tuning space mapping: the state of the art
TRANSCRIPT
Tuning Space Mapping: The State of the Art
Qingsha S. Cheng,1 John W. Bandler,1 Slawomir Koziel2
1 Simulation Optimization Systems Research Laboratory, Department of Electrical and ComputerEngineering, McMaster University, Hamilton, ON L8S 4K1, Canada
2 School of Science and Engineering, Reykjavik University, 101 Reykjavik, Iceland
Received 31 August 2011; accepted 20 December 2011
ABSTRACT: The electromagnetic (EM)-simulator-based tuning process for rapid microwave
design can combine EM accuracy with circuit-design speed. Our own approach is based on
the intuitive engineering idea of ‘‘space mapping.’’ In this article, we explain the art of
microwave design optimization through ‘‘tuning space mapping’’ procedures. We list various
appropriate types of models (called ‘‘surrogates’’). We demonstrate the implementation
of these surrogates through a simple bandstop filter. We provide application examples
using commercial simulation software. Our purpose is to help microwave engineers
understand the tuning space mapping methodology and to inspire new implementations and
applications. VC 2012 Wiley Periodicals, Inc. Int J RF and Microwave CAE 00:000–000, 2012.
Keywords: computer-aided design (CAD); engineering optimization and modeling; EM-based
design; tuning space mapping; surrogate models; design tuning
I. INTRODUCTION TO TUNING SPACE MAPPING
The use of accurate and computationally-intensive full-
wave electromagnetic (EM) simulation is now taken for
granted by the microwave industry. Although computers
become ever more powerful, microwave engineers are
faced with still larger design problems, tighter specifica-
tions, and shorter closure time requirements. Thus, EM-
based design remains challenging.
The EM-simulator-based tuning method [1, 2] is a fast
tuning or design approach that combines EM accuracy
with circuit speed. The method allows easy tuning and
instantaneous visualization of the EM-simulated responses
of a structure. The heart of the method is the hybridization
of EM simulation and circuit simulation in one structure
(a tuning model). It is achieved by replacing part of the
EM-simulated structure with equivalent circuit (physics-
based) models or by numerical approximations (predefined
physics-based models are preferable for their minimum
sample data requirements).
Space mapping [3] techniques also target fast but accu-
rate EM-based designs. They exploit surrogates to model
or to iteratively optimize a computationally-intensive EM-
simulated structure (‘‘fine’’ model). The surrogate models
themselves are constructed from a so-called ‘‘coarse’’
model, a less accurate but cheaper to evaluate representa-
tion of the fine model (e.g., an equivalent circuit).
The terms ‘‘tuning’’ and ‘‘space mapping’’ have been
mentioned together [4, 5] in a context where a coarse
model or surrogate is ‘‘tuned’’ within a space mapping
based algorithm. There, only circuit-based surrogates are
tuned. In this article, however, our tuning space mapping
algorithms combine the advantages of the EM-simulator-
based tuning concept with space mapping technology to
bring them to a new level in the realm of engineering
design and modeling. Tuning space mapping algorithms
exploit physics-based tuning models (EM-simulated mod-
els plus tuning components) as surrogates. The surrogates
approximate the full-wave EM simulation responses in an
iterative space mapping updating process, the same way
as do other space mapping approaches.
A tuning space mapping algorithm involves a fine
model (e.g., a full-wave EM simulation), auxiliary fine
models (fine models with tuning ports), a tuning model,
and a calibration scheme that sometimes exploits an extra
calibration (coarse) model. We may induce infinitesimal
tuning gaps (negligible compared with the wavelength)
into the fine model. We add tuning ports into the induced
gap edges or boundary of the fine model to form candi-
date auxiliary fine models. We insert or attach suitable
tuning elements to the tuning ports. Preferably, the tuning
elements are distributed circuit elements with physical
dimensions corresponding to those of the fine model. A
new structure bearing tuning elements is called a tuning
Correspondence to: Q. S. Cheng; e-mail: [email protected]
VC 2012 Wiley Periodicals, Inc.
DOI 10.1002/mmce.20621Published online in Wiley Online Library
(wileyonlinelibrary.com).
1
model or surrogate. After a simple alignment (also known
as parameter extraction or PE) procedure, we match the
tuning model with the fine model. Some of the fine-model
couplings are preserved (or represented through S-parame-
ters) in the tuning model. We normally obtain a good sur-
rogate of the fine model. In the next stage, the tuning
model is optimized by varying the design parameter val-
ues of the embedded tuning elements to satisfy given
design specifications. The obtained design parameter val-
ues constitute our next fine model iterate.
Figure 1 General tuning space mapping flowchart. [Color figure
can be viewed in the online issue, which is available at wiley
onlinelibrary.com.]
Figure 2 Type 0– tuning space mapping. Tuning ports are
created at the boundary of the fine model: (a) tuning of individual
components and (b) tuning of couplings between components.
[Color figure can be viewed in the online issue, which is
available at wileyonlinelibrary.com.]
Figure 3 Type 0 and 1 tuning space mapping. Tuning ports
are created inside the fine model. [Color figure can be viewed in
the online issue, which is available at wileyonlinelibrary.com.]
Figure 4 Fast tuning space mapping. The auxiliary fine model
is reduced to allow fast simulation. [Color figure can be viewed
in the online issue, which is available at wileyonlinelibrary.com.]
2 Cheng, Bandler, and Koziel
International Journal of RF and Microwave Computer-Aided Engineering/Vol. 000, No. 000, Month 2012
In this article, we review various tuning space mapping
approaches and categorize them into four main groups:
Type 0� [1, 6, 7], Type 0 [8–10], Type 1 [10–13], and
Type 2 [6, 14, 15].
II. GENERAL TUNING SPACE MAPPING ALGORITHM
We are concerned with the optimization problem
x�f ¼ argminx
U Rf ðxÞ� �
(1)
where Rf [ Rm denotes the response vector of a fine model
of the device of interest, U is a merit function (e.g., a mini-
max function or a norm), x is a vector of design parameters,
and xf* is the optimal solution to be determined.
The iteration of a typical tuning space mapping algo-
rithm consists of several steps: the simulation of the fine
and corresponding auxiliary fine models, the alignment of a
tuning model with the fine model (i.e., the updating of the
tuning model), the optimization of the tuning model, and
the calculation of the corresponding fine model design pa-
rameter values. In the ith iteration, first, based on data from
an auxiliary fine model (for example, a fine model with so-
called cocalibrated ports) at the current design x(i), an initial
tuning model Rt(i) is built on the auxiliary fine model em-
bedded with appropriate tuning elements. The current tun-
ing model is supplied with appropriate (initial) tuning pa-
rameters (parameter values t(i)) determined from microwave
engineering knowledge. We augment this tuning model
with certain other model parameters p (such as input map-
ping parameters and/or preassigned parameters) similar to
those found in the surrogate of the traditional space map-
ping techniques [16]. Our initial tuning model response
may not agree with the response of the original fine model
at x(i). We align these models by:
pðiÞ ¼ argminp
Rf ðxðiÞÞ � RðiÞt ðtðiÞ; pÞ
������ (2)
where Rt [ Rm denotes the response vector of the tuning
model. After p(i) is extracted, we optimize Rt(i) to have it
meet the design specifications w.r.t. the selected tuning pa-
rameters t, giving
Figure 5 Type 2 tuning space mapping. Multiple parts of the
fine models are simulated separately and combined with tuning
elements later in a circuit simulator. [Color figure can be viewed
in the online issue, which is available at wileyonlinelibrary.com.]
Figure 6 Simple bandstop microstrip filter: (a) geometry and
(b) coarse model in a circuit simulator [17].
Figure 7 Simple bandstop filter example: (a) Type 0� tuning
model, (b) Type 0 tuning model, and (c) Type 2 tuning model.
Tuning Space Mapping 3
International Journal of RF and Microwave Computer-Aided Engineering DOI 10.1002/mmce
tðiÞopt ¼ argmin
tU R
ðiÞt ðt; pðiÞÞ
� �(3)
The new design parameter values x(iþ1) can be calculated
either directly from x(i) þ tðiÞopt � t(i) (when the design pa-
rameters and tuning parameters have the same physical
meaning and units) or by substituting tðiÞopt, t
(i), and x(i) intoa suitable calibration model [9, 10]. Note that the current
initial tuning parameters t(i) are related to how the current
auxiliary fine model and tuning model are setup; their
values may or may not be directly related to the previous
optimal tuning parameter values tði�1Þopt . We show the
flowchart of a general tuning space mapping procedure in
Figure 1.
III. TYPES OF TUNING MODELS
As the tuning model is based on auxiliary fine models and
appropriate tuning elements, there are many ways to cre-
ate it. We now discuss various types of useful tuning
models of tuning space mapping algorithms.
Our first type is Type 0– tuning, where tuning ports
are added at the boundary of the EM model (fine) as
shown in Figure 2 to form an auxiliary fine model. The
auxiliary fine model is then simulated. The resulting
responses are equipped with tuning ports, where tuning
elements are attached. It is now a tuning model. We can
proceed using the algorithm described in Section II.
Type 0 and 1 tuning methods use internal tuning ports.
The auxiliary fine model is formed by inserting internal
ports into the fine model. The auxiliary fine model is then
Figure 8 Simple bandstop filter example: four different subtypes of Type 1 tuning model (Type 1a, 1b, 1c, and 1d (fast) tuning model).
Type 1a tuning can be used to tune an end section, in which the entire end section can be omitted in the auxiliary (EM) fine model. Type
1b tuning uses a complete fine model with gaps and tuning ports as the auxiliary fine model. Unused tuning ports exist in the tuning
model. In Type 1c tuning, the unused ports and the associated structure in the auxiliary fine model are omitted. In Type 1d tuning, the aux-
iliary fine model is collapsed based on the Type 1c auxiliary fine model.
TABLE I Setup of Simple Stopband Filter Tuning Modelsa
Type Tuning Model
Auxiliary
Fine Model
Momentum
Internal Port(s)
Sonnet em
Internal Port(s)
Sonnet em
Feedline Width
Type 0– Type 0– Type 0– Internal Cocalibrated 1 mm
Type 0 Type 0 Type 0 Single mode Cocalibrated 1 mm
Type 1a (I) Type 1a (unused ports are open) Type 0 Single mode Cocalibrated 1 mm
Type 1a (II) Type 1a Type 1a Single mode Cocalibrated 1 mm
Type 1b (I) Type 1b (unused ports are open) Type 1b Single mode Cocalibrated 1 mm
Type 1b (II) Type 1b (short the unused ports) Type 1b Single mode Cocalibrated 1 mm
Type 1b (III) Type 1b (ground the unused ports) Type 1b Single mode Cocalibrated 1 mm
Type 1c Type 1c Type 1c Single mode Cocalibrated 1 mm
Type 1d Type 1d Type 1d Single mode Cocalibrated 1 mm
Type 2 Type 2 Type 2 (partitioned) Single mode Standard n/a
a Type 1a, 1b, 1c, and 1d are all subtypes of Type 1. Their configurations are shown in Figure 8.
4 Cheng, Bandler, and Koziel
International Journal of RF and Microwave Computer-Aided Engineering/Vol. 000, No. 000, Month 2012
simulated to create multiport responses. We add tuning
elements to the tuning gaps (Type 0 tuning) or replace
fine model sections between the tuning gaps with tuning
elements (Type 1 tuning) to obtain a tuning model as
depicted in Figure 3. The fast tuning space mapping
approach is based on Type 1 tuning but uses a simplified
or reduced auxiliary model (Fig. 4).
Type 2 tuning breaks the fine model into pieces, which
differs from the Type 0–, 0, and 1. Each piece is simu-
lated separately in the EM simulator. The fine model
pieces are connected using tuning elements in a circuit
simulator to form a tuning model. See Figure 5.
We exhibit each type of tuning model using a simple
filter example shown in Figure 6a. We have one design
parameter, the stub length L. The goal is to find L so that
the center frequency of the filter is 5 GHz. The fine model
is simulated in the EM simulator. According to our space
mapping approach, instead of performing direct optimiza-
tion of the fine model, we want to use a fast surrogate
model instead. The traditional space mapping surrogate is
based on a coarse model shown in Figure 6b, which is a
circuit equivalent of the structure in Figure 6a and it is
implemented in a circuit simulator [17]. In tuning space
mapping, the coarse model serves as an initial design and/
or as a calibration model.
We start from the optimal solution of the coarse
model, which is 5.6329 mm. The fine model and auxiliary
fine model are simulated in two EM simulators [18, 19].
The tuning model is constructed using the EM simulation
results and tuning elements in a schematic design [17].
We set up various types and subtypes of tuning models of
the filter as shown in Figures 7 and 8. Table I summarizes
the setup of the tuning models. We can observe (Fig. 9)
that response misalignments between the tuning models
and the fine model vary depending on the tuning types.
The initial center frequency of the filter (fine model) is
4.896 GHz instead of the required 5.000 GHz. We cali-
brate the tuning elements so that the misalignments
between the fine and tuning surrogate models are reduced
(Fig. 10).
The next step is the surrogate model optimization in
which the length of the tuning surrogate model stub is
optimized to obtain the center frequency of 5 GHz. We
can see that almost all the variations of tuning model
have predicted a good result (Fig. 11) in a single iteration
(i.e., two evaluations of the fine model) of the tuning
space mapping algorithm. The responses are clustered at
the neighborhood of the target solution. The design can be
further improved by applying the second iteration of the
algorithm. Herein, we only show the results with an EM
simulator [18] as the fine and auxiliary fine model. With
the other EM simulator [19], similar results are achieved.
Table II summarizes the initial and after-PE matching
errors and the fine model errors and design variable values
after one iteration. Table III shows the number of ports
needed and the simulation metal area for each type.
Readers should not be discouraged by the seemingly
intricate types of tuning model of the synthetic bandstop
microstrip filter example. As an expert approach, building a
tuning model needs expert knowledge. For engineers famil-
iar with a particular device or component, creating a tuning
model based on one of the variations mentioned here
should not be difficult, as demonstrated in Section IV.
Figure 9 Simple bandstop filter example: responses of the tun-
ing models and the fine model [18]. [Color figure can be viewed
in the online issue, which is available at wileyonlinelibrary.com.]
Figure 10 Simple bandstop filter example: responses of the
tuning models after PE and the fine model [18]. [Color figure can
be viewed in the online issue, which is available at wileyonline
library.com.]
Figure 11 Simple bandstop filter example: responses of the fine
models after the first iteration [18]. [Color figure can be viewed in
the online issue, which is available at wileyonlinelibrary.com.]
Tuning Space Mapping 5
International Journal of RF and Microwave Computer-Aided Engineering DOI 10.1002/mmce
IV. EXAMPLES
We demonstrate in this section various tuning types using
microwave and radio frequency component designs.
A. Type 0� Tuning MethodSwanson and Wenzel [1] tuned a filter in a circuit simula-
tor using multiport S-parameter data and lumped capaci-
tors at the ports. They then optimize the combline filter
rapidly by mapping the ‘‘coarse’’ circuit model to the
‘‘fine’’ finite element method model. This optimization is
shown to converge in a single iteration, with a good start-
ing point. The method can be viewed as a Type 0� tuning
process (tuning elements added at the boundary). In
another Type 0� tuning example [6], a seven-post band-
pass filter is tuned by adding tuning elements (tuning
capacitors and inductors) to each resonator (Fig. 12). The
couplings between the resonators are also tuned by tuning
elements (transmission lines) added between the resona-
tors (Fig. 13). The tuning model is shown in Figure 14.
The after-tuning responses are shown in Figure 15.
B. Type 0 Tuning MethodWe formally introduced the tuning space mapping method
(Type 0 tuning) for microwave design optimization in [9].
We formulate the tuning space mapping concept and show
how it relates to the standard space mapping methodol-
ogy. The tuning model is created using engineering exper-
tise and knowledge of the design problem but also used
the efficiency of space mapping for translating the adjust-
ment of the tuning parameters into relevant updates of the
design variables. We illustrate Type 0 tuning through the
optimization of a high-temperature superconducting (HTS)
filter [9] in Figures 16a and 16b.
Figure 16a shows the auxiliary fine model constructed
by dividing the five coupled-line polygons in the middle
and inserting the tuning ports at the new cut edges. Figure
16b shows the tuning model implementation in the circuit
simulator [17] where the S22P data file (22 being the num-
ber of ports) of the simulated auxiliary fine model is
loaded. The circuit-theory coupled-line components and ca-
pacitor components are chosen to be the tuning elements
and are inserted into each pair of tuning ports. As all the
tuning elements are embedded in the small gaps in the aux-
iliary fine model, it is called a Type 0 tuning. Figure 17
shows the final response of the HTS filter using a Type 0
tuning space mapping algorithm. In [10], a narrowband 62
GHz interdigitated filter is designed using Type 0 tuning.
Thoma [8] implemented a simple hairpin filter tuning
space mapping design in an EM/circuit simulation suite
[20]. Short line segments are introduced to the EM model
for calibration. We categorize the method as Type 0,
as well.
TABLE II Simple Stopband Filter Tuning Example
Type
Initial Matching
Error
Matching
Error After PE
Fine Model
Objective Function
Value
Design
Variable
Value (mm)
Type 0– 0.632 0.003 1.492 5.529
Type 0 0.041 0.008 0.850 5.530
Type 1a (I) 3.501 0.241 5.063 5.522
Type 1a (II) 2.773 0.040 0.797 5.531
Type 1b (I) 0.282 0.151 3.205 5.526
Type 1b (II) 0.151 0.150 3.205 5.526
Type 1b (III) 0.220 0.169 3.205 5.526
Type 1c 0.173 0.152 3.205 5.526
Type 1d 0.376 0.152 3.665 5.536
Type 2 0.049 0.029 0.797 5.521
TABLE III Simple Stopband Filter Auxiliary Fine ModelComparison
Type Number of Ports Metal Area (mm2)
Type 0– 3 16.64
Type 0 4 16.65
Type 1a (I) 3 16.65
Type 1a (II) 3 13.5
Type 1b (I) 6 16.62
Type 1b (II) 6 16.62
Type 1b (III) 6 16.62
Type 1c 4 13.5
Type 1d 4 13.5
Type 2 4 (3 þ 1) 13.5 (12.5 þ 1)
Figure 12 Seven-post bandpass filter: adding tuning compo-
nents (self-impedance and self-capacitance) to the discrete ports
in each resonator [6] (Reproduced from Ref. [6]). [Color figure
can be viewed in the online issue, which is available at
wileyonlinelibrary.com.]
6 Cheng, Bandler, and Koziel
International Journal of RF and Microwave Computer-Aided Engineering/Vol. 000, No. 000, Month 2012
C. Type 1 Tuning MethodAn implementable microwave design framework is pre-
sented in [11]. In the framework, Type 1 (1b) tuning
space mapping is implemented. We alter an EM model by
replacing a section with suitable tuning elements. The
resulting tuning model is aligned with the original unal-
tered EM model. We then designate the aligned tuning
model as surrogate for design optimization purposes. The
Type 1 tuning space mapping framework is illustrated in
[11] using a simple microstrip line example, an open-loop
ring resonator bandpass filter and a low-
temperature cofired ceramic filter.
In [10], we show that the Type 0 and Type 1 arrange-
ments can be combined. A mixed Type 0 and Type 1b
tuning model of the open-loop ring resonator bandpass
filter is shown in Figure 18. The fine model is simulated
in the EM simulator, while the tuning model is con-
structed and optimized in the microwave circuit simulator
[17]. We divide the microstrip structure and insert cocali-
brated port pairs at the cut edges. Then, we simulate the
auxiliary EM structure with the ports and import the
resulting SNP data file (50 ports) as an SNP S-parameter
component into the circuit simulator. Equivalent circuit
microstrip lines (Type 0) are inserted, and microstrip
coupled-line and gap components (Type 1) replace sec-
tions of the structure in Figure 18. A new tuning model
is now available with tuning parameters [dL1 dL2 dL3 dL4S1 S2 g]
T mm.
As in [11], deviations between the tuning model and
fine model are compensated by calibrating the dielectric
Figure 13 Seven-post bandpass filter: adding tuning components (transmission line) to the discrete ports between the resonators [6]
(Reproduced from Ref. [6]). [Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]
Figure 14 Tuning model of the seven-post bandpass filter [6] (Reproduced from Ref. [6]). [Color figure can be viewed in the online
issue, which is available at wileyonlinelibrary.com.]
Tuning Space Mapping 7
International Journal of RF and Microwave Computer-Aided Engineering DOI 10.1002/mmce
constant and substrate height or length offsets of the tun-
ing elements. After compensation, the tuning model or
surrogate is seen as a better representation of the fine
model and is optimized by a circuit simulator [17] with
respect to the design parameters. The new design parame-
ters are then assigned to the fine model. The optimized
tuning model and the corresponding fine model responses
are shown in Figure 19.
D. Type 1d Tuning (Fast Tuning) MethodAlthough the tuning space mapping optimization process
typically requires only a few iterations to complete, the
simulation of the structure with a number of cocalibrated
ports (required to insert the tuning elements) is longer
than that of the original structure [13]. In [13], we show a
fast tuning space mapping (Type 1d) algorithm that
exploits a reduced structure (auxiliary fine model) with
fewer cocalibrated ports for creating the tuning model.
This results in reduction of the computational cost of the
optimization process. A third-order Chebyshev filter and a
coupled microstrip bandpass filter are verified and com-
pared with Type 1b tuning space mapping. Note that Type
1d tuning is a compact version of other Type 1 tuning
models.
Consider the third-order Chebyshev bandpass filter
[13] (Fig. 20a). Without the tuning ports, the evaluation
time of the fine model is 27 min. The simulation time
of the auxiliary fine model (with cocalibrated ports) in
Figure 20a is almost 11 h. In the case of the fast tuning
(Type 1d) algorithm, the tuning model is constructed
using the S-parameters of the reduced auxiliary fine model
along with appropriate tuning elements. Note that the
reduced auxiliary fine model (Fig. 20b) has a smaller
number of cocalibrated ports. Its simulation time is only
38 min, which is 17 times faster than for the structure in
Figure 20a.
The optimization results for fast tuning space mapping
(Type 1d) and another Type 1 tuning (Type 1b in this
case) are summarized in Table IV. The quality of the final
design is quite similar for both algorithms, which indi-
cates that it is indeed sufficient to simulate the reduced
Figure 15 Responses of the seven-post bandpass filter after tun-
ing [6] (Reproduced from Ref. [6]). [Color figure can be viewed in
the online issue, which is available at wileyonlinelibrary.com.]
Figure 17 HTS filter [9]: Fine model response (|S21| obtainedwith an EM simulator [19]) at the final design. [Color figure can be
viewed in the online issue, which is available at wileyonline
library.com.]
Figure 16 HTS filter: (a) physical structure [9] in which num-
bered tuning ports (port 3 to 22) and gaps (dashed line) are inserted
to create the auxiliary fine model; (b) tuning model [17]. [Color fig-
ure can be viewed in the online issue, which is available at
wileyonlinelibrary.com.]
8 Cheng, Bandler, and Koziel
International Journal of RF and Microwave Computer-Aided Engineering/Vol. 000, No. 000, Month 2012
auxiliary fine model to maintain the prediction capability
of the tuning model. On the other hand, the computational
cost is lower for fast tuning space mapping.
E. Type 2 Tuning MethodA tuning model can be constructed by simulating the EM
model sections separately and then connecting them with
the tuning components using a ‘‘cosimulation’’ process
[14]. This allows us to implement the tuning space map-
ping algorithm with any EM simulator. Response mis-
alignment between the original structure and the tuning
model is reduced using classical space mapping. In [14],
the Type 2 tuning space mapping algorithm is illustrated
through the design of two microstrip filters simulated in
an EM simulator [21].
Consider the coupled-line bandpass filter [22] shown in
Figure 21a. The fine model is simulated in the EM simu-
lator [21]. A schematic of the cosimulation-based tuning
model is shown in Figure 21b. Subsections marked black
are simulated in [21]. Due to symmetry, only two subsec-
tions need independent evaluation. The tuning model is
handled by the circuit simulator [17] (Fig. 21c).
The alignment procedure uses a vector consisting of
dielectric constants as well as substrate heights of the dis-
tributed circuit components corresponding to the design
variables. Figure 22 shows the fine model response after
the two iteration of the Type 2 algorithm.
For the sake of comparison, the filter was also optimized
using Matlab’s fminimax routine [23], a pattern search
Figure 18 The open-loop ring resonator bandpass filter realized in the circuit simulator [10] using mixed Type 0 and Type 1 (1b)
embedding. Type 1 tuning elements are in red circles. [Color figure can be viewed in the online issue, which is available at
wileyonlinelibrary.com.]
Figure 19 Responses after two iterations: the tuning model (—)
and the corresponding fine model (*). [Color figure can be viewed
in the online issue, which is available at wileyonlinelibrary.com.]
Figure 20 Third-order Chebyshev filter: (a) the geometry [13]
and the places (the dashed lines) for inserting the tuning ports for
the Type 1b algorithm. (b) the reduced structure [13] and the pla-
ces (the dashed lines) for inserting the tuning ports for the fast
tuning space mapping (Type 1d) algorithm.
Tuning Space Mapping 9
International Journal of RF and Microwave Computer-Aided Engineering DOI 10.1002/mmce
algorithm [24], as well as a space mapping algorithm
exploiting input, frequency, and output space mapping [25].
The results are shown in Table V.
It can be observed that both space mapping and gradi-
ent-based search fail to find a design satisfying the specifi-
cations. The design obtained using pattern search is slightly
better than that obtained by the technique described here;
however, the design cost is substantially higher.
In another example of Type 2 tuning [6], the author
tunes each resonator of a dual band duplexer for 400 MHz
using a tuning model comprised of connected submodels in
an EM/circuit simulation suite [20]. Tuning elements are
attached to the resonators (Fig. 23). Similarly, another sim-
ulation software vendor demonstrates the technique in
designing filters [26] using its EM [27] and circuit [28]
simulators.
V. DISCUSSION AND RECOMMENDATIONS
As demonstrated through examples, tuning space mapping
permits rapid design improvement with satisfactory
TABLE IV Third-Order Chebyshev Filter: Optimization Results [13]
Algorithm
Optimization Results Optimization Costa
Number of
Iterations
Specification
Error (dB)
Total
Time (h)
Equivalent
Cost (# of Fine Evaluations)
Fast tuning space mapping (Type 1d) 2 �1.7 1.7 3.4
Other Type1 (Type 1b) 2 �1.7 23.3 51.0
a Excluding the fine model evaluation at the initial design.
Figure 21 Cosimulation-based (Type 2) tuning model [14]: (a) a coupled-line bandpass microstrip filter structure [14], (b) its cosimula-
tion tuning model with black sections simulated using an EM solver [21] connecting designable tuning components, (c) circuit simulator
[17] implementation of the tuning model: S-parameters of the EM-simulated sections are stored in S3P and S4P data components SNP1 to
SNP6. Note that all the designable parameters (microstrip lengths, widths, and coupled-line gaps) are associated with the distributed circuit
components, which allows fast optimization of the tuning model. On the other hand, simulating parts of the filter using the EM solver
allows us to maintain good accuracy and predictability of the tuning model.
International Journal of RF and Microwave Computer-Aided Engineering/Vol. 000, No. 000, Month 2012
10 Cheng, Bandler, and Koziel
designs obtained after a few iterations (typically, one to
four). The variants of tuning space mapping mostly differ
in the construction of the tuning model, as well as in the
calibration procedure. The calibration procedure for Type
0 tuning can be quite complex and generally requires an
auxiliary calibration model. On the other hand, Type 0
tuning is probably the most robust technique because the
insertion of the tuning components is realized with mini-
mal or no disturbance to the fine model response. Also,
Type 0 tuning requires a smaller number of tuning ports
Figure 22 Coupled-line bandpass filter: the fine model
response at the final design obtained in two Type 2 iterations.
[Color figure can be viewed in the online issue, which is avail-
able at wileyonlinelibrary.com.]
TABLE V Coupled-Line Bandstop Filter:Cosimulation-Based Tuning (Type 2) VersusOther Optimization Approaches: Design Qualityand Computational Cost Comparison
Algorithm
Best Design
Found (dB)aDesign
Costb
Cosimulation-based
tuning
�1.3 3
Space mapping þ1.5 8
Matlab (fminimax) þ22 208
Pattern search �1.7 155
a Specification error at the final (optimized design).b Number of fine model evaluations.
Figure 23 Tuning model of a dual band duplexer for 400 MHz [6] (Reproduced from Ref. [6]). [Color figure can be viewed in the
online issue, which is available at wileyonlinelibrary.com.]
International Journal of RF and Microwave Computer-Aided Engineering DOI 10.1002/mmce
Tuning Space Mapping 11
than does Type 1 tuning, so the simulation time of the
‘‘cut’’ fine model is longer for Type 1 than for Type 0.
The major benefits of Type 1 tuning are simple cali-
bration and straightforward implementation. Also, Type 1
tuning allows us to tune cross-sectional parameters. The
downside is that a Type 1 tuning model has to be aligned
with the fine model as the insertion of the tuning compo-
nents requires the removal of substantial parts of the origi-
nal structure. For the same reason, the accuracy of the
Type 1 tuning model is not as good as that of the tuning
model in Type 0 tuning, so more iterations are normally
necessary to yield a satisfactory design.
Type 1d tuning (fast tuning) seems to inherit all the
advantages of Type 1 tuning while being computationally
more efficient. One difficulty, however, not observed for
the test problems described in the literature [12, 13], is
that using the reduced structure for complicated circuits
may result in further deterioration of the generalization
capability of the tuning model. This could lead to a larger
number of iterations required by the Type 1d tuning
algorithm.
Type 0� and Type 2 tuning are most found in the tun-
ing processes for waveguide structures. Type 2 tuning
simulates each submodel separately, which allows parallel
processing. However, care has to be taken since the cou-
plings between the submodels are not accounted for.
VI. CONCLUSIONS
We review various types of tuning space mapping proce-
dures and use a simple stopband filter to demonstrate their
differences. We show examples of these types of tuning
space mapping using published examples. Despite the
variations of tuning space mapping, it is generally robust
because misalignments can usually be compensated by
tuning elements. The main considerations for choosing
among the variations are the anticipated difficulties in the
implementations and the simulation cost of the auxiliary
fine model.
ACKNOWLEDGMENTS
This work was supported in part by the Icelandic Centre
for Research (RANNIS) Grant 110034021, and the Natural
Sciences and Engineering Research Council of Canada
under Grants RGPIN7239-11 and STPGP 381153-09, and by
Bandler Corporation. The authors thank Sonnet Software
Inc. for em, and Agilent Technologies, Santa Rosa, CA, for
making ADS available. They also thank F. Hirtenfelder,
CST, Darmstadt, Germany, for providing illustration
examples.
REFERENCES
1. D.G. Swanson and R.J. Wenzel, Fast analysis and optimization
of combline filters using FEM, IEEE MTT-S Int Microwave
Symp Dig, Boston, MA, July 2001, pp. 1159–1162.
2. J.C. Rautio, RF design closure—Companion modeling and tun-
ing methods, Proc IEEE MTT-S Int Microwave Symp Work-
shop: Microwave Component Design Using Space Mapping
Technology, San Francisco, CA, June 2006.
3. J.W. Bandler, Q.S. Cheng, S.A. Dakroury, A.S. Mohamed,
M.H. Bakr, K. Madsen, and J. Søndergaard, Space mapping:
The state of the art, IEEE Trans Microwave Theory Tech 52
(2004), 337–361.
4. Q.S. Cheng, J.W. Bandler, and J.E. Rayas-S�anchez, Tuning-
aided implicit space mapping, Int J RF Microwave CAE 18
(2008), 445–453.
5. A. Rasmita, Y.X. Guo, and A Alphones, Constrained explicit
knowledge embedded space mapping using circuit tuning based
on physical augmentation as parameter extraction, Asia Pacific
Microwave Conf (APMC), Singapore, Dec. 2009, pp. 1431–1434.
6. F. Hirtenfelder, Filter tuning techniques, CST User Group
Meeting, Lake Tegernsee, Germany, Sep., 2007.
7. A. Edquist, Filter design and Ansoft tools, HFSS Info Day,
Stockholm, Sweden, Sep. 22, 2009.
8. P. Thoma, Using coupled simulations to efficiently address
challenging EM problems, Int Workshop on Advances in
Modeling and Optimization of High Frequency Structures,
Reykjavik University, Iceland, Aug. 21–22, 2010.
9. S. Koziel, J. Meng, J.W. Bandler, M.H. Bakr, and Q.S.
Cheng, Accelerated microwave design optimization with tun-
ing space mapping, IEEE Trans Microwave Theory Tech 57
(2009), 383–394.
10. Q.S. Cheng, J.C. Rautio, J.W. Bandler, and S. Koziel,
Progress in simulator-based tuning—The art of tuning space
mapping, IEEE Microwave Mag 11 (2010), 96–110.
11. Q.S. Cheng, J.W. Bandler, and S. Koziel, Space mapping
design framework exploiting tuning elements, IEEE Trans
Microwave Theory Tech 58 (2010), 136–144.
12. S. Koziel, J.W. Bandler, and Q.S. Cheng, Design optimization
of microwave circuits through fast embedded tuning space
mapping, Proc 40th European Microwave Conf, Paris, France,
Sep./Oct. 2010, pp. 1186–1189.
13. S. Koziel, J.W. Bandler, and Q.S. Cheng, Tuning space map-
ping design framework exploiting reduced EM models, IET
Microwaves Antennas Propag 5 (2011), 1219–1226.
14. S. Koziel and J.W. Bandler, Fast design optimization of
microwave structures using co-simulation-based tuning space
mapping, Appl Comput Electromagn Soc J 26 (2011), 631–639.
15. S. Koziel, Robust optimization of microwave structures using
cosimulation-based surrogate models, Microwave Opt Tech-
nol Lett 53 (2011), 130–135.
16. S. Koziel, Q.S. Cheng, and J.W. Bandler, Space mapping,
IEEE Microwave Mag 9 (2008), 105–122.
17. Agilent ADS, Version 2009, Agilent Technologies, 395 Page
Mill Road, Palo Alto, CA 94304, 2009.
18. Agilent Momentum, Version 2009, Agilent Technologies, 395
Page Mill Road, Palo Alto, CA 94304, 2009.
19. Sonnet em, Ver. 12.54 Sonnet Software, North Syracuse, NY,
2009.
20. CST Microwave Studio, ver. 2009, CST AG, Bad Nauheimer
Str. 19, D-64289 Darmstadt, Germany, 2009.
21. FEKO, Suite 5.5, EM Software & Systems-S.A. (Pty) Ltd, 32
Techno Lane, Technopark, Stellenbosch, 7600, South Africa,
2009.
22. H.M. Lee and C.M. Tsai, Improved coupled-microstrip filter
design using effective even-mode and odd-model characteris-
tic impedances, IEEE Trans Microwave Theory Tech 53
(2005), 2812–2818.
23. MatlabTM, Version 7.6, The MathWorks, Inc., 3 Apple Hill
Drive, Natick, MA 01760–2098, 2008.
24. T.G. Kolda, R.M. Lewis, and V. Torczon, Optimization by
direct search: New perspectives on some classical and mod-
ern methods, SIAM Rev 45 (2003), 385–482.
International Journal of RF and Microwave Computer-Aided Engineering/Vol. 000, No. 000, Month 2012
12 Cheng, Bandler, and Koziel
25. S. Koziel, J.W. Bandler, and K. Madsen, A space mapping
framework for engineering optimization: Theory and imple-
mentation, IEEE Trans Microwave Theory Tech 54 (2006),
3721–3730.
26. R. Fernandes, The accuracy of HFSS with the speed of circuit
simulation, 10emes Journ�ees de Caract�erisation Microondes
et Mat�eriaux, Limoges, France, April, 2008 [Online]. Avail-
able at: http://jcmm2008.xlim.fr/hfss_presentation.pdf.
27. Ansoft HFSS Version 10.1, Ansoft Corporation, Pittsburgh,
PA, USA.
28. Ansoft Designer/Nexxim, Ansoft Corporation, Pittsburgh, PA,
USA.
BIOGRAPHIES
Qingsha S. Cheng was born in
China. He received the B.Eng. and
M.Eng. from Chongqing University,
China, in 1995 and 1998, respec-
tively. He received his Ph.D. from
McMaster University, Canada, in
2004. In 1998, he joined the Depart-
ment of Computer Science and Tech-
nology, Peking University, China. In 1999, he joined the
Department of Electrical and Computer Engineering,
McMaster University. Currently, he is a research engineer
in the Department of Electrical and Computer Engineer-
ing, McMaster University. His research interests are surro-
gate modeling, computer-aided design, modeling of micro-
wave circuits, software design technology, and
methodologies for microwave CAD.
John W. Bandler studied at Imperial
College and received the B.Sc.(Eng.),
Ph.D., and D.Sc.(Eng.) degrees from
the University of London, England,
in 1963, 1967, and 1976, respec-
tively. He joined McMaster Univer-
sity, Canada, in 1969. He is now a
Professor Emeritus. He was President
of Optimization Systems Associates Inc., which he
founded in 1983, until November 20, 1997, the date of
acquisition by Hewlett-Packard Company. He is President
of Bandler Corporation, which he founded in 1997. He is
a Fellow of several societies, including the Royal Society
of Canada. In 2004, he received the IEEE MTT-S Micro-
wave Application Award.
Slawomir Koziel received the M.Sc.
and Ph.D. degrees in electronic engi-
neering from Gdansk University of
Technology, Poland, in 1995 and
2000, respectively. He also received
the M.Sc. degrees in theoretical
physics and in mathematics, in 2000
and 2002, respectively, as well as the
Ph.D. in mathematics in 2003, from the University of
Gdansk, Poland. He is currently a Professor with the
School of Science and Engineering, Reykjavik University,
Iceland. His research interests include CAD and modeling
of microwave circuits, simulation-driven design, surro-
gate-based optimization, space mapping, circuit theory,
analog signal processing, evolutionary computation, and
numerical analysis.
International Journal of RF and Microwave Computer-Aided Engineering DOI 10.1002/mmce
Tuning Space Mapping 13