Progress In Electromagnetics Research, Vol.

127, 277295, 2012

EFFICIENT MODEL ORDER REDUCTION FOR FEM

ANALYSIS OF WAVEGUIDE STRUCTURES AND RESONATORS
G. Fotyga* , K. Nyka, and M. Mrozowski
Faculty of Electronics, Telecommunications and Informatics, Gdansk
University of Technology, 11/12, G. Narutowicza St., Gdansk 80-233,
Poland
AbstractAn efficient model order reduction method for threedimensional Finite Element Method (FEM) analysis of waveguide
structures is proposed. The method is based on the Efficient Modal
Order Reduction (ENOR) algorithm for creating macro-elements in
cascaded subdomains. The resulting macro-elements are represented
by very compact submatrices, leading to significant reduction of the
overall number of unknowns. The efficiency of the model order
reduction is enhanced by projecting fields at the boundaries of macroelements onto a subspace spanned by a few low-order waveguide
modes. The combination of these two techniques results in considerable
saving in overall computational time and memory requirement. An
additional advantage of the presented method is that the reducedorder system matrix remains frequency-independent, which allows
for very fast frequency sweeping and efficient calculation of resonant
frequencies. Several numerical examples for driven and eigenvalue
problems demonstrate the performance of the proposed methodology
in terms of accuracy, memory usage and simulation time.

1. INTRODUCTION
Due to the growing demand for efficient and accurate microwave
circuit simulation tools, model order reduction (MOR) techniques
have recently received considerable attention in computational
electromagnetics. The fundamental idea of MOR, developed originally
in context of Linear Time Invariant dynamical systems, is to replace the
Received 16 February 2012, Accepted 21 March 2012, Scheduled 16 April 2012
* Corresponding author: Grzegorz Fotyga (grzegorz.fotyga@eti.pg.gda.pl).

278

Fotyga, Nyka, and Mrozowski

original large set of state-space equations with a much smaller one in

such a way, that the transfer function derived from the resultant system
still provides satisfactory accuracy. Owing to significant decrease in the
number of unknowns, the main benefit arising from this operation is
the acceleration of computations, as well as the reduction of memory
demand.
Developed over a decade ago, techniques of reduced-order
modeling of large linear [13] and nonlinear [4] electronic circuits
paved the way for applications of MOR in analysis of electromagnetic
fields. Compact Pade approximation of the transfer functions of 3-D
electromagnetic systems in FDTD method, via Lanczos (PVL) and
PRIMA [2] algorithms, was proposed in [5]. This approach, along
with an alternative one based on ENOR algorithm [6], evolved into
the concept of macromodels (or macro-elements). They are generated
by applying model order reduction locally in selected subdomains that
require strong mesh refinement. The efficiency of the macromodeling
technique was demonstrated for finite difference methods in time or
frequency domain [68], as well as for the finite element method [9
11].
The idea of compact macromodels or macro-elements embedded in
the analyzed structure was generalized to the form that serves as a tool
for indirect domain decomposition [12, 13]. The entire computational
domain can be divided into many subdomains, which are thereafter
subject to independent model order reduction. Afterwards, the global
structure is reassembled from the reduced subdomains as a matrix
comprising coupled submatrices of macromodels [14, 15] or as a network
of subcomponents described by generalized admittance matrices [13].
There is a significant difference between these two approaches. The
former leaves the system matrix entries frequency-independent, while
in the latter approach, the system matrix entries become a non-linear
function of frequency.
Numerous efficiency enhancements of MOR techniques have
recently been proposed for various applications, e.g., low-frequency
response of multi-conductor transmission lines [16], ultra wideband
systems such as UWB antenna [17], or PEEC modeling using multipoint moment matching technique [18].
This paper addresses the problem of applying model order
reduction locally via macro-elements in a three-dimensional finite
element analysis of waveguide structures and resonators. It is based on
the general idea presented in [10], where 2-D scalar FEM method and
macro-elements reduced by the ENOR algorithm [3] were used. In this
reduction technique, selected parts of FEM matrices are compressed
one by one by projecting their relevant blocks onto a subspace spanned

Progress In Electromagnetics Research, Vol. 127, 2012

279

by an orthonormal basis, which is generated by the ENOR algorithm

for one, arbitrarily chosen frequency. Unlike in the method using the
PRIMA algorithm and the generalized admittance matrix [9, 13], the
resulting system matrices remain frequency-independent. This is an
important distinguishing feature of our approach, which implies two
fundamental advantages: it is sufficient to make the reduction only
once for wide frequency sweeps and, moreover, this approach can be
employed in eigenvalue problems in order to allow for their solving
solely, by means of standard techniques. So far we have demonstrated
the application of the new type of frequency-independent macroelements in 2-D FEM [10]. A direct extension of the 2-D method to
the 3-D analysis requires introducing 3-D vectorial FEM formulation.
This step is rather straightforward, however, the size of the macroelements in 3-D grows very rapidly, which may lead to efficiency issues
in some cases. It is caused by the fact that the computational cost
of the reduction is dependent on two factors: the size of the macroelement (which is a product of the reduction order) and the size of the
macro-element port (equal to the number of degrees of freedom of the
mesh at the macro-element boundary). Unlike in the 2-D case where
the size of 1-D macro-element boundary can be decreased to a very
small number of nodes by proper mesh coarsening, in 3-D problems
the number of elements at the 2-D boundary mesh is much bigger and
usually significantly affects the reduction performance. Geometrical
coarsening practically does not alleviate this problem, because a good
quality 3-D mesh requires large space for gradual decrease in the mesh
density within the macro-element regions.
In this paper we propose coarsening at macro-element ports by
replacing large FEM basis with a much smaller one consisting of a
few low-order modes. In other words, we decrease the size of an
individual macro-element by projecting the fields of its boundary onto
a subspace spanned by a reduced modal basis. A similar technique
was used in [19] to interface subdomains of FEM and mode matching
analysis. An alternative approach to modal decomposition and model
order reduction in FEM analysis was proposed in [13]. However, the
macromodels therein presented are based on generalized admittance
matrix representation of each subdomain and involve a nonlinear
function of frequency. The main advantage of our method is that no
frequency-dependent terms occur in the matrices of macro-elements,
thus enabling fast frequency sweeping in driven problems as well as
efficient analysis of resonators. In this respect our approach follows
the concepts presented in [20] for the FDTD method, and macromodels
with intermediate projection on functional bases proposed in [15, 24]
for finite-difference methods.

280

Fotyga, Nyka, and Mrozowski

2. THEORY
2.1. Three-dimensional Finite Element Formulation
Consider a source-free bounded 3-D domain . The electric field vector
Helmholtz equation is:
1
~ k02 r E
~ =0
E
(1)
r
~ is the electric field vector, k 2 = 2 0 0 is the wavenumber, is
where E
0
the angular frequency, 0 and 0 are, respectively, the permittivity and
permeability of the free space and r , r are the relative permittivity
and permeability of the medium, respectively. The weak formulation
of (1) is derived using the Galerkin method by multiplying the equation
with the vector basis functions w
~ and integrating the product over the
computational domain [22]:
ZZZ
ZZ
ZZZ
1
1
2
~
~ wdsk
~ = 0 (2)
w
~ Edv+
n
b E
~
w
~ r Edv
0
r
r

where denotes the boundary of , n

b is the outward-pointing unit
normal vector on [21, 22]. We assume the perfect electric conductor
surface (PEC) on all physical walls, therefore the uniform Dirichlet
boundary conditions are imposed on the tangential field components.
The remaining parts of the boundary are regarded as ports of the
structure, where the excitations are applied as nonuniform boundary
conditions [13, 23]. Employing the FEM discretization of using a
tetrahedral mesh and the first-order basis functions, we obtain the
following matrix equation:

K k02 M e = b,
(3)
which can be expressed as Ge = b. K and M are N N sparse
symmetric FEM matrices (stiffness matrix and mass matrix), N is the
total number of variables, e the vector of unknown coefficients of the
FE basis functions associated with mesh edges, and b is the excitation
vector. Once the Equation (3) has been solved, the reflection and
transmission coefficients can be calculated.
2.2. Model Order Reduction Algorithm
First, we shall present extension of model order reduction procedure,
proposed in [10] for 2-D, to 3-D vector problems. To this end a
simple waveguide structure (Figure 1) is considered. The waveguide is
terminated with two ports P1 , P2 , while the domain is divided into

Progress In Electromagnetics Research, Vol. 127, 2012

281

Figure 1. A basic waveguide structure divided into two subdomains:

1 and 2 . Planes P1 , P2 constitute external ports, while plane P3 is
an internal port.
two subdomains 1 and 2 by the interface P3 , henceforth referred
to as an internal port. This is a model problem for a general class of
3-D waveguide structures and resonators divisible into many cascaded
subdomains with transverse cross-section planes as interfaces. The
most appropriate locations for those subdomains are regions which
contain discontinuities, that require strong local mesh refinement. In
practice the size of the subdomain is selected in such a way that its
dimension does not exceed a wavelength at the center frequency. This
limitation implies that the macro-element order is not too large. The
internal ports should be placed in the uniform sections of the waveguide
between discontinuities. Their placement is arbitrary, however, moving
the ports closer to the discontinuities will increase the number of modes
required in the modal projection, which is discussed in the Section 2.3.
If consistent numbering is used for the edges at ports P1 , P2 , P3
and in subdomains 1 and 2 , the matrices and vectors from (3) split
into the following blocks:

KP 1
0
0
KTP 1,1
0

0
KP 2
0
0
KTP 2,2

T
K=
0
0
KP 3
KP 3,1 KTP 3,2

0
KP 3,1
K1
0
P 1,1
0
KP 2,2 KP 3,2
0
K2

T
MP 1
0
0
MP 1,1
0

0
MP 2
0
0
MTP 2,2

M=
0
0
MP 3
MTP 3,1 MTP 3,2

0
M
M
0
P 1,1
P 3,1
1
0
MP 2,2 MP 3,2
0
M2

282

Fotyga, Nyka, and Mrozowski

T
e = [ eP 1 eP 2 eP 3 e1 e2 ]
T
b = [ bP 1 bP 2 0 0 0 ]

(4)

The first three rows and columns of K and M represent the

elements located at the three ports of the structure (two external and
one internal), whereas the last two rows and columns correspond to
the two subdomains. The number of variables in each subvector of e
is: NP 1 , NP 2 , NP 3 , N1 and N2 . The subscripts of the blocks outside
of the main diagonal denote which subdomain is coupled with which
port. Since it is assumed that excitation is imposed on the external
ports, only the first two rows of the vector b are nonzero.
In order to compress 1 , 2 into macro-elements by means of
model order reduction (MOR) [10] using the ENOR algorithm [3],
Equation (4) have to be transformed to the following form:

1
b
Cs0 +
e = Be e0
(5)
s0
This represents the transfer function between the fields b
e (in a
macro-element) and e0 (elsewhere). After multiplication of (4) we
obtain five equations, however only two of them are used in the
reduction scheme, since there are only two subdomains:

KP 1,1 k02 MP 1,1 eP 1 + KP 3,1 k02 MP 3,1 eP 3

+ K1 k02 M1 e1 = 0
(6)

2
2
KP 2,2 k0 MP 2,2 eP 2 + KP 3,2 k0 MP 3,2 eP 3

+ K2 k02 M2 e2 = 0.
(7)
After substituting:
s0 = k0 , C1 = M1 , 1 = K1 ,

1
1
B1 =
KP 1,1 s0 MP 1,1
KP 3,1 s0 MP 3,1
s0
s0

(8)

C2 = M2 , 2 = K2 ,

1
B2 =
KP 2,2 s0 MP 2,2
s0

(9)

and

one obtains:

B1

B2

eP 1
eP 3
eP 2
eP 3

1
KP 3,2 s0 MP 3,2 ,
s0

1
= s0 C1 + 1 e1
s0

1
= s0 C2 + 2 e2 .
s0

(10)
(11)

Progress In Electromagnetics Research, Vol. 127, 2012

283

This form of equations is compliant with the ENOR algorithm.

By applying the algorithm to (10) and (11), so that block moments
expanded about the selected frequency s0 are matched [3, 10], one
obtains two sets of orthonormal vectors forming two orthonormal bases
V1 and V2 . The size of V1 and V2 is N1 NV 1 and N2 NV 2 ,
where NV 1 and NV 2 depend on the reduction order q and the number
of variables on the appropriate ports [10]:
NV 1 = (NP 1 + NP 3 ) q
(12)
NV 2 = (NP 2 + NP 3 ) q.
The minimum order of the reduced macro-element is affected
mainly by the following factors: the size of the macro-element
subdomain, the complexity of the structure embedded in the macroelement and the upper bound of the frequency band. V1 and V2
can be subsequently applied directly to perform orthogonal projection
of the blocks of the FEM Equations (4) that correspond to unknowns
in 1 , 2 . As a result, the reduced system Gv ev = b is obtained,
where:

Gv =

GP 1
0
0
GP 1,1 V1
0
0
GP 2
0
0
GTP 2,2 V2
0
0
GP 3
GTP 3,1 V1 GTP 3,2 V2

T
T
T
G1 V1
0
GP 3,1 V1
GP 1,1
0
V1
V1
T
T
T
0
V2 GP 2,2 V2 GP 3,2
0
V2 G2 V2

T
ev = [ eP 1 eP 2 eP 3 em1 em2 ]
and
T
em1 = V1
e1
T
em2 = V2
e2 .

(13)

(14)

T G V
T
Blocks V1
1 1 and V2 G2 V2 of size NV 1 NV 1 and
NV 2 NV 2 are referred to as macro-element matrices. They are of
a much smaller size than G1 and G2 in the original system, since
NV 1 N1 and NV 2 N2 . As a result, the number of variables
of the problem is significantly reduced, albeit the macro-element
matrices become dense. It should be noted that each macro-element is
created independently, so that the reduction process, which is a critical
contributor to the overall computational load, can be parallelized. An
important advantage is that no additional computations are necessary
to perform the parallel reduction of the subdomains. Unlike in [10],
where the macro-elements were surrounded by unreduced regions, this
formulation allows for reduction in all 3-D subregions that cover the
entire structure. The only subdomains of the original problem that
remain unchanged are the 2-D ports represented by sparse blocks GP k
in the system matrix Gv .

284

Fotyga, Nyka, and Mrozowski

In the ENOR algorithm the projection vectors V are calculated for

one arbitrarily chosen frequency s0 and remain frequency-independent.
Thus it is sufficient to perform the reduction only once for very wide
frequency range of analysis [3, 10]. This highly attractive property
of the V vectors leads to yet another important advantage of the
presented procedure. The reduced system (12) can be reverted to the
original form:

Kv k02 Mv ev = b,
(15)
where the reduced FEM matrices Kv , Mv are calculated in the
same way as Gv . Since the matrices Kv and Mv remain frequencyindependent, the proposed method can also be applied in calculation
of the resonant frequencies of the cavities. To this end one sets b = 0
and solves the resulting generalized eigenvalue problem. If one used
the segmentation and model order reduction approach adopted in [13],
finding the resonances would require solving a nonlinear equation,
which is a very time-consuming process.
Post-processing operations (such as computations of far field,
input impedance and currents) can be applied to the MOR solution
(15). To this end one has to project the field obtained using (15) back
onto the original space, using V.
2.3. Modal Projection
The efficiency of the reduction process depends on the following factors:
number of variables inside the macro-element subregion, the order of
reduction and the number of FEM mesh elements at the interface that
couples the macro-element with its surroundings. Whereas the first
two factors are implied by the required approximation accuracy of
the macro-element, the number of variables at the internal ports (12)
should be made as low as possible. Although in 2-D problems it can
be limited to less than 10 [10], these numbers are much bigger for
3-D structures, even if aggressive mesh coarsening is used. A large
number of variables at macro-element ports in 3-D affects not only the
reduction process but also the size of dense submatrices of the reduced
macro-elements, eventually leading to deterioration in the efficiency of
solving the resulting system.
One of the approaches to overcome this limitation is to apply the
projection at the interfaces of the macro-element subdomains before
applying the MOR formulation, as postulated in [15] and [24]. For
the model problem considered here (Figure 1) the interfaces of the
subdomains are the cross-sections of the waveguide denoted as ports
P1 , P2 , P3 . The distribution of the tangential electric field at the ports
can be expressed as modal expansion based on orthogonal waveguide

Progress In Electromagnetics Research, Vol. 127, 2012

285

TE and TM modes. These modes are used to project the fields at

the interfaces prior to creating a macro-element, thereby reducing
the number of variables at the interfaces. Although in the case of
a simple problem with a uniform rectangular waveguide the modal
basis is known analytically and is the same for each port, the proposed
procedure does not rely on such limiting assumption. For an arbitrary
shape of port that is transversal to the local section of the waveguide,
the modal basis is computed at each port separately as a solution of 2D vector FEM eigenvalue problem [25]. Once NP k unique eigenvalues
and eigenvectors have been calculated, the modal expansion of the
tangential electric field is:
E(k) =

N
Pk
X

(k) (k)

ai et,i ,

(16)

i=1
(k)

where k and i denote port and mode indices, respectively, ai is the

(k)
amplitude of the i-th mode and et,i is the vector of modal function
discretized on a 2-D FEM mesh. If the 2-D mesh at the port conforms
to the 3-D mesh in adjacent volumes, the resulting basis can be directly
used for modal projection of the fields at macro-elements ports.
eP k lowest
In practice, the expansion (16) can be truncated to N
modes without sacrificing the solution accuracy, since higher order
modes are heavily attenuated. The required number of modes depends
on the extent to which the discontinuities disturb the field distribution
in uniform sections of the waveguide. Usually the macro-element ports
are sufficiently distanced from the embedded discontinuities and most
of high-order modes, called localized [26] vanish before reaching the
ports. Only a few modes, called accessible [26], have to be taken into
eP k NP k . The required value of N
eP k depends on
account, thus N
the field distribution around the discontinuities which is not known
a priori. Therefore it can be found by successive evaluation of the
eP k .
structure response for the varying N
In order to apply the modal projection using the basis consisting
of a truncated set of eigensolutions of 2-D problems at internal and
external ports, one has to multiply appropriate blocks of matrices in
K (matrix M is treated in the analogous way) as follows:
e
K

eT

eP1
e TP 1 KTP 1,1
0
0
0
E
EP 1 K P 1 E
eP2
e TP 2 KTP 2,2
e TP 2 KP 2 E

0
0
E
0
E

e TP 3 KP 3 E
eP3 E
e TP 3 KTP 3,1 E
e TP 3 KTP 3,2
=
0
0
E
.

K
e
e
E
0
K
E
K
0
P 1,1 P 1
P 3,1 P 3
1
e P 2 KP 3,2 E
eP3
0
KP 2,2 E
0
K2

(17)

286

Fotyga, Nyka, and Mrozowski

As a result of modal projection, the unknown FEM coefficients

of the basis functions in vectors eP 1 , eP 2 and eP 3 are replaced with
eP k unknown
vectors eP 1 , eP 2 and eP 3 , each of them comprising N
(k)
amplitudes of the waveguide modes ai .
e = [e
eP 1

eP 2

eP 3

e1

e2 ]T

(18)

eP k NP k , k {1, 2, 3}, the modal projection significantly

Since N
reduces the number of unknowns at the ports of macro-elements.
The column ranks of orthonormal bases V1 and V2 as well as the
eV 1 and N
eV 2 decrease to the same
resulting sizes of macro-elements N
extent:

eV 1 = N
eP 1 + N
eP 3 q
N

(19)
eP 2 + N
eP 3 q.
eV 2 = N
N
For homogeneous ports, the modal expansion does not introduce
e and M
f
frequency-dependent terms, therefore the reduced matrices K
remain frequency-invariant.
Without prior knowledge of the response, the minimum order of
the model order reduction and the truncation of the modal expansion
can be determined by an iterative approach similar to that described
in [27]. At each iteration the model order or the expansion order
is increased and the response is compared with that obtained in the
previous iteration until convergence is achieved. Such a procedure is
fast and results in a very small time overhead.
3. NUMERICAL RESULTS
Three 3-D problems are investigated in order to test the performance
of the proposed method: a multiple-post filter, a dielectric-loaded filter
and a resonator consisting of two coupled dielectric-loaded cavities. All

Figure 2. Multiple-post filter. (a = 10.66 mm, b = 11.62, c = 1 mm,

d = 4.07 mm, e = 4.71 mm). The internal ports are depicted by the
dashed lines, the excitation ports by the dotted lines.

Progress In Electromagnetics Research, Vol. 127, 2012

287

calculations were performed using Matlab on a computer with an Intel

i7 processor and 8-GB RAM.
3.1. Example 1: Multiple-post Filter
In the first example the analyzed structure is the WR-62 waveguide
(15.799 mm 7.899 mm) loaded with nine metallic posts of square
cross-section (Figure 2). The posts are uniform in the y-direction,
which implies that the field does not vary in this direction, so in
fact, the structure could be analyzed in 2-D. However, the purpose
of this experiment was to examine the validity of the 3-D MOR
algorithm, therefore a 3-D formulation was applied. The structure
Table 1. Comparison of the numerical results of the multiplepost filter analysis for the standard FEM and the FEM with macroelements.
Number of unknowns

FEM

FEM with macro-elements

362,839

276

Reduction time

68 sec

Solution time per freq.

54 sec

0.0036 sec

Solution time for

the whole characteristic (73 fp)

66 min

0.262 sec

Figure 3. S-parameters of the multiple-post filter simulated using

HFSS, standard FEM and FEM with macro-elements.

288

Fotyga, Nyka, and Mrozowski

was discretized using 345,463 tetrahedral elements, which resulted in

362,839 FEM degrees of freedom. First, the standard first-order FEM
analysis without macro-elements was carried out in order to generate
reference results for further tests concerning effects of model order
reduction. The aim of the simulation was to compute the transmission
and reflection coefficients over the range of 15 to 17 GHz at 73 frequency
points (fp) assuming the TE10 mode excitation. The computation time
was approximately 54 sec for each frequency point, totaling to 66 min
for the whole characteristic (Table 1). The obtained results, shown in
Figure 3 are in very good agreement with those computed by means
of commercial FEM software Ansys HFSS [28].
Afterwards, the same structure was analyzed by means of the
proposed model order reduction algorithm applied to the same mesh
and FEM formulation as in the reference analysis. The computational
domain was split into five subdomains by placing 4 interfaces in the
middle between the groups of metallic posts. This division implies that
there are 6 ports of the macro-elements, including 2 excitation ports.
Fields at each port were projected onto modal subspaces spanned by
first 6 TE waveguide modes. The 4th order of reduction was used,
resulting in the size of each macro-element (the number of unknowns)
equal to 48. The overall number of unknowns was 276, which is
over three orders of magnitude less than in the FEM without macroelements. The reduction process takes from 11 to 16 sec for each macroelement, depending on the number of discontinuities embedded. After
the reduction, the computation time for one fp decreased from 54 sec
to as few as 0.0036 sec, which resulted in total 0.262 sec for all 73 fp

Figure 4. Four-pole dielectric-loaded cavity filter, t = 0.5 mm,

b = 9 mm. All other dimensions are given in [29]. Internal ports for
one section of the filter are depicted by the dashed lines in the inset.

Progress In Electromagnetics Research, Vol. 127, 2012

289

(Table 1). The resulting characteristics remain in excellent agreement

with the standard FEM formulation over the entire analysis bandwidth
(Figure 3). Counting together the total solution time and the reduction
time of all 5 macro-elements, the overall speedup is from 66 min to
69 sec, which means the complete calculation was about 57 times faster.
3.2. Example 2: Dielectric-loaded Cavity Filter
For the second example a more complex structure was chosen, namely a
four-pole dielectric-loaded cavity filter, shown in Figure 4. The purpose
of this test is to demonstrate how the proposed method copes with
the structures containing elements that significantly disturb the field
distribution. In particular, this test addresses the issue of placing
macro-elements near such discontinuities. Four resonators are coupled
by rectangular irises and contain large cylindrical dielectric pucks
(r = 30) on dielectric supports (r = 9). The input and output
waveguide is WR-75 (19.05 mm 9.52 mm). All other dimensions of
the filter are given in [29].
As in the first example, in order to generate reference results, the
filter was analyzed by means of the standard FEM formulation, using
798,175 degrees of freedom. Such a large problem resulted from strong
local mesh refinement required in big volumes, which was caused by
the presence of very high permittivity dielectric cylinders of relatively
large diameter. The scattering parameters of the filter were computed

Figure 5. S-parameters of the dielectric-loaded resonator filter

simulated using standard FEM, FEM with macro-elements and
compared to the measurements [29].

290

Fotyga, Nyka, and Mrozowski

Table 2. Comparison of the performance of the dielectric-loaded

cavity filter analysis for the standard FEM and the proposed algorithm
of FEM with macro-elements.
Standard FEM

FEM with macro-elements

Number of unknowns

798,175

550

Reduction time

570 sec

Simulation time per freq.

68 min

0.0144 sec

Simulation time for the

whole characteristic in 55 fp

400 min

0.79 sec

at 55 fp covering the bandwidth of 11 to 12 GHz. Good agreement with

experimental results reported in [29] is achieved, as shown in Figure 5.
The computation time for one frequency is approximately 68 min,
which results in about 400 min for the whole frequency response.
Upon setting the reference results, the filter was analyzed using
the proposed method. Five internal ports were placed in the middle of
rectangular coupling irises. This partitioning gives 6 subdomains-all
subject to the model order reduction, which turns them into 6 macroelements. The dominant source of field disturbances in the structure
are the dielectric cylinders. Since they excite higher modes closer to
macro-element ports than in the first example, longer modal expansion
is required. The first 10 waveguide modes were experimentally selected
for modal projection and found to provide sufficient accuracy of
approximation while at the same time yield radical decrease in the
number of unknowns at the ports. With the reduction order q = 4,
the overall number of unknowns in the whole domain decreases from
798,175 to just 550. The reduction process for each of the macroelements takes from 54 to 102 sec. The solution of the reduced system
and computation of the scattering parameters takes 0.0144 sec per each
frequency and 0.792 sec for the whole characteristic (55 fp). As in the
first example, one can observe very good agreement between the results
obtained in the standard FEM and FEM with macro-elements. The
comparison of the problem size and computation time for these two
methods is presented in Table 2. Even though the computation time
is dominated by the time of reduction (54 sec to 102 sec per one ME),
the overall speedup for the whole 55 point frequency response is from
400 min to 570 sec, which is an impressive number of 42 times.
3.3. Example 3: Dielectric-loaded Resonator
The purpose of the last example is to illustrate the capability of the
proposed method to solve eigenvalue problems in resonant frequency

Progress In Electromagnetics Research, Vol. 127, 2012

291

Figure 6. Two-cavity dielectric-loaded resonator based on the filter

presented in Figure 4, a1 = 6.91 mm, a2 = 7.93 mm, b = c = 9 mm
and t = 0.5 mm. Other parameters are identical as for the structure in
Figure 4. The internal port is depicted by the dashed line.
Table 3. Comparison of the numerical results of the dielectric-loaded
resonator analysis for the standard FEM, FEM with macro-elements
and HFSS.
HFSS Results
(GHz)

FEM without
ME (GHz)

FEM with
ME (GHz)

Error (%)
HFSS/FEM

Error (%)
FEM/FEM-ME

11.6857
11.8817

11.6303
11.8316

11.6303
11.8316

0.47
0.42

0.0065
0.0007

calculations. The analyzed structure is a resonator created from two

coupled cavities, as in the previous example. The geometry is detailed
in Figure 6. The discretization of the structure results in 218,283
FEM degrees of freedom. For the generalized eigenvalue problem and
the standard FEM formulation, which serves as a reference for the
proposed method, it takes 842 sec to compute the first two resonant
frequencies by means of the Implicitly Restarted Arnoldi Method
(IRAM).
The domain partitioning for applying the FEM with MOR is
accomplished by placing one interface in the middle of the coupling
iris between the two cavities, as in the previous example, which creates
two macro-elements. A satisfying agreement with the reference results
(Table 3) requires projecting the field at the interface onto a modal
subspace spanned by the first 11 waveguide modes and setting the
order of reduction in the two subdomains to q = 6. These numbers
are slightly bigger than in the filter example. It can be attributed
to the fact that for eigenvalue problems better approximation is

292

Fotyga, Nyka, and Mrozowski

needed, as here the field distribution is not constrained by fundamental

mode excitation present in driven problems. The total number of
unknowns of the resulting system is 143 and the reduction times for
the subdomains are 42.4 and 64.1 sec. The same eigensolver (IRAM)
was used for the reduced problem and the solution time proved to
be as short as 0,044 sec. To assess the overall efficiency of the FEM
with macro-elements the reduction time has to be taken into account,
but nonetheless the total speedup is as much as about 8 times. The
resonator was also simulated using the HFSS to validate the results
obtained with the presented method. The results are compared in
Table 3. It was observed that the error introduced by the reduction
was practically negligible. The results from standard FEM, as well as
those from the proposed MOR method, agree with the HFSS results.
The occurrence of small error of less than 0.5% can be attributed to
better FEM approximation in HFSS, in which higher-order curvilinear
elements and adaptive mesh refinement are employed.
4. CONCLUSION
An efficient model order reduction method for three-dimensional
FEM analysis of waveguide structures has been proposed. The
computational domain is divided into cascaded subdomains, in
which model order reduction is performed locally by means of the
ENOR algorithm. The resulting macro-elements are represented
by submatrices of significantly reduced size, which allows for faster
solution of the underlying system of equations. In order to additionally
improve the effectiveness of the proposed method, fields at macroelement interfaces are projected onto a modal subspace spanned by a
few low-order waveguide modes. Thus, by decreasing the number of
unknowns at macro-element ports, the reduction itself is considerably
accelerated and the resulting matrices become even more compressed.
An additional advantage of the presented method is that the reducedorder system matrix is frequency-independent, which allows for very
fast frequency sweeping and efficient calculation of resonant frequencies
without the need of nonlinear eigensolvers. Several numerical examples
for driven and eigenvalue problems have demonstrated very good
performance of the proposed methodology in terms of accuracy,
memory usage and simulation time.
ACKNOWLEDGMENT
The authors would like to thank Dr. A. Lamecki for providing a
simplified Matlab version of the finite element code which was used

Progress In Electromagnetics Research, Vol. 127, 2012

293

in the research reported in this paper and Dr. L. Kulas for his
valuable assistance and providing the Matlab code of the ENOR
algorithm. This work has been financed by MNiSzW under grant
5407/B/T02/2010/38.
REFERENCES
1. Celik, M. and A. C. Cangellaris, Simulation of dispersive
multiconductor transmission lines by Pade approximation via the
Lanczos process, IEEE Trans. Microwave Theory Tech., Vol. 44,
25252535, Dec. 1996.
2. Odabasioglu, A., M. Celik, and L. T. Pileggi, PRIMA: Passive
reduced order interconnect macromodeling algorithm, IEEE
Trans. Computer-Aided Design, Vol. 17, 645653, Aug. 1998.
3. Sheehan, B. N., ENOR: Model order reduction of RLC circuits
using nodal equations for efficient factorization, Proc. IEEE 36th
Design Autom. Conf., 1721, Jun. 1999.
4. Rewienski, M. and J. White, A trajectory piecewise-linear
approach to model order reduction and fast simulation of nonlinear
circuits and micromachined devices, IEEE Transactions on
Computer-Aided Design of Integrated Circuits and Systems,
Vol. 22, No. 2, 155170, Feb. 2003.
5. Cangellaris, A. C., M. Celik, S. Pasha, and L. Zhao, Electromagnetic model order reduction for system-level modeling, IEEE
Trans. Microwave Theory Tech., Vol. 7, 840850, Jun. 1999.
6. Denecker, B., F. Olyslager, L. Knockaert, and D. De Zutter,
Automatic generation of subdomain models in 2-D FDTD using
reduced order modeling, IEEE Microwave Guided Wave Lett.,
Vol. 10, 301303, Aug. 2000.
7. Kulas, L. and M. Mrozowski, Reduced-order models in FDTD,
IEEE Microw. Wireless Comp. Lett., Vol. 11, No. 10, 422424,
Oct. 2001.
8. Kulas, L. and M. Mrozowski, Reduced order models of refined
Yees cells, IEEE Microw. Wireless Comp. Lett., Vol. 13, 164
166, Apr. 2003.
9. Zhu, Y. and A. C. Cangellaris, Macro-elements for efficient FEM
simulation of small geometric features in waveguide components,
IEEE Trans. Microwave Theory Tech., Vol. 48, 22542260,
Dec. 2000.
10. Fotyga, G., K. Nyka, and L. Kulas, A new type of macro-elements
for efficient two-dimensional FEM analysis, IEEE Antennas and
Wireless Propagation Letters, Vol. 10, 270273, 2011.

294

Fotyga, Nyka, and Mrozowski

11. Lee, S.-H. and J. M. Jin, Fast reduced-order finite-element

modeling of lossy thin wires using lumped impedance elements,
IEEE Trans. Adv. Packag., Vol. 33, No. 1, 212218, Feb. 2010.
12. Kulas, L. and M. Mrozowski, Accelerated analysis of resonators
by a combined domain decomposition Model order reduction
approach, 34th European Microwave Conference, Vol. 2, 585588,
Oct. 14, 2004.
13. De la Rubia, V. and J. Zapata, Microwave circuit design by
means of direct decomposition in the finite-element method,
IEEE Trans. Microwave Theory Tech., Vol. 55, No. 7, 15201530,
Jul. 2007.
14. Kulas, L. and M. Mrozowski, Macromodels in the frequency
domain analysis of microwave resonators, IEEE Microw. Wireless
Comp. Lett., Vol. 14, No. 3, 9496, Mar. 2004.
15. Kulas, L., P. Kowalczyk, and M. Mrozowski, A novel modal
technique for time and frequency domain analysis of waveguide
components, IEEE Microw. Wireless Comp. Lett., Vol. 21, No. 1,
79, Jan. 2011.
16. Remis, R. F., An efficient model-order reduction approach
to low-frequency transmission line modeling, Progress In
Electromagnetics Research, Vol. 101, 139155, 2010.
17. Zhang, Z. and Y. H. Lee, An automatic model order reduction of
a UWB antenna system, Progress In Electromagnetics Research,
Vol. 104, 267282, 2010.
18. Song, Z., D. Su, F. Duval, and A. Louis, Model order reduction
for PEEC modeling based on moment matching, Progress In
Electromagnetics Research, Vol. 114, 285299, 2011.
19. Rubio, J., J. Arroyo, and J. Zapata, Analysis of passive
microwave circuits by using hybrid 2-D and 3-D finite-element
mode-matching method, IEEE Trans. Microwave Theory Tech.,
Vol. 47, No. 9, 17461749, Sep. 1999.
20. Mrozowski, M., A hybrid PEE-FDTD algorithm for accelerated
time domain analysis of electromagnetic waves in shielded
structures, IEEE Microwave Guided Wave Lett., Vol. 4, No. 10,
323325, Oct. 1994.
21. Jin, J. M., The Finite Element Method in Electromagnetics, 2nd
edition, IEEE Press, New York, 2002.
22. Pelosi, G., R. Coccioli, and S. Selleri, Quick Finite Elements
for Electromagnetic Waves, 2nd Edition, Artech House Antenna
Library, 2009.

Progress In Electromagnetics Research, Vol. 127, 2012

295

23. Ingelstrom, P., A new set of H(curl)-conforming hierarchical basis

functions for tetrahedral meshes, IEEE Trans. Microwave Theory
Tech., Vol. 54, No. 1, 106114, Jan. 2006.
24. Kowalczyk, P., L. Kulas, and M. Mrozowski, Analysis of
microstructured optical fibers using compact macromodels, Opt.
Express, Vol. 19, No. 20, 1935419364, 2011.
25. Lou, Z. and J. M. Jin, An accurate waveguide port boundary
condition for the time-domain finite-element method, IEEE
Trans. Microwave Theory Tech., Vol. 53, No. 9, 30143023,
Sep. 2005.
26. Stamatopoulos, I. D. and I. D. Robertson, Rigorous network
representation of microwave components by the use of indirect
mode matching, IEEE Trans. Microwave Theory Tech., Vol. 52,
No. 3, 935944, Mar. 2004.
27. Zhu, Y. and A. C. Cangellaris, Multigrid Finite Element Methods
for Electromagnetic Field Modeling, Wiley, New York, 2006.
28. ANSYS HFSS, 3D full-wave electromagnetic field simulation,
http://www.anasoft.com/products/hf/hfss/overview.cfm.
29. Alessandri, F., M. Chiodetti, A. Giugliarelli, D. Maiarelli,
G. Martirano, D. Schmitt, L. Vanni, and F. Vitulli, The electric
field integral-equation method for the analysis and design of a
class of rectangular cavity filters loaded by dielectric and metallic
cylindrical pucks, IEEE Trans. Microwave Theory Tech., Vol. 52,
No. 8, 17901797, Aug. 2004.