IEEE TRANSACTIONS ON MICROWAVE THEORY AND …mingyu/publications/P201305.pdf · 2013-11-27 · IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 61, NO. 5, MAY 2013 1839

IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 61, NO. 5, MAY 2013 1839

Hybrid Models for Effective Design and Optimizationof Large-Scale Multiplexing Networks

Ying Wang, Senior Member, IEEE, Shuqi Li, and Ming Yu, Fellow, IEEE

Abstract—In this paper, hybrid models for the design and opti-mization of large-scale multiplexers are presented. First, analysisbased on the low-pass prototype filter circuit model proves thatthe first cavity of a channel filter next to the manifold is criticalin the design of multiplexers. A hybrid model is proposed, com-bining a circuit model and the electromagnetic (EM) analysis ofthe coupling junction close to the manifold. The fact that only asingle junction requires EM simulation facilitates the introductionof neural networks to multiplexer design for the first time. Anotherformat of the hybrid model is then built by replacing the EM sim-ulation with a neural model for the output coupling to the firstresonator. The efficiency of the method is demonstrated using asix-channel multiplexer and a 13-channel multiplexer. Responsespredicted by the hybridmodels agree well with full EM simulationsand measurement results. The proposed hybrid models are provento be able to provide a high degree of accuracy in both in-bandand out-of-band frequency ranges and fast, enabling iterative op-timization involving all design parameters for large-scale multi-plexers.

Index Terms—Microwave filters, multiplexer design, multi-plexing networks, neural networks.

I. INTRODUCTION

A S AN essential part of satellite communication systems,multiplexing networks are composed of multiple band-

pass channel filters to combine the power outputs from the am-plifiers [1], [2]. The conventional design approach for a multi-plexer includes representing the channel filters with analyticalmodels, e.g., coupling matrices [3]. The interconnecting wave-guide junctions are modeled using scattering parameters, andthe waveguides between junctions and channel filters are repre-sented using the transmission line circuit model. It is thereforefeasible to carry out global optimization. Filter dimensions canbe obtained by simulating sections of the filter structure with anelectromagnetic (EM) simulator and relating each section withcircuit parameters in the analytical model [4], [5]. When a spu-rious mode is identified inside or close to a transmission band,filter dimensions are modified to shift the spurious mode away.

Manuscript received February 24, 2013; accepted February 26, 2013. Dateof publication March 29, 2013; date of current version May 02, 2013. Thiswork was supported in part by the Natural Sciences and Engineering ResearchCouncil of Canada, COM DEV Ltd., and the Ontario Centers of Excellence.Y. Wang is with the Department of Electrical, Computer and Software En-

gineering, University of Ontario Institute of Technology, Oshawa, ON, CanadaL1H 7K4 (e-mail: [email protected]).S. Li and M. Yu are with COMDEV Ltd., Cambridge, ON, Canada N1R 7H6

(e-mail: [email protected]).Color versions of one or more of the figures in this paper are available online

at http://ieeexplore.ieee.org.Digital Object Identifier 10.1109/TMTT.2013.2252921

Any remaining effects not identified with the analytical proto-type representation of channel filters are left to be dealt withduring the last manufacturing stage, the tuning stage. As thenumber of channel filters, frequency band coverage, and com-plexity of the filters increase, the damaging effect of spuriousmode interference can no longer be compensated this way, aswill be shown in a design example.Nonideal out-of-band effects need to be taken into account

and compensated during the design stage,which usually requiresfull EM analysis. However, there is limited number of reports onfull-wave optimization of an entire multiplexer due to the com-putational intensiveness. Full-wave optimization has been per-formedonmultiplexerswith channelfilters realizedusing rectan-gular waveguides and without tuning elements [6], [7]. General-izing tomore complicated channel filter configurations is largelyimpractical. EM analysis is usually limited to individual chan-nels. In [8], each channel filter is optimized and simulated usingan EM simulator. The EM -parameters sweep over the multi-plexer frequency band for each channel is then combined withthe circuit model of the interconnecting transmission lines. Onlypartialoptimization ispossible, i.e.,only themanifoldspacingbe-tween channels and the lengths of thewaveguides connecting thechannel filters to the manifold are optimized. In [9], the channelfilter is first simulated using an EM simulator to identify the spu-riousmode.An equivalent circuit of the spuriousmode is then in-cluded in themultiplexermodel in order to compensate for the ef-fectof thespuriousmode throughoptimization.Similar to [8],op-timization parameters are limited tomanifoldwaveguide lengthsand channelfilters other than the one carrying the spuriousmode.Later in [10], a dynamic inverter is developed by comparing the-parameters of a circuit model and those of the EM model of ashort-circuited cavity with the output iris, and then connected tothe rest of the filter circuit model to model the filter broadbandbehavior. The method is applied to a three-channel multiplexerwith a conventional dual-mode cavity filter with a TM110 spu-rious mode. As the wideband behavior is modeled with a simplecircuit model, it may not be valid for a different structure, suchas the side-coupled cavity filters, with different types of spuriousmodes. The model accuracy cannot be guaranteed generalizingto cases other than the examples given. In addition, it can takemore than 10 min per channel to perform a broadband EM simu-lation for a single cavity, as will be shown later. For a large-scalemultiplexer, it is clearly not suitable for iterative optimization. Inthemethods that have been reported, the out-of-band behavior ofa channel is usually assumed unchanged during multiplexer op-timization, which is generally not true.As an alternative modeling technique, neural networks have

been applied to the design of microwave filters [11]–[15]. How-

0018-9480/$31.00 © 2013 IEEE

1840 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 61, NO. 5, MAY 2013

ever, direct applications of neural networks to microwave filtersand multiplexers, i.e., the training of neural models using fullEM simulation of filters or multiplexers, are impractical. Themodel needs to be valid over a very wide frequency range, e.g.,2 GHz at -band, in order to model wideband spurious modeinterference. Collection of training data to cover the interestedinput parameter range for the large number of inputs and overa wide frequency band can be an overwhelmingly time-con-suming task, if at all possible. A different approach is to segmentthe filter and model the generalized scattering matrix (GSM) ofeach segment of the filter using neural networks [14], [15]. Thismethod is more suitable for simple filter structures with a min-imum number of variables, such as the -plane filters [15]. Forcomplex structures, a large number of modes is required to en-sure good accuracy. Using the conventional dual-mode filter asan example, more than ten modes are typically needed for thecircular waveguide port, resulting in a very large GSM at eachfrequency point over the entire multiplexer frequency range. Tomodel these EM data directly requires an enormous amount ofdata, with a huge number of neural network outputs. It is notonly difficult to develop accurate models, such a tremendouseffort is also proven to be unnecessary, as will be shown.In this paper, we present an efficient wideband channel filter

model allowing iterative design optimization involving all de-sign parameters, which is also simple to develop. It has been ob-served that the first cavity of a channel filter next to the manifoldis the most critical in the design of multiplexers [10], [16]. First,analysis based on the low-pass prototype circuit model is pre-sented to provide insights into this effect. A hybrid model is thenproposed, combining a circuit model and EM analysis of part ofthe filter structure, namely, the coupling junction close to themanifold, or the output coupling. As will be demonstrated, boththe in-band response and the out-of-band spurious mode be-havior of the channel filter are captured in the hybridmodel. Fur-thermore, the fact that only a single junction requires EM simu-lation facilitates the introduction of neural networks to the mul-tiplexer design for the first time. Neural networks are capable ofmodeling multi-dimensional nonlinear relationships accuratelyand fast. Only the -parameters of the output coupling junctionare modeled using neural networks. There is no need to modelthe GSM directly. The numbers of both input and output vari-ables are thus minimized, reducing the effort in the developmentof neural models tremendously. The resulting hybrid model isvery fast and the accuracy is comparable to EM simulationsfor both in-band and out-of-band responses. No prior knowl-edge or assumption of the spurious modes is required. Simula-tion and measurement results of a six-channel multiplexer anda 13-channel multiplexer are used to demonstrate the efficiencyof the method. Responses predicted by the hybrid model agreewell with full EM simulations and measurement results.

II. HYBRID MODELS FOR MULTIPLEXING NETWORKS

A. Multiplexer Model

Fig. 1 shows the multiplexer model. Channel filters are con-nected to the manifold through waveguide T-junctions, whichare characterized with pre-calculated -parameters using EMsimulation methods, such as the mode-matching method, over a

Fig. 1. -channel manifold-coupled multiplexer.

Fig. 2. Low-pass prototype ladder networks [17].

range of frequencies covering the bandwidths of all the channelfilters. The manifold waveguide is short-circuited in one end,and the other end is the output common port. Each channel filteris simulated using the hybrid models, replacing conventionalcircuit models, such as the coupling matrix [2], or -parame-ters obtained from EM simulation of the channel filters [8].

B. Circuit Model Analysis of Single Filters

The common parameter used in multiplexer design and op-timization is the common port return loss (CPRL), which de-pends only on the output port return loss of the channel filters,i.e., , as far as the filter parameters are concerned [2]. In thefollowing, we use the low-pass prototype to demonstrate that, inthe out-of-band frequency range, of channel filters is mainlydetermined by the first resonator. In other words, interferenceamong channels can be accurately modeled with a good modelof the first resonator.Fig. 2 shows a prototype low-pass filter with inverters ,

and , where , , and is thefilter order. The cutoff frequency . For one-pole filter,the matrix can be calculated by casting the three com-ponents as follows:

(1)

can be calculated using

(2)

WANG et al.: HYBRID MODELS FOR EFFECTIVE DESIGN AND OPTIMIZATION OF LARGE-SCALE MULTIPLEXING NETWORKS 1841

Since we are only concerned with the out-of-band interference,i.e., , let for . Equation (2) can thenbe simplified to be

(3)

For a two-pole filter,

(4)

Substituting the results in (1),

(5)

Similar to the approximation for the one-pole filter, letand (5) can be simplified as

(6)

becomes

(7)

which is the same as (3) for a one-pole filter.For an -pole filter,

(8)

It is can be shown that for ,

(9)Calculation of is similar to (7) and

(10)

which shows that is only a function of , and the fre-quency when .

Fig. 3. Coupling of a spurious mode.

Fig. 4. (a) Distributed circuit model for a four-pole filter with cross-coupling.(b) Output impedance inverter is replaced with the -parameters of the couplingstructure. (c) Hybrid filter model.

Fig. 3 demonstrates the coupling of a spurious mode. The re-flection coefficient is then changed to

(11)

where , represents the -parameter of the com-ponents enclosed in the dashed-line box as a result of the spu-rious mode.Transformation to bandpass filters from the low-pass net-

work is straightforward. Equations (10) and (11) show that inthe out-of-band frequency range the reflection coefficient(or with the spurious mode coupled to the first resonator)largely depends on the behavior of the first resonator of thefilter. Although the conclusion is drawn using a circuit modelwithout cross-couplings, it is applicable to cross-couplednarrow band filters, as will be demonstrated.

C. Channel Filter Model

Circuit models are fast and usually sufficient for the mod-eling of the in-band performance of a channel filter, especiallyfor narrowband applications such as those in communicationssatellites. In themeantime, an accurate model for the output cou-pling and the first resonator is critical for simulating out-of-bandperformance. Therefore, a hybrid model for channel filters com-bining an EM simulated output coupling and a circuit model is


Fig. 5. (a) Distributed circuit model for a classical four-pole dual mode filter[18]. (b) Hybrid filter model based on the dual-mode distributed model.

Fig. 6. (a) Side-coupled four-pole dual-mode pseudo-elliptic filter [19], [20].(b) Distributed circuit model for the side-coupled dual-mode filter.

presented for an accurate and efficient modeling of both out-of-band interference and in-band behavior.A general distributed model is adopted as the circuit model.

An example of a four-pole filter with cross-coupling is shown inFig. 4(a). Besides the fact that it can model the periodic behavior

Fig. 7. (a) Output cavity and iris of the side-coupled filter. (b) Port numbersof all modes of the GSM of the three-port junction with the reference planesdefined along , , and in (a).

of transmission lines and the model dispersion of the funda-mental mode, the EMmodel of the output coupling can be easilyincorporated. The output impedance inverter is on theman-ifold side and impedance inverter is on the input side of thechannel filter. The distributed model is modified by replacingthe output impedance inverter with the -parameters of the cou-pling structure, e.g., a coupling iris, as shown in Fig. 4(b). The-parameters are generated using an EM simulator, as will beexplained below. is the phase length introduced by the iris,which is to be subtracted from thewaveguide length between thechannel filter and the manifold. is the phase loading effect onthe cavity introduced by the iris, i.e., the electrical length of res-onator 1 ( ) needs to be reduced to . The resulting hybridmodel combining EM and circuit models is shown in Fig. 4(c).The hybrid filter model can be built using other circuit

models. For example, the distributed model for classicaldual-mode filters in [18] is shown in Fig. 5(a). For dual-modefilters, resonators 1 and 2 reside in the same cavity. and

are the admittance inverters that couple the two modes inthe same cavity, which are typically realized using couplingscrews. and are the impedance inverters for sequentialand cross coupling, respectively. The hybrid model shown inFig. 5(b) is constructed by replacing the output impedanceinverter with the -parameters of the output coupling iris,as shown in Fig. 4(b). Fig. 6(a) shows another example, aside-coupled circular waveguide dual-mode filter [19], [20].The input–output coupling and the coupling between resonators


Fig. 8. (a) CPRL and (b) transmission coefficients between the input port ofeach channel filter and the common port of all channels using coupling matrixcircuit models for channel filters.

are realized at the sides of the circular cavities. The distributedmodel is shown in Fig. 6(b). The hybrid model can be builtsimilarly. However, our results show that the differences inusing the distributed model in Fig. 6(b) and the general modelin Fig. 4 are minimal for the narrowband cases studied, whichdoes not justify the use of the complicated model in Fig. 6(b).Therefore, the general model in Fig. 4 is used in the following.

D. Modeling the Output Resonator

In this section, the output coupling is isolated, and the GSMof the coupling junction is generated using an EM simulator.The -parameters of the two-port network representing only thefundamental modes are then calculated by short circuiting allother modes. Without losing generality, we use an example forthe demonstration of the modeling approach.The side-coupled circular waveguide dual-mode filter, as

shown in Fig. 6(a), is used as the example. This type of filteroffers significant performance improvement and finds itsapplications in the satellite communications with extremelystringent mass, size, and thermal requirements. However, thedesign is more challenging compared to conventional longitu-dinal end-coupled filters due to its structural complexity.The output iris together with the cavity is redrawn in Fig. 7(a).

The only section of the filter that is EM simulated is the three-port junction with the reference planes defined along ,

Fig. 9. (a) CPRL and (b) transmission coefficients of all channels with mea-sured channel filter responses.

Fig. 10. Measured phase of channel 6, which diverges significantly fromthe coupling matrix circuit model response.

and . The structure is composed of a rectangular-to-rect-angular waveguide junction and a circular-to-rectangular side-coupled waveguide T-junction. The rigorous mode-matchingtechnique for simulating circular to multiple off-center rectan-gular side-coupled waveguide T-junctions in [20] is used for an-alyzing this structure. The numbers of modes along , ,and are assumed to be 1, , and , respectively, as an ex-ample. The resulting GSM is an matrix, where

. Port numbers of all modes of the GSM are shownin Fig. 7(b). Except for the two fundamental modes marked as


Fig. 11. Comparison of phase of channel filters obtained from the hybrid model, full EM simulation, measurement, and coupling matrix circuit model for:(a)–(f) channel 1 to channel 6, respectively.

Port 1 and Port 2, all other modes are short circuited by the topand bottom of the cavity. can therefore be transformed intothe -parameters for a two-port network. The reference planesfor all modes to be short circuited are first shifted by an appro-priate length: modes along are shifted by , the distancebetween reference plane and the bottom of the cavity, andmodes along are shifted by , the distance betweenand the top of the cavity, except for the fundamental mode. Theresulting matrix is obtained by

(12)

with (13), shown at the bottom of the following page.are wavenumbers for the corresponding

mode. Finally, the -parameters for the two-port networkare calculated by short-circuiting port 3 to port [17].It can then be plugged into the hybrid channel filter model inFig. 4.

E. Neural-Network Model of the Output Coupling

Modeling of the output coupling structure can be readily doneusing neural networks. The structure requiring EM simulation,i.e., the output coupling, is extremely simple. Using the exampleshown in Fig. 7, the inputs of the neural model are given by


, where is the frequency andis the length of the output coupling iris. As described in

Section II-D, there is no need to model the complete GSM di-rectly. The outputs of the neural network model are the -pa-rameters of the two-port network, with one fundamental modeon each port. Since , there are only six unknowns.During data collection, the only variables that require EM sim-ulations are the frequency and the iris length . The iriswidth is kept unchanged here. It can be added as another inputvariable if necessary. Alternatively, different sets of data canbe generated, each with different iris width, since typically thewidth is not varied continuously as the iris length [13]. OnceGSM data are available, the matrix reduction is performed bysimple analytical calculation, i.e., the variations of anddo not require additional EM simulations.In summary, the simple structure, minimum number of vari-

ables, and the straightforward calculation have significantly re-duced the effort required for data generation and model devel-opment, which is key to the application of neural networks toa design problem on such a large scale. In the following, Neu-roModeler [21] is used for neural-network model training andtesting.

III. EXPERIMENTAL RESULTS

The proposed method is applied to the analysis of asix-channel multiplexer from 10.9 to 11.8 GHz. Each channelfilter is a four-pole side-coupled circular waveguide dual-modefilter with 72-MHz bandwidth. The -parameters predictedusing coupling matrix circuit models are shown in Fig. 8, withFig. 8(a) showing CPRL and Fig. 8(b) showing transmissioncoefficients between the input port of each channel filter andthe common port. However, the plots in Fig. 9 with mea-sured channel filter responses show significant deviation fromthose predicted by the circuit model. The CPRL and channeltransmission coefficient, especially the highest frequencychannel (channel 1), are severely deteriorated, as shown inFig. 9(a) and (b). It has been proven in the laboratory thatthe multiplexer is not tunable, which shows the deficiency ofthe conventional design approach. It has been further proventhat, with each channel filter simulated using an EM simulatorand optimizing only the manifold waveguide spacings [8], themultiplexer is also not tunable.

Fig. 12. Comparison of CPRL with channel filter responses obtained from thehybrid model, full EM simulation, and measurement.

Fig. 13. Transmission coefficients of all channels with channel filter responsesobtained from the hybrid model.

A close look at the lowest frequency channel (channel 6)shows that a spurious mode far from the filter center frequencycauses the measured phase of to diverge significantly fromthe circuit model response, as can be seen in Fig. 10. A sim-ilar problem exists in channels 4 and 5 as well. These spu-rious modes of low-frequency channels severely damage the re-sponses of high-frequency channels.Using the method described, hybrid models of all channels

have been developed combining EM simulation of the outputcoupling structure with the rest of the filter represented by acircuit model. phase responses of channel filters obtained

. . .......

...

. . .

(13)


Fig. 14. Comparison of phase of channel filters obtained from the twotypes of hybrid models with the output coupling EM simulated and modeledwith neural networks for selected channels. (a) Channel 1. (b) Channel 6.

Fig. 15. Comparison of CPRL with channel filter responses obtained from thetwo hybrid models: with the output coupling EM simulated and modeled withneural networks.

TABLE ICOMPARISON OF CPU TIME FOR CHANNEL FILTER MODELS

Fig. 16. CPRL optimized using the hybrid model.

TABLE IICOMPARISON OF MANIFOLD DIMENSIONS OBTAINED USING THE CIRCUITMODEL AND THE HYBRID MODEL FOR THE SIX-CHANNEL MULTIPLEXER

Fig. 17. Isometric view (ProE) of the fabricated multiplexer.

from the hybrid model are compared with measurement resultsin Fig. 11 for all six channels. Comparison is also made withfull EM simulation of the entire filter for each channel. Goodagreement is observed among the hybrid model, measurement,and full EM results. As shown in Fig. 11, the coupling matrixcircuit model shows sufficient accuracy for the three high-fre-quency channels. However, the accuracy deteriorates signifi-cantly for the low-frequency channels. Fig. 12 shows the mul-tiplexer CPRL, with the channel filter responses from the hy-brid model, the full EM simulation, and the measurement re-sults, showing good agreements. Transmission coefficients ofall channels with channel filters responses obtained from the hy-brid model are shown in Fig. 13.


TABLE IIISIMULATION AND MEASUREMENT RESULTS OF THE MANIFOLD

DIMENSIONS FOR THE 13-CHANNEL MULTIPLEXER

Fig. 18. (a) CPRL and (b) transmission coefficients of all channels withchannel filter responses obtained from the hybrid model.

Furthermore, neural-network models are developed for theoutput coupling iris and built into the hybrid model. The GSMsare generated for the output iris shown in Fig. 7(a) with 12modes on each of the circular waveguide port, and subsequentlyreduced to the -parameters of the two-port network. Each-parameter is modeled separately to improve model accuracy.Total CPU time for training and testing data collection is lessthan 30 min on an Intel Core 2 Duo 2.2-GHz computer with3-GB RAM. The three-layer multilayer perceptron (MLP)

Fig. 19. Measurement results. (a) CPRL and (b) transmission coefficients ofall channels.

structure and quasi-Newton algorithm are used for the develop-ment of neural models. 36 hidden neurons were used. It took,on average, 50 min to develop each model, with training andtesting errors both below 1%. Fig. 14 compares the phase of

computed using the hybrid models for selected channels. Avery good match is shown between results with output couplingmodeled with the EM simulator and neural networks. Similaraccuracy has been achieved for other channels. The CPRLresponses also agree with each other for the two types of hybridmodels, as shown in Fig. 15.Table I compares the CPU time required to simulate a channel

filter using an EM model and the hybrid model. Full EM simu-lation using a mode-matching method for each channel takes35 min to cover the 1-GHz frequency range. If a finite-ele-ment-based EM simulator is used instead, the computation timewill be more than 1 h [12]. The hybrid model with EM simu-lation for the output coupling reduces the CPU time to 12 minper channel. However, it is still very difficult for iterative op-timization. Using the hybrid model with the neural model, onthe other hand, takes less than 1 s with comparable accuracy tocover the entire multiplexer frequency range. Optimization canthus be done within a few seconds.Fig. 16 shows that by using the hybrid models and involving

all variables in the iterative optimization, the severe spuriousinterference can be compensated, and the multiplexer can betuned to meet specifications. Table II compares the manifold


dimensions obtained using the conventional circuit model andthe hybrid model. Significant difference can be observed.The method is further validated by a 13-channel multi-

plexer from 11.7 to 12.5 GHz, which has been successfullybuilt and tested. Fig. 17 shows the isometric view (ProE) ofthe fabricated multiplexer. Each channel filter is a four-poleside-coupled circular waveguide dual-mode filter. The twohighest frequency channel filters have 72-MHz bandwidtheach, and the three lowest frequency channels have 54-MHzbandwidth each. All other channels have 33-MHz bandwidth.Waveguide WR75 -plane T-junctions are used. All channelsare simulated using the hybrid models. Table III lists all mani-fold dimensions. During tuning, the channel filter-to-manifolddistances have been adjusted slightly. The channel-to-channeldistances are kept unchanged, except for the distance betweenthe end channel and the short circuit. As shown in Table III,the difference between simulation and measurement is small.The CPRL and transmission coefficients of all channels withchannel filter responses obtained using the hybrid model areshown in Fig. 18(a) and (b), respectively. As can be seen, thesimulation results agree very well with the measurement resultsin Fig. 19.

IV. CONCLUSIONS

Hybrid models for the design and optimization of large-scalemultiplexers have been presented. Taking advantage of the factthat the cavity of a channel filter next to themanifold is critical inthe design of multiplexers, hybrid models are built combining acircuit model and the EM analysis or neural-network models ofthe coupling junction close to the manifold. Since only a singlejunction requires EM simulation, development of neural-net-work models is straightforward. The models are simple andhighly efficient, with very good accuracy in both in-band andout-of-band frequency ranges. The method is validated usinga six-channel multiplexer and a 13-channel multiplexer. Com-parisons of the responses of the hybrid models with those byfull EM simulations and measurement show good agreement.The method is applicable to other types of waveguide or trans-mission-line filters. Extension of the approach to dielectric fil-ters is possible in principle. However, due to the complexity offield distribution, such an extension may not be straightforward,which will be the subject of a future study.

REFERENCES

[1] C. M. Kudsia, R. Cameron, and W.-C. Tang, “Innovations in mi-crowave filters and multiplexing networks for communicationssatellite systems,” IEEE Trans. Microw. Theory Techn., vol. 40, no. 6,pp. 1133–1149, Jun. 1992.

[2] R. Cameron and M. Yu, “Design of manifold coupled multiplexers,”IEEE Microw. Mag., pp. 46–59, Oct. 2007.

[3] A. E. Atia and A. E. Williams, “Narrow-bandpass waveguide filters,”IEEE Trans. Microw. Theory Techn., vol. MTT-20, no. 4, pp. 258–265,Apr. 1972.

[4] W. Hauth, D. Schmitt, and M. Guglielmi, “Accurate modelling of nar-rowband filters for satellite communications,” in IEEE MTT-S Int. Mi-crow. Symp. Dig., 2000, vol. 3, pp. 1767–1770.

[5] M. Yu and Y.Wang, “Synthesis and beyond,” IEEEMicrow. Mag., vol.12, no. 6, pp. 62–76, Oct. 2011.

[6] M. Guglielmi, “Simple CAD procedure for microwave filters and mul-tiplexers,” IEEE Trans. Microw. Theory Techn., vol. 42, no. 7, pt. 1–2,pp. 1347–1352, Jul. 1994.

[7] L. Accatino and M. Mongiardo, “Hybrid circuit-full-wave computeraided design of a manifold multiplexers without tuning elements,”IEEE Trans. Microw. Theory Techn., vol. 50, no. 9, pp. 2044–2047,Sep. 2002.

[8] M. A. Ismail, D. Smith, A. Panariello, Y. Wang, and M. Yu, “EM-based design of large-scale dielectric-resonator filters and multiplexersby space mapping,” IEEE Trans. Microw. Theory Techn., vol. 52, no.1, pt. 2, pp. 386–392, Jan. 2004.

[9] D. Bariant, S. Bila, D. Baillargeat, S. Verdeyme, P. Guillon, D. Pacaud,and J. Herren, “Method of spurious mode compensation applied tomanifold multiplexer design,” in IEEEMTT-S Int. Microw. Symp. Dig.,Jun. 2002, pp. 1461–1464.

[10] P. Lenoir, S. Bila, D. Baillargeat, and S. Verdeyme, “Development ofan electromagnetic tool for the accurate synthesis of manifold mul-tiplexers: The dynamic inverter,” Int. J. RF Microw. Comput.-AidedEng., vol. 15, no. 3, pp. 278–285, May 2005.

[11] Q. J. Zhang, K. C. Gupta, and V. K. Devabhaktuni, “Artificial neuralnetworks for RF and microwave design-from theory to practice,” IEEETrans. Microw. Theory Techn., vol. 51, no. 4, pp. 1339–1350, Apr.2003.

[12] H. Kabir, Y. Wang, M. Yu, and Q. J. Zhang, “High-dimensionalneural-network technique and applications to microwave filter mod-eling,” IEEE Trans. Microw. Theory Techn., vol. 58, no. 1, pp.145–156, Jan. 2010.

[13] Y. Wang, M. Yu, H. Kabir, and Q. J. Zhang, “Effective design of cross-coupled filter using neural networks and coupling matrix,” in IEEEMTT-S Int. Microw. Symp. Dig., Jun. 2006, pp. 1431–1434.

[14] S. Bila, D. Baillargeat, M. Aubourg, S. Verdeyme, and P. Guillon, “Afull electromagnetic CAD tool for microwave devices using a finiteelement method and neural networks,” Int. J. Numer. Model., vol. 13,pp. 167–180, Mar. 2000.

[15] J. M. Cid and J. Zapata, “CAD of rectangular-waveguide -plane cir-cuits by segmentation, finite elements and artificial neural networks,”Electron. Lett., vol. 37, no. 2, pp. 98–99, Jan. 2001.

[16] A. Morini, G. Cereda, and M. Virgulti, “Re-configurable reciprocalmultiplexers (r-mux) for terrestrial radio links,” in Proc. 32nd Eur. Mi-crow. Conf., Milan, Italy, Sep. 2002, pp. 1–4.

[17] R. J. Cameron, C.M.Kudsia, and R. R.Mansour, Microwave Filters forCommunication Systems: Fundamentals, Design and Applications.New York, NY, USA: Wiley, 2007.

[18] S. Cogollos, M. Brumos, V. E. Boria, C. Vicente, J. Gil, B. Gimeno,and M. Guglielmi, “A systematic design procedure of classical dual-mode circular waveguide filters using an equivalent distributedmodel,”IEEE Trans. Microw. Theory Techn., vol. 60, no. 4, pp. 1006–1017,Apr. 2012.

[19] M. Yu, D. J. Smith, A. Sivadas, and W. Fitzpatrick, “A dual modefilter with trifurcated iris and reduced footprint,” in IEEE MTT-S Int.Microw. Symp. Dig., Jun. 2002, pp. 1457–1460.

[20] J. Zheng and M. Yu, “Rigorous mode-matching method of circular tooff-center rectangular side-coupled waveguide junctions for filter ap-plications,” IEEE Trans. Microw. Theory Techn., vol. 55, no. 11, pp.2365–2373, Nov. 2007.

[21] Q. J. Zhang, NeuroModeler. Dept. Electron., Carleton Univ., Ottawa,ON, Canada, 2004.

YingWang (M’05–SM’12) received the B. Eng. andMaster’s degrees in electronic engineering from theNanjing University of Science and Technology, Nan-jing, China, in 1993 and 1996, respectively, and thePh.D. degree in electrical engineering from the Uni-versity of Waterloo, Waterloo, ON, Canada, in 2000.From 2000 to 2007, she was with COM DEV Ltd.,

Cambridge, ON, Canada. As a Senior Member ofTechnical Staff, she was involved in development ofcomputer-aided design (CAD) software for design,simulation, and optimization of microwave circuits

for space applications. In 2007, she joined the Faculty of Engineering andApplied Science, University of Ontario Institute of Technology, Oshawa, ON,Canada, where she is currently an Assistant Professor. Her research interestsinclude RF/microwave CAD, microwave circuit design, and radio-wavepropagation modeling.


Shuqi Li received the B.Eng. degree from YangzhouUniversity, Yangzhou, China, in 2006, and theM.A.Sc. degree in electrical engineering fromthe University of Ontario Institute of Technology,Oshawa, ON, Canada, in 2011.In 2012, she joined COM DEV Ltd., Cambridge,

ON, Canada, as a Member of Technical Staff, whereshe has been involved in the design of microwave fil-ters and multiplexers for space applications.

Ming Yu (S’90–M’93–SM’01–F’09) received thePh.D. degree in electrical engineering from theUniversity of Victoria, Victoria, BC, Canada, in1995.In 1993, while working on his doctoral dissertation

part time, he joined COMDEVLtd., Cambridge, ON,Canada, as a Member of Technical Staff. He was in-volved in designing passive microwave/RF hardwarefrom 300MHz to 60GHz for both space- and ground-based applications. He was also a principal devel-oper of a variety of COMDEV Ltd.’s core design and

tuning software for microwave filters and multiplexers, including computer-aided tuning software in 1994 and a fully automated robotic diplexer tuningsystem in 1999. His varied experience also includes having been the Managerof Filter/Multiplexer Technology (Space Group) and Staff Scientist of Corpo-rate Research and Development (R&D). He is currently the Chief Scientist andDirector of R&D with COM DEV Ltd. He is responsible for overseeing thedevelopment of company R&D Roadmap and next-generation products andtechnologies, including high-frequency and high-power engineering, EM-basedCAD and tuning for complex and large problems, and novel miniaturizationtechniques for microwave networks. He is also an Adjunct Professor with theUniversity of Waterloo, Waterloo, ON, Canada. He has authored or coauthoredover 100 publications and numerous proprietary reports. He holds eight patentswith six pending.Dr. Yu was an IEEE Distinguished Microwave Lecturer (2010–2012). He

has been an IEEE Microwave Theory and Techniques Society (IEEE MTT-S)Filter Committee chair (MTT-8) since 2010 and was also chair of TPC-11. Heis an associate editor for the IEEE TRANSACTIONS ON MICROWAVE THEORYAND TECHNIQUES. He was the recipient of the 1995 and 2006 COM DEVAchievement Award for the development of computer-aided tuning algorithmsand systems for microwave filters and multiplexers. He holds a NaturalSciences and Engineering Research Council (NSERC) of Canada DiscoveryGrant (2004–2015) with the University of Waterloo.

Documents

IEEE TRANSACTIONS ON MICROWAVE THEORY AND …mingyu/publications/P201305.pdf · 2013-11-27 · IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 61, NO. 5, MAY 2013 1839