Application of Neural Networks in Space-Mapping Optimization of Microwave Filters

Ying Wang,1 Ming Yu,2 Humayun Kabir,2 Qi-Jun Zhang3

2 3

Faculty of Engineering and Applied Sciences, University of Ontario Institute of Technology, Oshawa, ON, Canada L1H 7K4 COM DEV Ltd., 155 Sheldon Drive, Cambridge, ON, Canada N1R 7H6 Department of Electronics, Carleton University, Ottawa, ON, Canada K1S 5B6

Received 3 May 2011; accepted 9 August 2011

ABSTRACT: In this paper, a design methodology combining coupling matrix representation of lters, neural models and space-mapping techniques is presented for further enhancement of optimization efcency of microwave lters. Neural models are developed for both initial dimension generation and design parameter sensitivity analysis. Combining neural models of lter substructures with space-mapping optimization, the total number of EM simulations of the complete lter structure is signicantly reduced. The improvement in efciency over conventional method is demonstrated using simulation and measurement results of both end-coupled and side-coupled waveguide dual-mode pseudo-elliptic lters. The total C CPU times for design and optimization are reduced by 50% to 70 %. V 2011 Wiley Periodicals,
Inc. Int J RF and Microwave CAE 22:159166, 2012.

Keywords: computer-aided design; electromagnetic optimization; filter design; neural networks; space mapping

1. INTRODUCTION

In microwave lter design, full electromagnetic (EM) simulation is essential to ensure rst pass success due to its superior accuracy. However, the computational intensiveness limits its applications in iterative design. For reduction of the cost of ne model (full EM) simulations, approximation models, or surrogate models, can be constructed to closely resemble the ne model response while being much faster to evaluate. Surrogate-based optimization methods are widely used for solving problems involving expensive function evaluation [1, 2]. For the applications in RF/ microwave design, the well-known space-mapping technique is such a surrogate-based optimization method [312]. Space mapping has been proven to be very effective in the optimization of different types of circuits, including lters, in which design optimizations involve combination of precise EM models and coarse/surrogate models. One of the unique features of space mapping optimization compared with other surrogate-based methods is that the coarse model in space mapping is physically meaningful, for examCorrespondence to: Y. Wang; e-mail: ying.wang@uoit.ca. DOI 10.1002/mmce.20572 Published online 30 December 2011 in Wiley Online Library (wileyonlinelibrary.com).
C V 2011 Wiley Periodicals, Inc.

ple, lumped or distributed circuit model, instead of being a pure black-box modeling, or curve-tting. Quality of the coarse model is essential for the performance of the space mapping optimization [10]. For the application of interest, namely cross-coupled microwave lters with narrow to medium bandwidth, the widely used coupling matrix representation of bandpass lters is adopted as a coarse model of the lter. The reason lies in the fact that coupling matrix has become the lter design tool of choice, especially for lters with complex coupling mechanisms. There exists a rich literature on coupling matrix synthesis for advanced cross-coupled microwave lters [1315]. In addition, each coupling value in the coupling matrix corresponds to a physical dimension of the lter structure. Combination of coupling matrix with space mapping has been proven to be effective for optimization of advanced microwave lters [12]. However, there is complexity caused by the fact that the coarse model and ne model parameters have different physical meanings. Therefore, initial values for the ne model cannot be simply a copy of the coarse model design parameters as in the original aggressive spacemapping method. Furthermore, the initial mapping is also unknown. With the choice of the coarse model and fullwave EM simulators as the ne model, the lter design can generally be divided into the following four steps. In the rst

159

160

Wang et al. with lter order, the developed neural models are applicable to lters of any order, showing great exibility. We demonstrate the proposed technique using two types of waveguide dual-mode pseudo-elliptic lters, namely endcoupled [21] and side-coupled [24, 25]. Both the types of lters are commonly used as narrow bandwidth bandpass channel lters for output multiplexing in communication satellites. As will be shown using simulation and measurement results, the method is highly efcient. The technique is readily applicable to other types of lters.
2. PROPOSED DESIGN AND OPTIMIZATION APPROACH

step, the ideal coupling values are synthesized and optimized to satisfy design specications. Initial design parameters are calculated realizing the ideal coupling values in the second step. These initial dimensions are in general not satisfactory especially for lters with higher orders or complex congurations. Optimizations are therefore necessary, which is the last step, that is, the fourth step, of the lter design. The trust-region aggressive space-mapping algorithm [11, 12] is used in the following. For preparation of optimizations, the step 3 is needed to nd out the initial mapping. In Ref. 12, before space-mapping optimizations, preliminary EM simulations are performed for sensitivity analysis in order to get the initial mapping. As shown in Ref. 12, even for a complex lter structure, the total number of optimization iterations is usually less than the number of design parameters. However, the number of preliminary EM simulations for sensitivity analysis is approximately the number of design parameters, which is the reason why initial EM simulations account for a signicant percentage of the entire simulation effort. This paper explores the application of neural networks in combination with the space-mapping optimization for further enhancement of lter design efciency. Neural network has been recognized as a useful alternative for device modeling [1619]. However, direct applications of neural networks to design and optimization of microwave lters are only feasible for simple structures [20] and are not practical in general for lters with complex coupling mechanisms, which require collection of massive training data. The authors have previously presented a method combining neural network and coupling matrix [21], which provides good initial design parameters for the second step of the lter design. Direct application of neural networks for steps 3 and 4, that is, replacing the EM simulator completely using neural models of the entire lter, is still difcult especially for complex lter structures. Techniques based on direct modeling of the generalized scattering matrix of each segment of the lter using neural models [22, 23] generally require enormous EM data and training effort. In Ref. 24, a new approach is presented where a mapping is formulated with the neural models of individual modules and circuit models. The required training data from EM simulation of entire lter is much reduced comparing to conventional neural network techniques. However, collection of EM data for the complete lter structure can still be challenging, because it requires optimization of each lter sample. Furthermore, the model needs to be re-established when lter order changes. This article aims at a signicant improvement of design efciency by practically removing the EM simulation effort required in the third step of the lter design. Both inverse and forward neural models are developed for circuit parameters for lter substructures. Besides initial dimensions, neural models are used to provide information of design parameter sensitivities. The total number of EM simulations is therefore signicantly reduced. Building on previous development of neural models, the improvement in efciency is obtained at practically no additional cost. Furthermore, as the models of substructures do not change

A. Overall Filter Modelling and Design Approach The space mapping optimization method as presented in Ref. 12 is adopted, which is briey reviewed here. Using the same notation in Ref. 12, we let vector m contain all nonzero coupling values in a coupling matrix and input/output couplings R1 and R2 and represent the coarse model parameters for space mapping. The coarse model response is Rc(m). The physical dimensions are ne model parameters, represented by vector d. The full EM simulation results using the values in d are the ne model outputs, that is, Rf(d). There exists a mapping between the vectors m and d, that is, the S-parameters calculated from the coupling values and from the EM simulation match. The mapping is represented by m pd: (1)

After the rst step in the lter design, the ideal coupling values (m*) are obtained and the response Rc(m*) satises design specications, for example, |S11| less than a certain level in given frequency range. The objective is then to nd the optimal design parameters d* corresponding to m*. The optimal design parameters can be determined by solving the nonlinear system of equations pd m 0: (2)

As in Ref. 12, the problem is solved iteratively using linear approximation of the mapping between vectors m and d. At the ith iteration, the linear mapping is given by m mi Ji d di ; (3)

where Ji is a matrix representing the Jacobian of the mapping at the ith iteration. The residue r(d) p(d)-m* is approximated by rdi s % rdi Ji s: (4)

In the ith iteration, the design parameters (di) are used in the EM simulator to obtain S-parameters. Coupling values (mi) are then extracted from the EM simulated S-parameters by enforcing kRf di Rc mi k < e (5)

In this step, the nonlinear least square optimization is used to extract the coupling vector mi. The residue ri

International Journal of RF and Microwave Computer-Aided Engineering/Vol. 22, No. 2, March 2012

Neural Networks in Optimization of Filters r(di) is the difference between mi and ideal coupling vector m*. The optimization step s is obtained by minimizing the l2-norm of the residual in (4). The design parameters are updated in the next iteration to be di s, which are subsequently used for EM simulation. Ji is also updated using Broyden formula as the optimization progresses. The iterations continue until ||ri|| becomes sufciently small. The original trust region aggressive space mapping algorithms do not require an initial set of ne model points. In the example given in Ref. 11 for the design of a high-temperature superconducting lter, the coarse model uses the empirical models of microstrip lines in a circuit simulation and the ne model uses full EM simulator. The vectors for both models contain the same design parameters, that is, coupled line lengths and gaps. When the coarse model and ne model parameters have same physical meanings, the initial value for ne model can be simply set as the optimum value of the coarse model and the initial value of the Jacobian J0 can be assumed to be I. However, it cannot be directly applied in the space mapping problem as dened herein, because the vector of ne model parameters (d) and the vector of coarse model parameters (m) have completely different physical meanings. It is obvious that these assumptions are erroneous, as d0 = m* and J0 = I in this case. The initial calculations, especially initial sensitivity analysis for J0, require signicant simulation overhead. Although the accuracy of initial Jacobian is not required to be extremely high because of the robustness of the space mapping algorithm, an erroneous initial value can negatively affect the optimization performance, for example, signicantly increase the number of iterations required for convergence or cause the optimization to fail. In the design example presented in Ref. 12 with 12 design parameters, seven initial EM simulations are required for sensitivity analysis, while the space-mapping optimization converges after only 10 iterations. Much effort has been made for fast and accurate calculation of initial dimensions, which can be achieved through proper applications of neural networks as proven in Ref. 21. Here, the efforts on collecting training and testing data are further exploited: neural models are developed to obtain both the initial value d0 and the initial value of the Jacobian J0. As a result, we present a complete design and optimization process with high efciency, in which EM simulation overhead is completely eliminated before space mapping optimization. For development of neural network models, the lter structure is decomposed into modules considering coupling mechanisms. Each junction module is calculated using EM simulator and coupling values are extracted from the EM simulated S-parameters. Such a modular approach signicantly reduces the cost of data collection and training effort, as the number of variables is much less than the overall lter. Furthermore, the developed neural models can be applicable even when lter order changes, as the coupling structures remain the same for different lter orders for many types of lters. Two sets of neural models are subsequently developed for each module: the inverse model [21], and the forward

161

Figure 1

Example of a six-pole dual-mode lter.

model. For the inverse models, model inputs are electrical parameters, such as coupling values, and the model outputs are physical dimensions. The inverse models are for initial value calculation, as the electrical parameters are given and the physical dimensions are unknowns. The model inputs of the forward models are physical dimensions and outputs are coupling values. The forward models are used for sensitivity analysis. For example, an input coupling probe can affect both coupling values and resonant frequency of the rst resonator. The relationship between the change in the dimension of the probe and coupling coefcients can be readily calculated using the forward model and used in J0. It is worth mentioning that development of the second set of neural models does not require additional EM data. Only exchanges of certain input and output parameters are needed. As they are forward models, development of neural models is straightforward. Furthermore, sensitivity analysis around the initial design parameters does not cost additional computation time, because the evaluation of neural models is extremely fast. Once the neural models of lter building blocks are developed, lter design can be performed in these steps: Step 1. Obtain ideal coupling matrix to meet specications. Step 2. Use the inverse neural network models of lter substructures to calculate initial design parameters using ideal coupling values as input. Step 3. Perform sensitivity analysis using the forward neural models in the vicinity of the initial design parameters. Obtain the initial Jacobian by nite difference. Step 4. Optimize the lter design using space-mapping until a preset accuracy criterion is met.

B. Example of Development of Neural Models and J0 Without losing generality, we use an example of six-pole end-coupled waveguide dual-mode lter as shown in Figure 1 to demonstrate the development of neural models and the calculation of J0. In this case the design parameters include I/O iris lengths (LR1 and LR2), internal iris lengths (L23, L14, and L45), tuning and coupling screw depths (L11, L22, L33, L44, L55, L66, L12, L34, and L56), that is, the vector of ne model parameters is d LR1 LR2 L11 L22 L33 L44 L55 L66 L12 L23 L34 L45 L56 L14 T : (6)

162

Wang et al. The tuning and coupling screws and the internal irises are modelled similarly. The neural model for initial value calculation is described in Ref. 21. The inputs and outputs of the neural model for sensitivity analysis, that is, the forward model, are given by (7) x D f 0 Lc Lt T y M12 Ph Pv T ; (11)

The coupling matrix is shown below. 2 6M 6 12 6 6 0 M6 6M 6 14 6 4 0 0 R1 ; R2 M11 M12 M22 M23 0 0 0 0 M23 M33 M34 0 0 M14 0 M34 M44 M45 0 0 0 0 M45 M55 M56 0 7 7 7 0 7 7 0 7 7 7 M56 5 M66 0 3

The corresponding vector m of coarse model parameters is given by m R1 R2 M11 M22 M33 M44 M55 M66 M12 M23 M34 M45 M56 M14 T 8 The lter structure can be decomposed into three types of waveguide junctions, namely I/O coupling irises, coupling/tuning screws and internal coupling irises. In the case of I/O irises, for calculation of initial dimensions the input vector xr and output vector yr are given by [21] x r D f 0 R T yr L Pv P h T ; (9)

where M12 is the coupling between the two orthogonal modes. Lc and Lt are the screw lengths. Coupling screw Lc is changed first to Lc DL and changes in all three output variables are recorded. For tuning screw Lt, it is assumed that it only affects one resonance frequency. For example, the horizontal screw can only change Ph. Any small variations in M12 and Pv are ignored. In doing so, the matrix Ji is kept sparse during this process. More than 34,000 EM data points are generated for development of neural models. The inputs and outputs of the internal iris forward model are x D f0 Lv Lh T y M23 M14 Ph Pv T (12) M23 and M14 are coupling values. Lv and Lh are the vertical and horizontal iris lengths, respectively. The number of EM data points is $37,000. The slot iris is trained separately, as shown in Figure 1 (L45 between the second and the third cavities), for which the inputs and outputs of the forward model are x D f 0 L T y M Ph Pv T (13)

where f0 is the lter center frequency, D is cavity diameter, R is the I/O coupling value, L is the iris length, Pv and Ph are the loading effect of the coupling iris on the two orthogonal modes in one cavity, respectively. For sensitivity analysis, the forward model is used. The input and output vectors are x D f 0 L T y R P v T (10)

The total number of data points is $3,000.

as a change in the length L affects the I/O coupling R and the resonance frequency of the first resonator, which changes as the phase length Pv varies. Ph is ignored because the effect is relatively small. Note that the only difference between the models in (9) and (10) is the exchange of variables R and L in the input and output vectors. Therefore, no additional EM data are required. Approximately 125,000 EM simulated data points have been generated covering the frequency range from 10.7 to 12.7 GHz and 1% bandwidth. Such a wide frequency range is sufficient for Ku band single-mode or dual-mode circular waveguide pseudo-elliptic filters, regardless of filter order, for applications in output multiplexing in communication satellites. The data collection is thus a one time effort, as long as the filter configuration remains unchanged. Note that the number of data samples depends on factors such as frequency range and bandwidth range and maybe reduced depending on the application. The neural model in (9) is used rst to obtain initial value of L. To obtain the initial value of the Jacobian, which is approximated by nite difference around the initial design parameters, the length of the iris is then changed to L DL and the variations in R and Pv are calculated using the neural model in (10).

A. Example of a Six-Pole End-Coupled Dual-Mode Filter The rst design example is a six-pole end-coupled waveguide dual-mode pseudo-elliptic lter. The lter is centered at 12.155 GHz, with 64 MHz bandwidth. The ideal coupling values are R1 R2 1:077 2 0 0:855 0 6 0:855 0 0:719 6 6 6 0 0:719 0 M6 6 0:16 0 0:558 6 6 4 0 0 0 0 0 0 3 0 0 7 7 7 0 7 7 0 7 7 7 0:87 5 0 (14)

0:16 0 0 0 0:558 0 0 0:614 0:614 0 0 0:87

As shown in Figures 2a and 2b, the EM responses based on initial dimensions obtained from inverse neural network models as discussed in Ref. 21 are not satisfactory. The rootmean-square (RMS) error between the ideal couplings and extracted couplings of the initial EM response is 0.40. The design is subsequently optimized using the technique described herein. Only two iterations are needed for the

International Journal of RF and Microwave Computer-Aided Engineering/Vol. 22, No. 2, March 2012

Neural Networks in Optimization of Filters

163

2 8.39 8.37 1.26 2.15 1.63 1.60 2.14 1.30 2.35 6.32 1.90 6.27 2.36 4.03

pared in Table III. For the conventional method, which also converges after two iterations, the total number of full EM simulation includes one initial response calculation, seven simulations for initial sensitivity analysis and two optimization iterations. In addition to the fact that the time required to compute initial dimensions is virtually removed, the new method eliminates the computation time required for initial sensitivity analysis, resulting in 70% reduction in total CPU time. The superiority of the proposed method is obvious. Furthermore when the lter order changes, there is no need to develop new neural models, as the models developed for lter building blocks are applicable to any lter order. B. Example of a Four-Pole Side-Coupled Dual-Mode Filter The second design example is a four-pole side-coupled waveguide dual-mode lter, centered at 11.6 GHz with 36 MHz bandwidth. The ideal coupling matrix is shown below. R1 R2 1:07 2 0 6 0:88 6 M6 4 0 0:10 3 0:88 0 0:10 0 0:73 0 7 7 7 0:73 0 0:88 5 0 0:88 0

Figure 2 EM responses during optimization process for the sixpole end-coupled dual-mode lter: (a) S21 and (b) S11. (c) Comparison of EM responses of the optimized design (second iteration) with measurement results.

(15)

optimization to converge, that is, the rms error of the extracted coupling values is reduced to be less than 0.02. To demonstrate progresses of optimization process, the EM responses of each iteration are compared in Figures 2a and 2b. Table I lists all design parameters during space mapping optimization and Table II compares rms error of the coupling values after each optimization step. The EM response of the last iteration is compared with measurement in Figure 2c, showing very good agreement. For comparison, the lter is designed and optimized using conventional method. The computation time is com-

The lter geometry is given in Figure 3 showing all design parameters. The vector of ne model parameters is d LR1 LR2 L11 L22 L33 L44 L12 L23 L34 L14 T (16)

TABLE II Error in Each Iteration during Space Mapping Optimization for the Six-Pole End-Coupled Filter Iterations RMS error of coupling values 0 0.40 1 0.091 2 0.019

International Journal of RF and Microwave Computer-Aided Engineering DOI 10.1002/mmce

164

Wang et al.

TABLE III Comparison of CPU Time for the Six-Pole End-Coupled Filter Conventional method CPU time for initial dimensions calculation Number of full EM simulations CPU time for each full EM simulation Total CPU time 7 min 10 14.5 min 152 min Proposed method <1 s 3 14.5 min 43.5 min

The lter can also be decomposed into three types of waveguide junctions although the structure is different from end-coupled lters. This type of lter is in general more difcult to design due to structural complexity. The initial EM responses and the responses of selected iterations are compared in Figures 4a and 4b, showing progress during the optimization. Six iterations are needed for the optimization to converge, as shown in the errors of the coupling values in Table IV. The optimum lter parameters are LR1 0.917 mm, LR2 0.916 mm, L11 0.111 mm, L22 0.243 mm, L33 0.245 mm, L44 0.120 mm, L12 0.184 mm, L23 0.919 mm, L34 0.185 mm, and L14 0.955 mm. The EM response of the last iteration agrees with measurement as demonstrated in Figure 4c. Conventional optimization method needs additional six full EM simulations for initial sensitivity analysis. Each full EM simulation based on mode-matching method takes 6 min. If nite-element-based EM simulator is used instead, which takes 45 min for each full EM simulation [24], the improvement in efciency using the proposed method will be even more signicant.
4. CONCLUSIONS

A complete lter design/optimization methodology, combining coupling matrix, neural models of lter substructures and space-mapping optimization, is presented. Design examples show that the total CPU times for design and optiFigure 4 EM responses during optimization process for the four-pole side-coupled dual-mode lter: (a) S21 and (b) S11. (c) Comparison of EM responses of the optimized design (sixth iteration) with measurement results.

mization are reduced by 5070% comparing with conventional method. The number of optimization iterations may vary for different designs, depending on structural
TABLE IV Error in Each Iteration during Space Mapping Optimization for the Four-Pole Side-Coupled Filter Iterations 0 1 2 3 4 5 6

Figure 3 A side-coupled circular waveguide dual-mode lter.

RMS error of 0.30 0.084 0.068 0.052 0.033 0.026 0.017 coupling values

International Journal of RF and Microwave Computer-Aided Engineering/Vol. 22, No. 2, March 2012

Neural Networks in Optimization of Filters complexity and accuracy of initial dimensions. However, the number of preliminary EM simulations and the number of iterations are typically comparable using conventional method. The total number of full EM simulations of entire lter structures is therefore signicantly reduced by eliminating the initial sensitivity analysis. Built on previous development of neural models, the much improved efciency is achieved with virtually no additional cost. Furthermore, the technique presented herein can be readily applied to design and optimization of other types of lters.
ACKNOWLEDGMENTS

165

This work was supported in part by Natural Sciences and Engineering Research Council of Canada, in part by COM DEV Ltd., and in part by the Ontario Centers of Excellence.
REFERENCES 1. A.J. Booker, J.E. Dennis, Jr., P.D. Frank, D.B. Serani, V. Torczon, and M.W. Trosset, A rigorous framework for optimization of expensive functions by surrogates, Struct Optim 17 (1999), 113. 2. N.V. Queipo, R.T. Haftka, W. Shyy, T. Goel, R. Vaidynathan, and P.K. Tucker, Surrogate-based analysis and optimization, Prog Aerosp Sci 41 (2005), 128. 3. M.H. Bakr, J.W. Bandler, K. Madsen, and J. Sndergaard, Review of the space mapping approach to engineering optimization and modeling, Opt Eng 1 (2000), 241276. 4. J.W. Bandler, Q.S. Cheng, S.A. Dakroury, A.S. Mohamed, M.H. Bakr, K. Madsen, and J. Sondergaard, Space mapping: The state of the art, IEEE Trans Microwave Theory Tech 52 (2004), 337361. 5. S. Koziel, Q.S. Cheng, and J.W. Bandler, Space mapping, Microwave Mag IEEE 9 (2008), 105122. 6. Q.S. Cheng and J.W. Bandler, Combining coarse and ne models for optimal design, Microwave Mag IEEE 9 (2008), 7988. 7. S. Koziel, J.W. Bandler, and Q.S. Cheng, Robust trust-region space-mapping algorithms for microwave design optimization, IEEE Trans Microwave Theory Tech 58 (2010), 21662174. 8. S. Koziel and J.W. Bandler, Space mapping algorithm with improved convergence properties for microwave design optimization, Int J RF Microwave CAE 20 (2010), 230240. 9. Q.S. Cheng, J.W. Bandler, S. Koziel, M.H. Bakr, and S. Ogurtsov, The state of the art of microwave CAD: EM-based optimization and modeling, Int J RF Microwave CAE 20 (2010), 475491. 10. S. Koziel and J.W. Bandler, Coarse and surrogate model assessment for engineering design optimization with space mapping, Microwave Symposium, 2007. IEEE/MTT-S International, 38 June, 2007, pp. 107110.

11. M.H. Bakr, J.W. Bandler, R.M. Biernacki, S.H. Chen, and K. Madsen, A trust region aggressive space mapping algorithm for EM optimization, IEEE Trans Microwave Theory Tech 46 (1998), 24122425. 12. M.A. Ismail, D. Smith, A. Panariello, Y. Wang, and M. Yu, EM-based design of large-scale dielectric-resonator lters and multiplexers by space mapping, IEEE Trans Microwave Theory Tech 52 (2004), 386392. 13. R.J. Cameron, C.M. Kudsia, and R.R. Mansour, Microwave Filters for Communication Systems: Fundamentals, Design and Applications, John Wiley & Sons, Inc.: Hoboken, NJ, 2007. 14. R.J. Cameron, Advanced coupling matrix synthesis techniques for microwave lters, IEEE Trans Microwave Theory Tech 51 (2003), 110. 15. R.J. Cameron, General coupling matrix synthesis methods for Chebyshev ltering functions, IEEE Trans Microwave Theory Tech 47 (1999), 433442. 16. P. Burrascano, S. Fiori, and M. Mongiardo, A review of articial neural networks applications in microwave computeraided design, Int J RF Microwave CAE 9 (1999), 158174. 17. Q.J. Zhang, K.C. Gupta, and V.K. Devabhaktuni, Articial neural networks for RF and microwave design-from theory to practice, IEEE Trans Microwave Theory Tech 51 (2003), 13391350. 18. J.E. Rayas-Sanchez, EM-based optimization of microwave circuits using articial neural networks: The state-of-the-art, IEEE Trans Microwave Theory Tech 52 (2004), 420435. 19. A. Patnaik and R.K. Mishra, ANN techniques in microwave engineering, IEEE Microwave Mag 1 (2000), 5560. 20. G. Fedi, A. Gaggelli, S. Manetti, and G. Pelosi, Directcoupled cavity lters design using a hybrid feedforward neural network-nite elements procedure, Int J RF Microwave CAE 9 (1999), 287296. 21. H. Kabir, Y. Wang, M. Yu, and Q.J. Zhang, Neural network inverse modeling and applications to microwave lter design, IEEE Trans Microwave Theory Tech 56 (2008), 867879. 22. S. Bila, D. Billargeat, M. Aubourg, S. Verdeyme, and P. Guillon, A full electromagnetic CAD tool for microwave devices using a nite element method and neural networks, Int J Numer Model: Electron Networks Devices Fields 13 (2000), 16780. 23. F. Nunez and A.K. Skrivervik, Filter approximation by RBFNN and segmentation method, IEEE MTT-S Int Microwave Symp Digest 3 (2004), 15611564. 24. H. Kabir, Y. Wang, M. Yu, and Q.J. Zhang, High-dimensional neural-network technique and applications to microwave lter modeling, IEEE Trans Microwave Theory Tech 58 (2010), 145156. 25. J. Zheng and M. Yu, Rigorous mode-matching method of circular to off-center rectangular side-coupled waveguide junctions for lter applications, IEEE Trans Microwave Theory Tech 55 (2007), 23652373.

BIOGRAPHIES

Ying Wang received the B. Eng. and the Masters degrees in electronic engineering from Nanjing University of Science and Technology, Nanjing, China, and the Ph.D. degree in electrical engineering from University of Waterloo, Waterloo, Ontario, Canada, in 1993, 1996, and 2000, respectively. From 2000 to 2007, she was with COM DEV, Cambridge,

Ontario, Canada, where she was involved in development of CAD software for design, simulation and optimization of microwave circuits for space application. In 2007, she joined the Faculty of Engineering and Applied Science, University of Ontario Institute of Technology, Oshawa, Ontario, Canada, where she is currently an Assistant Professor. Her research interests include RF/microwave computer aided design, microwave/millimeter-wave circuits design and radio wave propagation modeling.

International Journal of RF and Microwave Computer-Aided Engineering DOI 10.1002/mmce

166

Wang et al. Currently he is working at the R&D department of COM DEV where he is involved in modeling, design automation of electronic components for satellite applications. He received the NSERC industrial post doctoral research fellowship in 2010 for 2 years. He is the recipient of Senate medal at Carleton University for outstanding research achievement at doctoral level in 2009. Dr. Kabir published over 25 journal and conference papers. His research interests include RF/Microwave modeling, design and optimization. Qi-Jun Zhang received the B.Eng. degree from Nanjing University of Science and Technology, Nanjing, China, in 1982, and the Ph.D. degree in electrical engineering from McMaster University, Hamilton, Ontario, Canada, in 1987. From 1988 to 1990, he was with Optimization Systems Associates (OSA) Inc., Ontario, Canada. In 1990, he joined the Department of Electronics, Carleton University, Ottawa, Ontario, Canada, where he is currently a full Professor. His research interests are microwave CAD, neural network, and optimization methods for high-speed/highfrequency circuit design. He is a member on the editorial board of the IEEE Transactions on Microwave Theory and Techniques, the International Journal of Mathematical Modeling and Numerical Optimization, and the International Journal of Numerical Modeling. He is an Associate Editor for the International Journal of RF and Microwave CAE. He is a member of the Technical Committee on CAD (MTT-1) of the IEEE MTT Society. He is a Fellow of the Electromagnetics Academy, and a Fellow of the IEEE.

Ming Yu received the Ph.D. degree in electrical engineering from University of Victoria, Canada, in 1995. He is an IEEE Fellow. In 1993, he joined COM DEV, Cambridge, ON, Canada, where he is currently the Chief Scientist and Director of R&D. He is responsible for overseeing the development of company R&D Roadmap and next generation products and technologies, including high frequency and high power engineering, electromagnetic based CAD and tuning for complex and large problems, novel miniaturization techniques for microwave networks. He is also an Adjunct Professor with University of Waterloo, Waterloo, ON, Canada. He has authored or coauthored over 100 publications, numerous proprietary reports and holds eight patents with six more pending. Dr. Yu is IEEE MTT Filter committee Chair (MTT-8) since 2011. He is an Associate Editor of IEEE Transactions on Microwave Theory and Technique. He was the recipient of the 1995 and 2006 COM DEV Achievement Award. Humayun Kabir received the B.Sc. degree in electrical and electronic engineering from Bangladesh University of Engineering & Technology, Dhaka, Bangladesh, the MSEE degree from University of Arkansas, Fayetteville, AR, USA and the PhD degree in electrical and computer engineering from Carleton University, Ottawa, Canada in 1999, 2003, and 2009, respectively. Dr. Kabir worked at HiDEC, AR, USA as a research assistant from 2001 to 2003 and at COM DEV, ON, Canada in summer 2005.

International Journal of RF and Microwave Computer-Aided Engineering/Vol. 22, No. 2, March 2012