and 2.22%

at resonant frequencies 12GHz and 18GHz respectively. This allows a time-saving of 24%. Although a
time-saving of 6.3 minutes in this example may not be significant, a 24% time-saving from 2 days, I.e.

saving of about half a day of simulation run-time when more complex structures such as a human body or

a huge aircraft are modelled will prove to be quite beneficial.

In reality, the % errorsfor stability factorsgreaterthan3 as shownin table 5.1 may not be tolerablewhen
it comesto designinga microstrippatch.It is, however,importantto emphasizeherethat the accuracyof
the ADI-FDTD methodis very muchdependenton the structurebeing modelled.Therefore,whenapplied
to otherstructures,a stability factor of greaterthan 3 may be usedwhilst maintainingthe accuracywithin
a tolerablerange.

Note that the saving in computationalrun-timeis not directly proportionalto the stability factor usedas
computation in the ADI-FDTD method is muchmore complex than the conventional FDTD and involves
matrix inversions.
SimulatingCopperLayer In Alternating-DirectionImplicit Method 78

5.3.1.5 Input impedance

83

40

30-
20-
%%
E%
0.
%%
71 72 73 74 76 77 78 zo
.Gr
N -10ý. w*.-

Abouzahra's real --o- adl-fdtd real

-w- ----
Abouzahra's Imag
.... . ..... adl-fdtd Imag
-40-
41) -

Fig.5.14 : ComparisonbetweenAbouzahra'sresultandADI-FDTD with

stability factor I for real andimaginarypartsof input impedanceof the
patchantenna

Fig.5.14 showsthe comparisonbetweenAbouzahra'sresultsand the ADI-FDTD with stability factor I

for real and imaginarypartsof the input impedanceof the patchantenna. As discussedin chapter2, the
discrepancybetweenboth setsof data is due to the fact that in the publishedpaper,the microstrip is

assumedto have a constantcharacteristicimpedance,Z,, of 50 Q and an effective permittivity of 1.9 is

usedto calculatethe wavenumber,6, whereasin this researchwork, thesevalueshave beencalculated
usingdataobtainedfrom the simulation.In this way, the dispersivenatureof the microstrip is accounted
for in the simulateddata.

50

40

30-

20 convfdtd real
convfdtd Imag
....... adl real st I
N 0- -. -. -adi Imagst I
.17.1 7.2 7.3 7.4 'X. 5 7.6 7.7 7.8 7.2., -e
-10-

-20-

-30
freq (GHz)

Fig.5.15 : ComparisonbetweenconventionalFDTD andADl-FDTD with stability factor I

for real and imaginarypartsof input impedanceof the patchantenna
Simulating Copper Layer In Alternating-Direction Implicit Method 79

Fig.5.15 shows a good agreementbetween the results generated using the conventional FDTD and the
ADI-FDTD with stability factor 1. Figs. 5.16 shows a comparison between conventional FDTD and ADI.
FDTD results with stability factors 2. Again, when stability factor 2 and above are used, the data shift
towards the lower frequency as can be seen in Figs. 5.17,5.18 and 5.19.

50

40-

30-

20- convfdtd real

E
convfdtd Imag
adl real at 2
N 0- adl Imag st 2
7.1 7.2 7.3 7.4, "*Z.
5 7.6 7.7 7.8
N
-10

-20

-30
freq (GHz)

Fig.5.16 : ComparisonbetweenconventionalFDTD andADl-FDTD with stability factor2

for real andimaginarypartsof input impedanceof the patchantenna

50-

40-

30-

20- conv-fdtdreal
E
0 10 conv-fdtdImag
adi real sf. 3
.S
N 0-
X4 adi Imagsf: 3
.17.1 7.2 7.3 5 7.6 7.7
-10- 00

-20-

-30-
freq (GHz)

Fig.5.17 : ComparisonbetweenconventionalFDTD andADI-FDTD with stability factor

for real and imaginarypartsof input impedanceof the patchantenna
Simulating Copper Layer In Alternating-Direction Implicit Method 80

60

50

40
30
conv-fdtdreal
E 20
= conv-fdtdImag
0
c 10 adi real sf-.4
F4
0 adi Imagsf: 4
7.4 Y, 5 7.6 7.7-,? ý.S-- 7.9 - -8
-10
-20

-30
freq (GHz)

Fig.5.18: ComparisonbetweenconventionalFDTD andADl-FDTD with stability factor4

for real and imaginarypartsof input impedanceof the patchantenna

60-

50-

40-

30-
conv-fdtdreal
20 conv-fdtdImag
io--' adi real sf: 5
0- adi Imagsf: 5
7.1 7.2 13 7.4 Y,,5 7.6
-10-
-20-
-30-
fro q (GHz)

Fig.5.19 : ComparisonbetweenconventionalFDTD andADI-FDTD with stability factor 5

for real and imaginarypartsof input impedanceof the patchantenna
SimulatingCopperLayer In Alternating-DirectionImplicit Method 81

5.3.2 Simulation of a line-fed rectangular microstrip patch with three parasitic patches

r--------------------------------- 12.448mm

UlAy ........... I............................

...........................
.....
..... I" order hl ur
absorbing
boundary
16.000mm

91AY ...........
.................
..........
......
10 mm 0., 9mI
90AY ................

16.000mm

I ......
50AY ...........
. ..... -
-169>
12.448mm 8.169 m 12.448nim I
0 2.334rrm
I --
58Ax 126Ax
14Ax 46AT SOAX 112Ax
52Ax
47Ax 79Ax

Fig.5.20 : Line-fedrectangularmicrostrippatchwith thrce parasiticpatches

By modellingthe copperpatchas a materialwith an electric conductivityor 5.8 x 107S/m, we havebeen

able to model the line-fed microstrip
rectangular patchusing the ADI-FDTD method without introducing
instability or attenuationevenwhen CFL stability criterion is not observed.In order to validatethat tills
techniqueworks on a relatively bigger electricalobject, the line-red rectangularpatch Is surroundedby
The
threeparasiticpatches. plan view of the structureis shown in Fig.5.20 above.The meshparameters

andexcitationmethod are the same as that used in chapter2. Figs. 5.21 - 5.24 show transientresponses
for this patchcircuit.

5.3.2.1 Transientresponse

1.2

0.8

0.6
E
E 0.4 -convfdtd
0.2 I ....... adl-Md it I
N
Ui
0

-0.2
-0.4

-0.6
time (ps)

Fig. 5.21 : ComparisonbetweenconventionalFDTD andADl-FDTD with stability factor I

Simulating Copper Layer In Alternating-Direction Implicit Method 82

1.4
1.2
1
0.8
MOM at. I
0.6 edl-fdtdat, 2
E
W.-t adl-fdtdat. 3
adi-fdtdit 4
LU 0.2
edl-Md at. 5
0

-0.2
-0.4
-0.6
time (ps)

Fig. 5.22 : ComparisonbetweenADl-FDTD with stability factors I to 5 in stepsor I

Fig. 5.22 showscomparisonbetweenADI-FDTD with stability factorsof I to 5. Again, tile resultsare

completely stable. This shows that the ADI-FDTD can be applied successfullyon electrically large

objects and the stability of the system is still maintainedwhen the time step used In the algorithm Is

greaterthan the maximum allowable according to the CFL criterion. As tile stability factor Is Increased

and thus increasingthe numerical time-step,the numerical dispersion becomesmore significant its
expected.

1.2

1
0.8
0.6
E
E 0.4 con%tldtd
0.2 adl-fUldst I

0
50
-0.2
-0.4

-0.6
time (ps)

Fig.5.23 : Zoomed in comparisonbetweenconventionalFDTD and ADI-FDTD with stability

factor I

Figs. 5.23 - 5.25 showthe magnifiedview of the transientresponsesof the rectangularmicrostrip patch
with three parasiticpatches. From Fig. 5.25, it can be seenthat the transientresponsestarts to show
significant inaccuracyin the result when the stability factor is increasedbeyond3. This Inaccuracymay
be dueto the fact that the dielectricgapseparatingthe microstrippatcheshasbeenmodelledwith a single
mesh width. The accuracy of the model and hence the result can be improved by increasingmesh
resolutionin the dielectricgap.
Simulating Copper Layer In Alternating-Direction Implicit Method 83

1.2

0.8
0.6
E
E 0.4 ad-fdtdst I
0.2 "adi4dtds2
w
0

-0.2

-0.4

-0.6
time (ps)

Fig. 5.24 : Slight numerical dispersion is observed when stability factor 2 is used In tile
ADl-FDTD

1.4
1.2
1
0.8
_edl-fdtd St. I
0.6
E ....... adi-Idtdat 2
0.4 adl.fUldat. 3
UNJ 0.2 adi-fdtdat. 4
0 adl-fdtdat. 5
io
-0.2
-0.4
-0.6
time (PS)

Fig. 5.25 : More significant numericaldispersionwith stability factorsgreaterthan 3 and

inaccuracybeginsto showwith stability factorsbeyond3

5.3.2.2 Frequency response

02468 10 12 14 16 18 20
0i., III

-5

6-10 II
V convfdtd
adi-IdtdIt I
!L -15

-20

-25
freq (GHz)

Fig. 5.26 : Comparisonof conventionalFDTD with ADI-FDTD with stability ractor I

Simulating Copper Laycr In Aitcmating-Direction Implicit Method 84
Fig. 5.26 above shows the frequency responsesfrom the conventional FDTD and ADl-FDTD with

stability factor I agree almost perfectly with each other.

02468 10 12 14 16 18 20

-5

6-10
coni.,fdtd
adi-Idtdst 21
-15

-20

-25
freq (GHz)

Fig. 517: Comparison of conventional FDTD with ADl-FDTD with

stability factor 2

02468 10 12 14 16 18 20
011

.5

-10
conv-fdtd
adi-fdtdst 31
-15

-20

-25
freq (GHz)

Fig. 5.28: Comparison of convcntional FDTD with ADI-FDTD with stability factor 3

02468 10 12 14 16 18 20
0-

con%Fldtd

V) II....... adi-fdtd Sf 4

-20-

-25-

.WI
freq (GHz)

Fig. 519: Comparison of conventional FDTD with ADI-FDTD with stability factor 4
Simulating Copper Layer In Alternating-Direction Implicit Method 85

As mentionedearlier,the inaccuracyin the resultsstartto creepin whenthe stability factor is greaterthan

3. Fig. 5.28 showsthat when stability factor 3 is used,the resultssuffer from only numericaldispersion
due to the bigger time-stepused.However,when stability factor of 4 is used,the result shown in Fig,
5.29,suffer from both numericaldispersiondueto the increasedtime-stepusedand the Inaccuracyof tile
modeldue to insufficient meshresolutionwithin the dielectric gap separatingthe main microstip patch
andthe threeparasiticpatches.

5.3.2.3 Input impedance

70

60
50

40
conAtitcl real
E 30
convfdtd Imag
20 adl real at I
N adl Imag of I
10

0
4
-10
-20
freq (GHz)

Fig. 5.30 : ComparisonbetweenconventionalFDTD andADI-FDTD with stability factor 1 ror

real and imaginarypartsof input for
impedance the line-fed rectangularmicrostrippatchwith
threeparasiticpatches

70
60
50
40
30 -con%-fdtd real
----conwfdtd Imag
20
....... adl real at 2
R 10 imag 2
-. -. -adi of
0
4
-10
-20
-30
fro q (GHz)

Fig. 5.31 : ComparisonbetweenconventionalFDTD andADI-FDTD with stability factor2 for

real andimaginarypartsof input impedancefor the line-fed rectangularmicrostrippatchwith
threeparasiticpatches

Again, the discrepancyin the amplitudeof the input impedanceplots shown in Fig. 5.31 is due to the
insufficientmeshresolutionin the dielectricgapbetweenthe patches.
Simulating Copper Layer In Altemating-Direction Implicit Method 86

5.4 Conclusion

It hasbeenshownin this chapterthat modellingthe copperlayer in ADI-FDTD asa layer of materialwith

a finite electric conductivitytaken as 5.8 x1 07 S/m keepsthe systemcompletelystableevenwhen the
CFL stability criterion is violated.This is not true whenthe copperlayer is modelledas a perfectelectric
conductorasdiscussedearlierin chapter3.

Usingthis methodof implementingthe copperlayer,the resultsfor a line-fedmicrostriprectangularpatch

show favourable with
comparison published resultsand the resultsobtainedfrom the conventionalFDTD
method.Tables of accuracyagainststability factors used and the run-time comparisonbetweenthe
conventionalFDTD andADI-FDTD methodshavealsobeenpresented.

To further validatethe useof ADI-FDTD methodon an electrically large object, the microstrip patchis
modelledwith three parasitic to
adjacent
patches its three edges.The resultsshow stability when tested
with stability factors of up to 8. As expected,increasingthe stability factor increasesthe numerical
dispersionerror.

It hasto be mentionedthat the solepurposeof modellingthis relatively largecircuit is to validatethe use

of ADI-FDTD on electricallylargeobjects.As a result,the gapsbetweenthe patcheshavebeenmodelled
usinga singlemeshwidth. This hasintroducedsomeerrorsinto the resultsdueto inaccuratemodellingof
the resultsarestill stablewith stability factorsof up to 8.0.
the gaps.Nevertheless,

In orderto modelsmall gapsaccurately,a gradedmeshshouldbe employedin the computationaldomain

wherethe mesh size is reducedgraduallytowardsthe gapsand is maintainedsmall within the gaps.This
formspart of the suggestedfurtherwork.

As in any engineeringfeat, nothing comesfree and in this ADI-FDTD case,there is a trade-offbetween

accuracyand simulation run-time. ADI-FDTD allows us to violate the fundamentalCFL stability
criterion without causing instability in the system.By using the ADI-FDTD method insteadof the
conventionalFDTD method,potentiallythe simulationrun-time can be significantly reduced.However,
increasingthe time-stepalsoreducesthe accuracyof the simulationresults.This accuracyis dependenton
the structurebeing modelled.Therefore,for certainstructuresa significant reductionin simulationrun-
time will be possiblewhilst maintainingthe accuracywithin the requiredtolerablerange. In additionto
that, structurescontainingdiscontinuitiescanbe modelledusing fine meshsize without the constraintof
usinga correspondinglysmall time stepin the simulation.
ConclusionAnd FurtherWork 87

CHAPTER 6
CONCLUSION AND FURTHER WORK

6.1 Overall conclusion

The finite-differencetime-domain method has been studied and applied in the Cartesiancoordinate
rectangularmicrostripstructure.In order to removethe Courant-Friedrich-
systemfor a three-dimensional
Levy stability criterion that governsthe maximumtime-stepthat can be usedin the FDTD algorithmto
maintain the stability of the the
system, implicit
alternating-direction method is investigated. The main
contributionof this work has beenthe new methodof simulating the copper layer on a microstrip in a
Cartesiancoordinatesystemin the ADI-FDTD scheme.This allows the applicationof
three-dimensional
the ADI-FDTD methodto model any three-dimensionalstructurethat consistsof copper layersin the
structure.

Althoughthe ADI-FDTD methodhasbeenusedin the literature,mostanalysisand applicationhavebeen

cavities,in both homogeneousand inhomogeneous
performedon simplethree-dimensional media[1.18]
and [1.32]-[1.371.Very often two-dimensionalmodelshave beenusedto verify the algorithm [1.28] and
[1.29], when structuresother than the free-spacecavities were modelled, such as the parallel-plate
waveguidemodel in [1.14], a two-dimensional model was used and a lossy dielectric with electric

conductivity of 15.0 S/rn was included in the model. On anotheroccasion[1.38] a sheetof an infinite
groundplane was modelled using an electric conductivity of 20.0 S/m and in [1.25], a monopolewith a
thin dielectricwall with electricconductivityof 4.0 S/mwas modelled.

In [1.26], microstrip resonators and filters were modelled using the ADI-FDTD method in a graded mesh

and a perfect electric conductor (PEC) boundary condition was applied on the microstrip layer. However,
when a transmission line was modelled by implementing a perfect electric conductor boundary condition
on the strip in the ADI-FDTD method as discussed in chapter 3 in this research work, the result was
unstable. This phenomenon was later confirmed in [1.21] where it was reported that if the tri-diagonal
solver in the literature [2.6] was used to solve the ADI-FDTD method when modelling a microstrip line,
the result was not always stable. Subsequently, an alternative mathematical algorithm for solving the tri-
diagonal matrix in the ADI-FDTD method was reported in [1.21].

For the first time, this researchwork hasshownthat by simulatingthe copperlayer on the microstripas a
materialwith electric conductivity of 5.8 x 107 S/m (which is the electric conductivity of copper)in the
ADI-FDTD scheme,the numericalresultsare alwaysstableevenwhenthe tri-diagonal
three-dimensional
solver as proposedin [2.6] is used. A different tri-diagonal solver as reported in [1.21] is thus not
required. It has been shown that the AD-FDTD methodcan be used to model realistic problemsin
engineeringdesignwithout the needto put artificially high lossymaterialto maintainstability. In orderto
exploit the advantageousfeatureof the ADI-FDTD method,it is importantthat the ADI-FDTD method
ConclusionAnd FurtherWork 88

canbe successfullyimplementednot only on cavity structuresbut also on real practicalthree-dimensional

structuresthat may consistof striplinesandmicrostrips.

Another contribution of this researchwork is the proposednew modified ADl-FDTD methodwhich

introducesa factor f in the ADI-FDTD algorithm. Using this new method, microstrip lines can be
simulatedas perfectelectricwall boundarywherethe tangentialelectric fields on the microstriparesetto
zero and stability of the systemis still maintained.Although the resultsfrom the modified ADl-FDTD
methodshowa reductionin the amplitude,this techniqueis easyto implementand it is usefulasa quick
methodto obtain resonant
accurate frequency points.

The ADI-FDTD methodhascontributedenormouslyin the field of numericaleicctromagnetics. Whenin

the past,the computationalrun-time is restrictedindirectly by the CFL stability criterion, now with the
advent of ADI-FDTD scheme,computationalrun-time can be significantly cut down to make each
simulation a realistic, practical solution. However, nothing comes free. There is a trade-off between
simulationrun-time and accuracyof the simulationresults.ADI-FDTD gives us the flexibility of usinga
biggertime-stepthan that allowed by the fundamentalCFL stability criterion without causinginstability
in the system. By using the ADI-FDTD insteadof the conventionalFDTD method, potentially the
simulation run-time can be reduced.But increasingthe time-step also reducesthe accuracyof the
simulationresults.This accuracyis dependenton the structurebeing modelled.Therefore,for certain
structuresa hugereductionin simulationrun-timewill be possiblewhilst maintainingthe accuracywithin
the requiredtolerablerange. Structurescontainingdiscontinuitiescan be modelledusing fine meshsize
without the of
constraint usinga correspondinglysmall time stepin the simulation.

6.2 FurtherWork

6.2.1 Cylindrical coordinate system

The ADI-FDTD methodcanbe extendedto the cylindrical coordinatesystemto modelthree-dimensional
cylindrical structuresuch as the probe-fedcircular patch. As the diameterof the probe will be much
smallerthanthe circular patch,the applicationof the ADI-FDTD in this structuremeansthat the time-stcp
used in the algorithm will not be restrictedto the mesh size used to model the probe. Although, the
applicationof ADI-FDTD on cylindrical coordinatesystemhasrecentlybeenreported[1.31], thereis yet
to be any implementationof striplinesor microstripsin the cylindrical coordinatesystem.

6.2.2 Microstrips with slots and notches

Now that there is a method of implementing copper layer in the microstrip that is not dependenton the

accuracy of the tri-diagonal solver in the ADI-FDTD scheme, any three-dimensional structures with
microstrips, such as stacked array of microstrip antennas can be modelled with narrow gaps, slots and
notches where these discontinuities can be modelled with high spatial resolution by applying fine mesh
size without the prohibitive cost in computational time.
Conclusion And Further Work 89

6.2.3 Graded Mesh

In orderto model small gaps,slotsand notchesaccurately,higher meshresolutionis requiredaroundthe
vicinity of these To
discontinuities. avoid having fine meshthroughoutthe whole domain,
computational
the meshcan be gradedsuchthat the meshsize graduallydecreasesas it approachesthe discontinuities.
Sincein the ADI-FDTD method,the stability of the systemno longer dependson the CFL criterion,the
sametime stepwithin tolerablenumericaldispersion,canbe appliedto the whole computationaldomain.
 References 90

REFERENCES

Papers

F.A. Alhargan and S.R.Judah,"Reducedform of the Green's functions for disk and annular
rings", IEEETrans.MicrowaveTheoryTech.,vol.39, no.3, pp.601-604,March 1991.

[1.2] P.B.Johns,and R.L. Beurle, "Numerical solution of 2-dimensionalscatteringproblemsusing a

transmission-linematri)C',Proc.IEE, vol. 118,pp.1203-1208,1971.

[1.3] W.J.R.Hoeffer, "The transmission-linematrix method- theory and applications",IEEE Trans.,

vol. 33, pp. 882-892,1985.

[1.4) K. S.Yee, "Numericalsolutionof initial boundaryvalue problemsinvolving Maxwell's equations

in isotropicmedia," IEEE Trans.AntennasPropagat.,vol. AP-14,pp. 302-307,May 1966.

[1.5] D. M. Shcen, S.M. Ali, and M. D.Abouzahra, "Application of the three-dimensional finite-
difference time-domain method to the analysis of planar microstrip circuits", IEEE Trans.
Microwave Theory Tech., vol. 38, no. 7, pp. 849-857, July 1990.

[1.6] X.Zhang, and K.K. Mei, "Time-domain finite-difference approachto the calculation of the
frequency-dependentcharacteristicsof microstrip discontinuities", IEEE Trans. Microwave
TheoryTech.,vol. 36, no. 12,pp. 1775-1787,Dec. 1988.

[1.7] A.Taflove, and M.E.Brodwin, "Numerical solution of steady-stateelectromagneticscattering

problems using the time-dependentMaxwell's equations", IEEE Trans. Microwave Theory
Tech.,vol. 23, no. 8, pp. 623-630,Aug. 1975.

[1.8] P.Thoma, and T.Weiland, "A consistentsubgridding schemefor the finite difference time
domainmethod",InternationalJournalof NumericalModelling : ElectronicNetworks,Devices
andFields,vol. 9, pp. 359-374,1996.

[1.9] M.Okoniewski, E.Okoniewska,and M.A. Stuchly, " Three-dimensionalsubgriddingalgorithm

for FDTD", IEEE Trans.AntennasPropagat.,vol 45, no.3, pp. 422428, March 1997.

[1.10] S.S.Zivanovic, K. S.Yce, and K.K.Mei, "A subgridding method for the time-domainfinite-
differencemethodto solve Maxwell's equations",IEEE Trans. Microwave Theory Tech.,vol.
39, no. 3, pp. 471-479, March 1991.
References 91
[1.11] I. S.Kim, and W. J.R.Hoefer, "A local refinement algorithm for the time-domain finite-difference

methodusing Maxwell's curl IEEE

equations", Trans. Microwave Theory Tech., vol. 38, no. 6,

pp. 812-815,June1990.

[1.12] D.T.Prescott,and N.V. Shuley,"A methodof incorporatingdifferent sized cells into the finite-
differencetime-domainanalysistechnique",IEEE Trans. Microwave Theory Tech., vol. 2, no.
11,pp. 434436, Nov. 1992.

[1.13] D.W.Peaceman,and H.H.Rachford, Jr., "The numerical solution of parabolic and elliptic
differentialequations",J. Soc.Indust.Appl. Math, vol. 3, no.1, pp. 28-41, March 1955.

[1.14] T.Namiki, "A new FDTD algorithm basedon alternating-directionimplicit method", IEEE
Trans.MicrowaveTheoryTech.,vol. 47, no. 10,pp. 2003-2007,Oct. 1999.

[1.15] B. Engquist, and A.Majda, "Absorbing boundaryconditions for the numerical simulationof
waves",Math. Comp.,vol. 3 1, pp. 629-651,July 1977.

[1,16] G.Mur, "Absorbing boundaryconditions for the finite-differenceapproximationof the time-

domainelectromagnetic-fieldequations",IEEE Trans. Electromagn.Compat.,vol. 23 no. 4, pp.
377-382, Nov.1981.

[1.17] S.G.Garcia, T.Lee, and S.C.Hagness,"On the accuracy of the ADI-FDTD method", IEEE
AntennasandWirelessPropagationLetters,vol. 1, no. 1, pp. 31-34,2002.

[1.18] F.Zheng, Z.Chen, and J.Zhang, "Toward the development of a three-dimensional

unconditionallystablefinite-differencetime-domainmethod", IEEE Trans. MicrowaveTheory
Tech.,vol. 48, no. 9, pp.1550-1558,Sept.2000.

[1.191 A.P.Zhao, "Analysis of the numericaldispersionof the 2-D alternating-directionimplicit FDTD

method",IEEE Trans.MicrowaveTheoryTech.,vol. 50, no. 4, pp.1156-1164,Apr. 2002.

[1.20] M.J.Chang, L.C.Chow, and W.S.Chang, "Improved alternating-directionimplicit method for

solving transientthree-dimensionalheatdiffusion problems",Numerical Heat Transfer,Part B,
vol. 19,pp. 69-84,1991.

[1.21] A. P.Zhao, "Two specialnoteson the implementationof the unconditionallystableADI-FDTD

method",MicrowaveOpt TechnolLett, vol. 33, no. 4, pp. 273-277,May 2002.

[1.22] A. P.Zhao, "The influences of the time step on the numerical dispersion error of an
unconditionallystable3-D ADI-FDTD method:A simpleand unified approachto determinethe
maximumallowabletime steprequiredby a desirednumericaldispersionaccuracy",Microwave
Opt TechnolLett, vol. 35, no. 1, pp. 60-65,Oct. 2002.
References 92

[1.23] T.W.Lee, and Magness, "Wave sourceconditions for the unconditionallystableADI-FDTD

method",IEEE AP-S Digest,vol. 4, pp. 142-145,
July 2001.

[1.24] S.Garcia, T.Lee, and S.Hagness,"Accuracy limitations of the ADI-FDTD method due to
truncationerror", Proc.186'ACESConf., Monterey,CA, pp. 281-287,2002.

[1.25] T.Namiki, "3-D ADI-FDTD Method- Uncondionallystabletime-domainalgorithm for solving

full vector Maxwell's equations", IEEE Trans. Microwave Theory Tech., vol. 48, no. 10,
pp. 1743-1748, Oct. 2000.

[1.26] T.Namiki, and K. 1to,"Numerical simulationof microstrip resonatorsand filters usingthe ADI-
FDTD method",IEEE Trans.MicrowaveTheoryTech.,vol. 49, no. 4 pp. 665-670,Apr. 2001.

[1.27] T.Namiki, "Numerical simulationusing ADI-FDTD methodto estimateshieldingeffectiveness

of thin conductiveenclosures",IEEE Trans.Microwave Theory Tech., vol. 49, no. 6, pp. 1060.
1066,June2001.

[1.28] T.Namiki, K. Ito, "Investigation of numerical errors of the two-dimensional ADI-FDTD

method",IEEE Trans.MicrowaveTheoryTech.,vol. 48, no. 11,pp. 1950-1956,Nov. 2000.

[1.29] F.Zheng, Z.Chen, and J.Zhang,"A finite-differencetime-domainmethod without the Courant

stability conditions", IEEE Microwave Guided Wave Lett., vol. 9, no. 11, pp. 441-443,Nov.
1999.

[1.30] F.Zheng, and Z.Chen, "Numerical dispersionanalysisof the uncondionallystable 3-D ADI.
FDTD method", IEEE Trans. Microwave Theory Tech., vol. 49, no. 5, pp. 1006-1009,May
2001.

[1.31] C.Yuan, and Z.Chen, "A three-dimensionalunconditionallystable ADI-FDTD methodin the

cylindrical coordinatesystem",IEEE Trans.MicrowaveTheoryTech.,vol. 50, no. 10,pp. 2401-
2405,Oct.2002.

[1.32] M.Darms,R.Schuhmann,H.Spachmann,and T.Weiland,"Dispersionand asymmetryeffectsof

ADI-FDTD", IEEE Microwave and Wireless ComponentsLett., vol, 12, no.12, pp.491-493,
Dec.2002.

[1.33] S.Wang, and F.L.Teixeira, "An efficient PML implementationfor the ADI-FDTD method",
IEEE MicrowaveandWirelessComponentsLett., vol. 13,no. 2, pp. 72-74,Feb.2003.

[1.34] G.Liu, and S.D.Gedney,"Perfectly matchedlayer media for an uncondionallystable three-

dimensional ADI-FDTD method", IEEE Microwave and Guided Wave Lett., vol.10, no.7,
pp.261-263,July 2000.
References 93

[1.35] A.P.Zhao, "Uniaxial perfectly matchedlayer media for an unconditionally stable 3-D ADI-
FDTD method", IEEE Microwave and WirelessComponentsLett., vol. 12, no.12, pp. 497-499,
Dec. 2002.

[1.36] G.Lazzi, "Unconditionally stable D-H FDTD formulation with anisotropic PML boundary
conditions", IEEE Microwave and Wireless ComponentsLett., vol. 11, 4,
no. pp. 149-151,
Apr.2001.

[1.37] Z. Chen, and J.Zhang, "An unconditionally stable 3-D ADI-MRTD method free of the CFL

stability condition", IEEE Microwave and Wireless Components Lett., vol. 1 1, no. 8, pp. 349-351,

Aug. 2001.

[1.38] T.Namiki, "Unconditionally stable FDTP algorithm for solving Maxwell's

three-dimensional

equations",2000 IEEE MTT-S Int.Microwave Symp. Dig., pp.231-234, Boston,May 2000.

[1.39] J.P.Bercnger, "A perfectly matched layer for the absorption of electromagneticwaves",
J.Comput.Phys.,vol. 114,pp. 185-200,Oct. 1994.

[1.40] I. Wolff, "Finite difference time-domainsimulation of electromagneticfields and microwave

circuits", Int. Jour.
of MIMICAE, vol. 5, no. 3 pp. 163-182,1992.

[1,41] D.H.Choi, and W.J.

R.Hoefer,"The finite-differencetime-domain method and its applicationto

eigenvalueproblems", IEEE Microwave

Trans. Theory Tech., vol. 34, no. 12, pp. 1464-1470,
Dec. 1986.

[1.42] X.Zhang,J.Fang,and K.K.Mei, "Calculationsof the dispersivecharacteristicsof microstripsby

the time-domainfinite differencemethod",IEEE Trans.Microwave Theory Tech., vol. 36, no. 2,

pp. 263-267.Feb.1988.

[1.43] Z. Bi, K. Wu, C.Wu, and J.Litva, "A dispersive boundary condition for microstrip component

analysis using the FD-TD method", IEEE Trans. Microwave Theory Tech., vol. 40, no. 4, pp.
774-777, Apr. 1992.

[1.44] K.K.Mei, and J.Fang,"Superabsorption

-a method to improve absorbingboundarycondition",
IEEE Trans.MicrowaveTheoryTech.,vol. 40, no. 9, pp.1001-1010,Sept.1992.

[1.45] J.Fang,and D.Xeu, "Numerical errorsin the computationof impedancesby FDTD methodand
ways to eliminate them", IEEE Microwave and Guided Wave Lett., vol. 5, no. 1, pp. 6-8,
Jan.1995.
References 94

[1.46] G.Zheng, and K. Chen,"The studiesof cylindrical microstrip line with the FD-TD methodin
cylindrical coordinatesystem", IntJournal of Infrared Millimeter Wave, vol. 13, no. 9, pp.
1421-1431,1992.

[1.47] N. Dib, T. Weller, M. Scardelletti, and M. Imparato, "Analysis of cylindrical transmission lines the
finite-difference time-domain method", IEEE Trans. Microwave Theory Tech., vol. 47, no. 4,

pp. Apr.
509-512, 1999

[1.48] S.C.Ow and S.R.Judah, "Three-dimensionalFDTD based on modified altcrnating-direction

implicit method", Proceedingsof the WhInternationalSymposiumon Microwave and Optical
Technology,pp. 257-260,June2001.

121 Books

[2.1) K.C.Gupta, R.Garg, and R.Chadha,Computer-aideddesign of microwave circuits, Artech

House,Dedham,MA, 1981.

[2.2] G.F.Roach,GreensFunctions,Van NostrandReinholdCompany,1970.

[2.3] T.Itoh, Numericaltechniquesfor microwaveandmillimctcr-wave passivestructure,Wiley, New

York, 1989.

[2.4] G.D.Smith, Numerical solution of partial differential equations, Oxford University Press,
London, 1965.

[2.5] A.Taflove, ComputationalElectrodynamics: The finite-differencetime-domainmethod,Artech

House,MA, 1995.

[2.6] W.H.Press,S.A.Teukolsky,W.T.Vetterling,and B.P.Flannery,Numerical recipesin C- the art

of scientific computing,2nded.,CambridgeUniversityPress,New York, 1992,pp. 50-51.

[2.7] W.F.Ames,Numericalmethodsfor partial differentialequations,New York: Academic,1977.

[2.8] A. R.Mitchell, and D.F.Griffiths, The finite difference method in partial differential equations,
Wiley, New York, 1980.

[2.9] A.Taflove, Advances in ComputationalElectrodynamics: The finite-differencetime-domain

method,Artech House,MA, 1998.

[2.10] W.C.Chew,WavesandFields in Inhomogcneous

Media, IEEE Press,NY, 1995.

[2.11] R.F.Harrington,Time-HarmonicElectromagneticFields,McGraw-Hill, 1961.

 2ndOrder Accuracy of Central-Difference Approximation 95

APPENDIX Al
ND
2 ORDER ACCURACY OF CENTRAL-DIFFERENCE
APPROXIMATION

Consider a Taylor's series expansion of u (xi, t) about the time instancetn to the time instancet,, + At/2,
keeping the spacepoint fixed at xi :

At At I L-+-)2 1 (! Lt )3UI 03i

ul + +( + (A 1.1)
"ýýUj 2 _L
2 at 2t 2! at2 2)3! at
j. X, X,

Now, the Taylor's series expansion of u (xi, t) about the time instancetn to the time instancet, - At/2,
keeping the spacepoint fixed at xi is

11 (At)2 (T) I a, ul
At tUlat
OU
U 1 a2UI At),
+ (Al. 2)
+ ý-! -FtTI Tt
n2 2 at Tj 3 -31
Xi.t. Xi.t. Xi.t.
(1) - (2) gives :

(t. (At) 31 a3UIxi.

ItAt At
TIX = At + +
3 013
X, t.

Re-arranging(3), we get,

( At (t. At At), I aul

+ itTl
21 2) 2)3
oul X, X, xo,'.
At
(A 1.4)
At At
u+ 2
u tn _ 2J., 1
(At)' I a3;
_ 2)6 at I
At
x,.t.

Taking only the first RHS term of (4), the secondterm is the error term. Then,

Ix, (tn

u+U 2
aul
+
at] At
X,

[(AI)2]
where 0 is a shorthand notation for the remainder or error term, which approacheszero as the

square of the time increment. Equation (5) is referred to as a 2"d order accurate, central-difference
approximation to the first order time derivative of u.
 Tri-diagonal Matrix Equations For ADI-FDTD Method 96

APPENDIX BI
TRI-DIAGONAL MATRix EQUATIONs FOR
ADI-FDTD METHOD

Procedure I

I
U: CIAY)2
+1/2 E n+1/2 ,U
"
E, (i+112j-l, k) (1+112jk) 2+ + Exn+1/2 fl+112j+l, k)
jr dt
'41
(_ýýpL 2+(
lAy) Ay [Ey"fl+
"
-E, (i + 112j, k) I j+112, k)-Ey"flj+112, k) (i+ I j-112, k) + k)
Eyn(ij-112,
AX) -Eyn

2) (.
[H,"(i+112j+112, k)-H"(i+112j-112. UAy [H"(i+112jk+112)-H"(1+112jk-112)
ý-A,
-('uAyl )z( k)]+ 7AtAzjyyI

(BI. 1)

( )2]
A,
Ey*+' /2 (ij+112, k-l)-Ey+ 1/2 (ij+112, k)[2+ + Ey"+1/2 (Ij + 1/2, k+ 1)
Jjl
2
ru ('11,
c [E.,"(ij
-Ey" (ij+112, k) ýy-) + I, k +I/ 2) - E," rijk +I/ 2) - E,' (Ij + IJ-112) + En (tjk-112)
A t+ zI

ýA--
[H.' [Hn(!
- (ij+ 112,k + 1/2) - Hn(ij + 1/2, k -I/ 2)] + =At) +ll2j+112, k)-H, "(1-112J+112, k)I
At-)

(B 1.2)

[2
AxX)2]
ljk + 1/2) E»+112(iik + 1/2) + + Exll/2(1+ ljk + 1/ 2)

2
Ax ) [E,
- -E,' (ijk + 1/2) +( K;
" (I + 112jk + 1) ER(i + 112jk) Ex"(i 112jk + 1) + ER(I 112jk)
- - - -
At ,Az
[Hy'(1+112jk+112)-Hyn(i-112jk+112)1+ (,
UAX2 [H,' (ij 1/2.k 1/2) Hm(ij 112,k 1/2)]
-('uAx)
At XTAy , + + - - +

(B 1.3)
 Equations For ADI-FDTD Method 97
Tri-diagonal Matrix

Procedure 2

(- )2
NF/UEAZ + E,,
2+ ( -Idt n+'(i+ 112jk + 1)
E,""fl+ 112jk -1) - E,"+l# + 1/2jk) iI

lpýAz) 2
(-A'z ) [E,
"+1/2(i+ljk + 1/2) - E,"+1/2(ijk + 1/2) - E'+1/2 (i+ljk-112) + E,'0+1/2(ijk-112)
En+112(i + 112j, k) +
x At AX
'i;

(PAZ ) [H; L-- Z2 [H, H, "+1/2 (i+1/2j-1/2, k)]

+1/2(i + V2j, k + 1/2) - H; +112
(i + 1/2jk-1/2)] +(A. "+1/2(i + 1/2j + 1/2,k)
At Ady

(B 1.4)

Ey""(1- I J+112, k) -Ey"ý'(ij+112, k) 2++ Ey"+'(i+ I, J+112, k)

)I
0-1

-2
(xr,
M
&x) [E, +1/2
ýAx (i+1/2j+l, (i+112jk)
k) Ex"+112 Em+1/2 (i - I/ 2j + 1,k) + E"+1/2(1-1/ 2jk)
(ij + I/
-Ey"+112 2,k) +- FAy) xx
FAI [Hxn+112
[H,"+112(i+112J+I12, "+112 k)]+(*&ýý2) (ij I/ 2,k 1/ 2) (ij
Hx"+112 + I/ 2,k 2)]
ý-A, )
-rpAx) k)-H, (i-1/2j+1/2, rAFaTj + + - -I/

(Bl. 5)

FPEAY)2]
E,"*'(ij-l, k+112)-E, "*'(ij, k+112) 2+ ( + E,"+'(ij + I, k +I/ 2)

Fp [Ey"+1/2 I
EAY) Ay
En+1/2 (I + 112jk)
( +() (ij + I/ 2, k + 1) Ey"+1/2 (Ij + I/ 2, k) Ey"'1/2 (Ij k + 1) + Ey"+112(ij -I/2, k)
At Az - - -I/2,

(PAY) [H.,, [H;

+1/2 (ij +I/2, k + 1/2) H, " +1/2 (ij-1/2, k+1/2)] +) +1/2 (i+112jk+112). Hy"+112(i-112jk+112)]
T - ,

(B 1.6)
Tri-diagonal Matrix Equations For Modified ADI-FDTD Method 98

APPENDix B2
TRI-DIAGONAL MATRix EQUATIONs FOR
MODIFIED ADI-FDTD METHOD

Procedure I
)2( )I
E,n+"2(i+112j-l, k) -E,, '+1/2(i+ 112j, k) 2+ + En+"2(i + 1/2j + I, k)
(2-1 f)2
Idt

( '((2
(i + I/ 2j, k)
-En At )2
.,, _'f

Ay (f [E,,
+( (i+lj+112, k)-E"(ij+112, k)-E"(i+lj-112, k) + En Oj - 1/ 2,k)
AX) 2-f) yyy

Ay I [H"(i+112j+II2,
fu k)-Hn(i+112J-II2, k)
At 2-f)
(L &L11 f [Hn(i+112jk
+ , +112)_Hn(i + 112jk 1/2)]
(2 f )2 yy -
AtAZ
(B2.1)

)2(
E;n+112
(ij + I/ 2,k _ 1) Eyn+l/ 2(ij +1/2, k) 2+ + En+112(1,j 1/ 2,k+ 1)
(2 f)2 y+
Idt -

"AZ)2(
Alsi
3L'Uý I
-Ey"(ij+112, k) -
At (2- f)2

(Az (f [En
+ 0i + l, k + 1/2) - En Qk +1/2) - En (Ij +I, k - 1/2) + En Qk -1/2)
Ay) 2-f) z
(L-ýAz I [Hn
- 0i +I /2, k + 1/2) - H,',+1/2Q+I /2, k -1 /2)]
At 2-f) x
(EAZ f) [Hn(i
+ +1/2j+1/2, k)-Hz n6 -1/2j+1/2, k)]
(2 f)2 z
_ (B2.2)

)2(
1+E..
E.n+112(i-li, k+112)-En+112(ii, k+112) 2+ "41(i+ ljk +1/ 2)
z (2 A2
Idt -
)2(
"(i+112jk)
-E, At

[Exn
+f 0+ 1/2jk + 1) - En (i + 112jk) - Exn (i - I/ 2j, k+ 1) + E"x (i - I/ 2jk)
Az) 2-f)
(L41 ý2-f) I [Hy"
0+I/ 2jk + 1/2) - Hn (i - I/ 2jk + 1/2)
At y
(, )
2(f )xx[Hn(ij+112,
UAY k+112)-H"(ij-112,
+ k+112)]
AtAx (2 _ f)2
(B2.3)
 Tri-diagonal Matrix Equations For Modified ADI-FDTD Method 99

Procedure2

I( (ýPTCAZ) 2((2
I+E, )]
+'(i +112ik) UCAZ
E."+'(i+ 1/2ik - 1) - Eý", 2+ .
-ýL- '+'(i + 1/2j, k+ 1)
dt f)2
-

7-oAz
' 11,12
(i + 112j, k)
-E. At

f) [E,
' +1/2(i+ Ij, k+ 1/2) En+112(1j, k+ 1/2) Ez"-11/2(i+ljk-112) + E.."' 1/2 (ijk. 112)
. -
Ax 2-f
(EA-Z)( I [H
n+112
(i + 112j, k+ 1/2) -H n+1/2(i+1/2jk-1/2)
- At 2-f yy
(, JT2
&Z2
U, fz [H n+112
(i n+1/2
+
7)7) + 112j + k)
112, - H: 0+ 1/2j-1/2, k)]
WAY
(B2.4)

2
( )I
E'+'fl-lj+112, k)-E"+'(Ij+112, k) 2+ -Y
72-ýIf +En+'(i+I, J+112, k)
yy dt y

(EA"Otlllx), (I
n+112
(ij +I/2, k)
-E y (2-f)2

Ax ) [En,
+112(i+l / 2j + I, k) n+1120+I/ 2j, A) -E ,+1/2 0-112j+ I, k) + En+112(1-1/2jk)
-E xxxI
Ay 2-f
N 'i )z(ý [H
n+112
fl+112j +1/2, k) -H n+112
0-112J+112, k)
At -,
-7)

[HXn+112
+f (Ii +1/2, k +1/2) - Hxn+112(ij+1/2, k-1/2)]
AtAz (2-f)2
)
(B2.5)

[2 (. EpcjAy)2( )I
1
Eý+lflj -I, k +1/ 2) - E.."+'(ijk + 1/2) + ý (2 f)2
+ E,"+'(Ij + Lk +I/ 2)
A -
rF2
+1120+ 112ik) EAY
JUE,
-En (-3

Ay f) [Ey
+( n+1/2 fli +1 / 2,k + 1) En+112 (Ij
+ I/ 2,k) En+112 (IJ-1/2, k+l) + En+112 (Ij - 1/ 2,A)
Az 2-f yy
NI-y- )x(T--1f ) [Hx'+112
- At (ij+112, k+112) _ lln+112(ij-1/2, k+1/2)

(ýUAY2(f)
[Hym
+ +1/2(i+1/2jk+1/2) lln+112 (I
AlAx) (2-fy . y - 112jk +1 2)]
(B2.6)
Tri-diagonal Matrix Equations For ADI-FDTD Method With Electric Conductivity Term 100

APPENDix B3
TRI-DIAGONAL MATRix EQUATIONs FOR
ADI-FDTD METHOD WITH ELECTRIC CONDUCTIVITYTERM

Procedure I
2ý1+
FPCAY )4c
(I+ 1/2j, k) At )I
,
E,; +' 12(i + 1/2j I, k) - En+l /2(i + 1/2j, k) 2+( + En+1 12(1+ 1/2j + l, k)
-
'JI

(VPCAY )(2
c(i+112jk) At
"(i+112jk) I-
-E. 4c
At
(Ay) [Ey"
+ (i+lj + 112,k) - Ey"(ij + 1/2,k) - Eyn(I+IJ-1/2, k) + Ey"(Ij- 1/2,k)
Ax
)( [H, ) [Ily(i+112jk+112)-ify"(1+112jk-112)I
"(i+112J+II2, k)-H, "(i+ll2j-ll2, k)]+ PAY'
-('uAy
At AlAz

(B2.1)

u-,,Az (Ij
01 + 1/2,k) At )I
12 k 1)
Ey"" (Ij +I/2, - - 12(ij
Ey"+' +112,k) 2+ , + + 12(ij
Ey"+' + 1/2,k + 1)
it 4c

k) a (Ij + 1/2,k)
(ij
-Eyn + 1/2,
4e

2 [E(ij+I,
k+112)-E, "(ijk+112)-En, (/J+I, k-1/2) +E, "(Ijk -1/2)
y') z

L-Az ) [H" [/I,

(Ij + 1/2,k + 1/2) - Hn (Ij + 1/2,k - 1/2)] " (i + 1/2j + 112,k) I/, ' (I 1/2j + 1/2,k)]
At x - -

(B2.2)

12(i 12 UCAX)2(l fljk +1/ 2)

E -1jk+112)-E"+' (ijk+112) 2+ +a +E"+"2(i+ljk+112)
4t
dt 4e
,
2
I 'PCAX)4c(I
(ijk + 1/ 2) At
=-E"(ij, k+112) _a
At
(Ax) [En ]
+ ýý) (i+ 112jk +1) - En (l+ 1/2jk) - E," 0-1/ 2jk +1) + E.," (i-I 12jk)
xx
(LAt! ) [Hy"(1 ) [If,,
OAX2 '(Ij +I/2, k + 1/ 2) ll, " (Ij -I/2, k + 1/ 2)]
+ I/ 2jk + 1/2) -HI/y 2jk + 1/2)] +( rAtAy j - ,

(B2.3)
Tri-diagonal Matrix Equations For ADI-FDTD Method With Electric Conductivity Term 101

Procedure 2

(2
(i + 112jk) At )l
E""(i + 112jk - 1) - E,'+'(i + 1/2jk ) 2+ +, + E,' +'(I + 1/2jk + 1)
Of 4e

( ), (I (1+112jk) At
(I
-Exn+112+112jk) -a
At 4c

[E,
+(A-' "+112(i+ljk+112)-E, '+112(ijk+112)-E, '+112(i+ljk-112) + E' '+1/2(ijk-1/2)
Ar)
[Hn+112 /2 (i+112jk-112)] +
(1+112jk + 1/2) "+1/2(I + 1/2j-1/2, k)]
(1+ 1/2j + 112,k) 11,
y -Hy"+' &tAY)
At

(B2.4)

(f4t 2
A-X
UEE "1 (IJ+112, k)At )I
Ey"'(i-lj+112, k)-E"+'(ij+112, k) 2+ 1+ +Eym+'(I+I, J+112, k)
y 4c

ICA (ij +I/2, k) At

EPI+1/2(I k)
yj At
(yAr) [E, 1/2
+ t'-J "+ (i+112J+I, k)-E, '+112(i+112j. k)-E, "+'12(i-112J+I, k) + E,"+1/2(1-1/2jk)

(, UAX) [H, PAX2) [Ii, "+1/2(/J+1/2,

El ) "+112
(1+1/2j+1/2. k) "+1/2(I k)] +r,ý ) k+1/2)-il, '+1/2(ij+1/2, k-1/2)I
-H, -1/2j+1/2, Al

(B2.5)

()( 2
Es"+'(IJ-I, k+112)-E, ""(Ij, k+112) UCAY a(ij+112, k) At
2+ 1+ + En"(Ij + Lk + 1/ 2)
4c

(
(I-a(ij+112, k)At
--E, "+"(/+1/2j, k)
At 4c
[Eyll+1/2(/J+1/2,
+(Ay k+l)-E; +I,'2 (IJ+112, k)-E ; +1/2(Ij-
Az) I /2. k + 1) +E ; +'/z (ij - I/ 2.k)

(-ý Ay )) [H,
"+"2(ij+112, k+112)-H, "+112(IJ-112. k+112 ]+(P'ý'Y2) [11; +1/2(i+1/2jk+1/2)-Ily"+1/2(i-1/2j. k+1/2)I
- At (-Zl-m i

(B2.6)
Graphical illustration of implicit/explicit ADI-FDTD method 102

APPENDIX Cl
GRAPHICAL ILLUSTRATIONOF
IMPLICIVEXPLICIT ADI-FDTD METHOD

A'

E'+'

n+1/2
E

E
A
B
time
At /2 At /2

Fig. CIA

Using an explicit method or 2"d

order Runga-Kutta method to find electric field at time (n+I)At gives

E n+1 =En+ d El
At
dt
n+1/2
ýE
where is the gradient at E"+"2 illustrated in
dtln+1/2 Fig. C1.1 bythegradicntofAA'

In electromagnetic fields,
dEj IVX
It "I (C 1.2)
d n+1/2
n+ 112

Substituting (C 1 into (C 1.1), we get

-2)

ät
En+' = En + VX Hln+I/2
9

(Cl. 3) is the form of the conventionalFDTD explicit method.

Graphical illustration of implicit/explicit ADI-FDTD method 103

Using an explicit method to find E "112 from E' using forward difference gives:

At dE
E n+1/2 En+ (CIA)
2 dtý.
ý
dEj
where is the gradient at E" illustrated in Fig. CIA by the gradient of BB'.
dt
n

usingbackwarddifferencegives:
Usingan implicit methodto find E"+1from E'+112

At ýd
dE
E'+112 = E"+1 .T (C1.5)
tI
n+I
0dE
where is the gradientat En+1illustratedin Fig. CIA by the gradientof W.
dt In+1

Re-arranging(Cl. 5),

n+1/2 At dEI
E1 = E + (C1.6)
2 dtln+l

Combining(CIA) and(Cl. 6), we get

At At dEl
I-IPl
En+l = + lt., (CI.7)
2 dtln 2d
+,
or expressingthe time derivativeof electric field in (C 1.7)in termsof magneticfield givesus

En+1 t At (C1.8)
En + "In + VX "Ll
2tVX 2e

explicit implicit
term term
(CI. 8) is the form of the Crank-NicolsonFDTD implicit
method.
Expanding(C 1.8)for E, term we get:

ll, "(i+1/2j+1/2, k) -H: "fl+112j_-112, k)

At AY
E,"+l(i+112j, k)=E, n(i+112j, k) +
2s Ul
hIly"0+ 1/2j, k+ 1/2) fly" 0+ 1/2j, k-112)
-
(CI.9)
Hn+'(i+112J+112, k)-Il: n+'(i+112J-II2, k) (D
At AY
+'(i+1/2jk+1/2)-lln+'(1+1/2jk-1/2)
_y
Az

ADI-FDTD methodis in a similar form to (Cl. 9) but splitting it into two separateprocedureseachfor
half time-stepiteration. ProcedureI&2 are takenfrom RIIS terms as shownin (Cl. 9). This
successive
in the following terms for procedureI and2 of the ADI-FDTD method, repeatedhere from (3.1a)
results
and(3.4a)respectively.
 3hical illustration of im licit ADI-FDTD method 104

ýH "+1/2(i+
112i+112, k) - H.."+112(i+I 12j-112, k)
At AY
' (i + 112j, k) +
(i + 112j, k) = Eý',
2c H"(i+ 112jk + 1/2) - H"(i+ 112jk-112)
yy
AZ

(C1.1O)

Hz 112j + 1/2, k) - Hzn+II2 (i+ 1/2j-1/2, k)

ät AY
En+' (i. 112jk) =E n+112(i+112jk) + ý
2c Hyn+I(i+112jk+112)-Hy n+l(i+112jk-112)

AZ

(Cl.11)