1464

IEEE TRANSACTIONS ON MICROWAVETHEORYAND

TECHNIQUES,

VOL.

MTT-34, NO. 12, DECEMBER 1986

The Finite-DifferenceTime-Domain Method and its Application to Eigenvalue Problems

DOK HEE CHOI
AND WOLFGANG J. R. HOEFER, SENIOR MEMBER, IEEE

Abstract method time, While eigenvahre

This

paper describes domain Maxwells solutions

the application equations

of the finite-difference (3-D) in space and transform. method, the half

in the time
problems. and steady-state achieving time errors

to the solution

of three-dimensional are discretized as the TLM requires absence via Fourier

are then obtained (FD-TD) method include results, the

the scalar approach is limited to problems which lead to uncoupled modal solutions, i.e., TE and TM or LSE and LSM fields. The major this order storage technique reason for the large CPU memory resides in the basic 3-D TLM each electromagnetic 26 real memory network demand concept. of In each There-

the same accuracy and versatility and memory the TLM under identical

finite-difference-time-domain the CPU advantages boundary a microstrip over

less than of

simulation

conditions.

Other

to represent

component, parameters.

method

dielectric curves of

in the treatment

of 3-D inhomogeneous including are presented.

planar strnctnres, dispersion

3-D unit

cell requires

spaces, 12 for pulse a large number

such as microstrip.

Some numerical substrate,

and 14 for additional

on anisotropic

fore, each operation

on each node involves

I.

INTRODUCTION

T
of

HE TLM various

METHOD microwave

has been successfully circuit problems of the TLM

applied

to

of variables, requiring considerable computer CPU space and time. Furthermore, experience has shown that the number of iterations increases with the complexity of the structure of a finline under study. For example, the accurate analysis requires easily over 1000 iterations. Given these

for more than ten technique over

years. The special

advantages

other numerical methods are well illustrated by Johns and Beurle [1] in their original paper on the method. Since then, nique the several improvements by various solution authors and have been made to this techin order to enhance the accuracy CPU time and memory to the TLM method

massive requirements, numerical technique TLM result, ods. The approach a new

we have searched for an alternative that possesses the advantages of the resources. As a on both the and TLM methis proposed (FD-TD) was first based

but needs fewer computer algorithm

economize

finite-difference-time-domain FD-TD method

space [2][5].

Mariki

[7] has extended

analyze anisotropic media, and Saguet and Pic [4] as well as A1-Mukhtar and Sitch [5] have employed a graded mesh to make the algorithm faster and more efficient. Although the graded mesh algorithm far more reduces memory than space requirements, it demands iterations

formulated

by Yee [6],

and has been applied extensively to scattering and coupling problems with open boundaries [9][15], i.e., to the solution between magnetic the TLM of deterministic this method problems. We noted the similarity method, which has Since of the lines, solution of the electrodomain. and the TLM problems is based

the original method for equal frequency resolution [5]. This is especially obvious in the case of three-dimensional (3-D) simulations where the grade ratio N requires an additional N 2 1 iterations. These requirements of large computer resources may critically limit the applicability of the TLM technique. Thus, Saguet [16] has proposed a simplified node which reduces the number of variables to be processed and stored at each node by one third. However, this modification increases the velocity error. A further reduction in computational instead potential expenditure has been proposed of the original vector solution, solution using a scalar by the authors [8]; we obtain a scalar However,

been widely

used in the numerical method on

eigenvalue

in the time

the computation

impulse response of a large mesh of transmission much unwanted information is usually generated.

We have therefore developed a novel procedure which increases the numerical efficiency of the time-domain approach without sacrificing its advantages. The method differs from the classical FDTD method in the assignment of initial field values and the application of the Fourier transform to the time-domain solution. In the foIlowing, we will describe this method problems. and its application to some typical microwave II.

3-D network.

YEES ALGORITHM have [6] solved to been solve expressed in 3-D and other finite-dif(2-D) scattering workers

Manuscript received March 22, 1986; revised July 8, 1986. This work was supported in part by the National Science and Engineering Research Council of Canada. The authors are with the Department of Electrical Engineering, University of Ottawa, Ottawa, Ontario, Canada KIN 6N5. IEEE Log number 8610560.

Maxwells ference wave problems form scattering

equations by Yee

two-dimensional

problems. been

Subsequently, by Taflove

have

0018-9480/86/1200-1464$01.00

01986

IEEE

CHOI

AND

HOEFER:

FDTD

METHOD

AND

EIGENVALUE

PROBLEMS

1465

[9]-[15].

We will

adopt further

Yees original equations. system, to include

algorithm Also, the

for

the ex-

three-dimensional tend the concept In bolic Maxwells a rectangular

Maxwells

we will media.

anisotropic

coordinate

source-free hyperE*

equations

can be written

as first-order

equations d~/dt dH/dt = IAXI d~/dx = IBXI d~/dx + IAYI d~/dy + IBYI d~/dy + IA,l dH/dz
(l-l,j, kl

+ IB,I d~/dz

(1)
h,j,k) ,

where ~=

(EX, EY, E=)f, ~=

(HX, HY, H=)f, and

0 1 o
Ilt,l = ,;;

6;:

01
<;:
()

1-,;.
.

0]

Here,

CXX, CYY, and 6== are the diagonal tensor. All CXX, CYY, and C== by pXx,

[1
000
Fn(i, increment.

o
()

Fig. 1.

The position of the field components m Yees mesh.

E;+l(i+l/2,

j,k) =EJ(i+l\2, j,k)+s/fXX(i+l/2, +1/2, j+l/2, j1/2, j, k -1/2) j, k +1/2)] (2d) k) k) j,k)

elements pYY, and

of the by . H~+l/2(i [ H:+l/2(i+l/2, + H~+l/2(i+l/2, HJ+l/2(i+l/2, E~+l(i, j+l/2, =E~(i, k) j+l/2, k)+s/cYY(i, j+l/2, j+l/2, k +1/2) k-l/2) j+l/2, j+l/2, k) k)] (2e) j+l/2, k) pZZ

permittivity replacing

BX, BY, and B= are expressed

components, which are the diagonal elements of the permeability tensor. Then each of these scalar equations can be expressed in finite-difference form. Following Yees nomenclature, any function of space and time is discretized
1

j,k~=F(i

Ax, jAy,

kAz,

nAt) and At is of ~ ~

where Ax = A y = Az = Al is the space increment the time and ~ By positioning and ~ on the mesh as depicted at alternate nents of Maxwells HJ+/2(i, j+l/2, equations. k+l/2) j+l/2, j+l/2, j+l/2, j+l, k)+ k+l/2)] k+l/2)

the components

. [H+112(i, -~:+112n(i,

in Fig. 1 and evaluating

half time steps, we obtain

the compo-

+ HZ+l/2(il/2, --H:+lz2(i+l/2, EY+l(i, j,k+l/2) =E~(i, j,k+l/2)+s/tZZ(i,

=H;-i2(i, +s/PXX(i, E;(i, -E:(i, H;+112(i+l/2, = HJ-/2(i +s/PYY(i E~(i,

k+l/2)[E~(i, E:(i, j,k+l/2)

j+l/2,

k+l)

j,k+l/2) j,k+l/2)

o H~+l/2(i+l/2, [ (2a) - HJ+l/2(i + H;+l/2(i, - HJ+lz2(i, +1, j,k+l/2) j,k) (2b) where the stability of light. such that In -1/2, j-1/2, j+l/2j

j, k +1/2) k +1/2) k +1/2)] (2f)

j, k +1/2) +1/2, +1/2, j, k +1/2) j,k+l/2)[EJ(i

factor s = c At /Al, E and impedance

and c is the velocity H are normalized of space is unity.

j,k+l/2)+EJ(i+l/2, j, k +1)] j+l/2, k) 2k)k) +1/2, j+l/2, j,k)+Ef(i, j+l/2, k)]

these expressions, for stability

the characteristic

E~(i+l/2, H:+l/2(i+l/2, =H:-/2(i+2j+l+ +s/Vzz(i E~(i+l/2, E~(i+l,

The condition

of (1) in free space is [10] (3)

s<l/JT.

k)[E~(i+l/2, j+l/2, k)

j+l,

k)

III.

BOUNDARY

CONDITIONS

(2C)

So far, a space-time mesh has been introduced and Maxwells equations have been replaced by a system of finite-difference equations. Difficulties arise when the do-

1466

IEEE TRANSACTIONS ON MICROWAVETHEORYAND

TECHNIQUES,

VOL.

MTF34,

NO, 12, DECEMBER

1986

main in which the field must be computed is unbounded. Since no computer can store an unlimited amount of data, a special which so-called conditions technique absorbing must be used to limit computation or soft boundary the domain in the numerical is made, by introducing conditions. These by Taylor et al. [9], who

have been described

use a simple extrapolation Brodwin [10], who simulate averaging angles of Lee [12] from the boundary radiation Although

method, and by Taflove and the outgoing waves and use an

process in an attempt to account for all possible propagation of the outgoing waves. Kunz and use the radiation condition at a large distance center of the scatterer to obtain an absorbing condition. Mur [13] employs a second-order condition to improve the accuracy of the results. these schemes have been used in scattering boundary so far. probcondi-

lems in the past, no ideal reflection-free tion has been proposed

However, in the formulation of eigenvalue problems, only hard boundaries usually represented by conducting wallsoccur. At these boundaries, the tangential electric and the normal magnetic field components are mainwall
JE,I

(a)

tained

at zero. For example,

on a perfectly

conducting

in the plane

i = 1 (see Fig. 1) E;(l, j+l/2, j,k+l/2) j+l/2, k) =0 =0 k+l/2)=o two, but its
G05 067
AI!IA 0.11 0.15

for all n

E;(l, { H;(l,

(the third condition is implicit in the previous implementation reduces numerical errors). IV.
In plane most wave scattering is injected in eigenvalue vector coordinates these cases, should to ringifig to start INITIAL problems, VALUES an impulsive

(b) or sinusoidal Fig. 2. (a) Field distribution of an empty rectangular resonator obtained with the FDTD method. (b) Output spectrum obtained with both the FDTD and TLM methods under identicaf conditions

at the beginning problems, not

of the computation. the direction and of the on

However, propagation the space In that size

is usually and the in the

known

depends that

on the logical radial

eigenvalue choice direction. with is

is to be isotropic spatial to the due the to space the first expeapis is in and

medium

is replaced

with a 3-D uniform

mesh. To solve the

found. pulse pulse mesh lattice.

an

propagates not and way

The respect errors

envelope

be wide

enough

accumulate

numerical

overshoot A better field and rienced proximate therefore field equivalent impulse which tral-domain V. The pattern

as it propagates the computation mode

through

system of equations (2) in this mesh, initial values must be assigned first as described in the previous section. For a rectangular-type resonator, a simple sinusoidal function is an appropriate choice for the dominant mode eigenvalue. As n increases, the discrete time functions for ~ and ~ fields evolve towards the steady state which is characterisIn this way, the tic of the desired mode in the geometry.

is to estimate in the structure The of the idea

distribution then choose researcher modal able for

of the desired the field initial usually has

value

accordingly. a good in

evolution of all six field components is obtained simultaneously at discrete time points n At. The final steady-state field distribution may be calculated by taking the time average of the time-domain solution at each mesh point. Thus, the steady-state solution is given by F(iO, jO, kO) = ~lF(iO,
n

distributions an educated of a TLM basis

a structw-e This with

to make

guess of the steady-state procedure a weighted in the specto the way mesh similar functions

a particular

eigenmode.

to the excitation distribution,

jO, kO) l/N.

(4)

and is somewhat appropriate

one chooses

approach. OUTLINE OF THE NUMERICAL of the FDTD method PROCEDURE will be discussed

This simple procedure is another advantage quires mode. In two simulations

to obtain the final field distribution over the TLM method, which refor finding the fields of a given is a can

application

eigenvalue

problems, function,

the steady-state from which

solution

using a rectangular

resonator

as an example.

A continuous

time-harmonic

the eigenvalues

CHO1

AND

HOEFER:

FDTDMETHOD

AND

EIGENVALUE

PROBLEMS

1467

a=l b=

2dl 6dI

a=20mm b.10m c=15mm d=4mm S=lmn E r = 2.22

(a)

c= 8dl S= 2dl E,=3.75

l------=
(b) a=20dl b= 6dl
C=

Fig. 4.

Finline cavity

8dl

a =5dl ha----i c b =4dl

C =3dl

TABLE II
RESONANT FINLINE FREQUENCIES CAVITY IN FIG. OF THE UNILATERAL 4, OBTAINED WITH

c,=

16

VARIOUS

METHODS

(c) i

Fig. 3.

Three-dimensional

inhomogeneous resonators ahalyzed in this study.

GHz

TABLE I
COMPARISON FREQUENCIES METHODS OF RESULTS FOR THE NORMALIZED WITH THE TLM RESONANT AND FD-TD FOR THE

A1/A

OBTAINED

UNDER

IDENTICAL

SIMULATION

CONDITIONS

GEOMETRIES IN FIG. 3
Fig.3 Mode .s/a 1/6 TRM 0.05220 TLM (CPU time) FD-TD (CPU time) -.1 . . .. 0.0516 (144) I 0.0517 (51)

14E...H==:lIIl{~::til-oo{?~?!5)d)d
b) Hybrid ! 0.0278 0.0405 (357) 0.0278 (1 17) c) Hybrid (357) 0.0405 (1 17)

...

~p
0.0 0.1

0.2

0.3

0.4

Fig. 5.

Dispersion

be extracted method S(~) Both

by discrete Fourier

transform,

as in the TLM the CPU time

diagram of unilateral finline with the cross-sectional geometry given in Fig. 4.

=~F(iO, n factor

jO, kO)exp(j2rsnj). s and the number of iterations

(5) n

spent

by

the

equivalent

TLM

program

the stability affect

under identical conditions, including the initial excitation distribution. Furthermore, while the TLM procedure requires 22 real memory stores per 3-D node in an isotropic dielectric, the FDTD method requires only seven real memcmy stores per node. Fig. 3(a), (b), and (c) shows structures for which solutions have been computed with this method. The dominant resonant frequencies of these structures are given and compared with the TLM solution in Talble I. The inhomogeneous rectangular cavity of Fig. 3

strongly

the spectral response.

In order to test this algorithm for validity, it has been applied to a simple rectangular cavity with sides 12 Al x 6 Al x 8 Al. We have assumed a dominant TEIOI mode in the initial value assignment. The time-domain solution is given in Fig. the numerical with 2(a). Discontinuous field figures are due to error caused by the finite-difference form of and TLM methods responses obtained under identical con-

(2). Fig. 2(b) compares the frequency

the FDTD ditions. hundred

(b) and (c) illustrates the capability of this algorithm to solve hybrid field problems. The number of nodes chosen in each problem solution. Furthermore, of a finline is the same as that employed we have computed in the TLM frequency by Saguet

The responses are not distinguishable. Five iterations have been used. The peak of the solu-

the resonant

tion is located at A1/X = 0.0750 in both methods. The exact analytical solution is 0.07511. Even though a small numb& of meshes is used in this algorithm when compared with the scattering problems in [9]-[13], it is noted that the accuracy in the solution of the eigenvalue problems is better than that of the scattering order of magnitude. VI.
NUMERICAL

resonator

(Fig. 4.) treated previously

[16]. Results are compared in Table II, which includes a value obtained with the spectral-domain method by Saguet. Fig. 5 shows the dispersion the same cross section, characteristics as obtained of a finline with our with method.

problems

by one

Resulks calculated with our spectral-domain program are also shown in the same figure. In order to compare convergence of both time-domain methods, solutions obtained after every fifth iteration are drawn in Fig. 6. The results show virtually identical convergence. To show the versatility of this method, the characteristics c~f a microstrip resonator on anisotropic substrate are

RESULTS

We have applied this technique to most of the examples described in the TLM literature and obtained practically identical results. The method requires less than one-half of

1468

IEEE

TRANSACTIONS

ON MICROWAVE

THEORY

AND

TECHNIQUES,

VOL.

MTT-34,

NO

12,

DECEMBER

1986

F! GHz) 11.5 +

t
---- TLM

FD-TD

los~mo 2CQ

400

6(X)

8~

Number

of

iterations

Fig. 6. Stability and convergence of the TLM and FDTD methods as a function of the number of Iterations. The solution of the finline problem in Fig. 4 N represented.

t-~+
= 9.4 xx = 22 = 11.6 YY Fig. 7. Microstnp

h/Al
w/Al

=
=

3 3 0.5 ml

Al

cavity on anisotropic substrate,

TABLE III
DOMINANT BY BOTH RESONANT THE TLM AND FREQUENCIES THE OBTAINED

FDTD

METHODS

order to illustrate this process, frequency spectra obtained with the TLM method for the two cases where c is equal to 10 Al are shown in Fig. 8. The solution obtained with the FDTD method is also drawn in the same figure. As expected, the latter solution is located exactly between the two values obtained with the TLM method. This clearly illustrates the accuracy and convenience of the FDTD method in such situations. Fig. 9 shows the dispersion characteristics of the microstrip which has the same cross section as that in Fig. 7. Again, both methods give very similar results except at higher frequencies, where the discretization errors associated with both methods become more pronounced, and their differences are more visibIe.

computed in the last example, shown in Fig. 7. Several different resonant frequencies obtained by changing the length c are tabulated in Table III. It is well known [17] that the TLM simulation of 3-D inhomogeneous planar structures involves dielectric interface ambiguity. The best way to resolve this error is to employ two dielectric substrate thicknesses differing by one Al. In our case, 3 Al and 4 Al are used. The final result is obtained by taking the average of the solutions obtained for these two values. In

VII.
The 3-D proposed new

CONCLUSIONS of the FD-TD practically method to

application gives

eigenvalue

problems

the same

results

CHOI

AND

HOEFER:

FDTD

METHOD

AND

EIGENVALUE

PROBLEMS

1469

Fig. 8. Frequency spectra for the microstrip problem in Fig. 7, showing the effect of dielectric interface error in the TI.M solution.

GHz

[4]

16

[5]

12

[6]

[7]

8
[8]

v
,, .

Tech., vol. MTT-28, pp. 743-746, July 1980. P. ISaguet and E. Pie, Le maillage rectangulaire et le changement de maille clans la methode TLM en deux dimensions, Electron. Lett., vol. 17, no. 7, pp. 277-279, Apr. 1981. D. A. A1-Mukhtar and J. E. Sitch, Transmission line matrix method with irregularly graded space, Proc. Inst. Elec. Eng., vol. 128, no. 6, pp. 299-305, Dec. 1981. K. S. Yee, Numerical solution of initial boundary value problems involving Maxwells equations in isotropic media, IEEE Trans. Antennas Propagat., vol. AP-14, pp. 302-307, May 1966. G. E. Mariki, Analysis of microstnp lines on inhomogeneous anisotropic substrates by the TLM numerical technique) Ph.D. thesis, Univ. California, Los Angeles, CA, June 1978. D. Choi and W. J. R. Hoefer, The simulation of three dimensional wave propagation by a scalar TLM model, IEEE MTT-S Synrp. Di&, 1984, pp. 70-71, C. D. Taylor, D. H. Lam, and T. H. Shumpert, Electromagnetic puke scattering in time varying inhomogeneous media, IEEE Trizns. Antennas Propagat., vol. AP-17, pp. 586-589, Sept. 1969.

,*

.
--:

.. :

isotropic

substrate

E,=9.4

TLM \ , anisotropic : FDTD em= e,?

[9]
medium

9.4, EV=l I.6 1.0 P

o. Fig. 9.

0.2

0.4

0.6

0.8

Dispersion diagram of the anisotropic microstrip in Fig. 7.

as the TLM However, are field solution. typically

method the overall

under identical CPU half those

simulation needed in

conditions. the TLM which Further mesh

time and storage requirements reside in the ease with in planar structures.

less than

Other

advantages error

distributions

can be computed,

and in the elimination losses, variable procedure.

of dielectric

boundary

efforts are being made to implement size, and nonlinearity in the FDTD
REFERENCES

[1] P. B. Johns and R. L. Beurle, Numerical

[2]

[3]

solution of two-dimensionaf scattering problems using a transmission line matrix, Proc. Inst. Elec. Eng., vol. 118, no. 9, pp. 12031208, Sept. 1971. S. Akhtarzad and P. B. Johns, Solution of Maxwelfs equations in three space dimensions and time by the TLM method of numerical anafysis~ Proc. Inst. Elec. Eng., vol. 122, no. 12, pp. 1344-1348, Dec. 1975. Y. C. Shih and W. J. R. Hoefer, The accuracy of TLM analysis of finned rectangular waveguides~ IEEE Trans. Microwave Theory

A. Taflove and M. E. Brodwin, Numerical solution of steady state electromagnetic scattering problems using the time dependent Maxwells equations, IEEE Trans. Microwave Theory Tech., vol. Ml_T-23, pp. 623-630, Aug. 1975. [11] R. Holland, Threde: A free-field EMP coupling and scattering code, IEEE Trans. Nucl. Sci., vol. NS-24, pp. 24162421, Dec. 1977. K. S. Kunz and K. M. Lee, A three-dimensional finite-difference 1121 solution of the extemaf response of an aircraft to a complex transient EM environment: Part 1The method and its implementation, IEEE Trurrs. Electromagn. Compat., vol. EMC-20, pp. ~1~1 328-33, May 1978. G. Mur, Absorbing conditions for the finite-difference approximation of the time-domain electromagnetic-field equations, IEEE Trans. E[ectromagn. Compat., vol. EMC-23, Nov. 1981. and K. R. Umashankar, A hybrid moment [14] A. Taflove method/finite difference time domain approach to electromagnetic coupling and aperture penetration into complex geometries, IEEE Trans. Antennas Propagat., vol. AP,30, pp. 617-627, July 1982. [15] K. R. Umashankar and A. Taflove, A novel method to analyze electromagnetic scattering of complex objects, IEEE Trans. Electromagn. Compat., vol. EMC-24, pp. 397405, Nov. 1982. [16] P. Saguet, Le maillage paralle14pip4dique et le changement de maille clans la mdthode TLM en trois dimensions, Electron. .Mt. vol. 20, no. 5, pp. 22224, Mar. 15, 1984. [17] W. J. R. Hoefer, The transmission-line matrix methodTheory IEEE Trans. Microwave Theory Tech., VOL and applications, MT1 33, no. 10, pp. 882-893, Oct. 1985. [10]

1470

IEEE

TRANSACTIONS

ON MICROWAVE

THEORY

AND

TECHNIQUES,

VOL.

MTT-34,

NO.

12, DECEMBER 1986

Dok Hee Choi was born in Seoul, Korea, in 1949

and graduated from Sogang University, Seoul, in 1973. From 1973 to 1977, he served as a radar officer in the Korean air force. He received the M. SC.degree in electrical engineering from Concordia University, Montreaf, in 1981 and the Ph.D. from the University of Ottawa, Ottawa, in 1986. From 1977 to 1978, he worked for the Korean Institute of Science and Technology as a microwave engineer involved in quality control in the manufacture of radar systems. F~om 1981 to 1984, he was employed by Canadian Astronautics Ltd., as a microwave engineer specializing in electronic warfare system design and millimeter-wave VCO development. He joined the Spacecraft Division of Telesat Canada, Ottawa, in 1986, and is currently developing numerical models to simulate communications satellite antennae. He is also involved in satellite transponder front-end design. His special interests include millimeter-wave component design and the application of numericaf techniques in applied electromagnetic.

J. R. Hoefer (M71-SM78) was born in Urmitz/Rhein, Germany, on February 6, 1941. He received the diploma in electrical engineering from the Technische Hochschule Aachen, West Germany, in 1965, and the D.-Ing. degree from the University of Grenoble, France, in 1968. After one year of teaching and research at the Institut National Polytechrrique de Grenoble, he joined the Department of Electrical Engineering, the University of Ottawaj Canada, where he is currently a professor. HIS sabbatical activities included six months with the space division of the AEG-Telefunken in Backnang, West Germany, in 1976, six months with the Institut National Polytechnique de Grenoble, France, in 1977, and one year with the Space Electronics Directorate, the Communications Research Centre in Ottawa, Canada, during 1984/85. His research interests include microwave and millimeter-wave circuit design, numerical techniques for solving electromagnetic field problems, and microwave measurement techniques. Dr. Hoefer is a Registered Professional Engineer in the Province of Ontario. Wolfgang