ultilevel green’s function interpolation method: an ... multilevel...characteristics of the...
TRANSCRIPT
Forum for Electromagnetic Research Methods and Application Technologies (FERMAT)
1
Abstract—An efficient algorithm for fast field
evaluation, namely multilevel Green’s function
interpolation method (MLGFIM), has been successfully
applied to solve electromagnetic problems for both low
frequency and full-wave. Further refinements, including
new interpolation scheme, QR factorization, and FFT
technique, have been applied to accelerate this algorithm
and extend the problem scale that MLGFIM can solve. In
this paper, we present a review of different techniques in
MLGFIM for the fast evaluations of large-scale
electromagnetic problems. The applications of simulating
periodic structures and layered mediums have been
accelerated by MLGFIM as well. Some numerical
examples are given in this paper to demonstrate the
accuracy and efficiency of this algorithm.
Index Terms—Method of Moments (MoM), Green’s
Function Interpolation, Fast Solver, Parallel Computing,
FFT.
I. INTRODUCTION
Calculation of antenna radiation pattern, analysis of radar
cross section (RCS), and extraction of circuit parameters are
often performed using integral equation-based computational
techniques, e.g., the method of moment (MoM) [1]. However,
the computational requirements for MoM are very high, and
various fast numerical methods have been proposed to
accelerate the computation, for instance, multilevel fast
multipole algorithm (MLFMA) [2]-[4], pre-corrected fast
Fourier transform (FFT) [5],[6], adaptive integral method
(AIM) [7],[8], and sparse matrix canonical grid (SMCG)
[9],[10]. The first method uses multipole expansion to
approximate the far field interactions, while the later three
methods apply FFT for fast field evaluation. Although these
methods calculate the matrix-vector multiplication indirectly
and thus significantly reduce the computational complexity,
each of the methods has its own weaknesses. The FFT-based
methods are most suitable for densely-packed structures, in
which the unknowns are associated with most of the grid
points. The PFFT approach employs the bi-linear
interpolation method, and it is not suitable for analysis of
full-wave electromagnetic problems as bi-linear interpolation
is not accurate for the rapidly-varying kernels in full-wave
applications. On the other hand, the multipole-based
algorithms depend on the type of Green’s function used for
the specific problems and they become cumbersome when
dealing with structures embedded in a multilayered medium.
When the simulated structure changes, for instance, insert an
additional layer of dielectric substrate, the problem has to be
reformulated.
Some interpolation-based fast algorithms have the potential
to address many of the aforementioned weaknesses. In [11],
Brandt developed a multilevel interpolation approach, in
which a softened kernel and phase compensation is employed
when the kernel is oscillatory. Non-uniform grid (NG) [12],
[13] and multilevel NG [14],[15] algorithms, as kernel
independent algorithms, are proposed for fast field evaluation.
With the technique of phase and amplitude compensations,
NG-based algorithms sample the compensated field using a
non-uniform polar grid. The NG and multilevel NG methods
achieve computational complexities of O(N3/2) and O(NlogN),
respectively. Adaptive cross approximation (ACA) [16],[17]
employs interpolation approach directly on low-rank
sub-matrices denoting the far field interaction, and thus this
method is not difficult to implement. The H2 matrix-based
method is proposed in [18],[19]. It applies multilevel
interpolation with Lagrange interpolation function and Tartan
grid.
Another kernel-independent interpolation algorithm, viz.
multilevel Green’s function interpolation method (MLGFIM),
has been developed to solve quasi-static [20]-[22] and
full-wave electromagnetic problems [23],[24]. The
application of an octa-tree structure and non-requisite of a
softened kernel distinguish the MLGFIM with the algorithm
in [11]. For full-wave problems, instead of using Lagrange
Multilevel Green’s Function Interpolation
Method: An Interpolation-based Solver for
Efficient Analysis of Large-scale
Electromagnetic Problems
Peng Zhao(1,2), Da Qing Liu(2) and Chi Hou Chan(2)
(1) Electronics and Information College, Hangzhou Dianzi University, Hangzhou, China(2) State Key Laboratory of Millimeter Waves, City University of Hong Kong, Hong Kong
SAR, China
(Email: [email protected])
*This use of this work is restricted solely for academic purposes. The author of this work owns the copyright and no reproduction in any form is permitted without written permission by the author.*
Forum for Electromagnetic Research Methods and Application Technologies (FERMAT)
2
interpolation as in the quasi-static problems, radial basis
function (RBF) interpolation is employed. Lagrange
interpolation method can not be directly applied for the
approximation of the full-wave Green’s function, unless
phase compensation technique is incorporated [25]. The RBF
interpolation has excellent approximation properties [26], and
we find that if the object is not extremely large, the full-wave
Green’s function can be interpolated without the
compensation technique, which is different from the
NG-based algorithm. Although ACA is flexible for solving
problems with a complex medium, it is difficult to enhance
the efficiency by using the translational invariance
characteristics of the Green’s functions in some problems, e.g.
the multilayered planar problem. In addition, the multilevel
structure of MLGFIM also differs from ACA. Compared with
the H2 matrix-based method, the applications of RBF
interpolation and staggered Tartan grid (STG) make the
number of interpolation points in MLGFIM increases more
slowly. This is because the interpolation points at the coarser
levels do not coincide with those at the finer levels in the
MLGFIM, viz., the volumetric density of interpolation point
at the coarser levels is lower than that at the finer levels.
Moreover, some novel techniques used in MLGFIM, e.g. QR
factorization [27] and hybrid quasi-2D/3D multilevel
partitioning approach [24], distinguish itself from other
interpolation-based methods. In addition, the applications of
peer-level and lower-to-upper-level interpolations make the
MLGFIM very efficient for the analysis of large-scale
problems, and the computational complexity of 3-D full-wave
MLGFIM is O(NlogN) [23].
Although so many fast algorithms have been developed,
there are still many challenges in the electromagnetic
simulations. For instance, electromagnetic enhancement of
surface enhanced Raman scattering (SERS) attributed to the
enhanced E-fields in the vicinity of plasmonic nanostructures
is the major approach of SERS enhancement [28],[29]. When
the exciting electromagnetic radiation approaches the
resonance frequency, the enhancement factor which is
proportional to the fourth power of the gain in the
electromagnetic field becomes very difficult to evaluate. To
accurately calculate the electromagnetic field, the
nanoparticles must use deep sub-wavelength discretization.
Due to interpolation-based algorithm, MLGFIM is capable of
dealing with the low-frequency and multibody structure. In
addition, some thin strip with high dielectric constant has been
used to promote the magnetic emission in quantum emitters
[30],[31], and we found when the emitter is close to the
scatters, the near-field is difficult to be accurately computed
as well. The high-contrast material makes the whole structure
extremely in-homogenous, and thus deep sub-wavelength
mesh is also required. Moreover, the analysis of metasurfaces,
e.g. holographic metasurfaces [32],[33] which excites surface
waves with the desirable field distributions, is challenging.
For many of the holographic structures, the size and shape of
each sub-scatter which stuck on a substrate may be different,
so the computational burden is very large. This problem can
be released by adopting multilayered medium Green’s
function, and MLGFIM is able to be implemented based on
this complex Green’s function.
In this paper, various techniques for the MLGFIM to solve
electrically large electromagnetic problems are reviewed. The
MLGFIM used for accelerating the matrix-vector
multiplication is introduced first. QR factorization technique
is employed to compress the matrix storage. In order to make
the MLGFIM more efficient, different radial basis functions
(RBFs) are compared and interpolation grid patterns are
designed so that the MLGFIM could be accelerated by
applying newly proposed interpolation schemes. Moreover,
the technique of fast Fourier transform (FFT), which has been
combined with MLFMA [34],[35], could also be applied to
reduce the translation time in MLGFIM. The parallelized
MLGFIM versions based on OpenMP and message passing
interface (MPI) techniques have been realized to make this
fast solver more efficient for the large-scale electromagnetic
problems. Since the MLGFIM is a kernel-independent
method, it is easy to use MLGFIM for the analysis of periodic
structures and multilayer problems. Numerical results
demonstrate the good accuracy and efficiency of the above
implementations of the MLGFIM.
II. DESCRIPTION OF MLGFIM
For the 3-D free-space aperiodic problems, it is always
described by electric-field integral equation (EFIE),
magnetic-field integral equation (MFIE) and combined-field
integral equation (CFIE). With boundary conditions imposed
on the surface of the body, another integral equation, i.e.
Poggio-Miller-Chang-Harrington-Wu-Tsai (PMCHWT)
integral equations, could be used to analyze 3-D composite
metallic and dielectric bodies with surface discretization.
Using Galerkin method, the aforementioned integral
equations are converted to N×N dense matrix equations, viz.,
* *A x b (1)
where * can represent EFIE, MFIE, CFIE or PMCHWT. For
simplicity and clarity, we use EFIE as an example. The entries
in N×N matrix EFIEA and N×1 vector
EFIEb are written as:
2
0 0
[ ( ) ( )
1( ) ( )] ( , )
i jij i j
s s
i j
A ds ds
G
j r j r
j r j r r r
(2)
and
0
4( ) ( )
i
inc
i is
ib ds
j r E r (3)
where r , r , , 0, 0, ( )j r , ( , )G r r and ( )inc
E r are the
observation point, source point, angular frequency, free space
permittivity and permeability, basis function, interaction
function, and incident electric field, respectively.
To apply MLGFIM, (2) is decomposed into scalar form as:
x y z d
ij ij ij ij ijA A A A A (4)
where
# # #( ) ( ) ( , )
# , or,
i jij i j
s sA ds ds G
x y z
j r j r r r (5)
Forum for Electromagnetic Research Methods and Application Technologies (FERMAT)
3
and
2
0 0
1[ ( ) ( )] ( , )
i j
d
ij i js s
A ds ds G
j r j r r r (6)
Each scalar part ijA in (5) and (6) has the same form as:
( ) ( ) ( , )i j
ij i js s
A ds ds G r r r r (7)
where ( )i r and ( )j r are related to the weighting function
and the basis function, respectively.
n3
n2
m2 m2 m2 m2
m3 m3
m3 m3
m3 m3
m3 m3
m3 m3
m3 m3
m2 m2 m2 m2
m2 m2
m2 m2
(a)
m2
n2
rr
r r
o
(b)
Fig. 1. (a) 2-D pictorial representation of 3-D multilevel division in which
the interaction list of box n2 at level 2 are marked with m2, and the interaction
list of box n3 at level 3 are marked with m3. (b) 2-D pictorial representation of
3-D locations of the interpolation points (white circles) and interpolated
points (black circles) are in the boxes m2 and n2.
In [24] and [36], it is proven that PMCHWT and CFIE
which are associated with both EFIE and MFIE can also be
decomposed into the same form as (7). The interaction
functions ( , )G r r corresponding to the EFIE and MFIE can
be expressed as:
( )
( )
3
(for EFIE)
( , ')
1 (for MFIE)
ik
ik
e
Ge
ik
r r
r r
r rr r
r rr r
(8)
For the far-field interactions, instead of directly calculating
the function ( , )G r r , MLGFIM use interpolation method in a
multilevel tree to quickly approximate this function. The
multilevel tree is constructed by enclosing the object into a
box (level 0) and recursively dividing the large box into
sub-boxes. Box ml is the neighbor/interaction list of box nl at
level l if ml is near/well separated from nl, and their parents are
neighbors of each other, as shown in Fig. 1(a). If box ml is the
interaction list of box nl at level l, the interaction function
between these two boxes could be obtained by interpolation
method, as shown in Fig. 1(b), viz.:
, ,
1 1
, ,
;
, ,
l l
l ll l
l l l l
K K
m i n j
i j
m i n j
T
m m n n
GG
G
r r r r r r
r r
(9)
where,lm i
r and ,ln jr are the i-th and j-th interpolation points in
field box ml and source box nl, ,lm i r and ,ln j
r are the i-th
and j-th interpolation functions, respectively. Kl is the number
of interpolation points at level l.
Moreover, using the interpolation method, 1 ,l lm nG
and
1,l lm nG
can also be interpolated using ,l lm nG [20], viz.:
1 1
1 1
; ; ;
; ; ;
l l l l l l
l l l l l l
m n m m m n
T
m n m n n n
G C G
G G C
(10)
where 1l lm m ,
1l ln n , and matrix 1;l lm mC
is named as
lower-to-upper-level interpolation matrix. If the box ml is
interaction list of box nl are at level l, ( , )G r r at the finest
level L can be obtained using (9) and (10), viz.:
1 11 1
1 1 11
; ; ;
; ; ; ; ;
( , ) ( ) ( )
( ) ... ... ( )
l l l l l ll l
l l l l l lL L L LL L
TT
m m m n n nm n
T TT
m m m m m n n n n nm n
G C G C
C C G C C
r r r r
r r
(11)
Substituting (11) into (7) gives:
1 1 11
1 1 11
; ; ; ; ;
; ; ; ; ;, ,
[ ( ) ( )] ... ...
[ ( ) ( )]
= ( ) ... ... ( )
l l l l l lL L L LLi
Lj
l l l l l lL L L LL L
T TT
m m m m m n n n n nij i ms
j ns
T TT
m m m m m n n n n nm i n j
A ds C C G C C
ds
T C C G C C Z
r r
r r
r r
(12)
It should be mentioned that if box mL is the interaction list
of box nL at level L, there is no lower-to-upper-level
interpolation matrix in (12), and it becomes the conventional
peer-level interpolation. Subsequently, the sub-matrix
L Lm nA can be rewritten as:
1 1 11
11
,1 1
,2 2; ; ; ; ;
,
,1 1 ,2 2 ,
; ; ; ;
( )
( )... ...
( )
( ) ( ) ( )
...
L
L
l l l l l lL L L LL L
L m mL L
L L L n nL L
l l l l l lL LL
T
m
TT T
mm m m m m n n n n nm n
T
m N N
n n n N N
m m m m m n n nm
T
TA C C G C C
T
Z Z Z
T C C G C
r
r
r
r r r
1 1 ;...L L L
T TT
n n nC Z
(13)
whereLm
N andLnN denote the number of unknowns in boxes
mL and nL, respectively. From (13), it is observed that only the
Forum for Electromagnetic Research Methods and Application Technologies (FERMAT)
4
interpolation matrixes LmT and
LnZ at level L should be
calculated. Assuming there are PL boxes at level L, the
MLGFIM algorithm can be finally shown as Fig. 2.
:
( 1,2, , )
LL L
T
nn n
L L
S Z x
n P
1 1
1
;
( 1, 2, , 2)
l l l l
l l
n n n n
n n
S C S
l L L
2 2 2 2
2 2
;
interaction list of
:m m n n
n m
B G S
11
;
interaction list of
; ( 3, , )
l l l l
l l
ll l
m m n n
n m
T
mm m
B G S
C B l L
1
( 1,2, , )
L Lm m
L L
b T B
m P
Step 2:
Step 1:
Step 3:
Step 4:
Step 5: Step 6:
Upward Pass
Upward PassDownward Pass
Downward Pass
1 ;
neighbors of
:L L L
L L
m n n
n m
b b A x
Fig. 2. Flow chart of the implementation of MLGFIM.
Considering the interaction function matrix ;l lm nG
represents interactions between the interpolation points of two
well separated boxes, it is in low rank and it could be
factorized with QR factorization technique as:
;; ; l ll ll l
m nm n m nG Q R (14)
where ;l lm n
Q and ;l lm nR are the Kl × rl unitary matrix and the rl
× Kl upper triangular matrix. rl is the numerical rank of ;l lm nG ,
and the condition rl << Kl always holds. Instead of directly
store the Green’s function matrix ;l lm nG which is a Kl × Kl
full-matrix, we use (14) to compress the Green’s function
matrix and it also brings a great advantages in the computation
of step 3 and 4 of Fig. 2 [23].
III. THE INTERPOLATION SCHEME FOR MLGFIM
The accuracy and efficiency of MLGFIM tightly depend on
the interpolation scheme for the approximation of interaction
function. RBF interpolation method has been proven to be an
appropriate approach for the interpolation of interaction
function [23]. A lot of efforts have been made on searching a
better RBF interpolation scheme to improve the efficiency of
MLGFIM. Generally, an RBF i iR r r is a
continuous function that depends only on the distance
between the data (interpolated) point r and the center
(interpolation) point ri. Considering an approximation
function f (r) in an influence domain that has a set of K
arbitrarily distributed nodes with corresponding
values 1{ }K
i if r , the RBF interpolation reads:
1
1 2
1 2
( )
K
T
K
f R R R
f f f
r
r r r
(15)
where the entries of matrix are expressed as
ij j i r r .
There are many types of RBFs that can be used as the
interpolation function. Table 1 shows three types of RBFs that
have been used in MLGFIM. Parameter c in these expressions
is named as the shape parameter, and it determines the shape
of the interpolation function. In [37], it has proven that GA
RBFs can obtain better interpolation accuracy than IMQ
RBFs which have been used in [23] with the same number of
interpolation points, because they have better convergence
behavior. As a new class of oscillatory RBFs, BE RBFs have
shown their good interpolation performance in [38], and have
been applied to further enhance the efficiency of MLGFIM as
well [39].
Table 1. Definitions of Some Typical Radial Basis
Functions
Infinitely smooth functions R
Inverse multiquadric (IMQ) 2 2
1
1 c R
Gaussian (GA) 2exp c R
Bessel (BE)
/2 1
/2 1, 1, 2,
v
v
J cRv
cR
grid element
x
yz
(a)
grid element
x
yz
(b)
x
yz
grid element
(c)
Fig. 3. 2-D pictorial representation of (a) original STG, (b) modified STG
and (c) Boundary clustered STG (square points, hollow and solid dots are
volume center, surface center and vertex points in a grid element,
respectively).
Besides, the distribution of interpolation grid also affects
the interpolation efficiency of Green’s function. Tartan grid,
as a conventional grid pattern which is always used in the
Lagrange interpolation, just distributes interpolation points
evenly spaced in the source/field boxes, so the total number of
Forum for Electromagnetic Research Methods and Application Technologies (FERMAT)
5
interpolation points is K = K1D3. Since the RBF method allows
us to use scattered point distribution, STG which outperforms
Tartan grid in accuracy is used in MLGFIM. Three types of
STGs have been applied as the grid patterns which are shown
in Fig. 3. The original STG (Fig. 3(a)) consists of Tartan grids
and the points in the volume center of grid element of which
vertexes are formed by adjacent Tartan grids [23]. And the
modified STG, as shown in Fig. 3(b), consists of Tartan grids
and the points in the center of the sides of the grid element
[37]. Hence the total number of interpolation points for the
original STG and modified STG are K1D3+(K1D-1)3 and K1D
3+3K1D(K1D-1)2, respectively. It has been proven that with
similar number of interpolation points, the results using the
modified STG are more accurate than those using the original
STG [37]. This is because the condition number of matrix
generated by the modified STG is better than the original
STG. However, it is pointed out that an accurate method on
high-degree uniform grid sets leads to large errors near the
boundaries [40],[41]. So if the boundary errors can be
suppressed, the interpolation accuracy will be further
enhanced. Boundary clustered STG is generated by making
the points denser near the boundary [42]. In 1-D, the
Chebyshev-Lobatto nodes in the closed interval [-1, 1] are
expressed as:
1
1
2cos , 0,1, 1
1 2s D
D
sh s K
K
(16)
where hs is the location of s-th interpolation point. However, it
has been proven that the 3-D non-uniform grid generated by
simply clustering the interpolation points toward boundary
using (16) in x, y and z directions cannot improve the
interpolation accuracy [42]. This is because this grid helps to
reduce the boundary interpolation errors but it induces large
interior errors as well. In order to balance the boundary and
interior errors and make interpolation points still symmetric to
the center of interval, the 1-D grid pattern in the closed
interval [0, l] is generated as:
1
1
1 11
1
121 cos , 0,1,
2 1 2 2
2 2 ( 1) / 2 11 cos , , 1
2 1 2 2
D
D
s
D DD
D
Kl ss
K
h
s K Kls K
K
(17)
In function (17), the parameter is introduced to adjust the
degree of boundary clustering. The interpolation points, as
approaches 0, are distributed at l/2. Increasing will shift the
interpolation points toward the boundaries, and it improves
the resolution at the boundary but lowers the resolution in the
interior box. By adopting function (17) in x, y and z directions,
we obtain a boundary clustered Tartan grid with K1D3
interpolation points. Then, the boundary clustered STG is
generated by adding 3K1D(K1D-1)2 interpolation points to the
center of the sides of the rectangular block (grid element), as
shown in Fig. 3(c).
The maximum interpolation errors using three different
interpolation schemes are compared in Table 2. The first two
schemes employ the same interpolation grid (viz., original
STG), and the threshold of the maximum error for the second
scheme is set as 0.05. With the same number of interpolation
points, the maximum errors using IMQ RBFs are always
larger than those using BE RBFs, as shown in Table 2. And as
the edge length of the box increases, the method using BE
RBFs becomes more accurate than IMQ RBFs. The third
scheme adopts a new interpolation grid pattern, viz., boundary
clustered STG, to further enhance the accuracy and efficiency
of the interpolation. With the same error threshold, the
method using boundary clustered STG requires less number
of interpolation points than the original STG. Moreover, the
interpolation accuracy using the third scheme, even with less
number of points, is better than the other two. And it is also
observed that if we employ the third scheme, much more
number of interpolation points is saved as the size of box
increases. The efficiency of MLGFIM can be dramatically
enhanced by adopting the third method.
Table 2. Comparison of accuracy using different kinds of interpolation schemes.
Interpolation
scheme
Box
length ()
Number of
interpolation
points
Optimized
in (17)
Optimized
shape
parameter c
Maximum
error
IMQ RBFs
and original
STG
1 189 2.10 0.00681
2 559 0.80 0.0731
3 1241 0.69 0.131
4 2331 0.51 0.296
BE RBFs and
original STG
1 189 6.00 0.00632
2 559 3.20 0.0174
3 1241 2.35 0.0295
4 2331 1.705 0.0498
BE RBFs and
boundary
clustered STG
1 172 0.75 6.90 0.00439
2 365 0.97 3.21 0.0125
3 666 0.96 2.10 0.0291
4 1099 0.83 1.60 0.0465
Forum for Electromagnetic Research Methods and Application Technologies (FERMAT)
6
IV. MLGFIM FOR PERIODIC STRUCTURES AND
MULTILAYERED PROBLEMS
As a kernel independent algorithm, MLGFIM can also be
used for the fast analysis of other complex applications, e.g.,
periodic structures and multilayer problems. The Floquet
theorem allows limiting the analysis of periodic structures to
that of a unit cell; however, the number of unknowns can still
be large with there are fine features exist in the unit cell.
For periodic structures, instead of interpolating free-space
Green’s function, we should interpolate periodic Green’s
function. Ewald’s transformation [43],[44], which
efficiently combines both spatial and spectral formulations
of the periodic Green’s function, is applied to speed up the
evaluation of periodic Green’s function. Besides, since the
unit cells are actually connected together, the periodic
boundary condition must be taken into consideration. Half
basis functions at the connection boundary between two
adjacent unit cells have to be introduced to ensure the
boundary currents continuous.
After that, we can use MLGFIM to accelerate the analysis
of periodic structures. There are some modifications in the
definition of neighbor and interaction list when applying
MLGFIM for these structures, and they are defined as [45]:
box m is the neighbor/interaction list of box n when box m is
near/well separated from box n and the images of box n, and
parent box of box m is also near parent boxes of box n and
images of box n simultaneously. Fig. 4 shows the neighbor
and interaction lists of box n at level 3. In order to implement
MLGFIM easier, the box of interaction list which is not
around box n (but around the image of box n), e.g. box m1 in
Fig. 4, should be moved to the place of box m2 (image of m1)
which is around box n.
m1
n
image
n
image
n
image
n
m2
ra
rb
Fig. 4. 2-D representation of a 3-D multilevel division in which the
neighbor (white colored) and interaction (slash) lists of box n at level 3 are
shown.
In order to address the problem of the value of optimum
shape parameter c shift during the interpolation, RBF-QR
method is applied for the interpolation of periodic Green’s
function [46]. As the periodicity of one structure changes,
the interpolated function (viz., periodic Green’s function)
changes, and consequently the optimum value of shape
parameter c in RBF should be retested. Because the RBF-QR
method is insensitive to the value of c, it is not required to
retest the shape parameter when the interpolated function
changes. By using expansion with Chebyshev polynomials
and QR-factorization to the Gaussian RBFs, a better
conditioned interpolation matrix with its condition number
insensitive to c is generated [46],[47]. With RBF-QR
interpolation, MLGFIM becomes more robust for the
analysis of periodic structures.
By interpolating the multilayered Green’s function,
MLGFIM has been applied to solve the problems of
multilayered medium structures. An improved discrete
complex image method (DCIM) which uses a new sampling
path directly goes through the branch point is adopted for the
evaluation of the multilayered Green’s function [48]. This
method helps to capture the contribution to the lateral wave
more accurately. The property of planar structures of
multilayered problems demands only 2-D interpolation
scheme (viz., the interpolation points are only distributed at
the interfaces of every two layers), while 1-D interpolation
scheme is used for the vertical via holes only at the finest
level [49]. We will later show that this interpolation
approach is very efficient for the analysis of multilayered
medium problems.
V. OTHER APPROACHES TO COMBINE WITH MLGFIM
Given that the translation occupies most of the
computation time in the interpolation procedure of
lower-to-upper-level, it is possible to further enhance the
efficiency of MLGFIM by reducing the translation time.
Although the interpolation performance of 3-D Tartan grid is
not as good as the STG-based grid patterns, it generates a
3-D Toeplitz translation matrix G which could use FFT to
accelerate the matrix-vector multiplication. Moreover, the
difference in the interpolation efficiency between 2-D
Tartan grid and STG-based grid patterns is smaller than that
of the 3-D case, and hence the MLGFIM-FFT can be used to
accelerate the solution, especially the solution of multilayer
problems where 2-D interpolation scheme is adopted.
For the multilevel tree structure, there are two ways to
implement the FFT-accelerated translation procedure. In
[50],[51], the multilevel tree will be truncated at certain
level, and the pre-corrected FFT approach is applied at this
level. A global Toeplitz matrix is generated and the
interactions among all boxes are efficiently calculated by
performing only one FFT. However, a pre-corrected near
field evaluation is required for this method. So the
translation (also shown in Fig. 2) is rewritten as a
convolution form, viz.:
1;
neighbors of
l l l l
l l
m m n n
n m
B FFF FFT G FFT S G S
(18)
where G is comprised of Green’s function matrix ;l lm nG
and S consists of vectorlnS . Below the truncated level, the
lower-to-upper-level interpolation as shown in Fig. 2 could
be carried out as well [50].
The second way is based on local boxes, so it generates a
sub-domain Toeplitz matrix. Consequently, sub-domain
FFT is applied to accelerate the translation. Pre-correction is
Forum for Electromagnetic Research Methods and Application Technologies (FERMAT)
7
not required because the sub-domain Toeplitz Green’s
function matrix is formed by two well separated boxes [52].
Hence, instead of (18), the translation could be expressed as:
;
interaction list of
1;
interaction list of
l l l l
l l
l l l
l l
m m n n
n m
m n n
n m
B G S
FFF FFT G FFT S
(19)
where ;l lm nG is a vector extracted from the Toeplitz matrix
;l lm nG . The FFT is not implemented only at one level, but at
many high levels unless the number of interpolation points is
too small and direct matrix-vector multiplication will be
more efficient than the FFT. With the second method, no
truncated level exists, so the lower-to-upper-level
interpolation is implemented from the top level to the bottom
level to make the MLGFIM-FFT more efficient.
In addition, parallel computing could also be applied to
accelerate the solution. Parallelized MLGFIM using OpenMP
[50] and MPI [36] have been developed. For achieving high
efficiency, we must make workload balanced and that could
be accomplished by assigning approximately the same
number of boxes at each level into each processor. Different
from the OpenMP parallelization on a share-memory
computer system, the MPI parallelization on a
distributed-memory computer must take into account of the
communications among different processors as well. If the
upward or downward pass shown in Fig. 2 is implemented
through different processors, a large amount of data
communication is required. Hence, each root box (at level 2)
and all its descendants should be assigned to the same
processor [36]. Besides, in order to reduce the data
communication in the computation of near field, the
neighbors of each finest-level box are duplicated if they are
not assigned to the same processor.
Parallelized MLGFIM on a distributed-memory computer
system can be developed by modifying the steps of the
original MLGFIM. First, the matrix-vector multiplication
LL Lnn nS Z x is calculated at the finest level l = L with its own
sub-tasks. Second, the vectors 1 1
1
;l l l l
l l
n n n n
n n
S C S
for each
box assigned to this processor are computed from level l = L -
1 to 2. Then the translation from the source box to the
interaction list is computed. Data communication is required
in this step, and we use MPI_Allreduce command to collect
the vectors lnS from each processor, and distribute it to all
processors. After that, the vectors
11; ;
interaction list of
l ll l l l l
l l
T
m mm n n m m
n m
B G S C B
in the corresponding
boxes for the processor from level l = 3 to L are computed.
Finally, the matrix-vector multiplication L L
L
T
m mm nearb b T B is
performed and the vector Lmb is also collected and distributed
to all processors.
Certain factors should be considered when examining the
efficiency of parallel computing. Intuitively, a parallel
program leads to a shorter execution time for a given
computation task, which can be expressed by a speedup and
efficiency factors as:
p s ps n t t (20)
and
100%sp
p p
te n
n t
(21)
where ts and tp are serial execution time using one processor
and parallel execution time using np processors, respectively.
If the efficiency factor approaches 100%, this indicates that
the parallel computing is very efficient and the
communication/computation ratio is very small.
VI. NUMERICAL EXAMPLES
In order to demonstrate good performance of the proposed
method, some examples are given as follows. The generalized
minimal residual (GMRES) iteration method with a relative
error norm of 10-3 is employed, and the inner loop of GMRES
contains 100 matrix-vector multiplications.
x
y
o
w1
dd
w2
t
x
yz
o
(a) (b)
0 20 40 60 80 100 120 140 160 180
(degree)
-110
-100
-90
-80
-70
-60
-50
-40
-30
RC
S(d
Bs
m)
MLGFIMMoM
(c)
Fig. 5. Plane wave impinges normally on a fishnet-type array. (a)
Geometry. (b) Unit cell. (c) Bistatic RCS with polarization.
In the first example, MLGFIM is applied to plane wave
scattering from a fishnet-type array with 10×10 elements, as
shown in Fig. 5(a). The geometrical parameters of the array
are as follows: w1 = 10 m, w2 = 17.5 m, d = 115m, t = 9 m.
The patterns on the top and bottom surfaces of the substrate
are made of the same shaped PEC, and the relative
permittivity of the substrate is 2.25. A plane wave with
electric field parallel to the x-axis propagates along –z-axis,
and the center frequency is 0.5 THz. PMCHWT integral
equation is used for this structure to obtain the surface current.
The bistatic RCS with polarization is calculated using the
5-level MLGFIM, and compared with MoM, as shown in Fig.
5(b). From this figure, it is observed that these two results
agree very well.
Forum for Electromagnetic Research Methods and Application Technologies (FERMAT)
8
x y
z
Improved MLGFIM
Original MLGFIM
0
45
90
18
0
13
5
27
0
22
531
5
4 1 -2 -5 -8
-27 -21 -15 -9 -3 3 9 15 21 27
y (cm)
-27
-21
-15
-9
-3
3
9
15
21
27
z (
cm
)
Poyting vector (kW/m^2)
1000+
875 to 1000
750 to 875
625 to 750
500 to 625
375 to 500
250 to 375
125 to 250
0 to 125
(a) (b) (c)
Fig. 6. A dipole in front of an 11×11 dielectric sphere array. (a) Geometry. (b) Polar plot of the directivity versus in dBi when= 90o. (c) Distribution of
Poynting vector in yz plane.
Table 3. Comparison of CPU Time and Memory Requirement for Simulating the 11×11 Dielectric Sphere Array
Method
Maximum
electrical
length(d)
The
number of
levels
The number
of
unknowns
CPU time for per
matrix–vector
multiplication
Memory
requirement
Original MLGFIM 16.87 6 144,020
96.3 s 8.7 GB
Improved MLGFIM 21.4 s 3.1 GB
In the second example, a radiation problem that a dipole in
front of an 11×11 dielectric sphere array (r = 2.25) is
investigated, as shown in Fig. 6(a). The radius of the sphere is
2 cm and the distance between the centers of two adjacent
spheres is 5 cm. A Hertzian dipole which works at the
frequency of 6.25 GHz is located 10 cm in front of the array
center and placed along the z-axis. In this example, ten grids
per wavelength d (where 0d r
) are used to discretize
this structure, so that the number of unknowns for equivalent
electric and magnetic currents is 144,020. The polar plot of
the directivity versus when= 90o is shown in Fig. 6(b).
Two methods, i.e., original MLGFIM (IMQ RBF and original
STG) and improved MLGFIM (BE RBF and boundary
clustered STG), are used for the simulation. Good agreement
can be observed between these two results. The distribution of
Poynting vector in the plane of x = –20 cm is also calculated,
as shown in Fig. 6(c). The computational time and memory
requirement for this example are shown in Table 3. Compared
with the original method, only 22.2% CPU time and 35.6%
memory are required for the improved one.
(a)
0
45
90
135
180
225
270
315
-35
-20
-20
-5
-5
10
10
25
25
Double Positive Medium
Single Negative Medium
Double Negative Medium
(b)
Forum for Electromagnetic Research Methods and Application Technologies (FERMAT)
9
104
105
106
Number of unknowns (N)
1
10
100
CP
U t
ime
(s
)
1
10
100
Impoved MLGFIM
Original MLGFIM
2.7e-5*NlogN
7 levels
6 levels
5 levels
(c)
104
105
106
Number of unknowns (N)
0.1
1
10
Me
mo
ry r
eq
uir
em
en
t (G
B)
0.1
1
10
Improved MLGFIM
Original MLGFIM
1.6e-2*N
7 levels
6 levels5 levels
(d)
Fig. 7. Plane wave scattering from a dielectric prolate spheroid. (a)
Geometry. (b) Bistatic RCS with polarization at 300 MHz. (c) CPU time.
(d) Memory requirement.
Next, an example is provided to show the CPU time and
memory requirement with different number of unknowns. A
dielectric prolate spheroid is considered, as shown in Fig.
7(a). The long axis of the prolate spheroid is 6 m, whereas the
short axis is 1.5 m. A plane wave with a frequency of 300 MHz
impinges on the prolate spheroid along –x-axis. The plane
wave scattered by prolate spheroids composed of double
positive (r = 2.25, r = 1.0), single negative (r = –2.25, r =
1.0) and double negative (r = –2.25, r = –1.0) medium are
studied. Fig. 7(b) compares the bistatic RCS with
polarization for these three cases. The forward scattering is
enhanced significantly due to negative permittivity and
permeability. The CPU time for performing each
matrix-vector multiplication and the corresponding memory
requirement are shown in Figs. 7(c) and (d). In these two
figures, about 20 grids per wavelength are used to discretize
the prolate spheroid, and the frequency varies from 150 MHz
to 660 MHz. Three types of numbers of levels are used for the
calculation of the structure, and the original MLGFIM cannot
simulate the structure when the number of unknowns is larger
than 5×105 because of the memory limitation. From Fig. 7(c)
and (d), it is observed that the CPU time and memory
requirement of the improved MLGFIM are always smaller
than those of the original one. And when the number of
unknowns increases to 1,089,224, the CPU time for
performing one matrix-vector multiplication and the memory
requirement for the improved MLGFIM are 128.7 seconds
and 17.68 GB, respectively. In addition, it is also observed
that the improved MLGFIM approximately obeys a
computational complexity of O(N log N) and memory
complexity of O(N).
x
yz
3 mm
1.8 mm
0.6 mm
3 mm
1.8 mm
(a) (b)
-200 -150 -100 -50 0 50 100 150
Bistatic angle (degree)
-120
-100
-80
-60
-40
-20
0
RC
S (
dB
sm
)
Parallelized MLGFIM
Serial MLGFIM
(c)
Fig. 8. Plane wave scattering from a metallic scatterer. (a) Geometry. (b)
Unit. (c) Bistatic RCS with polarization.
Table 4. Comparison of CPU Time and Memory
Requirement
Number
of levels
Number of
processors Speed-up
Efficiency
(%)
Save
time (%)
7
1 1.0 100 0
2 1.98 99.2 49.6
4 3.66 91.5 72.4
8 6.51 81.3 84.6
To show the effect and accuracy of parallelized MLGFIM,
we consider the plane wave scattering from a complicated
metallic scatterer, as shown in Fig. 8(a). This structure is
composed of orthogonal square rings, and four square rings
orthogonally connect with one ring to generate a unit, as
shown in Fig. 8(b). The outer length, inner length and height
of each ring are 3 mm, 1.8 mm and 0.6 mm, respectively. On
the boundary of this structure, there exists only half rings, and
hence the total structure contains 6 orthogonal rings in each
direction. The structure is investigated at 300 GHz and
illuminated by a plane wave propagating along the
–x-direction, with the electric field polarized in the
z-direction. After the structure is discretized, the total number
of unknowns obtained is 1,134,768. The bistatic RCS result
Forum for Electromagnetic Research Methods and Application Technologies (FERMAT)
10
calculated by the serial MLGFIM and parallelized MLGFIM
using 8 processors are compared in Fig. 8(c). Good agreement
is observed from this figure. Furthermore, the speed-up factor
and efficiency factor are demonstrated, as shown in Table 4.
According to this table, we observe that the computational
time is reduced significantly with the increase of the number
of processors. The efficiencies of the parallelized MLGFIM
are larger than 80% for all cases. The CPU time is reduced by
84.6% after calculating this problem with eight processors,
simultaneously.
r=3.0
4.83 mm
5 m
m
5 mm
d
0.17 mm
0.17 mm
0.027 mm
x
y
(a)
10 15 20 25 30 35 40
Frequency (GHz)
-30
-25
-20
-15
-10
-5
0
Re
fle
cti
on
Co
eff
icie
nt
MLGFIMReference [47]Reference [48]
d=3.0 mm
d=2.5 mm
(b)
Fig. 9. FSS comprising two centered square loops on a dielectric substrate.
(a) Side and front view of the unit cell. (b) TM reflection coefficient with
different values of d.
In the fifth example, MLGFIM is used to analyze a periodic
structure, in which two centered square loops are etched on a
dielectric layer (r = 3). Fig. 9(a) shows the geometry of the
frequency selective surface (FSS), and it is illuminated by a
uniform plane wave incident from the normal direction.
Periodic Green’s function interpolation and periodic
boundary conditions are used for this structure. The results for
different dimensions d of the inner loop are calculated by
MLGFIM and compared with the simulation results in [53]
and experimental results in [54]. RBF-QR method is adopted
as the interpolation approach, so that the MLGFIM becomes
more robust for the analysis of periodic problems. From Fig.
9(b), it can be seen that our results are closer to the
experimental results compared with the simulation results in
[53]. And we also find that the second resonance shifts toward
a lower frequency band with the increase of the dimension d.
wl
ra
rb
x
y
(a)
0 10 20 30 40 50 60 70 80 90-70
-60
-50
-40
-30
-20
-10
0
E-f
ield
(d
B)
8 by 8
16 by 16
32 by 32
64 by 64
(b)
10000 100000 1E+006
Number of unknowns
0.001
0.01
0.1
1
10C
PU
tim
e (
se
co
nd
)
0.01
0.1
1
10
Me
mo
ry (G
B)
MLGFIM
MLGFIM_FFT
MLGFIM
MLGFIM_FFT
(c)
Fig. 10. Plane wave scattering from a patch array. (a) Geometry. (b)
Normalized E-field on xz plane. (c) CPU time and memory requirement.
In the sixth example, MLGFIM is applied to analyze some
patch arrays with planar multilayer substrate, in which
sub-domain FFT is combined to accelerate the calculation. A
plane wave with the electric field parallels to the y axis
incidents from the normal direction. The parameters of the
microstrip patch array are l = 49 mm, w = 41 mm, ra = 90
mm,rb = 82 mm, as shown in Fig. 10(a). The thickness and
dielectric constant of the substrate are 1.59 mm and 2.2,
respectively. The working frequency is set as 2.2GHz, and the
normalized electric fields on xz plane for 8×8, 16×16, 32×32,
and 64×64 arrays are shown in Fig. 10(b). In Fig. 10(c), the
efficiencies, using MLGFIM and MLGFIM-FFT, for
analyzing these arrays with different number of unknowns are
compared. For this example, 2-D interpolation method is
applied, and hence the CPU time and memory requirement are
both much smaller than those of the above examples which
use 3-D interpolation. From Fig. 10(c), it is observed that the
CPU time and memory requirement using MLGFIM-FFT are
always smaller than those using MLGFIM alone. With the
Forum for Electromagnetic Research Methods and Application Technologies (FERMAT)
11
increase of the array size, the reductions of time and memory
become more significant.
x
y
Feeding
line
(a)
ra
rb
wl
t1
t2
(b)
-100 -80 -60 -40 -20 0 20 40 60 80 100
(degree)
-60
-50
-40
-30
-20
-10
0
Ra
dia
tio
n (
dB
)
co-pol
cr-pol
(c)
-100 -80 -60 -40 -20 0 20 40 60 80 100
(degree)
-60
-50
-40
-30
-20
-10
0
Ra
dia
tio
n (
dB
)
co-pol
cr-pol
(d)
Fig. 11. Radiation of a 32×32 parallel fed antenna array. (a) Geometry. (b)
Unit. (c) Radiation pattern on H-plane (= 0º). (d) Radiation pattern on
E-plane (= 90º).
Finally, the MLGFIM-FFT method is used to calculate the
radiation pattern from a 32×32 patch array with planar
multilayer substrate. The structure of this array is shown in
Fig. 11(a). And this structure is composed by eight 4×4 unit in
one direction which is given in Fig. 11(b). The parameters of
the microstrip patch array are l = 41 mm, w = 49 mm, ra = 100
mm, rb = 100 mm, t1 = 5.0 mm, t2 = 1.5 mm, and the thickness
and relative dielectric constant of the substrate are 1.59 mm
and 2.2, respectively. The working frequency is 9.42 GHz and
the number of the unknowns of the metallic layer is about
310,000. The radiation patterns in both E- and H-plane are
plotted in Fig. 11(c) and (d). A very narrow beam is formed as
expected. The FFT will be applied at level 2 and 3. Because
the structure is electrically large and complex, the
convergence of the iteration process will be a big problem. To
ensure convergence, the number of the spanning vector in
GMRES is set to 1,000. The CPU time per iteration is about
1.8 seconds which is still very efficient, and the memory
requirement is 11 GB due to the large number of spanning
vector.
VII. CONCLUSIONS
The development and improvement of the efficient fast
algorithm, namely MLGFIM, is reviewed. Objects in
free-space, periodic and multilayer structures are analyzed
using this fast simulation algorithm. Various techniques,
including: BE RBF interpolation functions, boundary
clustered STG pattern, FFT technique and parallelization
approaches, are used to enhance the efficiency of the
MLGFIM. Seven examples are given to validate the accuracy,
and demonstrate the improvement of the MLGFIM after using
the above new techniques.
REFERENCES
[1] R. F. Harrington, Field Computation by Moment
Methods, Piscataway, N.J.: IEEE, 1993.
[2] J. M. Song and W. C. Chew, “Multilevel fast-multipole
algorithm for solving combined field integral equation of
electromagnetic scattering,” Microw. Opt. Tech. Lett.,
vol. 10, pp. 14–19, Sep. 1995.
[3] J. M. Song, C. C. Lu, W. C. Chew, and S. W. Lee, “Fast
Illinois solver code (FISC),” IEEE Antennas Propag.
Mag., vol. 48, pp. 27–34, Jun. 1998.
[4] W. C. Chew, J. M. Jin, E. Michielssen, and J. M. Song,
Fast and Efficient Algorithms in Computational
Electromagnetics. Boston, MA: Artech House, 2001.
[5] J. R. Phillips and J. K. White, “A precorrected-FFT
method for electrostatic analysis of complicated 3-D
structures,” IEEE Trans. Comput.- Aided Des. Integr.
Circuits Syst., vol. 16, no. 10, pp. 1059–1072, Oct. 1997.
[6] H. Hu, D. T. Blaauw, V. Zolotov, K. Gala, M. Zhao, R.
Panda, and S. S. Sapatnekar, “Fast on-chip inductance
simulation using a precorrected- FFT method,” IEEE
Trans. Comput.-Aided Des. Integr. Circuits Syst., vol. 22,
no. 1, pp. 49–66, Jan. 2003.
[7] E. Bleszynski, M. Bleszynski, and T. Jaroszewicz,
“AIM: Adaptive integral method for solving large-scale
electromagnetic scattering and radiation problems,”
Radio Sci., vol. 31, pp. 1225–1251, Sep.-Oct. 1996.
Forum for Electromagnetic Research Methods and Application Technologies (FERMAT)
12
[8] F. Ling, C. F. Wang, and J. M. Jin, “An efficient
algorithm for analyzing large-scale microstrip structures
using adaptive integral method combined with discrete
complex-image method,” IEEE Trans. Microw. Theory
Tech., vol. 48, pp. 832–839, May 2000.
[9] C. H. Chan, C. M. Lin, L. Tsang, and Y. F. Leung, “A
sparse-matrix/canonical grid method for analyzing
microstrip structures,” IEICE Trans. Electron, E80-C,
pp. 1354–1359, 1997.
[10] S. Q. Li, Y. X. Yu, C. H. Chan, K. F. Chan, and L.
Tsang, “A sparse-matrix/canonical grid method for
analyzing densely packed interconnects,” IEEE Trans.
Microw. Theory Tech., vol. 49, no. 7, pp. 1221–1228, Jul.
2001.
[11] A. Brandt, “Multilevel computations of integral
transforms and particle interactions with oscillatory
kernels,” Comput. Phys. Commun., vol. 65, pp. 24–38,
1991.
[12] A. Boag, E. Michielssen, and A. Brandt, “Nonuniform
polar grid algorithm for fast field evaluation,” IEEE
Antennas Wirel. Propag. Lett., vol. 1, pp. 142–145, 2002.
[13] A. Boag, “A fast iterative physical optics (FIPO)
algorithm based on nonuniform polar grid interpolation,”
Microw. Opt. Technol. Lett., vol. 35, pp. 240–244, Nov.
2002.
[14] Y. Brick and A. Boag, “3D multilevel non-uniform grid
algorithm for fast field evaluation,” presented at IEEE
Antennas Propagation Int. Symp., Albuquerque, NM,
Jul. 2006.
[15] A. Rotbart and A. Boag, “A multilevel non-uniform grid
algorithm for acceleration of volumetric integral
equation based solvers for analysis of arbitrarily shaped
dielectric radomes,” presented at IEEE Int. conf. Microw.
Commun. Antennas Electron. Syst. (COMCAS), Tel
Aviv, Israel, Nov. 2011.
[16] M. Bebendorf, “Approximation of boundary element
matrices,” Numer. Math., vol. 86, no. 4, pp. 565–589,
2000.
[17] S. Kurz, O. Rain, and S. Rjasanow, “The adaptive
cross-approximation technique for the 3-D
boundary-element method,” IEEE Trans. Magn., vol.
38, no. 2, pp. 421–424, Mar. 2002.
[18] W. Hackbusch and S. Boerm, “H2-matrix approximation
of integral operators by interpolation,” Appl. Numer.
Math., vol. 43, pp. 129–143, 2002.
[19] W. Chai and D. Jiao, “An H2-matrix-based
integral-equation solver of reduced complexity and
controlled accuracy for solving electrodynamic
problems,” IEEE Trans. Antennas Propag., vol. 57, no.
10, pp. 3147–3159, 2009.
[20] H. G. Wang, C. H. Chan, and L. Tsang, “A new
multilevel Green’s function interpolation method for
large-scale low-frequency EM simulations,” IEEE
Trans. Comput. Aided Des. Integr. Circuits Syst., vol. 24,
pp. 1427–1443, 2005.
[21] H. G. Wang and P. Zhao, “Combing multilevel Green's
function interpolation method with volume loop
bases for inductance extraction problems,” Prog.
Electromagn. Res., vol. 80, pp. 225–239, 2008.
[22] P. Zhao and H. G. Wang, “Resistances and inductances
extraction using surface integral equation with
the acceleration of multilevel Green function
interpolation method,” Prog. Electromagn. Res., vol. 83,
pp. 43–54, 2008.
[23] H. G. Wang and C. H. Chan, “The implementation of
multilevel Green’s function interpolation method for
full-wave electromagnetic problems,” IEEE Trans.
Antennas Propag., vol. 55, pp. 1348–1358, 2007.
[24] Y. Shi, H. G. Wang, L. Li, and C. H. Chan, “Multilevel
Green’s function interpolation method for scattering
from composite metallic and dielectric objects,” J. Opt.
Soc. Am. A, vol. 25, pp. 2535–2548, 2008.
[25] L. Li, H. G. Wang, and C. H. Chan, “An improved
multilevel Green’s function interpolation method with
adaptive phase compensation for large-scale full-wave
EM simulation,” IEEE Trans. Antennas Propagat., vol.
56, pp. 1381–1393, 2008.
[26] M. D. Buhmann, Radial basis functions: theory and
implementations, Cambridge U. Press, 2003.
[27] H. G. Wang, C. H. Chan, L. Tsang, and V. Jandhyala,
“On sampling algorithms in multilevel QR factorization
method for magnetoquasistatic analysis of integrated
circuits over multilayered lossy substrates,” IEEE Trans.
Comput.-Aided Des. Integr. Circuits Syst., vol. 25, no. 9,
pp. 1777–1792, Sep. 2006.
[28] H. Ko, S. Singamaneni, and V. V. Tsukruk,
“Nanostructured surfaces and assemblies as SERS
media,” Small, vol. 4, no.10, pp. 1576–1599, 2008.
[29] B. B. Xu, Y. L. Zhang, W. Y. Zhang, X. Q. Liu, J. N.
Wang, X. L. Zhang, D. D. Zhang, H. B. Jiang, R. Zhang,
and H. B. Sun, “Silver‐coated rose petal: green, facile,
low‐cost and sustainable fabrication of a SERS substrate
with unique superhydrophobicity and high
efficiency,” Adv. Opt. Mater., vol. 1, no. 1, pp. 56–60,
2013.
[30] B. Rolly, B. Bebey, S. Bidault, B. Stout, and N. Bonod,
“Promoting magnetic dipolar transition in trivalent
lanthanide ions with lossless Mie resonances,” Phys.
Rev. B, vol. 85, no. 24, 245432, 2012.
[31] G. Boudarham, R. Abdeddaim, and N. Bonod,
“Enhancing the magnetic field intensity with a dielectric
gap antenna,” Appl. Phys. Lett., vol. 104, no. 2, 021117,
2014.
[32] B. H. Fong , J. S. Colburn , J. J. Ottusch , J. L. Visher
and D. F. Sievenpiper, “Scalar and tensor holographic
artificial impedance surfaces,” IEEE Trans. Antennas
Propag., vol. 58, no. 10, pp.3212–3221, 2010.
[33] P. Genevet and F. Capasso, “Holographic optical
metasurfaces: a review of current progress,” Rep. Prog.
Phys., vol. 78, no. 2, 024401, 2015.
[34] J. De Zaeytijd, I. Bogaert, and A. Franchois, “An
efficient hybrid MLFMA-FFT solver for the volume
integral equation in case of sparse3D inhomogeneous
dielectric scatterers,” J. Comput. Phys., vol. 227, no. 14,
pp. 7052–7068, 2008.
[35] J. M. Taboada, M. G. Araujo, J. M. Bertolo, L. Landesa,
F. Obelleiro, and J. L. Rodriguez, “MLFMA-FFT
parallel algorithm for the solution of large-scale
problems in electromagnetics,” Prog. Electromagn.
Forum for Electromagnetic Research Methods and Application Technologies (FERMAT)
13
Res., vol. 105, pp.15–30, 2010.
[36] P. Zhao, D. Q. Liu, and C. H. Chan, “A parallelized
multilevel Green’s function interpolation method for
open and closed objects,” Int. Symp. Antennas Propagat.,
Dec. 2014.
[37] Y. Shi and C. H. Chan, “Comparison of interpolating
functions and interpolating points in full-wave multilevel
Green’s function interpolation method,” IEEE Trans.
Antennas Propagat., vol. 58, pp. 2691–2699, 2010.
[38] B. Fornberg, E. Larsson, and G. Wright, “A new class of
oscillatory radial basis functions,” Comput. Math. Appl.,
vol. 51, pp. 1209–1222, 2006.
[39] Y. Shi and C. H. Chan, “Improved 3D full-wave
multilevel Green's function interpolation method,”
Electronics Letter, vol. 47, pp. 174–175, 2011.
[40] B. Fornberg, T. A. Driscoll, G. Wrights, and R. Charles,
“Observations on the behavior of radial basis function
approximations near boundaries,” Comput. Math. Appl.,
vol. 43, pp. 473–490, 2002.
[41] B. Fornberg and J. Zuev, “The Runge phenomenon and
spatially variable shape parameters in RBF
interpolation,” Comput. Math. Appl., vol. 54, pp.
379–398, 2007.
[42] P. Zhao, D. Q. Liu, Y. Shi and C. H. Chan, “Improving
multilevel Green's function interpolation method with a new
interpolation grid,” J. Electromagn. Waves Applicat., vol.
27, no. 15, pp. 1892–1901, 2013.
[43] P. P. Ewald, “Die berechnung optischer und
elektrostatischer gitterpotentiale,” Ann. Phys., vol. 64,
pp. 253–287, 1921.
[44] I. Stevanovic, P. Crespo-Valero, K. Blagovic, F.
Bongrad, and J. R. Mosig, “Integral-equation analysis of
3-D metallic objects arranged in 2-D lattices using the
Ewald transformation,” IEEE Trans. Microw. Theory
Tech., vol. 54, no. 10, pp. 3688–3697, Oct. 2006.
[45] Y. Shi and C. H. Chan, “Multilevel Green’s function
interpolation method for analysis of 3-D frequency
selective structures using volume/surface integral
equation,” J. Opt. Soc. Am. A, vol. 27, pp. 308–318,
2010.
[46] P. Zhao, “Study of multilevel Green’s function
interpolation method for efficient analysis of large-scale
electromagnetic problems,” PhD Thesis, City University
of Hong Kong, 2013.
[47] B. Fornberg, E. Larsson, and N. Flyer, “Stable
computations with Gaussian radial basis functions,”
SIAM J. Sci. Comput., vol. 33, pp. 869–892, 2011.
[48] D. Q. Liu, P. Zhao, and C. H. Chan, “An improved
discrete complex image method based on a special
sampling path,” IEEE Trans. Antennas Propag., vol. 62,
no. 7, pp. 3858–3864, 2014.
[49] H. G. Wang, H. Li, D. Q. Liu and X. H. Yu, “A quasi-three
dimensional Multi-Level Green's function interpolation
method for multilayered structures,” Int. Conf.
Electromagn. Adv. Appl. (ICEAA), pp.125–128, Sydney,
2011.
[50] Y. Shi and C. H. Chan, “An OpenMP parallelized
multilevel Green's function interpolation method
accelerated by fast fourier transform technique,” IEEE
Trans. Antennas Propagat., vol. 60, no. 7, pp.
3305–3313, Jul. 2012.
[51] K. Yang and A. E. Yilmaz, “An FFT-truncated
multilevel interpolation method for multiscale integral
equation analysis,” APS/USNC/USRI Rad. Sci. Meet.,
2011.
[52] M. M. Li, R. S. Chen, H. X. Wang, Z. H. Fan, and Q. Hu,
“A multilevel FFT method for the 3-D capacitance
extraction,” IEEE Trans. Comput.-Aided Des. Integr.
Circuits Syst., vol. 32, no. 2, pp. 318–322, Feb. 2013.
[53] M. Bozzi and L. Perregrini, “Analysis of multilayered
printed frequency selective surfaces by the
MoM/BI-REM method,” IEEE Trans. Antennas
Propag., vol. 51, no.10, pp. 2830–2836, 2003.
[54] R. J. Langley and E. A. Parker, “Double square FSS’s
and their equivalent circuit,” Electron. Lett., vol. 19, no.
17, pp. 675–677, Aug. 1983.
Peng Zhao received his B.Eng degree
and M.Phil degree in electronic
engineering department both from
Zhejiang University, China, in 2006 and
2008, respectively. He earns his Ph.D.
degree in electronic engineering from
City University of Hong Kong, China, in
2014. He is currently a faculty member in
the CAD laboratory at Hangzhou Dianzi
Univeristy. His research interest focuses on computational
electromagnetics.
Da Qing Liu received the B.S. degree in
optical science from Wuhan University of
Science and Technology, Wuhan, China,
in 2008 and the M.S. degrees in electronic
engineering from Zhejiang University,
Hangzhou, China, in 2011. He is
currently pursuing his Ph.D. degree in
Electronic Engineering at City University
of Hong Kong, China. His research
interests include computational electromagnetic and antenna
design.
Chi Hou Chan received the Ph.D.
degree in electrical engineering from the
University of Illinois, Urbana, IL, USA,
in 1987.
From 1989 to 1998, Dr. Chan was a
faculty member in the Department of
Electrical Engineering at the University
of Washington, Seattle, WA, USA. In
1996, he joined the Department of Electronic Engineering,
City University of Hong Kong, and was promoted to Chair
Professor of Electronic Engineering in 1998. From 1998 to
2009, he was first Associate Dean and then Dean of College of
Science and Engineering. He also served as Acting Provost of
the university from July 2009 to September 2010. He is
currently the Director of State Key Laboratory of Millimeter
Waves, Partner Laboratory in the City University of Hong
Forum for Electromagnetic Research Methods and Application Technologies (FERMAT)
14
Kong. His current research interests include computational
electromagnetics, millimeter-wave circuits and antennas, and
terahertz science and technology.
Dr. Chan received the U.S. National Science Foundation
Presidential Young Investigator Award in 1991 and the Joint
Research Fund for Hong Kong and Macao Young Scholars,
National Science Fund for Distinguished Young Scholars,
China, in 2004. He received outstanding teacher awards from
the Department of Electronic Engineering, CityU in 1998,
1999, 2000, and 2008. He is the General Co-Chair of ISAP
2010, iWAT2011, and iWEN 2013.