You are on page 1of 4

382

IEEE ANTENNAS AND WIRELESS PROPAGATION LETTERS, VOL. 14, 2015

Automated Reduced Model Order Selection


Micha Rewieski, Member, IEEE, Grzegorz Fotyga, Adam Lamecki, Member, IEEE, and
Micha Mrozowski, Fellow, IEEE

AbstractThis letter proposes to automate generation of


reduced-order models used for accelerated -parameter computation by applying a posteriori model error estimators. So far,
a posteriori error estimators were used in Reduced Basis Method
(RBM) and Proper Orthogonal Decomposition (POD) to select
frequency points at which basis vectors are generated. This letter
shows how a posteriori error estimators can be applied to automatically select the order of the reduced model in second-order Model
Order Reduction (MOR) methods. Three different error estimators
are investigated and compared in order to arrive at a new MOR
scheme that is fast, reliable, and fully automated. The effectiveness
of the proposed approach is verified by very high accuracy of the
computed scattering parameters ( -parameters) for an example
of a waveguide filter over a prescribed frequency band.
Index Termsa posteriori error estimator, model order
reduction, -parameter computation.

I. INTRODUCTION

ODEL order reduction (MOR) is a powerful technique


of reducing the numerical cost of spectral analysis of
linear time-invariant (LTI) dynamical systems. It uses projection of the original large set of state-space equations onto a
much smaller one. In the moment-matching MOR, the projection space is constructed so that the value and derivatives (moments) up to a certain order of the transfer function derived from
the resultant projected system and the original one are matched
at a certain frequency (expansion point). MOR has been shown
to yield substantial speed gains in RF circuit design for interconnects analysis [1] and also in computational electromagnetics for the finite element method (FEM) [2] and the finite-difference time-domain (FDTD) method [3]. MOR has also been
applied for selected regions in mesh-based techniques (FDFD,
FDTD, FEM) [4][6] to reduce the numerical cost of problems
involving very fine local meshes.
Spectral analysis (frequency sweep) of large dynamical
systems can also be accelerated using the reduced basis method
(RBM) and proper orthogonal decomposition (POD). In both
these techniques, solution snapshots taken at different frequencies are found, and a projection subspace is constructed
from these snapshots. RBM and POD do not have the moment-matching feature of MOR, but have been shown to provide excellent speedups in wideband FEM simulations [7][9].

One advantage of MOR over RBM and POD is that not only
is a certain number of moments matched, but also projection
space can be found with a single matrix factorization. On the
other hand, one of the key and long-standing problems of MOR
based on moment-matching is finding a reliable criterion for
selecting the order of the reduced model that would guarantee
desired accuracy of the model. Automated model order selection has been available only for certain specific problems
such as
circuits (modeled by first-order systems) and
used conservative a priori model error estimates or heuristic
convergence detection [10]. However, no practical model order
selectors have been proposed for moment-matching-based
MOR of second-order systems used to accelerate computation
of -parameters. Yet, for macromodeling methods based on
RBM and POD, a few a posteriori model error estimators
have been developed for second-order electromagnetic systems
discretized using FEM [7][9], [11]. Their key advantage is
that they all inexpensively compute the model error estimators by exploiting the knowledge of the approximate solution
found using the macromodel. So far, those estimators were
used only in the context of RBM and POD to automate or
optimize selection of frequency points used to generate the
projection basis vectors. However, the discussed estimators
were never applied in moment-matching-based second-order
MOR methods such as SAPOR [12]. The goal of this letter is to
investigate how the available a posteriori error estimators can
be used to automate second-order MOR, and establish which
of the estimators is most effective for the problem of generating reduced-order models used for high-accuracy accelerated
-parameter computations.
II. FORMULATION OF ELECTROMAGNETICS PROBLEMS
TERMS OF SECOND-ORDER SYSTEMS

IN

To introduce MOR, RBM, and POD algorithms, let us consider FEM applied to the vector Helmholtz equation. Dielectrically loaded cavity including ports that excite the electromagnetic field may be described using the following weak E-field
formulation [13], [14]:

(1)
Manuscript received August 22, 2014; accepted October 16, 2014. Date of
publication October 23, 2014; date of current version February 05, 2015. This
work was supported by the under Polish National Science Centre under Agreement #2012/07/B/ST7/01241.
The authors are with the Department of Electronics, Telecommunications and
Informatics, Gdask University of Technology, 80-233 Gdask, Poland (e-mail:
mrewiens@eti.pg.gda.pl).
Color versions of one or more of the figures in this letter are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/LAWP.2014.2364849

where
is a volume enclosing the modeled system,
is
a vector testing function,
is the electric field, and are
relative permittivity and permeability, respectively,
is the
wavenumber, is the angular frequency,
is the number of
ports,
is the surface of the th port,
is the number of
modes excited at the th port,
is the amplitude of the th
mode (excited at port ), is a vector normal to surface , and

1536-1225 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

REWIESKI et al.: AUTOMATED REDUCED MODEL ORDER SELECTION

383

is the tangential magnetic field for the th mode at port .


If finite-element (FE) discretization is applied, both and
are expanded as
(2)
where
are FEM vector interpolation functions. Substituting
the above expansion to (1) yields the second-order system
(3)
system matrices,
where and are
is the vector of weights that specify the electric field in the
cavity, is the input vector of amplitudes for excitation modes
(
) port selection
at ports, and is the
matrix, whose columns are expressed as
(4)
If one defines the complex frequency parameter
, then system (3) transforms into
(5)
Applying normalization
and
,
where
is the port impedance (at some given frequency ),
and specifying the output at the ports in terms of amplitudes of
the voltage waves
yields
(6)

related to the reduced-order vector


through the projection
matrix :
.) Assuming unit input wave
amplitudes, the discrete analogue of (8) is the residual error
matrix
(9)
, and the
where matrix
residual vector corresponds to the excitation of the th mode
at the th port.
In RBM and POD, the frequencies for projection basis construction are selected based on a greedy algorithm that requires
repetitive evaluation of error estimators. The error estimator
can be related to the residual (9) in several ways, but since
the direct computation of (9) at many frequency points is timeconsuming, the resulting expressions cannot involve high-order
space computation.
In [8], the error estimator is a norm of the residual error function (8), which can be computed (in the dual space) as
(10)
and is a positive-definite matrix with
where
elements
. Reference [8]
describes in detail how the above estimator can be computed
efficiently. A simplified version of the above estimator has been
proposed in [11] and computes Euclidean norm of the residual
error
(11)

III. A POSTERIORI ESTIMATORS FOR REDUCED-ORDER


MODEL ERRORS
Applying MOR, RBM, or POD to (6) yields a reduced-order
system
(7)
, and
are the
where
reduced-order system matrices, is the
projection matrix,
and
is the reduced-order internal state vector of order .
Although the projected system looks identical for all three
approaches, what sets these methods apart is the way the projection bases are constructed. In MOR, the projection vectors
are constructed so that the moments of the original and projected systems at a selected frequency point are matched. On
the other hand, the moment-matching property, which deemphasizes the importance of the internal states in favor of the system
inputoutput behavior, is absent in RBM and POD. As proposed
in [7][9] and [11], the projection basis in the latter methods
is generated from snapshot solutions found at several frequency
points selected by using the computed a posteriori model error
estimators. The starting point for all the considered error estimators is a residual error function defined as

Finally, the third approach proposed in [7] defines a goal


function, geared toward -parameter error estimation, based on
computing the residual current interaction with the port modal
fields. This function that describes the impact of the residual
current (excited by the th mode at the th port) on the th
mode of the th port is specified as

where
is the residual field due to excitation of mode at
port ,
is the electric field for the th mode for the th port,
and is a port impedance normalization for port . Expanding
the residual field in terms of the vector basis functions [cf. (2)
and (4)] yields
(13)
Using the above gives a matrix formula for the goal function
involving all the ports and all the modes
(14)

(8)
where
is the approximate electric field computed using the reduced-order model (7). (The weights
are

is a diagonal port
where
impedance normalization matrix. Finally, the error estimator is
computed as the normalized goal function
(15)

384

IEEE ANTENNAS AND WIRELESS PROPAGATION LETTERS, VOL. 14, 2015

Fig. 1. MOR with automated order selector.

Fig. 3. Values of
in the inset).

and

for the fourth waveguide filter example (shown

Fig. 2. ErrEst: fast error estimator.

It can be shown that the above estimator can be evaluated inexpensively by performing computations solely in the reducedorder state space.
IV. MOR WITH AUTOMATED ORDER SELECTION
The a posteriori error estimators presented in Section III have
been originally proposed to guide an algorithm for selecting
frequency points for solution snapshots. They can, however,
be used also as stopping criteria for the model order reduction
process by computing the error estimates over a specified frequency range
and requiring that those estimates do
not exceed a prespecified tolerance.
The entire scheme with automated order selection based on
an arbitrary moment-matching MOR technique and one of the
three fast error estimators outlined in Section III is summarized
in pseudocodes in Figs. 1 and 2.
V. NUMERICAL VERIFICATION OF THE ERROR ESTIMATORS
A series of numerical tests were performed that investigate
the quality of the three different estimators for reduced model
errors presented above. The goals of the tests were to compare
accuracy of the error estimators and to determine the robustness
and usability of the error estimators as automated model order
selectors.
To this end, reduced-order models were generated for a
fourth-order E-plane metal inset WR90 waveguide filter shown
in Fig. 3. The inset lengths equaled 5, 15, 15, 15, and 5 mm. The
inset widths equaled 2 mm. The four waveguides (resonators)
connecting the insets were each 15 mm long. The reference
filter response computed using FEM model is shown in Fig. 3.
In this letter, (block) SAPOR [12] moment-matching MOR
algorithm has been used since it provides a numerically stable
method for generating reduced-order models for second-order

Fig. 4. Estimated and actual absolute error for


.
reduced model of order

values computed using a

systems, and it also readily extends to second-order systems


with losses.
Fig. 4 compares the three error estimates with the actual error
for
values computed with a reduced-order model of order
over the specified frequency range. (The actual error has
been computed using the reference -parameter values computed with the full-order model.) First, note that the reduced
model provides excellent accuracy over the frequency range,
as confirmed by the values of the actual error. It is also apparent that the goal-oriented error estimator
(15) provides
a much more accurate approximation of the actual -parameter error as compared to the two other estimators (10) and (11)
based on residual norms over the entire frequency range.
Further experiments investigated accuracy of the error
estimators for different sizes of the reduced-order models.
Fig. 5 shows the maximum value of each error estimator over
the frequency range as a function of the model order . Also,
the actual maximum error for the -parameters computed
with the reduced-order models is shown. It follows from the

REWIESKI et al.: AUTOMATED REDUCED MODEL ORDER SELECTION

385

VI. CONCLUSION
This letter proposed a method for automated selection of
the model order during MOR process, based on computing
a posteriori error estimates. Three different error estimators, so
far used only in RBM and POD, were considered jointly with
SAPOR reduction algorithm as stopping criteria. Numerical
tests indicate that applying one of the estimators (the goal-oriented estimator) yields a reliable and fully automated model
order selector. It is shown that the proposed automated MOR
approach generates reduced models that very accurately approximate -parameters of the initial system over a prescribed
frequency range. The described automated MOR algorithm
can also be readily extended to systems that include losses.
Finally, the goal-oriented estimator can be easily adapted
to reduction algorithms based on the finite difference method
(e.g., [5] and [6]) and applied in FDTD analysis.
Fig. 5. Estimated and actual absolute maximum error for -parameters versus
order of the used reduced model.
TABLE I
FINAL ORDER OF THE REDUCED MODEL SELECTED BY DIFFERENT
STOPPING CRITERIA. THE REFERENCE OPTIMAL ALGORITHM
USED EXACT -PARAMETER ERROR IN ITS STOPPING CRITERION

graph that while for lower orders of the reduced model, all
three estimates reasonably approximate the actual error, for
higher orders, the estimates (10) and (11) substantially diverge
from the actual error. Only the goal-oriented estimator (15)
consistently follows the actual -parameter error.
This result indicates that error estimates (10) and (11) unfortunately cannot provide robust stopping criteria for the model
order reduction process. Only the estimator (15) can be used to
automate the model order selection. This is confirmed by the results from Table I, which shows the orders of the reduced models
produced by Algorithm 1 for different values of the target tolerance (tol parameter) when different error estimates are applied. As expected, only the goal-oriented estimator provides
a usable stopping criterion for different tolerances, while the
other two estimators lead to false nonconvergence of the MOR
process. The table also shows the order of the reduced model
selected using an optimal model MOR algorithm with stopping criterion based on actual, exact -parameter error. (This
algorithm clearly cannot be used in practice since it requires an
a priori knowledge of the exact solution.) This result confirms
that Algorithm 1 with goal-oriented error estimator yields reduced models of orders that are close to the optimal orders for
the considered test structure and frequency range.

REFERENCES
[1] A. Odabasioglu, M. Celik, and L. Pileggi, PRIMA: Passive reduced
order interconnect macromodeling algorithm, IEEE Trans. Comput.Aided Design, vol. 17, no. 8, pp. 645653, Aug. 1998.
[2] S.-H. Lee and J. Jin, Fast reduced-order finite-element modeling of
lossy thin wires using lumped impedance elements, IEEE Trans. Adv.
Packag., vol. 33, no. 1, pp. 212218, Feb. 2010.
[3] A. Cangellaris, M. Celik, S. Pasha, and L. Zhao, Electromagnetic
model order reduction for system-level modeling, IEEE Trans. Microw. Theory Tech., vol. 47, no. 6, pp. 840850, Jun., 1999.
[4] G. Fotyga, K. Nyka, and M. Mrozowski, Multilevel model order reduction with generalized compression of boundaries for 3-D FEM electromagnetic analysis, Prog. Electromagn. Res., vol. 139, pp. 743775,
2013.
[5] B. Denecker, F. Olyslager, L. Knockaert, and D. De Zutter, Generation of FDTD subcell equations by means of reduced order modeling,
IEEE Trans. Antennas Propag., vol. 51, no. 8, pp. 18061817, Aug.
2003.
[6] L. Kulas and M. Mrozowski, A fast high-resolution 3-D finite-difference time-domain scheme with macromodels, IEEE Trans. Microw.
Theory Tech., vol. 52, no. 9, pp. 23302335, Sep. 2004.
[7] W. Wang, G. Paraschos, and M. Vouvakis, Fast frequency sweep of
FEM models via the balanced truncation proper orthogonal decomposition, IEEE Trans. Antennas Propag., vol. 59, no. 11, pp. 41424154,
Nov. 2011.
[8] V. De La Rubia, U. Razafison, and Y. Maday, Reliable fast frequency
sweep for microwave devices via the reduced-basis method, IEEE
Trans. Microw. Theory Tech., vol. 57, no. 12, pp. 29232937, Dec.
2009.
[9] M. Hess and P. Benner, Fast evaluation of time-harmonic maxwells
equations using the reduced basis method, IEEE Trans. Microw.
Theory Tech., vol. 61, no. 6, pp. 22652274, Jun. 2013.
[10] M. Celik, L. Pileggi, and A. Odabasioglu, IC Interconnect Analysis.
Norwell, MA, USA: Kluwer, 2002.
[11] V. De La Rubia, Reliable reduced-order model for fast frequency
sweep in microwave circuits, Electromagnetics, vol. 34, pp. 161170,
2014.
[12] B. Liu et al., Block SAPOR: Block second-order arnoldi method for
passive order reduction of multi-input multi-output RCS interconnect
circuits, in Proc. Asia South Pacific Design Autom. Conf., 2005, vol.
1, pp. 255259.
[13] J. Rubio, J. Arroyo, and J. Zapata, Analysis of passive microwave
circuits by using a hybrid 2-D and 3-D finite-element mode-matching
method, IEEE Trans. Microw. Theory Tech., vol. 47, no. 9, pp. 14,
Sep. 1999.
[14] J.-S. Wang and N. Ida, Eigenvalue analysis in electromagnetic cavities
using divergence-free finite elements, IEEE Trans. Magn., vol. 27, no.
5, pp. 397881, Sep. 1991.

You might also like