You are on page 1of 8

110 IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 44, NO.

1, JANUARY 1996

tion
teraction
Jean-Pierre Berenger

Abstract-The perfectly matched layer, PML, is a new tech- Vacuum-Layer Interface Perfect Conductor
nique developed for the simulation of free space with the finite-
difference time-domain (FDTD) method. This paper deals with
the application of this technique to the solution of wave-structure
interaction problems. It is shown that the absorbing layer can be
set very close to the structure, provided some specifications are
satisfied, with the consequence that the computational require-
ments of the PML technique are far shorter than those of the
other methods of free-space simulation.
A EY
I. INTRQDUCTIQN
0 Hz
N A PREVIOUS paper [l], we have described the PML
technique, a new technique of free-space simulation to be
used with the finite-difference time-domain method [2], [3]. c3 I E;
The numerical experiments reported in [ 11 have shown that this
technique is very efficient for absorbing the electromagnetic Y
waves and then for solving unbounded problems. Among such
problems, the wave-structure interactions are the most usual X
applications of the finite-difference method, in electromagnetic L = 0 112 1 3J2 2 512 3
compatibility and radar cross-section computations. The aim Fig. 1. k g h t side of a domain surrounded by a PML layer
of this paper is to get a detailed insight into the implementation
of the PML technique when dealing with such important random, either the results may be erroneous or the computa-
applications. tional requirements may be needlessly great. For this reason
For solving interaction problems with the finite-difference we have analyzed the numerical reflection produced by PML
method, various techniques have been used in the past to layers. The parameters goveming this reflection have been
absorb the outgoing waves, such as the matched layer [4], [5] found, allowing an optimum PML layer to be specified in such
or the one-way approximation of the wave equation [6]-[8]. a way that accurate solutions can be obtained with computa-
To obtain satisfactory solutions, it is well known that these tional requirements as small as possible. This work has been
absorbing boundaries must be set some distance from the scat- mainly based on the observation and analysis of numerical
tering structure with the result that most of the computational experiments, in both 2-D and 3-D cases, solving 3-D problems
domain is a surrounding vacuum. In [ 11, wave-structure inter- being our final objective. For simplicity and compactness, only
action computations have been reported for two-dimensional the 2-D case is considered in the present paper, nevertheless
problems. Correct solutions were computed with the PML some comments concerning the 3-D case are provided.
layer set only two cells from the scatterer, showing that the
computational domain may drastically reduce while using the 11. THE PML TECHNIQUE
PML technique. Unfortunately, due to the presence of some The PML technique consists in surrounding a computational
numerical reflection, such an achievement cannot be obtained domain with the absorbing layer described in [ 11. In this layer,
using any PML layer. As observed in [l] and in other compu- and for the TE case, the magnetic component H , is split into
tations, when the PML layer is set very close to the scatterer, two subcomponents H,, and Ifzy(Fig. 11, and the following
the results are exact only if the thickness of the layer equals equations replace the Maxwell equations
at least a certain value which appears to be directly related to
the problem to be solved. The consequence is that the PML
technique is not an efficient nor a reliable method of free-space
simulation, as long as the relation between layer thickness and
problem of concern is not clarified. If the layer is chosen at

Manuscript received December 20, 1994; revised September 18, 1995.


The author is with the Centre d’Analyse de DBfense, 16 bis, Avenue Prieur
de la CBte d’Qr, 94114 Arcueil, France.
Publisher Item Identifier S 0018-926X(96)00626-6.

0018-926X/96$05.00 0 1996 IEEE


BERENGER PERFECTLY MATCHED LAYER FOR THE FDTD SOLUTION OF WAVE-STRUCTURE INTERACTION PROBLEMS 111

where the electric and magnetic conductivities satisfy the usual


matching impedance condition ( a / & ,= a*/po). With such a
layer and adequate sets of conductivities, theoretically there
is no reflection at the vacuum-layer interface. Nevertheless,
the layer is not without reflection, since the outgoing waves
are reflected by the perfectly conducting conditions set on its
outer boundary and they can return into the vacuum. So for
a plane wave, an apparent reflection is defined [1] which is Y
a function of two parameters, the layer thickness S and the
conductivity profile a ( p ) , where p is the distance from the
interface (Fig. 1). This conductivity a is either ax in a layer X At = 0.1 ns
to x, as in Fig' Or in a layer 'Orma' to 9' with Fig. 2. Computational domain for wave-structure interaction problems.
such notations the reflection factor at an angle of incidence 0 is
For linear or parabolic profiles, inserting (4) into (9) and taking
R(0) = [R(O)]"""e (2) into account (5),we obtain the following FDTD conductivities,
with at index L = 0 , 1 / 2 , 1 , 3 / 2 , . . . (Fig. 1)

In principle, one can obtain reflection factors as short as g,(L > 0) =an(0)[(2L + l),+l - (2L - l ) n + l ] . (11)
required by increasing the thickness S or the conductivity
a ( p ) , or both. In practical computations, a small amount of Similarly, for the geometric progression (6), we have
numerical reflection occurs which depends on 6 and a ( p ) .
The experiments reported in [l] were performed with linear
and parabolic conductivities varying from zero at the interface
to am at the outer side of the layer. Such profiles and their
reflection R(0) are the following ( n = 1 for linear, n = 2
for parabolic)
OF RESULTS
111. ANALYSIS
a(p) =am(!)n ( n = 1 or 2 ) (4) COMPUTED BY THE PML TECHNIQUE
We will consider the two-dimensional problem of Fig. 2.
~ ( 0=) e-(2/(n+'))(om'/Eoc) (5)
The scatterer is a perfectly conducting plate of zero thickness,
These PML layers defined by their number of FDTD cells N , a having an infinite width and a length of 20 FDTD cells. The
letter L for linear or P for parabolic, and the normal reflection computational domain is bounded by a PML layer located Nd
R(0) expressed as a percentage, were denoted respectively as cells from the plate, and the incident wave is produced by a
PML(N-L-R(0)) and PML(N-P-R(0)). In the present paper, Huyghens surface one cell from it. Two incident waves will
we will also consider layers with conductivities increasing be used, the following unit step and Gaussian pulse
geometrically. Denoting by Ax the spacial increment of the
E ( t ) =Einc(l - e P t ) (time in ns) (14)
FDTD mesh and 00 the conductivity in the vacuum-layer
interface, such profiles will be ~ ( t= )~ ~ ~ ~ e - ( ~ (time
- 3 )in~ns). (15)

a ( p ) = ao(gl/AX)P (6) We will focus on the normal electric field at the end of the plate
(A in Fig. 2), and the magnetic field at its center (B in Fig. 2).
so that the conductivity increases by the factor g from one The fields are maximum at these points and representative
cell to the next. Inserting (6) into (3), for a N-cell layer the of what is observed on the whole surface of the plate. For
normal reflection is then each wave, a reference solution has been computed within a
domain of 250 by 250 cells so that the boundaries were about
R(0) = ,-(2/&oc)((gN-1)/ 1ng)goA". (7)
five structure lengths from the plate.
Such layers will be denoted as PML(N-Gg-R(0)). Using (7) For distances layer-plate varying from two to twenty cells,
the conductivity a0 can be obtained from N , g, R(0) Fig. 3 shows the fields computed at points A and B, with
both incident waves. The layer used in this computation was a
four-cell layer having a linear conductivity and a 1% normal
reflection R(O), that is a PML(4-L-1) layer. It appears that
In all computations in this paper, the conductivities at the the fields computed with the Gaussian pulse (15) are exact
mesh points were implemented as the average value in the cell for any layer-plate distance. Solving this case does not require
around the index location. At index L (Fig. 1) any vacuum surrounding the scatterer. For the unit step (14),
1 YP(L)+AX/Z the current on the plate is correct, but the electric field is not.
.(U) du. (9) As the plate-layer separation decreases, it is correct during
a time of about 40 ns, and then differs from the reference.
112 IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 44, NO. 1, JANUARY 1996

REFERENCE
REFERENCE ,__..... ML(4-L-I) Nd = 20 CELLS
Nd = 20 CELLS ML(4-L-1) Nd = 10 CELLS
Nd = 10 CELLS -..... ML(4-L-1) N d = 5 CELLS
...... N d = 5 CELLS ......... ML(4-L-1) N d = 2 CELLS
4 ......... I( ONE-WAY Nd = 20 CELLS
__-- A ONE-WAY Nd = 10 CELLS
.e
. 3

t
0

.
r

c 2
3

GAUSSIAN PULSE 1

W
I 0
1 3 10 30 100 300 loo0 I
TIME (NS) 1 3 10 30 100 300 1000
TII
I uJIE (NS)

2.0
- REFERENCE Fig. 4. Electric field at point A computed using a four-cell classical matched
___ N d = 2 0 CELLS Layer set 2-20 cells from the plate and the second-order one-way technique
Nd = 10 CELLS
Nd= 5 CELLS set 10 and 20 cells from the plate.
10

I 8 , v , I
1 3 10 30 100 300 1003
TIME (NS)

Fig 3 . For the two incident waves, electric and magnetic fields at points A
and B computed using a four-cell linear PML layer set 2-20 cells from the
20-cell plate L I
1 3 10 30 100 300 I000 3000 10000
TIME (NS)
So, with the PML(4-L-1) layer, the PML technique fails to
solve the problem without vacuum around the scatterer. The Fig. 5. For the unit-step incident wave, electric field at point A, computed
plate-layer separation must be on the order of the plate length, using various linear and parabolic layers set two cells from the plate.
as with the classical matched layer [4], [5] or the one-way
approximation [6]-[7], whose results for the same plate and The results of Fig. 5 can be predicted by supposing that
incident waves are shown in Fig. 4. Nevertheless, comparing the parameter which governs the time of validity is the
Fig. 3 with Fig. 4, a large improvement is observed when conductivity a,(O) used in the first row of the PML layer.
using the PML technique, since with the other techniques the More precisely, we will see that this time increases inversely
results are only correct for less than 10 ns. We will show as on(0).Considering the first three layers of Fig. 5, as R(0)
below that the time of validity of the PML technique can be increases from 1% to lo%, (10) predicts that a,(O) decreases
increased, and we will determine what parameters of the layer by a factor of 2, and as R(0)decreases from 1% to 0.01% that
have to be adjusted to achieve such an objective. ~ ~ ( increases
0 ) by a factor of 2. Such variations are in good
Let us first consider Fig. 5 giving the electric field computed agreement with the ratios of the times of validity observed
using various linear layers set two cells from the plate. The in Fig. 5 (times on the order of 20, 40, 80 ns). Similarly,
first three layers show what occurs when varying the normal from PML(4-L-1) to PML(8-L-1), (10) predicts that an(0)is
reflection R(O), with a four-cell linear layer. Starting with reduced by a factor of 4, and from PML(4-L-1) to PML(4-P-1)
the PML(4-L-1) layer and reducing R(0) to 0.01% does not by a factor of 8. In both cases, that is in agreement with the
improve the results, on the contrary the time of validity is enhancement of the times of validity observed in Fig. 5, since
reduced to about 20 ns. Conversely, increasing R(0) to 10% with the PML(8-L-1) and PML(4-P-1) layers these times are
increases this time to 80 ns, but an error occurs in the first respectively about 150 and 300 ns. So, the FDTD conductivity
peak before 10 ns. This error is due to the reflection from the a,(O) governs the time of validity, reducing ~ ~ ( is0the ) way
perfectly conducting wall ending the domain (Fig. 1). Similar to increase this time. Below, we will state more precisely how
results can be observed with any 2-D or 3-D scatterer. In all these quantities depend on each other.
cases, R(0) on the order of 1% is sufficient to absorb the Let us suppose that the absorbing layer is a usual conductive
reflection from the outer conducting plane, and reducing R(0) medium of conductivity a,(O). The index of refraction of the
below this value reduces the time of validity. Thus, in practice, layer would be
R(0)on the order of 1%appears to be an optimum value. Such
a 1% reflection has been used in all computations in this paper.
The last three cases of Fig. 5 show that the time of validity
is enhanced by using a thicker layer, or a parabolic layer in The low frequencies (n >> I) would be reflected by the
place of a linear one. vacuum-layer interface while the high frequencies ( n = 1)
BERENGER: PERFECTLY MATCHED LAYER FOR THE FDTD SOLUTION OF WAVE-STRUCTURE INTERACTION PROBLEMS 113

...................
REFERENCE
......... PML(4-L-1)
PML(4-P-1)
PML(8-L-1)
PML(8-P-1)
*. ,

0 -
O Y
1
0.01 0.03 0.1 0.3 1 3 10 30 100 300 1 3 IO 30 100 300 1000 3000 10000
FREQUENCY (MHz) TIME (NS)

Fig. 6 . Frequency contents of the electric field at point A, computed using Fig. 7. For the unit-step incident wave, electric field at point A, computed
linear and parabolic layers set two cells from the plate. using various geometric layers set two cells from the plate.

would penetrate without reflection into the layer. The cutoff


frequency f, of such a selective reflection is

PML(8-03.16-1)
(17)
s-. *, A
In the time domain this cutoff frequency would allow the
reflection from the interface to be small, up to a cutoff time
on the order of
~~

v \
tc - 1 27r€ - 1 1
f, a,(O) 1.8 1O1O a,(O)’
(18) o -
1
p G = q ~ l ~ l
With the PML(4-L-1), PML(4-P-1), PML(8-L-1), and PML(8-
P-1) layers, we have computed the t, times given by (10) and
(18). They are shown in Fig. 5 by means of flags (for each
flag the vertical mark is set at the corresponding time). One
can see that the cutoff times t, are very close to the times of
validity of the computed results. In each case, t, is located just where
before the curve crosses the reference solution for the last time. 1 - ia,/aow
Fig. 6 confirms in the frequency domain what is observed in wz =
1- i a ; / p o w .
Fig. 5. The electric field normalized to the incident one was
obtained using the Gaussian pulse (15) and then performing Theoretically, the layer is matched so that w, = 1 and T = 0.
a Fourier transformation. Frequency f, from (17) is shown Numerically, in the first row the electric conductivity is a,(O)
by flags in each case. As expected, at frequencies higher than while the magnetic conductivity a:(0) is not involved in
f, the results are close to the reference, while at frequencies the numerical process. Supposing a :(O) equal to zero, the
lower than f, they are in error. reflection would be
So, we have observed that the first row of the PML layer
produces some reflection, as if this row were the first one of 1 - J l - ian(0)/&ow
’r =
a conductive medium. Such a reflection originates in the shift 1 + J1 - ia,(O)/eow.
Ax/2 between the computational points of the FDTD grid. For
instance, an outgoing wave (ICy, H,) traveling along IC and This is the reflection factor at an interface with the conductive
striking the right side of the domain (Fig. 1) is first absorbed medium of index (16), at normal incidence.
by the electric conductivity axin the first row of the layer, and A second set of numerical experiments is reported in Figs. 7
secondly by the magnetic conductivity a : in the second row. and 8 for conductivities increasing as geometrical progres-
So it is not surprising that the matching impedance condition, sions. Times t, and frequencies f c obtained from (12), (17),
which physically means an equal absorption for the electric (18), are shown for the first four layers. The results agree with
and magnetic fields, is not perfectly achieved by the numerical the reference during times in agreement with (18) predictions.
process. Another interpretation of the numerical reflection can The last layer denoted as PMLm(4-G4.64-1) is a PML(4-
be found considering the theoretical reflection factor at an G4.64-1) layer for which we have set o n ( 0 ) = 0,so that the
interface between a vacuum and a PML layer. At an interface first conductivity experienced by an outgoing wave is the mag-
normal to IC, from [l] this factor is netic conductivity oG(1/2). As expected, at low frequencies
the sign of the numerical reflection is reversed, the layer acting
as a magnetic conductive medium whose cutoff frequency is
given by (17) with 0:(1/2)/po in place of ~ , ( O ) / E O .
114 IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 44, NO. 1, JANUARY 1996

So the experiments of Figs. 5-8 have shown that in order to Hi Ei


solve an interaction problem during a time D,, the cutoff time v
(18) must be greater than D , and then the FDTD conductivity
a,(O) in the first row of the layer must satisfy the condition
27rq
a,(O) < -.
D,
For a given number of cells N and a normal reflection R(O),
Fig. 9. Scatterers used to evaluate g and 0 , and locations of the test points.
one might think that reducing ~ ~ ( can 0 be) obtained at will by
increasing n in profiles (4) or g in profiles (6). Unfortunately,
to remove the numerical reflection the absorbing layer must
as thin as possible if the conductivity grows according to the
satisfy a second constraint. Before the cutoff time, or above
the cutoff frequency, in Figs. 7 and 8 we can observe some maximum rate of increase allowed by the constraint (23). This
oscillations in the results, especially with the PML(4-ClO- 1) means that the ratio of successive conductivities does not vary
layer. Such experiments, and others to be shown later, show from the interface to the outer boundary. Such a profile of
that the geometric progression g cannot increase arbitrarily, Conductivity is a geometric progression like (6). Since with
over a certain value the results become invalid. Such a (6) a,(L +1/2)/an(L) equals from (13), the ratio g of
phenomenon is the result of numerical reflections produced (6) must be chosen so that
by the rows in which the FDTD computational points are set.
Starting from a frequency lower than f c , we first observe the
& =s. (24)
reflection from the first row, then the reflection from the second Thus, for solving interaction problems the geometric progres-
row for f greater than f,, then the reflection from the third row sion (6) is the optimum profile of conductivity. By using
for f greater than a frequency given by (17) with a;(1/2)/po another profile, for instance a parabolic one, the ratio of
in place of on(0)/ E O , and so on. This explains the oscillations successive conductivities would depend on the location in
in the numerical reflection. Depending on the frequency the the layer so that at some locations either it would not satisfy
reflection occurs either from an electric conductivity or from (23) or it would be needlessly small resulting in a needlessly
a magnetic conductivity, resulting in an alternating sign of thickness of the layer.
the error. This phenomenon is very visible with geometric To implement the optimum PML layer, two parameters
conductivities, but it is also present in Figs. 5 and 6. With must be set, g and N . We will show below some numerical
parabolic conductivities, the troghs before the times t, are due experiments that we have performed to empirically evaluate S
to the reflection from the second row of conductivity 0:(1/2). and therefore g. Let us rewrite (22) as
With linear conductivities, the troghs are shallower because the
ratios an(l/2)/a,(0) are smaller (from (lo), 4 with linear, 8
with parabolic).
In summary, for solving wave-structure interaction prob-
lems, the PML layer must satisfy two constraints. First, the where 0 is a parameter that we will evaluate. From (18) and
FDTD conductivity a,(O) is bounded by (22) where D, is the (25) we note that t , = OD,. Since (12) can be rewritten as
duration of the computation, secondly the ratio of successive
conductivities must be lower than a value that we will denote
by S. This last constraint can be written as
the number of cells N is, as a function of g and 0

The parameter S depends on the problem to be solved. We


will estimate it empirically in the next section.
We can note that the duration of the computation D , is
Iv. THE PML TECHNIQUE
FOR
embedded within a logarithmic term, so that N increases
WAVE-STRUCTURE
INTERACTIONS smoothly with D,. This is the result of using a conductivity
As shown in the previous section, the reflection R(0) increasing geometrically.
must be on the order of I%, and conditions (22) and (23) The parameters g and 0 have been evaluated by means of
must be satisfied. To achieve an optimum implementation, the some numerical experiments, with the scatterers of Fig. 9. In
profile of conductivity in the layer has been selected with all computations, the distance between the scatterer and the
the intention of satisfying these conditions with a number PML layer was two cells. Fig. 10 presents some results, for
of cells N as small as possible. This has been done by three plates of lengths 20, 100, and 300 cells. We have plotted
considering that the conductivity in the layer must grow from the two components of the electric field at the end of the plates
a small value in the interface (22) to a maximum value on and the magnetic field at their center (test points in Fig. 9).
the outer boundary, this maximum being mainly determined For each plate, the results were computed with layers having
by the required reflection R(0).From this, the PML layer is various geometric progressions g, but the same cutoff times
BERENGER. PERFECTLY MATCHED LAYER FOR THE FDTD SOLUTION OF WAVE-STRUCTURE INTERACTION PROBLEMS 115

4.0
_ _ _ PML(4-G12.16-1)
5 - _ _ _PML(5-G6.78-1) tc = 4000 ns 3.5
PO-CELL PLATE - _ _ _ _ PML(6-G4.66-1)
........ 3.0 !- - REFERENCE 7 I
4
2.5
I
z U PML(11-G2.13-1)
;3
b
. 2.0

1.5
........ PML(14-G1.77-1)

.
me
w
2 1.o

0.5
1
0.0 I//I I
L e I
0 1 3 10 30 100 300 1000 3000 10000
TIME (NS)
1 3 10 30 100 300 loo0 3OOO loo00
TIME (NS)

10

- 100-CELL PLATE

e
? 6
I
b
e 4
.
iii
W
2
PML(11-G2.13-1)

0 1 3 10 30 100 300 1000 3000 10000


TIME (NS)
1 3 10 30 100 300 loo0 3000 10000
TIME (NS)
Fig. 11. Electric field on a 100-cell square and a 100-cell cross.
18

16
300-CELLPLATE
14 those obtained on the 20-cell plate with a 2-cell plate-layer
z 12 separation. As a consequence, one may think that the same
P
1 10 g could be used for all plates, provided the ratios of plate-
layer distance to plate length are the same. But this method
would not yield an optimum implementation, to optimize the
W
PML(17-G1.76-1)
computational cost one should set the PML layer as close as
........ PML(20-G1.60-1)
possible to the scatterer. For this reason we have chosen to
2
set the PML layer two cells from the scatterer, despite the
1
3 10 30 100 300 1000 3000 10000 30000 inconvenience of having to evaluate g as a function of its
TIME (NS)
length.
Fig. 10. Electric and magnetic fields on 20-cell, 100-cell, and 300-cell plates Fig. 11 shows the electric field computed on other 100-cell
computed using various geometric progressions. scatterers, a square and a cross (Fig. 9). Comparing to the 100-
cell plate, for a given value of g, the oscillations are a little
smaller with the square and far smaller with the cross. In this
(18). It appears first that reducing the progression g reduces the last case, that is due to the presence of a certain amount of
oscillations of the electric field on the plates, and secondly that vacuum around the scatterer, since the computational domain
the progression g required to obtain a correct result depends was a square with 104 by 104 cells of vacuum. So, the plates
on the length of the considered plate. As the length increases, appear to yield the more severe constraint on the parameter
g must be reduced. For the three plates, 4.66, 2.13, and 1.76 g. Using g determined with plates insures correct results with
could be viewed as suitable values for g. other scatterers.
So, the value of g insuring a correct result depends on the Beside the above visual evaluation, we have used another
length of the scatterer. Actually, the parameter governing g is method for the determination of g. For plates of lengths 20,
the number of cells of the scatterer. As we have experienced, 30,40, 60, 100, 175, and 300 cells, we have performed a set of
multiplying the FDTD cell by 10 or 100 does not modify the computations withi layers chosen in such a way that the cutoff
results. For a given g the oscillations are unchanged. This is times (18) are equal to l O O O T s , where Ts is the resonance
due to the fact that the scatterer-layer distance is multiplied by period of the plates. The magnitude of the oscillations from
the same factor, and so the ratio of the scatterer-layer distance the numerical reflection have been measured during an interval
to the scatterer length is unchanged too. This suggests that this of time in which both the physical oscillations are damped
ratio is the actual parameter governing g . Some experiments and the average value of the computed result does not depart
have confirmed such an hypothesis. For instance with the from the reference. Denoting by (tl ,t z ) this interval, such
layer set 10 cells from the 100-cell plate, the results look like conditions may be written as tl >> TS and t 2 << t,. We used
116 IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 44, NO. 1, JANUARY 1996

0.8 TABLE I
I OF D,/Ts
NUMBEROF CELLSOF THE LAYERAS A FUNCTION

I
NUMERICAL ERROR 3 %
NUMERICAL ERROR 1 %

a = 1.7842 p=-l.a433 y=0.17749

1 Ns=100 g= 1.97 I 11.81 I 15.21 1 18.62 I


I Nr=300 E= 1.95 I 13.63 1 17.09 I 20.55 I
1 Ns= 300 g = 1.60 I 18.43 I 23.33 I 28.22 1
0.1 J
1.2 1.4 1.6 1.8 2.0 2.2 2.4 2.6
log (W for a desired time D, is given by (27) in which 0 has to be
Fig. 12. Geometric progression g as a function of the number of cells of the
set to 10.
plate, for two numerical errors. As a final remark, an interesting form of the number of
cells (27) can be found by expressing it as a function of
the resonance period Ts.For a scatterer of N s cells, with
tl = 10 Ts and t 2 = t,/10 so that the considered interval was 2NsAz = C T S ,from (27) we have
l 0 T s to 100Ts. Denoting as Emaxand E,,, the maximum
and minimum of the field at a given point during this interval, . (30)
we have defined an error as
Emax - E m m Table I gives the number of cells computed by (30) for
E r r o r = 100
Emax + Em,, ’
(28) scatterers of Iengths 20, 100, and 300 cells, the two values of
g from (29), R(0) = 0.01,O = 10, and three ratios D,/Ts.
For each plate, this error has been computed for a set of values For instance, computing the field on a 100-cell scatterer during
of g corresponding to the integer values of N in (26). By 100 periods requires a layer having 11 cells using g = 2.56.
interpolating between these values, we have then obtained the
values of g for given percentages of error. For the test point 2 V. CONCLUSION
of Fig. 9, the results of such a procedure are plotted in Fig. 12,
for 3% and 1% errors. These results are very well fitted by The primary objective of this work was the determination
the parabolas whose coefficients are included in the figure. of the conditions that must be satisfied by PML layers lo-
Therefore, the geometric progression g can be obtained for a cated very close to scattering structures, in order to remove
plate of length N s cells by the formula the numerical reflection produced by such layers or at least
make it negligible. By means of numerical experiments, we
have shown that this reflection is mainly governed by the
conductivity in the vacuum-layer interface and the rate of
where a , io, y are the coefficients in Fig. 12. increase of the conductivity in the layer. The reflection can be
The second parameter to be determined is the ratio 0 of remove by setting these parameters equal to adequate values
the cutoff time t , (18) to the computational duration D,. which depend on the duration of the computation and the size
Considering the 100-cell plate (Fig. lo), with all the layers of the scattering structure. From these observations, we have
whose cutoff times are 4000 ns, the results depart from the determined an optimum PML layer which allows a desired
reference just before 1000 ns. This suggests that 0 must be accuracy to be obtained with computational requirements as
on the order of 10 to insure an exact result during the time small as possible. In such a layer the conductivity increases
D,. The other two plates confirm such a value. For instance as a geometric progression whose ratio g can empirically be
with the 300-cell plate whose cutoff times are 20000 ns, the evaluated as a function of the size of the scatterer, and whose
results do not depart before 2000 ns. With the square and thickness N depends on two other empirical parameters, the
the cross of Fig. 1I , the conclusions are the same, the results theoretical normal reflection of the layer R(O),and a margin
are correct up to t,/10 and then they begin to depart from factor denoted by 0.The order of magnitude of R(O),g, 0
the reference. Similar results have been observed with 3-D insuring a good accuracy in the results when using a scatterer-
realistic scatterers. Thus, a suitable value for 0 is 10. PML layer separation of only two cells have been evaluated
In summary, for solving interaction problems the optimum for typical 2-D scatterers.
profile of conductivity is the geometric progression defined Although only 2-D experiments have been shown in this
by (6) and (8). For 2-D canonical scatterers, the empirical paper, the 3-D case has been treated along with the 2-D
parameters R(O),g,and 0 have been estimated by means of one, since implementing the PML method in 3-D computer
numerical experiments. For scatterers of lengths 20, 100, and codes was our final objective. In 3-D, the key parameters
300 cells, g equals to 4.66, 2.13, and 1.76 insures correct of the layer are also the conductivity in the interface and
results. Another estimation of g is given by formula (29) with the rate of increase of the conductivity, so that the best
the coefficients of Fig. 12. The optimum value of the normal profile of conductivity is also a geometrical progression with
reflection R(0) is about 0.01 (l%), and the number of cells empirical parameters on the same order of magnitude as those
of the layer that insures results without numerical reflection determined in the 2-D case.
BERENGER: PERFECTLY MATCHED LAYER FOR THE FDTD SOLUTION OF WAIIE-STRUCTURE INTERACTION PROBLEMS 117

In principle, for solving any given problem the heuristic [3] A. Taflove, “Review of the formulation and applications of the finite-
analysis of the numerical reflection presented in this paper difference time-domain method for numerical modeling of electromag-
netic wave interactions with arbitrary structures,” Wave Motion, vol. 10,
allows the PML layer to be easily selected as a function of pp. 547-582, Dec. 1988.
the size of the scattering structure and the duration of the [4] J. P. BCrenger, “Calcul de la diffraction 9 l’aide d’une mithode aux
computation. However, an inconvenience to this method may diffkrences finies,” Actes du colloque CEM, CNFRS-URSI, pp. ~ 2 1 x 2 6 ,
TrCgastel, France June 1983.
be seen in the fact that the optimum PML layer depends on [5] R. Holland and J. Williams, “Total-field versus scattered-field finite-
three empirical parameters, R(O),g, and 0. In practice, such a difference: A comparative assessment,” IEEE Nucl. Sci., vol. 30, no. 6,
dependance is not a great inconvenience, for two reasons. First, pp. 45834588, Dec. 1983.
[6] B. Engquist and A. Majda, “Absorbing boundary conditions for the
for a desired accuracy in the results the required empirical numerical simulation of waves,” Math. Comput., vol. 31, no. 139, pp.
parameters does not depend much on the geometry of the 629-651, July 1977.
[7] G. Mur, “Absorbing boundary conditions for the finite-difference
scatterer, in 3-D as in 2-D, so that using the values of R(O),g approximation of the time-domain electromagnetic field equations,”
and 0 determined in this paper insures very good results with IEEE Electromagn. Compat., vol. 23, no. 4, pp. 377-382, Nov.
most 2-D and 3-D structures. Second, the thickness N of the 1981.
[8] R. Higdon, “Numerical absorbing conditions for the wave equation,”
PML layer varies smoothly with the empirical parameters, Math. Comput., vol. 49, no. 179, pp. 65-90, July 1987.
due to the fact that these parameters are embedded within [9] J. P. Btrenger, “Etude prCalable 9 l’utilisation des differences finies
logarithmic terms. As a result a safety margin can be included pour le calcul de la propagation d’ondes radioClectriques. Mise au
point d’une nouvelle cavitC anCchoYde numkrique,” Note Technique
in the parameters, with a relatively small increase of the DGA/ETCA/CAD/352, Apr. 1992.
computational requirements. Thus, the presence of empirical [ 101 -, “A perfectly matched layer for free-space simulation in finite-
parameters is only a small inconvenience in comparison to that difference computer codes,” Annales des Tklkcommunicutions, vol. 5 1,
no. 1-2, Jan. 1996.
experienced when using the methods [4]-[8] with which the [ 1I] -, “Three-dimensional perfectly matched layer for the absorption
scatterer-boundary separation is an empirical parameter whose of electromagetic waves,” J. Comput. Phys., submitted for publication.
variations may result in a dramatic increase of the number
of cells in the computational domain. The consequence is
that the PML method is more easy to use than the methods
[4]-[8], with computational requirements substantially less. A J.-P. Berenger received the Maitrise de Physique
from UniversitC de Grenoble, in 1973, and
few examples of 3-D computations illustrating these comments the Dipl6me d’Ing6nieur from &ole SupCrienre
can be found in [9]-[ 111, and a more detailed discussion, along d’Optique de Paris, in 1975.
with numerical computations related to EMC applications will With the DCpartement h d e s ThCoriques of the
Centre d’Analyse de DCfense, from 1975-1984,
be the material of a future paper. his areas of interest were the propagation of
waves and the coupling problems related to the
REFERENCES nuclear electromagnetic pulse. During this period,
he contributed to popularizing the finite-difference
[l] J. P. BCrenger, “A perfectly matched layer for the absorption of time-domain method in France. In 1984, he moved
electromagnetic waves,” J. Comput. Phys., vol. 114, no. 2, pp. 185-200, to the DCpartement NuclCaire where he was involved in the development
Oct. 1994. of simulation software Since 1989, he has held a position as expert on the
[2] K. S. Yee, “Numerical solution of initial boundary value problems electromagnetic effects of nuclear events while pursuing works to improve
involving Maxwell’s equations in isotropic media,” ZEEE Antennas the methods and codes used to predict the propagation of waves in such
Propagat., vol. 14, no. 3, pp. 302-307, May 1966. disturbed conditions.

You might also like