close

Вход

Забыли?

вход по аккаунту

?

Efficient design and optimisation techniques for planar microwave filters

код для вставкиСкачать
Efficient Design and Optimisation
Techniques for Planar Microwave
Filters
by
S0ren F. Peik
A Thesis
presented to the University of Waterloo
in fulfilment of the
thesis requirement for the degree of
Doctor of Philosophy
in
Electrical Engineering
Waterloo, Ontario, Canada, 1999
©Spren F. Peik 1999
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
1
*
1
National Library
of Canada
Bibliotheque nationale
du Canada
Acquisitions and
Bibliographic Services
Acquisisitons et
services bibliographiques
395 W ellington Street
Ottawa ON K1A 0N4
Canada
395, rue W ellington
Ottawa ON K1A 0N4
Canada
Your file Votre reference
ISBN: 0-612-92046-1
O ur file Notre reference
ISBN: 0-612-92046-1
The author has granted a non­
exclusive licence allowing the
National Library of Canada to
reproduce, loan, distribute or sell
copies of this thesis in microform,
paper or electronic formats.
L'auteur a accorde une licence non
exclusive permettant a la
Bibliotheque nationale du Canada de
reproduire, preter, distribuer ou
vendre des copies de cette these sous
la forme de microfiche/film, de
reproduction sur papier ou sur format
electronique.
The author retains ownership of the
copyright in this thesis. Neither the
thesis nor substantial extracts from it
may be printed or otherwise
reproduced without the author's
permission.
L'auteur conserve la propriete du
droit d'auteur qui protege cette these.
Ni la these ni des extraits substantiels
de celle-ci ne doivent etre imprimes
ou aturement reproduits sans son
autorisation.
In compliance with the Canadian
Privacy Act some supporting
forms may have been removed
from this dissertation.
Conformement a la loi canadienne
sur la protection de la vie privee,
quelques formulaires secondaires
ont ete enleves de ce manuscrit.
While these forms may be included
in the document page count,
their removal does not represent
any loss of content from the
dissertation.
Bien que ces formulaires
aient inclus dans la pagination,
il n'y aura aucun contenu manquant.
Canada
Reproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
I hereby declare that I am the sole author o f this thesis.
I authorise the University o f W aterloo to lend this thesis to other institutions or individuals for
the purpose o f scholarly research.
I further authorise the University o f W aterloo to reproduce this thesis by photocopying or by
other means, in total or in part, at the request o f other institutions or individuals for the purpose
o f scholarly research.
ii
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
The University o f W aterloo requires the signatures o f all persons using or photocopying this
thesis. Please sign below, and give address and date.
iii
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
Abstract
The topic o f this thesis is the efficient and accurate synthesis o f arbitrarily shaped m icrowave
filters. Currently, filter circuits are synthesised using an iterative optim isation in conjunction
w ith an electrom agnetic (EM ) simulator. This so-called direct EM optim isation has tw o m ajor
drawbacks. It does not generally attain the optim al solution, but rather a sub-optim al solution. In
addition, direct E M optim isation is com putationally very expensive. In this thesis, we propose
four novel techniques to overcom e these drawbacks.
First, we propose the application o f a genetic algorithm (GA) to guide the optim iser towards the
optim al solution. GA’s identify the optim al filter layout even where gradient-based m ethods fail.
We enhance a generic GA in tw o ways. O ur algorithm allows the user to im pose arbitrary con­
straints. O ur G A also provides the user w ith vital data about the correlation and the sensitivity of
the param eters. W ith these enhancem ents, our im proved GA generally attains better-perform ing
circuits than gradient search m ethods at sim ilar com putational expenses.
Second, w e extend the Cauchy m ethod for fast frequency sweeps to a m ulti-dim ensional Cauchy
method, w ith respect to both frequency and physical dim ensions. Hence, our new algorithm can
be used as a substitute for neural networks. The proposed m ulti-dim ensional C auchy m ethod
accurately and inexpensively com putes circuit m odels from arbitrary topologies. It m inim ises the
com putational cost o f the optim isation process. Exam ples dem onstrate that— w ithout sacrificing
accuracy— the optimisation tim e is reduced by several orders o f magnitude.
Third, for tackling large filter circuits, we introduce a hybrid optimisation schem e based on cou­
pling m atrix alteration. This new technique com bines an accurate— but expensive— com plete
electrom agnetic analysis with an inexpensive— but less accurate— decom posed analysis. First,
we derive tw o coupling m atrices o f the filter: one from the response found by com plete anal­
ysis, and one from the response found by decom posed analysis. We then com pare the two
coupling m atrices and apply this knowledge to elim inate inaccuracies in the decom posed circuit
m odel. O ur hybrid technique attains the com putational speed o f the decom posed analysis, but
the accuracy o f the full analysis. The technique can optim ise filter circuits for w hich direct EM
optim isation w ould require excessive com puter resources.
Fourth, we create a technique for optim ising non-resonant filter structures, such as transversal fil­
ters. Transversal filters m ay include up to several hundred design param eters, m aking direct EM
optim isation im possible. An optim isation o f such filters becom es feasible using the tim e-dom ain
response rather than the frequency-dom ain response. O ur technique isolates the fault locations
o f the circuit one after the other by tim e-dom ain reflectometry. Hence, the optim al circuit paiv
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
ram eters can be found one at a time. The tim e-dom ain optim isation presented in this thesis can
optimise a transversal filter w ithout a restrictive upper lim it on the num ber o f param eters.
Several exam ples throughout this thesis illustrate that the proposed techniques result in a consid­
erable gain in design reliability and a significant reduction o f com putational cost when com pared
to conventional m ethods. One chapter is exclusively dedicated to the design and the m easure­
m ent of various m icrowave devices for superconductive satellite systems using our new tech­
niques. The m easured results are in excellent agreem ent w ith the theory.
v
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
Acknowledgements
I would like to thank Professor Y. L. Chow, Professor R.R. M ansour and P rofessor A. SafaviNaeini who agreed to be my supervisors. Their encouragem ent and advice during the course
o f this thesis are deeply appreciated. They also generously supported m e in overcom ing the
countless adm inistrative hurdles that resulted from w ith m y unusual situation w ith one supervisor
in Hong Kong, one in C am bridge, and one in W aterloo, and m yself working full-tim e at C O M
DEV, Cambridge.
I take special pleasure in thanking all m y colleagues at C O M D EV International and C O M D EV
Space Group, for providing m e w ith an excellent environm ent to pursue my research. This
thesis could not have been com pleted w ithout the technical help of the C O M D EV R& D staff.
In particular, I would like to thank Raafat M ansour, Bill Jolley, Tony Rom ano, Shen Ye, and
Ke-Li Wu for the countless stim ulating discussions on m y thesis project.
Last, but not least, I thank m y wife, Gudrun, who often m issed me on the w eekends and late
evenings during the course o f this study.
The preparation o f this thesis was largely supported by C O M DEV International, Cam bridge,
Ontario. The synthesis techniques and hardware described in this thesis are the result o f research
program s supported in part by the D efence A dvanced Research Projects Agency (DARPA), the
Canadian Space Agency (CSA), the Canadian D epartm ent o f National D efence (DN D), and
C O M D EV under N A SA Co-operative A greem ent NCC3-517.
vi
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
To my parents
vii
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
Contents
1 Introduction
1
1.1
M otivation for This Thesis and Proposed S o l u t i o n s ..................................................
2
1.2
Thesis Organisation
5
..........................................................................................................
2 Available Filter Design Techniques
2.1
6
Standard F ilter D esign T e c h n iq u e s ..................................................................................
2.1.1
....................................................................
8
2.2
Enhanced Filter Optim isation A l g o r i t h m s ....................................................................
13
2.3
Lim itations o f Available M ethods and O ur Proposed S o l u t i o n s .............................
15
2.3.1
R obustness and Cost o f G radient C a l c u l a t i o n ...............................................
16
2.3.2
D iscrete M eshes and Interpolation E r r o r s ......................................................
17
2.3.3
Com putational E x p e n s e s ......................................................................................
18
2.4
Application o f Standard M ethods
6
C onclusion
............................................................................................................................
3 Genetic Algorithms in Microwave Circuit Design
19
21
3.1
B a c k g r o u n d ............................................................................................................................
21
3.2
A lgorithm Outline and A daption to Circuit S y n t h e s i s ................................................
22
3.3
M icrowave-Specific Extensions
25
.....................................................................................
3.3.1
Im plem enting Physical C onstraints
.................................................................
26
3.3.2
Param eter Space V is u a lis a tio n ...........................................................................
27
viii
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
3.4
E x a m p l e s ..............................................................................................................................
29
3.4.1
K a-B and Stripline R ejection Filter D e s ig n ......................................................
29
3.4.2
W aveguide Cavity F ilter D e s i g n ........................................................................
33
3.5
O ptim isation Perform ance o f GA’s ..................................................................................
34
3.6
Conclusion
36
..........................................................................................................................
4 Multi-Dimensional Cauchy Method
4.1
37
M ulti-D im ensional In te r p o la tio n ....................................................................................
38
4.1.1
N eural N e tw o r k s ....................................................................................................
38
C auchy M e t h o d ....................................................................................................................
42
4.2.1
O ne-D im ensional Cauchy M e t h o d ....................................................................
43
4.3
Adaptive S a m p l i n g ............................................................................................................
46
4.4
M ulti-dim ensional Cauchy M e th o d ..................................................................................
46
4.4.1
Recursive Cauchy M e t h o d ...................................................................................
46
4.4.2
M ulti-D im ensional Cauchy E x p a n sio n ..............................................................
52
Num erical R e s u lts ................................................................................................................
53
4.5.1
M icrostrip Line Im p e d a n c e ...................................................................................
53
4.5.2
R ecessed Line Fed M icrostrip A n t e n n a ..........................................................
53
4.5.3
N arrow -Band 3-Pole F i l t e r ...................................................................................
56
4.2
4.5
4.6
C onclusion
...........................................................................................................................
5 Hybrid Optimisation Using Coupling Matrix Alteration
5.1
5.2
57
61
B a c k g r o u n d ...........................................................................................................................
62
5.1.1
Space M a p p in g ........................................................................................................
66
M -Param eter Correction......... .............................................................................................
67
5.2.1
C oupling M atrix R e p re s e n ta tio n ........................................................................
68
5.2.2 Coupling M atrix o f Three-Pole F i l t e r ..............................................................
71
ix
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
5.2.3
M -Param eter C orrected O p tim is a tio n .............................................................
72
5.3
N um erical E x a m p le ............................................................................................................
74
5.4
C om parison o f Com putation T i m e ................................................................................
83
5.5
Conclusion
84
..........................................................................................................................
6 Optimisation by Time-Domain Reflectometry
6.1
Theory
6.1.1
.................................................................................................................................
88
Transform ation into T im e -D o m a in ....................................................................
90
6.1.3
Pinpointing o f Fault L ocation
...........................................................................
91
6.1.4
Elim ination o f F a u l t s .............................................................................................
93
6.1.5
N on-U niqueness o f Fault E lim in a tio n .............................................................
95
6.1.6
Elim inating M ultiple Faults
...............................................................................
95
Pulse T r a c k i n g ........................................................................................................
97
6
.1.2
.1.7
Ideal F ilter Synthesis and R eal Filter Perform ance
87
.....................................
6
6.2
Tuning o f F ilter H ardw are
...............................................................................................
97
6.3
Optim isation and Tuning o f N arrow -Band R esonant Filter C i r c u i t s .....................
99
6.4
N um erical R e s u lts ................................................................................................................
101
6.4.1
103
6.5
7
86
C om parison o f Com putation T i m e s .................................................................
C onclusion
...........................................................................................................................
105
Hardware Verification
106
7.1
K a-B and R ejection F i l t e r ..................................................................................................
106
7.2
3-dB Pow er D iv id e r.............................................................................................................
107
7.3
Butler-M atrix B e a m f o r m e r ...............................................................................................
109
7.4
C ryogenic S w itc h e s .............................................................................................................
115
7.4.1
HTS C - S w i t c h ........................................................................................................
115
7.4.2
HTS H igh Isolation S w itc h ...................................................................................
118
7.5
Six-Pole N arrow -Band F i l t e r ...........................................................................................
118
7.6
C o n c lu s io n s ...........................................................................................................................
122
x
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
8
Summary and Future Work
124
8.1
R eco m m en d atio n s.................................................................................................................
126
8.2
Final R e m a rk s ........................................................................................................................
127
8.3
Future W o r k ...........................................................................................................................
128
A Theory of Space Mapping
129
B Coupling Matrix Algebra
132
B .l
C oupling C apacitor Representations for B and-Stop Filters
...................................
132
B .2
D uality o f R e p re s e n ta tio n s ...............................................................................................
134
B.3
Coupling Values are Pure R eal N u m b e r s ........................................................................
136
C Validity of Rational Function Representation
C .l
D istributed E l e m e n t s ..........................................................................................................
D Circuit Layouts
137
138
140
xi
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
List of Tables
3.1
Transfer Param eter S 21 S p e c ific a tio n s ...........................................................................
30
3.2
Starting Values for Optim isation o f Folded Stub Filter
...........................................
34
5.1
Specifications for Narrow B and HTS F ilter
................................................................
74
5.2
C PU Tim e C om parison for D ifferent Optim isation S tr a te g ie s .................................
84
6.1
C om parison o f C PU Tim es for Conventional and Tim e-dom ain Optim isation
7.1
C om parison o f B ranch-C oupler T e c h n o lo g ie s .............................................................
107
8.1
Recom m ended Optim isation M ethods for D ifferent Filter Types
126
.
..........................
xii
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
103
List of Figures
1.1
Analysis Versus Synthesis
................................................................................................
2
2.1
D esign Procedure for Iterative Optim isation o f M icrowave Circuits
...................
7
2.2 Coupling o f Circuit Sections in a C ascaded and Com plete M icrostrip Circuit
2.3
(Side V i e w ) .....................................................................................................................
9
Folded Two-Pole F i l t e r .................................................................................................
9
2.4 Schem atic Circuit for Two Pole Filter
2.5
...........................................................................
Layouts U sed for Electrom agnetic S im u la tio n .......................................................
10
11
2.6 (a) Filter R esponse C alculated from Circuit Theory, (b) Full-EM Sim ulation, (c)
C ascaded E M Sim ulation, and (d) Full-EM Sim ulation From M odified Circuit .
2.7
Space M a p p i n g ...............................................................................................................
2.8 Objective Function W ith Two Local M axim a
12
15
.............................................................
16
2.9 Planar C apacitor on a D iscrete G eom etry G rid o f Sonnet e m ...........................
17
2.10 C om putation Tim e as a Function o f C ircuit Com plexity........................................
18
2.11 Planar N on-U niform Transm ission Line Filters
3.1
..........................................................
19
Planar Circuit W ith Three Param eters and Its R epresentation as Individual W ith
Three Chrom osom es (Binary C o d e d ).......................................................................
23
3.2 Population o f Four Individuals W ith Three Chrom osom es E a c h .......................
23
3.3 Outline o f Genetic A lg o rith m ......................................................................................
24
3.4 O pen Stripline Stub, Sm ith Charts for 20 GHz and 30 G H z ..............................
27
xiii
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
3.5
L ayout and Param eters o f Four-Pole D ual M ode F ilte r ............................................
28
3.6
Population A fter 20 Generations. Num bers are in mils
28
3.7
L ayout o f Stripline R ejection Filter, G ray=B efore optim isation, Black=A fter
O ptim isation
.........................................
........................................................................................................................
30
3.8
R esponse o f Stripline F ilter Before and A fter GA O p tim is a tio n ...........................
32
3.9
L ayout o f Four-Pole R ectangular Iris-Coupled W ave-Guide Filter Show ing all
Adjustable Param eters
......................................................................................................
3.10 Optim ised R esponse o f W aveguide F ilter L ayout
33
....................................................
33
L ayout o f M icrostrip L ine F i l t e r ...................................................................................
34
.........................................................................
35
3.13
Decay o f E rror Function for Optim isation M e t h o d s ................................................
35
4.1
Neural N etw ork M odel o f M icrowave C i r c u i t ...........................................................
39
4.2
Planar Three-Pole F ilter
.................................................................................................
40
4.3
Neural N etw ork Approxim ation and Exact EM Sim ulation for R andom Input
3.11
3.12 Return Loss o f O ptim ised Stub Filter
V e c t o r ......................................................................................................................................
4.4
41
Neural N etw ork A pproxim ation and E xact E M Sim ulation for an Optim ised Cir­
cuit L a y o u t ...........................................................................................................................
41
4.5
Parallel R esonator and Input I m p e d a n c e .....................................................................
43
4.6
Interpolation o f Four-Pole Filter Response, a) Linear, b) Cubic Spline, c) C auchy
45
4.7
Adaptive Sam pling Procedure
47
4.8
Sam ple Locations for Two-dim ensional and Three D im ensional Param eter Space
48
4.9
Steps in Recursive C auchy M ethod in Two D im e n s io n s .........................................
50
4.10 Inverted Tree Representation for Recursive Cauchy M e th o d ..................................
51
.......................................................................................
4.11 Im pedance o f M icrostrip L ine as a Function o f e r and w / h using M ulti-D im ensional
4.12
R ational Function E x p a n s io n ...........................................................................................
54
M icrostrip A ntenna w ith R ecessed line f e e d ..............................................................
54
4.13 R eturn Loss o f A ntenna Geom etry p \= 24.5 m m , p 2 - 1.4 m m , p:>= 4.2 m m
C om puted by M ulti-D Cauchy M odel and Full E M -S im u la tio n ............................
xiv
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
55
4.14 Effect o f Inset location on A ntenna’s R eturn L o s s ....................................................
56
4.15 L ayout o f Three-pole F ilter Show ing Param eters p i to p 4 ......................................
56
4.16 Interpolated S n Param eter U sing M ulti-D im ensional Cauchy M e t h o d
58
4.17 S ii Param eter W ithout Using M ulti-D im ensional Cauchy M ethod Showing the
Sam pled Points Above and B e l o w .................................................................................
58
4.18 R esponse o f O ptim ised Three-Pole F ilte r .....................................................................
59
4.19 M onte-C arlo A nalysis o f O ptim ised Three-Pole f i l t e r .............................................
60
5.1
Training and A pplication Phase s ......................................................................................
63
5.2 R esponse o f Circuit L ayout p \ and p 2 ...............................................................................
64
5.3 R eal Part o f A 5 n ( f ) for L ayout p i and L ayout p 2 (Left), and Im aginary Part of
A S 'n (f) for L ayout p \ and L ayout P 2 (Right)
............................................................
65
..............................
65
................................................
66
5.4 A djusted S-Param eter S'* u ( / ) for L ayout p 2 U sing [AS 1](p i)
5.5 Training and Application U sing Space M apping
5.6
Three-Pole Filter, a) C ascaded (Decom posed) Layout, b) Com plete Layout
. .
5.7
Cascaded and C om plete Sim ulation R esults o f Three-Pole Filter.C learly Seen
67
in the Com plete Sim ulation is a Transm ission Zero at f=3.96 G H z D ue to Stray
C oupling Betw een R esonator O ne and T hree................................................................
68
5.8
D irect C oupled R e s o n a to r s ..............................................................................................
68
5.9
Inductive C oupled R esonators
.......................................................................................
69
5.10 Training and Application o f Coupling M atrix C o r r e c t i o n .......................................
72
5.11 O ptim isation Including Correction o f C ascaded R e s p o n s e .......................................
73
5.12 L ayout o f Six-pole F i l t e r .................................................................................................
74
5.13 C ascaded Six-pole F i l t e r ..................................................................................................
75
5.14 Definition of Variable Param eters in Single R e s o n a to r..............................................
75
5.15 R esponse o f Initial F ilter L a y o u t ....................................................................................
76
5.16 R esponse o f Optim ised F i l t e r ...........................................................................................
77
5.17 A dditional Couplings in Full-EM Sim ulation
78
............................................................
xv
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
5.18 R esponse o f (G rid-Snapped) Filter Layout p geo3 using Fine M odel and C oarse
M o d e l ......................................................................................................................................
78
5.19 R esponse o f F ilter L ayout p g e o 2 using C orrected C oarse M odel and U ncorrected
Coarse M odel. M arked in Gray is the Forbidden A rea o f the Specifications.
. .
80
5.20 R esponse o f Filter W ith L ayout p g e 0 4 O ptim ised U sing Corrected C oarse m odel
81
5.21 Com parison: R esponse o f Filter by C ascaded C orrected Sim ulation and Full-EM
Sim ulation at N ext G rid-Snap C alled L ayout p geo5
82
6.1
N on-U niform Transm ission L i n e ...................................................................................
6.2
Circuit L ayout U sing Circuit Theory and Realisation in M icrostrip Topology and
88
Their R espective R e s p o n s e ...............................................................................................
89
6.3
Frequency R esponse o f Ideal and Real C i r c u i t ...........................................................
90
6.4
Com putation o f Tim e-dom ain From Frequency-D om ain Electrom agnetic Sim u­
lation
6.5
......................................................................................................................................
91
Excitation Pulses in Tim e-D om ain a) Gaussian b) Differential Gaussian c) G aus­
sian Step d) Sine M odulated G aussian Pulse ...............................................................
91
6 .6
D esired R esponse s f f s (f) and A ctual R esponse s f f ( i ) in Tim e-D om ain . . . .
92
6.7
Exciting Gaussian Pulse and D ifference o f D esired and Actual Response;
<1
Denotes the Tim e E lapsed B etw een Excitation and Getting the Echo o f the Fault
L o c a tio n ..................................................................................................................................
93
6 .8
Optim isation Param eters at Fault L o c a tio n ..................................................................
94
6.9
O ptim ised Circuit L ayout
94
..............................................................................................
6.10 D esired Frequency R esponse and Frequency R esponse of M odified C ircuit . . .
95
6.11 Three D ifferent M odifications o f the L ayout Yielding the D esired Response: a)
Extending the width at Fault; b) Shortening the Line with Im pedance Zo ; c)
A djusting C ircuit | Away From F a u l t ..........................................................................
96
6.12 Tim e Sequence o f Pulse Travelling Through a C i r c u i t .............................................
98
6.13 B and-pass W indow F u n c t i o n ..........................................................................................
99
6.14 Tim e-D om ain R esponse o f Tuned N arrow -Band Filter
.........................................
6.15 Tim e-D om ain R esponse o f N arrow -Band Filter W ith R esonator three D etuned
xvi
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
100
100
6.16 Planar N onuniform Transm ission Line Low-Pass F i l t e r ..........................................
101
6.17 Frequency-D om ain response 5 n F rom Circuit Theory (Desired Response), FullEM Sim ulation (Actual R e s p o n s e ) .................................................................................
102
6.18
Tim e-D om ain R esponse s n (< ) From Circuit Theory and raw F u l l - E M
102
6.19
Variation o f Frequency and Tim e-Dom ain R esponse by Changing ru4 .................
103
6.20 Tim e-D om ain R esponse o f
sn(f)
from Circuit Theory and Corrected Circuit
U sing Full EM and TD R O p tim is a tio n ..........................................................................
104
6.21 Frequency-D om ain R esponse S n From Circuit Theory, Full-EM , Corrected
Full-EM Sim ulation
.........................................................................................................
104
7.1
B lock D iagram o f a Typical R elays S a t e l l it e ..............................................................
107
7.2
Sim ulated (Thin Line) and M easured R esponse (Fat Line) o f Stripline F ilter .At
the Top 5 n [d B ] and at the B ottom
5
i 2 [ d B ] ...............................................................
108
7.3
Com parison o f L-B and Hybrids
...................................................................................
108
7.4
L ayout o f H y b r i d ...............................................................................................................
109
7.5
Photograph o f the Folded H ybrid in C om parison W ith One C ent C o i n .............
110
7.6
Full E M Sim ulated R esponse o f O ptim ised H y b r i d ................................................
110
7.7
M easured R esponse o f O ptim ised H y b r i d ..................................................................
110
7.8
B eam Generation by M ultiple Fixed B eam form ing N etw ork
Ill
7.9
Schem atic B utler M atrix Beam form er, The X ’s Represent 90° Hybrids W ith a
...............................
Pow er Split Ratio o f 3dB /3dB ............................................................................................
Ill
7.10
L ayout o f 8 x 8 M atrix in H o u s in g ...................................................................................
112
7.11
Photograph o f 8 x 8 B utler M atrix B eam form er
113
.......................................................
7.12 R eturn Loss o f 8 x 8 M atrix for B eam Ports 1R, 2R, 3R, 4R Sim ulated (left), and
M easured (right)
7.13 Insertion Loss o f
.................................................................................................................
8 x8
113
M atrix W ith B eam -Port Input 1R, sim ulated (left), and
m easured (r ig h t)....................................................................................................................
114
7.14 Phase Taper o f 8 x 8 M atrix Sim ulated and M e a s u re d ................................................
114
7.15
114
Phase Distribution o f 8 x 8 m atrix (Sim ulated and M easured)....................................
xvii
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
7.16
Radiation Pattern (Sim ulated from M easured D a t a ) ................................................
115
7.17
Four-For-Two Redundancy Sw itch A rrangem ent o f A N IK D (see [141]) . . . .
115
7.18
C -sw itch Operation M o d e s ..............................................................................................
116
7.19
L ayout o f C -Sw itch . Inset Shows One o f the Eight D.C. Decoupled Stub Lines.
116
7.20 Photograph o f Conventional C o-A xial C -Switch, HTS C-Sw itch and One D ollar
C o i n .........................................................................................................................................
117
7.21
Subsection o f C -Sw itch Showing the four Optim isation Param eters
.................
117
7.22
Sim ulated (Left) and M easured (Right) Perform ance o f C - s w itc h ........................
118
7.23 L ayout o f Four PIN D iode Sw itch and B low -U p o f the Diodes Including D.C.
7.24
B ias N etw ork and R F C h o k e s ...........................................................................................
119
Photograph o f H igh-Isolation Sw itch W ith L i d ...........................................................
119
7.25 Sim ulated and M easured R esponse o f The High-Isolation Integrated H TS/PIN
D iode S w i t c h ........................................................................................................................
120
7.26
L ayout o f Six-Pole Filter in H o u s i n g ............................................................................
121
7.27
Photograph o f the six-Pole F i l t e r ...................................................................................
121
7.28
Sim ulated and M easured Frequency R esponse o f Six-Pole F ilter W ithout Tuning
123
A .l
Illustration o f Space M apping
130
B .l
C apacitive C oupled Resonators Z
................................................................................
132
B.2
Lossy Parallel Resonator C i r c u i t ...................................................................................
134
B.3
Duality o f C oupling N etw orks, On the left C oupling by Lum ped E lem ent Induc­
.......................................................................................
tive T-Network, on the Right Coupling by Capacitive 7r-N etw ork .........................
135
C .l
T-N etwork and 7r - N e t w o r k ..............................................................................................
137
C.2
Tangents o f x and Pade A p p ro x im atio n .........................................................................
139
D .l
L ayout o f LC Low -Pass C ir c u it.......................................................................................
141
D.2
L ayout o f sub-section o f six-pole f i l t e r .........................................................................
142
D.3
L ayout o f Six-Pole filter
143
.................................................................................................
xviii
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
Chapter 1
Introduction
"In scien ce one tries to tell p eople, in such a w ay as to be un­
dersto o d by everyone, som ething that no one ever knew before.
But in poetry, i t ’s the exact o p p o site.", Paul Dirac
M icrowave filters are the backbone o f any m odem m icrowave com m unication system. D esign
and m anufacturing o f small, lightw eight, high-perform ing filters is a m ulti-billion dollar industry
nowadays. Consequently, a lot o f effort is put into the precise and fast synthesis o f microwave
filter circuits.
The approach o f filter synthesis has changed over the years. In the early days o f m icrowave engi­
neering in the 1930’s, filter design was perform ed exclusively by formal synthesis. This approach
uses a ladder netw ork o f lum ped im pedances to com pose a filter from a prototype function that
represents the desired response. W here lum ped elem ents are im practical, distributed elem ents
can approxim ate them. The continued refinem ent o f form al synthesis gave rise to w hat is now
com m only called Filter Theory.
The progressive m iniaturisation o f filters that becam e possible w ith the invention o f planar struc­
tures in the 1950’s, has m ade form al synthesis increasingly difficult. Form al synthesis cannot
take fringe fields, parasitic elem ent coupling, etc. into account. These effects can adversely
affect a circuit’s perform ance, especially at very high frequencies beyond the X-band. Further­
m ore, form al synthesis is restricted to regularly shaped layouts. A dvanced filters, however, m ay
dem and irregularly shaped circuit elements.
An alternative m ethod to form al synthesis was developed in the 1960’s. The m ethod originates
from the fact that synthesis m ay not be possible for a certain circuit layout, but a circuit analysis
is feasible. Figure 1.1 shows the definitions o f analysis and synthesis. Synthesis derives a circuit
1
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
2
CHAPTER 1. IN TRO D U C TIO N
structure from a given response function. Analysis, on the other hand, calculates the response o f
a given filter structure.
^ ^ y n th e s i^ ^
Response
,
Analysis
Circuit Structure
Figure 1.1: Analysis Versus Synthesis
If analysis is possible, but form al synthesis is not, w e can still synthesise the filter by an iter­
ative optim isation process, wherein a circuit layout is m odified and analysed repeatedly until
the specifications are satisfied. T e m e s and C a l a h a n in their classical paper [11] in 1967
were am ong the first to advocate the use o f such an iterative optim isation process to design m i­
crowave circuits. The draw back o f this m ethod is that, in general, this iterative process is rather
tim e-consum ing. W ith only very sim ple com puters available in the 1960’s, the analysis— and
therefore the optim isation— were restricted to very small circuits.
The situation changed rapidly with the advent o f powerful digital com puters in the early 1970’s.
N um erical m icrowave circuit analysis reached undream t-of capabilities. U sing finite-elem ent
m ethods, it becam e possible to derive an accurate response o f alm ost any irregularly shaped
microwave structure. W ith these sophisticated analysis techniques on hand, the synthesis of
m icrowave circuits using iterative optim isation gained m uch popularity in the m icrowave com ­
munity.
Today, num erical m ethods are the m ain tools for analysis o f m icrowave circuits. M any com ­
m ercial program s on the m arket tackle any kind o f analysis problem that involves microwave
structures. Optim isation tools are either part o f the package, or they can be purchased separately.
1.1 Motivation for This Thesis and Proposed Solutions
The design by num erical iterative optim isation m ight be the m ost adopted tool in microwave
circuit design, but it is not the panacea o f m icrowave engineering. An application o f the iterative
optim isation synthesis technique is accom panied by the following problems:
1. Present optim isation schemes do not guarantee to find the optimal circuit layout subject
to the circuit specifications. The search m ay get trapped in a sub-optim al, unsatisfactory
solution. The optim isation process is then said to be non-robust.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CH APTER 1. INTRO D U CTIO N
3
2. The optim isation process requires repeated analysis o f the circuit for evaluating the cir­
cuit’s perform ance. This repetitive analysis step can com putationally be extrem ely expen­
sive. W ith today’s techniques, expenses can often be reduced only by sacrificing accuracy.
3. The m ore com plex the circuit, the m ore apparent the problem s described in 1. and 2. b e­
com e. M edium -sized synthesis problem s already put today’s m icrowave design software
to the limit. Even larger circuits can only be analysed and optim ised by decom position,
i.e. by splitting them into several sub-circuits. This approach has the draw back o f losing
accuracy in the simulations.
4. C om m ercial circuit sim ulation tools do not allow an arbitrary scaling o f the circuit’s ge­
ometry. Rather, the dim ensions can only be changed in finite steps in order to fall on
the boundaries o f a fixed mesh. This lim itation is necessary to achieve acceptable per­
form ance o f the simulator. The optimal circuit geom etry m ay not be representable by a
layout snapped to this discrete grid. H ence, the true optim um cannot be found.
This thesis proposes a num ber o f novel techniques to avoid the problem s outlined above. R o­
bustness is im proved by switching to a novel class o f optim isation algorithm s, the so-called
genetic algorithm s. We are am ong the first to apply these algorithm s to filter synthesis prob­
lems. To yield better perform ance, w e m odify the standard genetic algorithm when applying
it to filter synthesis problem s. Especially, we allow the user to include knowledge about the
circuit’s physical constraints. As shown in this work, such constraints cannot be incorporated
into conventional optim isation schemes. The m easures taken result in novel robust optim isa­
tion algorithm s for m ore reliable m icrowave circuit designs. O ur algorithm s are m eant to be a
supplem ent to existing optim isers rather than a substitution.
T he problem o f large com putational expenses and too m any optim isation variables is solved
by several actions: A m ulti-dim ensional Cauchy m ethod, hybrid optim isation and tim e-dom ain
optim isation. These techniques m ay be used individually or combined.
O ur m ulti-dim ensional Cauchy m ethod is a novel extension o f the one-dim ensional Cauchy
m ethod for frequency response interpolation. The m ulti-dim ensional m ethod interpolates with
respect to both frequency and geom etrical param eters. The technique can precisely derive the
response o f any arbitrary filter geom etry from a small set o f pre-calculated response samples.
W hen applied to the optim isation o f filter circuits, the m ulti-dim ensional Cauchy m ethod makes
the evaluation o f the circuit perform ance (the analysis step in the optim isation) very inexpensive.
The new interpolation scheme also allows us to avoid the problem o f finite m esh sizes in the
circuit sim ulation software, as response sam ples can be com puted from any circuit layout. A p­
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 1. INTRO D U CTIO N
4
plying the new interpolation scheme, we can use a very rough m esh for fast sim ulation and still
achieve the desired accuracy.
For large optim isation problem s, the database concept becom es impractical, because it requires
a vast num ber o f circuit samples. The circuit m ust be calculated from sm aller decom posed sub­
sections and the overall response found by network-theory. Unfortunately, this decom position
introduces errors. Parasitic couplings, w hich always exist in the real circuits, are cut off. As a
result, the response calculated from the decom posed circuit is different from the response cal­
culated from the com plete circuit. In the past, this error introduced by the decom posed analysis
had to be accepted for optimisation, as there was no practical way to perform an optim isation on
the com plete circuit.
We develop here a novel technique which provides the speed o f the decom posed analysis with
the accuracy o f the com plete circuit analysis. The m ethod, called hybrid optim isation, perform s
one circuit sim ulation using both the decom posed circuit and the com plete circuit. From their
respective responses the coupling param eters are derived. The differences betw een coupling
param eters describe the parasitic couplings betw een the filter resonators, w hich are present in
the com plete circuit but m issing in the decom posed circuit. It is shown that we can adjust the
responses derived from subsequent decom posed circuit analyses to include the parasitic coupling
if we use the difference coupling matrix. The result is an accurate m odel with the com putational
cost o f the decom posed analysis.
O ur hybrid optim isation is specialised to resonant band-pass and band-stop filters. The technique
cannot optim ise other filters, such as low-pass filters and transversal, i.e. non-resonant, filters.
This problem can be tackled with another innovation we introduce here, nam ely optimisation
using time-dom ain reflectometry. The tim e-dom ain response reveals different characteristics of
the circuit than the frequency-dom ain response. This well-known fact is exploited by timedom ain reflectom etry (TD R) m easurem ents. Here, we adapt the TDR concept for a tim e-dom ain
based synthesis o f m icrowave circuits. W ith our new technique, we can pinpoint one fault after
the other and fix (i.e. optim ise) them one at a time. D ue to non-uniqueness, w e often even have
several choices for optim ising the circuit. Applying the new technique to filter optimisation,
we can split an optim isation problem with N param eters into N problem s with one parameter,
which m akes the overall optim isation m uch easier to handle.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 1. IN TRO D U C TIO N
1.2
5
Thesis Organisation
C hapter 2 gives a review o f the currently available filter design techniques. Standard m ethods
as w ell as enhanced algorithm s are briefly explained and many references given. The chapter
pinpoints the problem s inherent in the standard techniques as well as enhanced m ethods, and
outlines our ideas for overcom ing these problems.
C hapter 3 proposes the im plem entation o f genetic algorithm s in the filter design process. The
new approach im proves the robustness o f the optim isation process significantly. We outline
how we m odify the generic genetic algorithm to boost perform ance. Genetic algorithm s are
also com pared to traditional optim isers with respect to robustness and com putational costs. The
concepts are dem onstrated by several examples.
Chapter 4 introduces a novel, accurate concept o f response interpolation that uses a m ulti­
dim ensional Cauchy method. The novel interpolation schem e perm its the fast calculation o f
any m odified circuit response from response sam ples calculated a-priori. As dem onstrated by
various exam ples, the concept gives the designer a fast and hassle-free optim isation, param eter
plots, and M onte-C arlo analyses.
C hapter 5 develops the idea o f hybrid optimisation. By hybrid optim isation we understand the
optim isation o f filter structures utilising two different circuit m odels, nam ely a coarse, fast m odel
and a fine, accurate m odel. R epresenting both m odels by their respective coupling m atrices, we
succeed in correcting the coarse m odel so that the accurate response for any geom etry can be
derived w ithout further fine m odel simulations. The novel m ethod achieves the com putational
speed of the coarse m odel w ith the accuracy o f the fine model.
O ur next new contribution, the filter optim isation by tim e-dom ain reflectometry, is introduced
in C hapter
6
. We recom m end tim e-dom ain optim isation particularly for the synthesis o f non-
uniform transm ission line filters, e.g. transversal filters. The chapter gives a detailed explanation
o f the theory underlying tim e-dom ain reflectom etry and applications.
C hapter 7 lists and describes all hardware built throughout the course o f this project. We verify
the perform ance o f m icrow ave circuits synthesised by our new techniques on several pieces o f
hardware. Am ong others, w e test a superconductive B utler m atrix beam form er, several cryo­
genic active switch circuits, and a superconductive planar six-pole narrow-band filter.
F our appendices provide m athem atical details and derivations. Furtherm ore, the appendices
contain technical draw ings o f the m anufactured hardware. A list of references can be found at
the end.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
Chapter 2
Available Filter Design Techniques
"I sa w that it was great, greater, an d greatest, w ith prodigious
p o ssib ilities in its power. I w as determ in ed to m aster the book...
It took m e several years before I cou ld understand as much as I
p o ssib le could. Then I set M axw ell aside an d fo llo w e d m y own
course. A nd I progressed much m ore quickly.", Oliver H eaviside
on M axw ell’s book "Treatise on Electricity and Magnetism"
O ver the years, many techniques have been reported in the literature for designing m icrowave
filters. In this chapter, w e briefly review currently used filter design techniques and the draw ­
backs associated with them. A dvanced techniques that w ere developed to avoid these drawbacks
are the topic o f this chapter’s second section. These m ethod, even though very powerful, still
have some limitations. R esolving these lim itations is the topic o f this thesis. A t the end o f this
chapter, we present a num ber o f new ly developed techniques that enable us to design arbitrarily
shaped large filter circuits in an accurate, reliable, and inexpensive way.
2.1
Standard Filter Design Techniques
The starting point o f microwave filter design is the form al synthesis o f lum ped and distributed
elem ent netw orks using circuit theory. W e can derive filter circuits for alm ost any desired filter
response using circuit theory as described in [27, 29, 12]. Various aspects o f form al filter design
techniques are the topic o f a vast num ber o f publications, e.g. [14, 23, 25, 28], The standard
w ork [12] o f
M
a t t h a e i,
Y
oung
and
Jo n e s
presents an exhaustive discussion o f form al filter
design with m any supporting design tables and charts.
6
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 2. A V A IL A B L E FILTER D E SIG N TECHNIQUES
7
O ver the years, m any specialised design techniques for specific filters have been developed.
Filters with finite transm ission zeros for better perform ance are developed in [1 6 ,1 7 ,1 8 ], am ong
others. C oupled resonator representations o f bandpass filters are discussed in [26, 30, 31, 32],
The so called transversal filters, first proposed by K allm ann in [109], are presented in [19, 20]
and others. On the other hand, m any papers have been published with a focus on either circuit
topology (e.g. W aveguide filter design [21], planar filters [22], superconductive filters [140]) or
filter purpose, e.g. space applications[15].
However, it has been w ell-recognised that pure form al filter synthesis techniques have their lim i­
tations. T he m ethod can only synthesise filters using the ideal, basic, regularly shaped elem ents,
nam ely inductors, capacitors, and transm ission lines. Filter layouts including arbitrarily shaped
elem ents are not feasible. Furtherm ore, not every desired filter response can be achieved using
the form al synthesis.
In order to liberate the design process from these restrictions, an iterative circuit design by opti­
m isation was suggested in [11, 39, 40]. The idea originates from the fact that any circuit perfor­
m ance can be analysed, but not necessarily synthesised. The solution is an iterative optim isation
process. W e m odify and analyse the circuit layout repeatedly until the circuit’s response satisfies
the specifications. The flow o f the process is outlined in the chart o f Figure 2.1. O ptim isation
techniques are discussed in-depth in the literature. There are basically three different ways o f
searching the param eter space: G radient search m ethods, direct search m ethods, and guided
random m ethods. For m ore inform ation, the reader m ay refer to [61, 81, 62, 65,
66
].
^ Start ^
Initial D esig n
Circuit T op ology
Parameter
M odifications
Circuit
A nalysis
Requirements
. Satisfied ? ;
I yes
Q rn T )
Figure 2.1: D esign Procedure for Iterative Optim isation o f M icrow ave Circuits
The filter design using iterative optim isation is only lim ited by the ability to analyse the circuit.
Initially, circuit analysis was done exclusively on netw orks o f lum ped and distributed elem ents
w hose analytical responses are available. Later, arbitrary elem ents w ith responses derived from
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 2. A V A IL A B L E FILTER D E SIG N TECHNIQUES
8
m easurem ents or heuristic form ulas w ere included. The actual circuit analysis is carried out
using netw ork theory. The technique is exhaustively studied in [39] and [44]. W ith the aid o f
today’s com puters, we can also include full electrom agnetic (so-called fu ll E M ) sim ulated circuit
elem ents, see [42]. F or m ore inform ation on full EM sim ulation m ethods, refer to [36, 34, 35,
33].
The ultim ate application o f netw ork theory based analysis is decom posed fu ll E M simulation
(also called cascaded fu ll EM ). The circuit is split into several sub-sections which are com puted
separately using full E M sim ulation m ethods, and the overall response is com puted from circuit
theory. D ecom position is used for exam ple in [42],
Versatility and speed o f the com bined netw ork analysis/optim isation technique have led to their
widespread application in com puter aided design (CAD) software o f m icrowave circuits. The
m ost prom inent software packages using this approach are Touchstone [145], O SA 90 [147], and
H P-M DS [146]. The m ethod has its drawbacks: N o interaction betw een non-adjacent circuit ele­
m ents can be considered. In contrast to the real circuit, non-adjacent coupling, o f the sub-circuits
is not considered. These effects include stray quasi-static field coupling as well as coupling by
box-m odes and surface-m odes inside the housing. Figure 2.2 illustrates the possible couplings
o f a decom posed m icrostrip circuit versus the com plete circuit in a housing.
It becom es clear, that highest accuracy is achieved with an optim isation using com plete fu ll EM
simulation. C om plete sim ulation m eans that w e com pute the response from a single full EM
sim ulation o f the circuit as one block. The com plete full E M simulation m odels the actual cir­
cuit m ost closely. Consequently, an iterative optim isation using the com plete full E M sim ulation
results in the m ost accurate design. The com plete EM (also called direct) optim isation is ap­
plied for exam ple in [43, 49, 41, 46]. The m ethod’s handicap lies in its extrem e com putational
dem ands. C om plete full EM sim ulations take hours or even days, even for relatively sm all-size
circuits.
2.1.1
Application of Standard Methods
The differences betw een the analysis m ethods, nam ely heuristic m odels+network-theory, de­
composed EM+network theory, and complete EM, are illustrated by considering a two-pole
filter. A typical two-pole line filter is shown in Figure 2.3.
W e can sim ulate this filter with circuit com posed o f com ponents based on heuristic m odels
from e.g. [39], The schematic diagram o f the circuit from elem entary com ponents is shown in
Figure 2.4.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 2. A V A IL A B L E FILTER D E SIG N TECHNIQUES
9
C ascad ed C ircuit
’V/V/V/V///V/V//V/V<V/\V/V/VV/V/V/V
;\v
S u b -C ircu it 1
S u b -C ircu it 2
A ir
M etallisation
'i m
tk im
M
m
%
R eal C ircuit
Wave-Guide Modes
S pace Waves
o
Ouasi-Static E/H Fields
Surface Waves
Figure 2.2: Coupling o f C ircuit Sections in a C ascaded and Com plete M icrostrip C ircuit (Side
View)
Figure 2.3: Folded Two-Pole Filter
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 2. A V A IL A B L E FILTER D E SIG N TECHNIQUES
10
OV22
H5SUB5TRATE
200
SUBST-LaAl
ER-24
MUR-1
C0ND-1.0E+306
ROUGH-O un
TANO-O.O
nil
H-20 n il
EQUATION LJ101-(10<40. K45)
H-7 n il
L-L res n i I
SUBST-LaAl
EQUATION L_H 12 -( 10<45. 0<45)
EQUATIONL res-(120<130.2KlB0)
SUBST-LaAl
H-7 ni
QF9
HSCRNR
W P10
H5TL
SUBST-LaAL
Ka"i!,
avw
CHP4
H-7 n iI
SUBST-LaAl
H
1
:S-7 n il
T-LaAl
1
HSOC
CHP43
HSOC
H -7 n i 1
O F13
SUBST-LaAl
L-L NI2 n il
>{
SUBST-LaAl
L-L.M01 n i l
CHP17
N SOM
H
5-7 n il
an
: H-7 n i 1
J02 iTH
H-7 n i l
L-18 n i l
SUBST-LaAl
HSOC
KSCTL
SUBST-LaAl
L - O n il
H-7 n i 1
HSOC
H-7 n
L-50 n i l
L-0 n il
H-7 n i l
SUBST-LaAl
f a SrflR
PQRTNUM-1
R-50.0 OH
JX-0.0 OH
AGROUND
i
s
m
rs.
H-7 mi 1
L-L res n i l
SUBST-LaAl
3 i
CHP23
«CS«
CHP33
CHP32
cm?
SUBST-LaAL
H-7 m il
L-18 n i l
OF47
H-7 n i l
HSCTL
S-7 n il
Q P36
CfWB
H-7 n il
S-7 n il
H-7 n il
ono
H50C
L-0 n il
H-7 mil
SUBST-LaAl
OP24
H51L
H-7 n iI
L-100 n il
SUBST-LaAl
.SMR
PORTNUM-2
R-50.0 OH
JX-0.0 OH
AGROUND
F igure 2.4: Schem atic Circuit for Two Pole Filter
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 2. A V A IL A B L E FILTER D E SIG N TECHNIQUES
11
The schematic circuit is optim ised for 4 G H z centre frequency, -15 dB return loss, and a 0.5%
bandw idth. The response o f the optim ised filter using circuit theory is shown in Figure 2.6 (a).
The response is verified by a full EM analysis. Three layouts are used for the sim ulation as
shown in Figure 2.5: a com plete layout (Layout 1), a cascaded circuit (Layout 2), and a layout
with a flipped second resonator (Layout 3).
Layout 1
Layout 2
Layout 3
Figure 2.5: Layouts U sed for Electrom agnetic Sim ulation
The response param eters S u and
£21
for the three sim ulations are shown in Figure 2.6 (b) to (d).
T he circuit theory sim ulation is also shown in Figure 2.6 (a). W e clearly see that the response of
the full E M sim ulation o f L ayout 1 does not m atch the circuit theory sim ulation. A com parison
betw een the sim ulations o f the cascaded circuit o f L ayout 2 and the m odified circuit o f Layout
3 shows a m uch closer resem blance o f the response with the theoretical results. The response is
shifted only slightly in frequency.
All layouts are identical from the circuit theory point o f view. So why are the responses so
different? The reason for this lies in the differences in the separation o f the resonators. M utual
coupling betw een the resonators o f layout 1 is m uch stronger than in layout 2 or 3, because the
resonators are facing each other. In layout 3, the resonators are spatially separated, whereas
in layout 2, they are com pletely decoupled from parasitic coupling. R F energy flows m ainly
through the ports from one resonator to the next. This condition is violated in layout 1. Energy
flows also via electrom agnetic coupling from resonator to resonator. Consequently, the response
from layout
1
differs dram atically from the theoretical response.
Obviously, the optim al layout o f an iteratively optim ised circuit is only as good as the underlying
analysis m ethod. A lthough the filter is optim ised for 4 GHz, the use o f the heuristic m odels have
led to an actual design w ith a centre frequency o f 4.02 GHz due to the m odel errors. Further,
bandw idth and return loss o f the actual circuit are slightly different from the response derived
using circuit theory. This clearly dem onstrates, that we cannot blindly trust the circuit theory
m odelling. Accurate synthesis is not always feasible. In order to gain accuracy, w e have to use
full EM models.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CH APTER 2. A V A IL A B L E FILTER D E SIG N TECHNIQUES
3.9
3.9
12
GHz
freq
freq
GHz
(b) L ayout 1
(a) C ircuit-Theory
Trace2«dB(FULLEM_DESIGN_AFS.,S(1,1])
Tr ace4-dB(FULLEM_DESIGN_AFS. . SC1,2])
Tracel-dB (S C l,m
Trace6-dB(Stt,2])
s
<
/
\
/
/'
\
/
\
1=1
1
(c) L ayout 2
Trace3*dBCDATAFROMHALFS. .SCI,1 ])
Trace5*dB(DATAFR0MHALFS..Stl,2 ])
3.9
3.9
s
\
V
freq
freq
S
GHz
GHz
N
.
[_ _ ]
1
A
freq
freq
4.1
4.1
GHz A
GHz B
(d) L ayout 3
Tr ace7-dB(FULLEMJJESIGNjb i r ror _AFS. . S[1,11)
Trace8*dB(FULLEM_DESIGN_mirror_AFS..Stl,2 ])
Figure 2.6: (a) Filter R esponse C alculated from Circuit Theory, (b) Full-EM Sim ulation, (c)
C ascaded EM Sim ulation, and (d) Full-EM Sim ulation From M odified Circuit
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 2. A V A IL A B L E FILTER D E SIG N TECHNIQUES
13
To sum m arise, w e can classify the available techniques into four categories:
1. Form al synthesis using circuit theory
2. Iterative optim isation using heuristic m odel, circuit theory analysis, and netw ork theory
3. Iterative optim isation using circuit m odels based on a full electrom agnetic sim ulation or
m easurem ents
4. Iterative optim isation using com plete full electrom agnetic sim ulation (direct optimisation)
The boundaries betw een the m ethods are not rigid and the different m ethods can be com bined.
A ll o f these m ethods are in widespread use by m icrowave engineers. W hich technique to choose
is often determ ined by a trade-off betw een required accuracy and the perm itted com putational
costs. Typically, the engineer derives an initial design using m ethod 1, and then switches to
m ethods 2 and 3 for a refinem ent o f the optim al design. In the last step, the design is verified by
a com plete E M sim ulation. However, even w ith today’s fast com puters, a direct optim isation of
large circuits (i.e. m ethod 4) is not feasible because o f excessive com putational costs.
W ithout question, the goal is an optim isation w ith the accuracy of the full E M sim ulation and
speed o f the circuit theory approach. In the follow ing section, we sum m arise w hat has been
reported in the literature to com e closer to this goal.
2.2
Enhanced Filter Optimisation Algorithms
T he basic technique o f iterative optim isation can be im proved by two strategies. R ecalling the
flow chart o f Figure 2.1, we can either im prove the perform ance o f the analysis step, or upgrade
the param eter m odification algorithm. The general im provem ent of the analysis step is a big— if
not the biggest— research field in m icrowave theory. These im proved techniques are beyond the
scope of this thesis. It is sufficient to say that they indirectly accelerate the optim isation process.
In the follow ing sections, we concentrate on the review o f m ethods whose im m ediate goal is to
im prove the optim isation process.
Response Database and Interpolation: D uring the optim isation process, identical or alm ost
identical circuit layouts are sim ulated repeatedly. Consequently, it is useful to store pre­
vious sim ulation results in a response database, as proposed by B a n d l e r in [43]. W ith
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
14
CHAPTER 2. A V A IL A B L E FILTER D E SIG N TECHNIQUES
a sufficiently large database, we can also interpolate the response o f unknown circuit lay­
outs from the know n saved results. This requires a m ulti-dim ensional interpolation. In the
literature linear and quadratic interpolation has been proposed, e.g. in [43].
Cauchy Method and Adaptive Sampling:
M i l l e r et. al. suggested in [76, 77, 82] to apply
the Cauchy m ethod to the frequency response. The Cauchy m ethod is a very effective
interpolation schem e based on rational functions. U sing this m ethod, a handful o f fre­
quency sam ples can often predict the exact frequency response over a wide frequency
range. D h a e n e et. al. showed in [82] that adaptive sam pling can reduce the required
num ber o f frequency sam ples even further. In M iller’s work, however, the m ethod is re­
stricted to frequency response interpolation. W e give a discussion o f the Cauchy m ethod
in C hapter 4. There, we also extend the one-dim ensional Cauchy m ethod to a m ulti­
dim ensional Cauchy method.
Neural Networks: A very pow erful interpolation schem e is a technique known as neural net­
works as described in [120, 122, 119]. G u p t a in [126, 127] and Z h a n g in [130, 131]
were am ong the first to apply neural networks to microwave analysis. Neural networks
com pute a response from a known set o f response samples. Z H A N G and G UPTA im proved
the standard netw ork perform ance by introducing so-called knowledge-based neural net­
works (K B -N N ) [132, 133, 134, 128, 127]. K B -N N use heuristic m icrowave m odels as
activation functions (neurons) rather than the standard function. IN [136] W u et. al. pro­
posed a technique for input-output reversal o f neural networks. In doing so, the network
can be used not only for analysis but also for synthesis. M ore inform ation on neural nets
is given in C hapter 4.
Space Mapping and Hybrid Optimisation:
B a n d l e r at al.
developed a technique called
space m apping in [72]. The m ethod is a hybrid optim isation scheme. The idea behind
hybrid optim isation is to use a coarse circuit m odel, e.g. decom posed m odel, within the
optimisation loop, and a fine m odel, e.g. a com plete EM , for verification and adjustm ents
of the coarse model. Space m apping establishes a relationship betw een the param eter
spaces o f a coarse m odel and o f a line m odel by m apping the two param eter spaces as
illustrated in Figure 2.7. We briefly outline the theory o f the m apping algorithm in A p ­
pendix A. The algorithm has been updated w ith several im provem ents [73].
Circuit Extraction: YE and
M a n s o u r in [45] developed a technique for including the stray
coupling o f real circuits into the decom posed analysis. The sub-circuits are connected
by additional tw o-port netw orks w hich em ulate the m issing stray coupling. The netw ork
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 2. A V A IL A B L E FILTER D E SIG N TECHNIQUES
Parameter
Space of
Coarse
Model
15
Parameter
Space of
fine
Model
Figure 2.7: Space M apping
elem ents are derived using circuit param eter extraction.
Adjoint Network Method: M ost optim isation techniques are based on the knowledge o f the
gradient o f the objective function, i.e. its derivatives with respect to the geom etrical pa­
rameters. W hen using full E M sim ulators, the derivatives m ust be calculated numerically,
i.e. by finite-difference m ethods. A
llesa n d r i
et. al. showed in [48] that the gradient
can be com puted by an adjoint netw ork m ethod that is well known in circuit theory ana­
lysis. However, the m ethod is limited. The optim isation is only applicable to simulation
m ethods and topologies w here the analytical gradient o f the Y-matrix o f the sub-circuit
can be found.
Order Recursive Gaussian Elimination: W hen designing a decom posed planar circuit by the
m ethod o f m om ents (see [33]), w e often sim ulate very sim ilar subsections. It can be
shown that there is a considerable overlap o f data o f the various sub-circuit simulations.
In [5 0 ] , M i s r a et al. proposed the order recursive Gaussian elim ination algorithm to
exploit the data-overlap. The m ethod is very limited, because w e need access to the the
matrix form ulation in the sim ulation software and we m ust reform ulate the algorithm for
each specific layout.
2.3
Limitations of Available Methods and Our Proposed Solutions
Section 2.2 has shown that m any efforts have been m ade to im prove the standard optim isation
algorithm s. The suggested techniques are a big step towards an accurate and fast optimisation.
Nevertheless, m ost o f these techniques are only applicable to specific circuit layouts or suffer
from other severe restrictions. The follow ing sections discuss in m ore detail the known prob­
lem s with robustness, com putational expenses, discrete geom etry grids, and large num ber of
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 2. A V A IL A B L E FILTER D E SIG N TECHNIQUES
16
optim isation variables. W ithin each section, we outline our proposed solution to the problem
discussed.
2.3.1
Robustness and Cost of Gradient Calculation
The traditional gradient optim isers [63] are often unable to find the best solution, because their
search gets trapped in a sub-optim al solution (local m inim um ). For a typical objective function
as shown in Figure 2.8, the gradient-guided optim iser always follows the direction o f steepest
ascent. Thus, the param eter vector p is always guided to the closest peak in the objective func­
tion. Once the optim iser com es close enough to the local peak, the gradient-based algorithm
cannot jum p to another peak. The algorithm is called non robust, because it fails to find the
global maxim um.
G lobal Maximum
Local Maximum
Figure 2.8: Objective Function W ith Two Local M axim a
N one of the available techniques addresses the problem s with robustness. It is left up to the user
o f the optim isation software to w atch the optim isation process carefully and to intervene in case
the optim iser gets stuck.
The problem can be fixed by using a robust optim isation algorithm which searches the param eter
space globally. Genetic algorithm s (GA ) are o f this type. As shown in [95], their search is not
stalled by the presence o f local m axim a. To our knowledge, genetic algorithm s have never
been applied to the synthesis o f microwave filter. In this thesis, we propose to em ploy genetic
algorithm s in a robust m icrowave filter synthesis. C hapter 3 dem onstrates how we use genetic
algorithm s in connection with an EM -sim ulator for optimisation. As shown there, w e add some
extensions to the standard GA to boost its optim isation perform ance and to extract intriguing
circuit insights.
A nother advantage o f GA’s is that they do not require any derivative inform ation. W hen using
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 2. A V A IL A B L E FILTER D E SIG N TECHNIQUES
17
direct full E M optim isation, the com putation o f the derivatives is very expensive. These costs
are avoided by switching to genetic algorithm s.
2.3.2
Discrete Meshes and Interpolation Errors
F or better perform ance, com m ercially available circuit sim ulators like Sonnet em [150] use a
discrete m esh for the circuit layout. All vertices o f the circuit layout m ust fall on equally spaced
grid-points. Figure 2.9 shows a typical grid used by Sonnet em. Due to the restrictions to the
grid, the capacitor gap can take only discrete values. Thus the capacitance cannot be freely
chosen. The circuit response o f layouts w ith geom etries “off-the-grid” m ust be interpolated.
HI—II
lh
Figure 2.9: Planar C apacitor on a D iscrete G eom etry Grid o f Sonnet em
L inear interpolation, cubic spline interpolation, and neural networks have been recom m ended
for such geom etries. These schem es w ork very reliably as long as the interpolated functions
are only weakly non-linear. N onetheless, the available m ethods have difficulties handling highly
non-linear functions, in particular, functions with finite poles. The Cauchy m ethod does not
suffer this problem , but has the lim itation that it is applicable to the frequency interpolation,
only.
To overcom e these problem s, we propose the m ulti-dim ensional Cauchy method. This novel
m ethod is capable o f handling several variables, as well as reliably interpolating highly-nonlinear
functions. The concept and exam ples are discussed in detail in Chapter 4.
Besides its interpolation accuracy, the Cauchy m ethod can significantly reduce the C PU tim e
required during a circuit optimisation. Since the m ethod derives all responses from a-priori cal­
culated response sam ples, the com putational costs for the circuit analysis in the optim isation
loop are m inim al. The com putation o f the sam ples is a solitary event, and could be carried out
over-night. Once the response sam ples are available, circuits can be analysed alm ost instanta­
neously. N ote however, that in practice the m ulti-dim ensional Cauchy m ethod is in practice for
m edium size problem s only. F or circuits w ith m any optim isation variables, the advantages o f
quick analysis are counterbalanced by the cost o f generating the response database. In order to
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 2. A V A IL A B L E FILTER D E SIG N TECHNIQUES
18
address large scale circuits, w e have to decom pose the circuit first. The sub-sections can then be
m odelled with the m ulti-dim ensional Cauchy method.
2.3.3
Computational Expenses
The m assive need for com puter resources is the m ost serious disadvantage o f optim isation by
com plete full E M sim ulations. The com putational expenses can be dramatic. F or example,
a single full electrom agnetic sim ulation o f an
8
x 8 beam form er outlined in Section 7.3 takes
already 10 CPU -hours on a state-of-the-art H ew lett-Packard K-class machine. For problem s o f
that order, we cannot do m ore than one or tw o sim ulations a day. To optim ise this circuit we
need many o f such sim ulations. H ence, direct optim isation is im possible in a reasonable tim e
frame.
The main problem is that with increasing com plexity o f the circuit, the required C PU tim e in­
creases exponentially. Figure 2.10 shows the typical relationship between circuit com plexity and
com putation time. A t a certain complexity, the lim its o f the underlying software and hardw are
are reached. The num ber o f calculations becom es im m ense, and com putation tim e increases
alm ost boundless. Sonnet Software [150] refers to this lim it as The Wall. The Wall is represented
by the dotted line in Figure 2.10.
Computation
Time
Complexity
Figure 2.10: C om putation Time as a Function o f Circuit Complexity.
Consequently, it is very im portant to lim it the circuit com plexity to stay below the wall. To
reduce the complexity, the circuit can be decom posed as discussed earlier. The decrease in
com putational expenses from using decom position is paid for by a decrease in accuracy. For
many applications, the errors thus introduced are unacceptably high. The exam ple o f the twopole filter responses given in Figure 2.6 dem onstrates this clearly.
Space m apping, as suggested to overcom e this problem , reduces the costs o f the optim isation
with a hybrid optim isation algorithm. As outlined in C hapter 5, however, space m apping fails
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 2. A V A IL A B L E LILTER D E SIG N TECHNIQUES
19
on circuits where the coarse m odel, i.e. the decom posed analysis, cannot describe the stray
couplings.
In this thesis, w e introduce a new hybrid optim isation concept. The m ethod builds on the direct
coupled resonator representation. W e can describe coupled resonator filters com pletely by their
coupling matrix. The key o f our new technique is to use an additional coupling m atrix in the
decom posed analysis to represent all stray couplings. This concept and exam ples are presented
in C hapter 5. The m ethod, however, is lim ited to resonant band-pass filter structures.
F or non-resonant and low-pass/high-pass filter structures, w e offer an alternative optim isation
technique. Such filters are often realised as non-uniform transm ission line filters such as shown
in Figure 2.11. A direct optim isation o f large non-uniform transm ission line filters is not feasible,
no m atter w hich o f the available techniques is used. U p to now, these filters are designed exclu­
sively by form al synthesis. D irect optim isation o f the frequency response cannot be perform ed,
due to the vast num ber o f circuit param eters.
Figure 2.11: Planar N on-U niform Transm ission Line Filters
We show in C hapter
6
that direct E M optim isation is feasible when w e utilise the tim e-dom ain
response. The tim e-dom ain response o f a microwave circuit reveals other properties than the
frequency-response does. These characteristics can be exploited for the circuit optimisation.
C hapter
6
introduces a new technique for non-uniform transm ission line filters. This new tech­
nique is based on on tim e-dom ain reflectometry.
2.4
Conclusion
This C hapter has reviewed the m ost com m only used filter synthesis techniques. The techniques
can be classified into tw o m ajor groups: form al synthesis and iterative optimisation. Form al
synthesis derives the filter layout analytically from filter theory. Iterative optim isation is usually
perform ed with tw o steps in a loop: the analyses o f the circuit by num erical sim ulation and the
m odification o f the circuit param eters.
The num erical sim ulation is done by one o f circuit theory, cascaded electrom agnetic analysis,
or com plete electrom agnetic analysis. The pros and cons o f the available sim ulation m ethods
were discussed. To sum m arise, the m odels are either accurate or com putationally fast, but never
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 2. A V A IL A B L E FILTER D E SIG N TECHNIQUES
20
both. The decom posed full E M analysis o f circuits is a technique often applied to com prom ise
betw een accuracy and speed.
The second step o f the optim isation loop— the m odification o f the param eters— is done with
the goal to m axim ise the circuit’s perform ance. W e can update the param eters with various
available algorithm s. Nowadays, a gradient-based (steepest descent) search technique is applied
m ost often. The draw back o f this m ethod is, however, that it is not generally robust, i.e. it
m ight stop searching the param eter space after getting trapped in a sub-optim al solution. In the
next chapter, w e propose the application o f genetic algorithm s to filter optim isation. Genetic
algorithm s search the param eter space by sim ulating evolution.
Besides the choice o f the right optim isation algorithm , the key to an efficient m icrowave circuit
optim isation is the developm ent o f accurate and fast circuit models. A prom ising approach is to
m ove the expensive evaluations o f the circuit m odel out o f the optim isation loop. A ll required
circuit responses are com puted a priori and saved in a database.
Since the database cannot
include all possible param eter vectors, the response com putations in the optim isation loop are
interpolated from values that do exist in the database. In C hapter 4 w e propose an accurate
m ulti-dim ensional interpolation scheme, the m ulti-dim ensional Cauchy m ethod.
A s m entioned in this chapter, circuit theory provides excellent com putation speed, whereas full
electrom agnetic m odels achieve superior accuracy. The tw o m ethods can be com bined into an
hybrid optim isation technique. The hybrid optim isation uses the fast circuit theory m odel in the
optim isation loop. To increase the accuracy o f the circuit theory model, w e include a correction
term . The correction term is derived from a com parison o f one full-EM sim ulation with its
corresponding circuit theory simulation. In C hapter 5, we discuss available hybrid optimisation
m ethods and derive a new technique for hybrid optim isation. O ur new technique derives the
correction term from the differences in the the coupling m atrix o f the circuit theory m odel and
the full E M model.
Since our hybrid optim isation m ethod is lim ited to resonant band-pass filter structures, we advo­
cate an alternative optim isation algorithm for non-resonant and low-pass filter structures. Using
the tim e-dom ain, rather than frequency dom ain response, w e can optim ise such circuits m ore
efficiently. The theory and exam ples are given in C hapter 6 .
All hardware built and m easured throughout this thesis is described separately in C hapter 7. We
do not only present filter circuits, but also other m icrowave devices such as switches and antenna
beamform er. As dem onstrated, the above m ethods are also applicable to the synthesis o f such
devices. The m easured responses o f the devices agree with the predicted responses calculated
w ith our new techniques.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
Chapter 3
Genetic Algorithms in Microwave
Circuit Design
"Research is w hat I am doing i f I don 7 know w hat 1 am doing ",
W em her von Braun
In this chapter, we bring the concept o f evolutional com putation [87] in context with the synthesis
o f m icrowave circuits. The algorithm s and techniques presented provide an additional optim i­
sation m ethod when others fail or perform poorly. B y no m eans are the algorithm s intended to
substitute for the conventional m ethods entirely.
The background and history o f genetic algorithm s are outlined in Section 3.1. An algorithm
itself is explained in Section 3.2, w ith strong em phasis on the application to m icrowave circuits.
In Section 3.3, w e describe the features we add to the algorithm to im prove its applicability to
the design o f m icrowave circuits.
3.1
Background
Genetic algorithm s [95] search an optim um by sim ulating biological evolution. They take their
m o tiv a tio n fr o m th e m e c h a n is m o f n a tu ra l s e le c tio n , a ls o k n o w n as “ su r v iv a l o f th e fitte st” . In
natural selection, individuals com pete for the opportunity to reproduce. Individuals with new
features com e into the population when existing individuals exchange genetic m aterial during
reproduction and when m utations occur.
Genetic algorithm s (GA’s) work differently from traditional optim isation m ethods in four ways:
21
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 3. G ENETIC ALG O R ITH M S IN M ICROW AVE CIRCUIT D E SIG N
22
• GA’s w ork w ith a coding o f the param eter set, not on the param eters them selves
• GA’s search from a population o f points, not from a single point
• GA’s use payoff (objective function) inform ation, instead o f derivatives
• GA’s use probabilistic transition rules, instead o f determ inistic rules
The technique o f evolution program m ing leads to the follow ing properties of GA’s:
• GA’s are very robust, as shown in [95]. T hat m eans that they do not get trapped in local
m inim a w hile searching for param eter values that m inim ise the perform ance function.
• GA’s do not require derivatives o f the perform ance function as m ost other m ethods do.
C alculating the approxim ations o f derivatives, usually by sensitivity analysis [39], is very
tim e consum ing.
• GA’s are very versatile. There exist virtually no restrictions o f the perform ance function
and the constraints. The evaluation function m ay not be continuous. The constraints may
be non-linear [90].
GA’s are already being used com m ercially for scheduling and vehicle routing problem s [8 8 ].
Since a few years, they have also been applied increasingly in engineering design problem s.
S c h a f f e r et al.
[89] design digital filters using GA’s. T a n a k a et al. [91] apply GA’s to
inverse electrom agnetic problem s.
All o f the above-listed properties are also o f value for the optimisation o f m icrowave circuits.
We usually know little about the error functions. They are occasionally discontinuous and may
have many local m inim a. Derivatives o f the error function are m ost often unknown and can only
be found by a tedious num erical approxim ation. This is our motivation for applying GA’s to the
optimisation o f m icrow ave circuits.
3.2
Algorithm Outline and Adaption to Circuit Synthesis
The term inology w e use here is a m ix from the areas o f com puter science and genetics. Since
the algorithm is to be applied in microwave circuit optim isation, the m eaning o f the term s in the
microwave application is also explained.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 3. G ENETIC A LG O R ITH M S IN M ICROW AVE CIRCUIT D E SIG N
23
We refer to m icrowave circuits in evolutional program m ing as individuals. The variable param ­
eters, e.g., line widths and lengths, form their data structure or genetic structure. We assum e that
the genetic structure can be described by a bit string, i.e. list o f 0 ’s and l ’s. In genetics, these
strings are called chromosomes, and the individual bits are called genes. For instance, consider
a planar circuit with three param eters as seen in Figure 3.1. The three param eters are the length
o f the stub, width o f the stub, and its distance from port one.
b in a ry c o d in g
M IC
in d iv id u a l
p2
K-M
1101
p i = 13 m m ■=> 1101
p2 = 8 mm
1000
p3 = 1 1 m m !=> 1011
1000
10111
chrom osom e 1
chrom osom e 2
chrom osom e 3
Figure 3.1: Planar Circuit W ith Three Param eters and Its R epresentation as Individual With
Three C hrom osom es (Binary Coded)
The three param eters are coded in three chrom osom es (bit strings) o f length 4 bits. The string
length is chosen based on the required accuracy. Here, an individual with the chrom osom e set
{1101,1000,1011} represents a circuit w ith p i = 13 mm , p 2 =
8
m m a n d p 3 = 11mm.
Traditional optimisers simulate, evaluate the error function and adjust the layout on one circuit
at a time. GA’s w ork differently. They evaluate the error functions for a group o f individuals
(circuits) characterised by different values o f the chrom osom es. Such a group o f individuals is
called a population. A population o f four individuals for our exam ple is illustrated in Figure 3.2.
chromosomes = parameter
l\
l.iifr*
individual
one
specific
circuit
genes = binary digits
coding parameter value
population=
set of several
different circuits
Figure 3.2: Population o f Four Individuals W ith Three C hrom osom es Each
One simulation step o f the population is called a generation. During the sim ulation the algorithm
assigns every individual a fitness value. The higher the fitness value o f an individual, the better
the individual m atches the specifications. In our example, the fitness value is calculated from the
reciprocal o f the m ean-square-error betw een the response o f any specific circuit and the desired
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 3. G ENETIC ALG O R ITH M S I N M ICROW AVE CIRCUIT D E SIG N
24
response.
Given the fitness values o f a population o f individuals, the algorithm sim ulates natural selection
and reproduction to obtain the next generation. Individuals in the next generation m ay have
different chrom osom es. These are obtained by the GA’s three basic operations:
reproduction, in w hich individuals from one generation are selected for the next generation,
crossover, in which genetic m aterial from one individual is exchanged with genetic m aterial of
another individual, and
m u ta tio n , in w hich the genetic m aterial o f an individual is altered.
C ircuit 1
Circuit 2
Circuit 3
p2
C ircuit 4
■*
K-N
X
.__
p i = 13 mm - 1101
p2 =
p3 = 11 mm - 1011
oS §
? I
iE
55
o
Individual
«55
pl = 11 m m = 1011
p2 = 3 mm = 0011
p3 = 10 mm —1010
Individual
pl = 4 mm = 0100
p2= 13 mm = 1101
p3 = 0 mm = 0000
Individual
Jh
pl = 9 mm - 1001
p2 = 11 mm = 1011
p3 - 15 mm =1111
Individual
Generation
1101
1011
Reproduction
1011
Crossover
111:
1001
1011
Vt VV
VT V ▼
y 11 t
1001
1000
1001
Mutation
r
▼ t
▼
Generation 1 1 1 1
n+1
%
O
F ig u r e 3 .3 : O u tlin e o f G e n e tic A lg o r ith m
Figure 3.3 shows the three operations being used to obtain a new generation from an existing
population. In this exam ple, the population o f any generation consists o f 4 individuals with 3
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 3. GENETIC A LG O R ITH M S IN M ICROW AVE CIRCUIT D E SIG N
25
param eters each. F or clarity, we show the m odification o f one param eter (chrom osom e) only.
The rem aining param eters are changed correspondingly.
The tilted gray num bers indicate the fitness values o f the individuals, obtained from an EM sim ulation o f each specific circuit. Arrows with sm ooth lines indicate copying; arrows with
squiggly lines indicate mutation.
The selection o f individuals that are allow ed to reproduce is done random ly w ith probabilities
based on their respective fitness values. In Figure 3.3, we note that individual 3 with the chrom o­
som e ( 0 1 0 0 ) is not allow ed to reproduce and individual
2
( 1 0 1 1 ) which has a very high fitness
value is allow ed tw o offsprings.
The crossover corresponds to tw o individuals exchanging genetic m aterial.
Conventionally
speaking, crossover introduces som e perturbation into the param eters.
C rossover is controlled by a preset probability. In typical im plem entations, it occurs w ith prob­
ability 0.5 to 0.6. C rossover random ly chooses a pair o f individuals, and exchanges substrings
o f their genetic m aterial. In our circuit optim isation, this corresponds to a slight change in the
layouts o f tw o specific circuits.
Mutation is carried out by creating a new chrom osom es in w hich each bit is copied from the old
without change w ith probability 1 —p mut, and a bit is toggled w ith probability p mut • GOLDBERG
[95] suggests a mutation rate p m ut o f 0.03, but often it is necessary to experim ent until one finds
values that work w ell for a given application.
To search for an optim al solution, w e repeat the evolutional sim ulation process for som e num ber
o f generations and take the fittest individual in the final generation as our solution. W e can
stop the process after a fixed num ber o f generations and pick the best solution, or execute the
algorithm until a satisfactory solution by preset criteria is achieved.
In this section, we studied a sim ple genetic algorithm governed by three operations: reproduc­
tion, crossover, and m utation. A dvanced operators have been tried with the goal o f im proving
the perform ance o f this sim ple settings. In [9 5 ] G o l d b e r g exam ines the m ost com m only used
advanced operators. H e finds that their use generally does im prove the efficiency, i.e. speed and
robustness, o f genetic algorithm s.
3.3
Microwave-Specific Extensions
The algorithm discussed in Section 3.2 is com m on to any genetic-algorithm -based optimiser.
The algorithm works satisfactorily when applied to m icrowave circuit optim isation. N everthe­
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 3. GENETIC ALG O R ITH M S I N M ICROW AVE CIRCUIT D E SIG N
26
less, the m ethod can be significantly improved. In this section we dem onstrate the versatile
possibilities to adapt GA’s to specific applications.
3.3.1
Implementing Physical Constraints
It is rare to find an engineering problem w ithout constraints. W hen physical considerations
indicate that the optim um lies within a certain region o f the param eter space, the algorithm
should take advantage o f this knowledge. C onstraints in m icrowave circuit design can take a
variety o f form s. They include upper and low er bounds on param eters and non-negativity of
network elem ents. Constraints can also be m uch m ore com plex. For exam ple, there could be a
specific region in the param eter space, or param eter com binations, w hich are favourable or not
acceptable.
We can often derive additional constraints w hen we include our knowledge about the specific
circuit topology. Any knowledge can generally help to pinpoint the optimal circuit layout faster
and m ore reliably. This knowledge can be, for instance, past experience or insights on the
theoretical behaviour o f the circuit.
Gradient search algorithm s are lim ited to the im plem entation o f fairly sim ple constraints, how ­
ever. B ackground inform ation, such as experience, insights, and com plex constraints are diffi­
cult, if not im possible, to im plem ent in the rigid gradient-based optim isation schemes.
Genetic algorithm s are different. They pick points from the param eter space guided by the law
o f the survival o f the fittest, rather than following a tedious gradient calculation. Picking a point
p that represents a particular circuit is com putationally alm ost free. The selected point can be
checked for coherence before the circuit is simulated. M any tests can be run on p before its
objective function is calculated. These tests include, but are not lim ited to, sim ple and com bined
constraints, validation o f conform ance with rules set up for the specific circuit, and obsoleteness
o f circuits known to be ill-perform ing. Once a param eter vector is identified to be ill-perform ing,
it can be elim inated from the gene-pool w ithout further com putational expenses. The algorithm
sim ply picks a new point from the rem aining population and goes on.
L et us illustrate the procedure on the follow ing example. Suppose, w e want to optim ise an open
strip-line stub as shown in Figure 3.4. The goal o f the optim isation is that the input im pedance
Z rn is zero at 20 G H z (short) and approaches infinity (open) at 30 GHz.
Transm ission line theory tells us that this goal is theoretically achieved for an im pedance ratio
o f 3:1 for the narrow to the wide line, and an electrical total length o f one-half wavelength at 30
GHz. The sm ith charts in Figure 3.4 explain why: The open end at point A m oves to point B,
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 3. GENETIC ALG O R ITH M S IN M ICROW AVE CIRCUIT D E SIG N
27
due to the line length o f the narrow line. The point B is transform ed into B’ due to the step. The
w ide line moves the point B’ to the point C. As seen, at 20 GHz, C ends up at Z =
0
(short) and
at 30 GHz, the im pedance at C is infinite (open).
Figure 3.4: Open Stripline Stub, Sm ith C harts for 20 GHz and 30 GHz
A conventional optim iser can only im plem ent sim ple bounds on the param eters w \, W2 , h , and
12- The im plem entation o f a constraint prescribing that the im pedance ratio o f the lines be close
to 3:1 is very difficult w ith conventional optimiser. For genetic algorithm s, this constraint is
sim ple to include. W ith a single line o f code, w e can check the ratio for every individual and sort
out the ones that do not conform with our constraint before optimisation.
W hat constraints are acceptable in the algorithm is only lim ited by the ability to im plem ent them
in com puter code, in our case FO R TRA N code. An exam ple o f an insight-guided search is given
in Section 3.4.1.
3.3.2
Parameter Space Visualisation
In the previous section, w e dem onstrated how the algorithm profits from inform ation that the
user gives to the program . The opposite also applies: The user can gain inform ation provided
by the algorithm. It is characteristic o f genetic algorithm s to search the param eter space m ostly
around areas with high fitness values.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 3. GENETIC A LG O R ITH M S I N M ICROW AVE CIRCUIT D E SIG N
28
W hen we plot all individuals ever generated in the param eter space, we can m ake out areas
with high population density. Such areas indicate high fitness o f the individuals located within.
W e can use this inform ation to gain insights into the circuit’s perform ance and to guide our
optim isation further.
To illustrate the technique, we take a look at a four-pole dual m ode filter with three param eters
shown in Figure 3.5. The filter is optim ised for a 2% passband at 4 GHz. A population plot of
all generated individuals is shown in Figure 3.6.
Figure 3.5: L ayout and Param eters o f Four-Pole D ual M ode Filter
Best Individual:
pl=0.438
p2=8.9
p3=-9.2
-10
Pl
F ig u r e 3 .6 : P o p u la tio n A fte r 2 0 G e n e r a tio n s. N u m b e r s are in m ils
We can see that the filter perform s best at values o f param eter p i around 0.5 mil. Param eter p \
does not seem to be correlated w ith param eter p 2 - Param eter p 2 , by contrast, varies widely, but
does show some crowding around p 2 = 9 m ils and p 2 = -2 mil.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 3. GENETIC ALG O R ITH M S IN M ICROW AVE CIRCUIT D E SIG N
3.4
29
Examples
In the following exam ples, several m icrowave circuits are optimised. The exam ples dem onstrate
the capabilities o f genetic algorithm s in optim isation o f microwave filter circuits. The perfor­
m ance o f the GA is com pared with that o f gradient-search algorithm.
3.4.1
Ka-Band Stripline Rejection Filter Design
The K a-band filter presented here is part o f a satellite system currently under developm ent at
SPAR Canada. The satellite system receives signals from a ground station in one frequency
band f up, am plifies this signal, and transm its it back to earth in another frequency band f downThe am plitude o f the received signal is typically about 100 dB sm aller than the am plitude of
the transm itted signal. H ence, if the transm itted signal— even in very sm all am ounts-reach the
am plifier input, it would disturb the system. The system m ight even start oscillating at the
transm itting frequency. Thus, it is im portant to block any signals of the transm itting frequency
at the am plifier input by a band stop filter.
The given satellite system operates at the following frequencies:
R eceiver centre frequency :
R eceiver band width:
Transm itter centre frequency:
Transm itter band width:
29.5
GHz
1.0
GHz
20.0
GHz
1.0
GHz
In order to ensure reliable operation o f the amplifier, the filter transfer function has to satisfy the
following specifications:
A ttenuation o f transm itter signal
:
|
(19. 5 . . . 20.5G H z)|
<
-70 dB
A ttenuation o f receiver signal
:
|*S,2 i ( 2 9 . 0 . . . 30.0G H z)|
>
-0.2 dB
The substrate values are pre-determ ined as follows:
Circuit structure
: Stripline technology
D ielectric constant er
: 3.0
Substrate height
: 2 x 10 m ils equiv. 0.254 m m
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 3. G ENETIC ALG O R ITH M S IN M ICROW AVE CIRCUIT D E SIG N
The layout is first designed with the circuit-based program
to u c h sto n e
30
[145]. In order to
obtain the desired response, we propose the filter layout seen in Figure 3.7. Four stubs branch
off a transm ission line. Each stub has a step to a higher im pedance at h alf o f its length. The stubs
follow the open-short input im pedance behaviour discussed in Figure 3.4. An EM -sim ulation of
the layout shows that the filter design needs further optimisation, because its actual response
does not satisfy the specifications.
In our filter optim isation process, eight param eters are adjustable. These are the lengths and
widths o f the stub lines, and the lengths o f the transm ission lines betw een the stub lines. All
param eters (pi to p§) are shown in the sketch o f Figure 3.7.
91
[mm]
90
89
88
87
86
85
84
[m m ] 2 2
Figure 3.7: L ayout o f Stripline R ejection Filter, G ray=B efore optim isation, B lack=A fter O pti­
m isation
R ecalling the term inology o f Section 3.2, we translate this situation into the GA settings: The
param eters are coded into 10 bit binary strings. The population is set to 5 individuals, and the
GA com putes ten generations. The fitness is calculated from an error function based on six
frequency points as shown in Table 3.1.
band
stop band
pass band
frequency
/ 1 = 19.5 GHz
f 2 = 20.0 GHZ
/ 3 = 20.5 G H Z
/ 4 = 29.0 GHz
f 5 = 29.5 GHz
/ 6 = 30.0 GHz
S 21 spec.
0
0
0
1
1
1
Table 3.1: Transfer Param eter S 2 1 Specifications
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 3. G ENETIC ALG O R ITH M S IN M ICROW AVE CIRCUIT D E SIG N
31
As an objective function, w e use the m ean-square-error function
2
E ( p i , . . . ,ps)
=
\ei ( p i ,
•••, p s ) \ 2 ,
(3
.1
)
2= 1
where e\ and e 2 are the errors in the stop and pass band, respectively, i.e.
n= 3
el
=
E l ° - | ^ ( / n , P t , . . , P 8 )||
(3-2)
n— 1
71 = 6
e2
=
X ) I1 - \s 2 i ( f n , P i , - - - , P s ) \ \ ■
(3-3)
n = 4
The optim isation starts w ith the layout shown in gray in Figure 3.7, , which also shows the
param eters p l to p 8 .
Analysing the filter geometry, one notices that the filter works best when the ratio o f the line
im pedances o f the narrow and wide stub lines §£(see Figure 3.7) is around 3:1. This knowledge
can be im plem ented into the genetic algorithm. The search in the param eter space is then biased
towards this ratio. This kind o f non-linear objective is difficult to im plem ent in traditional search
methods.
Further, due to the available etching technology, the m inim al line w idth is constrained to be 1
mil. In the genetic algorithm approach, constraints like this can be easily considered. Invalid
geom etries, as the ones with too narrow lines, sim ply “die” , even before they are analysed.
F or the optim isation, a m icro-genetic algorithm [97] is used in connection with a m om entm ethod EM -sim ulator.
The results o f our optim isation are shown in Figure 3.8. A fter optim ising the transfer param eter
621
yields exactly the desired response, as defined in Table 3.1.
Since the search is guided by physical insights as described above, the optim iser adjusts the
circuit layout within a m ere 350 function calls. By com parison, the sam e optimisation w ithout
constraints takes about 600 function calls. The gradient search optim isation cam e to a similarly
optim ised circuit within 600 function calls, too. O ne function call here refers to the calculation
o f the response at a single frequency point. The optim ised layout is shown in black in Figure
3.7.
Altogether, the optim isation takes approxim ately 30 m inutes on an H P-700 series U N IX work­
station. A fter optim isation, the transfer param eter S 21 satisfies the desired 60 dB isolation at 20
GHz as shown in Figure 3.8. M easured results for this filter are given in C hapter 7.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 3. GENETIC A LG O R ITH M S IN M ICROW AVE CIRCUIT D E SIG N
befo re o p tim izatio n
after optim izatio n
f[G H z ]
[d B ]
19.5 - 20.5 G H z
2 9 .0 - 3 0 .0 G H z
-10
-20
-30
befo re o p tim izatio n
-40
after o p tim izatio n
-50
-60
-70
f[G H z ]
Figure 3.8: R esponse o f Stripline Filter Before and A fter GA O ptim isation
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 3. GENETIC ALG O R ITH M S IN M ICROW AVE CIRCUIT D E SIG N
3.4.2
33
Waveguide Cavity Filter Design
This exam ple dem onstrates the optim isation o f an K u-band wave guide cavity filter. The four
cavities o f the filter are iris-coupled as seen in Figure 3.9. The optim isation param eters p i to
p5 are the iris widths and the cavity lengths. The optim isation goal is a -25 dB pass band at 15
GHz with 3% bandw idth. The analysis is perform ed by the full electrom agnetic m ode-m atching
program
em
_
m ux
[153].
P.
M
P5
N M
P5
H M
P.
MM
H
Figure 3.9: L ayout o f Four-Pole R ectangular Iris-Coupled W ave-Guide F ilter Showing all A d­
justable Param eters
As seen from the frequency response in Figure 3.10 the genetic algorithm obtains an overall
better solution after approxim ately 500 function calls. The gradient search needs 220 function
calls in this exam ple, but does not find the global m inim um . The optim isation stops when the
algorithm is trapped in a local m inim um.
[dB]
-10
-20
-25
-30
G ra d ie n t O p tim ise r
-35
------------- G A O p tim ise r
-40
14
1 4.2
14.4
14.6
14.8
15
15.2
15.4
15.6
15.8
16
f [G H z ]
Figure 3.10: O ptim ised R esponse o f W aveguide Filter Layout
The quality o f the optimal solution o f the gradient-search algorithm depends strongly on the
starting point given. A nother optim al response is found, if another starting point is chosen. By
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 3. GENETIC A LG O R ITH M S I N M ICROW AVE CIRCUIT D E SIG N
34
contrast, the GA always yields the optim al response shown in Figure 3.10. The initial param eters
m ay affect the convergence speed but do not affect the final optimal solution.
3.5
Optimisation Performance of GA’s
In order to com pare the GA’s convergence speed m ore precisely with that o f a traditional gradientsearch optimiser, a standard problem is tackled by both optimisers: A folded stub m icrostrip filter
as shown in Figure 3.11 is optimised.
m>[t2
■
j
Figure 3.11: L ayout o f M icrostrip Line Filter
All line widths are fixed at 4.8 mil. The substrate is 5 m ils thick with a dielectric constant o f 9.9.
The design param eters p i to p3 are chosen as shown in Figure 5. F or the gradient search two
different sets o f start values as shown in Table 3.2 are used. The starting values are chosen the
sam e as in [43].
p l=
P2 =
p3=
Start Set 1
Start Set 2
74 mils
62 mils
73 m ils
61 m ils
1 2 m ils
13 mils
Table 3.2: Starting Values for Optim isation o f Folded Stub Filter
Figure 3.12 presents the results o f the optimisation. It is seen that the GA finds the best solution
overall, w hereas the gradient search (starting point set
1
used) yields a sub-optim al solution.
Figure 3.13 shows the decay o f the error with respect to the num ber o f function calls. The GA’s
error function decays faster than that o f the gradient search, no m atter which o f the tw o starting
points is chosen.
The decay rate o f the gradient search strongly depends on the starting point chosen. M oreover,
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 3. GENETIC ALG O R ITH M S IN M ICROW AVE CIRCUIT D E SIG N
S21
[dB]
-10
-20
-30
G enetic
A lgorithm
-50
-60
G radientSearch
-70
-80
f [GHz]
Figure 3.12: R eturn Loss o f Optim ised Stub Filter
25
Gradient Search
Starting Point #2
20
15
G ra d ie n t S e a rc h
S ta rtin g P o in t #1
10
5
G e n e tic
Algorithm
o
100
120
140
function calls per frequency point
Figure 3.13: D ecay o f E rror Function for Optim isation M ethods
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 3. G ENETIC A LG O R ITH M S IN M ICROW AVE CIRCUIT D E SIG N
36
the gradient search does not find the overall solution for either starting point. By contrast, the
GA’s error function decays slowly to the overall m inim um.
3.6
Conclusion
Genetic algorithm s have been shown to be effective in the optim isation o f m icrowave circuits.
Robustness m akes them a good choice for many synthesis tasks. GA’s are particularly effec­
tive for finding optim al solutions in high-dim ensional param eter spaces, as they do not require
derivatives.
Although searching by evolution for the best solution is generally known to be slow, it has been
shown that genetic algorithm s can som etim es be even faster than directed optim isation m ethods
such as gradient search.
Especially, when physical insights are included in the algorithm s, and/or when various con­
straints and objectives m ust considered, the optim isation by “survival o f the fittest” is often the
best available optim isation method.
However, the superior optim isation speed o f the GA over the gradient search cannot be gener­
alised. The decay rate o f the error function depends in both m ethods on many param eters, e.g.
starting value o f the gradient search, param eter range.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
Chapter 4
Multi-Dimensional Cauchy Method
"If it keeps up, man w ill atrophy all his lim bs but the push-button
finger", Frank L loyd Wright
A num ber o f software packages are now com m ercially available for Electro-M agnetic (EM )
sim ulation o f m icrowave circuits. These packages provide reasonably accurate results, allowing
designs to be im plem ented in a cost effective way. These packages, however, are typically
very com putation intensive. It has been well recognised that the CPU tim e and m em ory space
required to sim ulate fully integrated m icrowave circuits, using these E M sim ulators, far exceed
the capabilities o f today’s com puter workstations.
O ver the past years, there has been a strong interest am ong researchers to avoid this problem
by using neural netw orks [130], space m apping [72] and param eter extraction [45]. The use of
C auchy’s m ethod has also been proposed in [76, 77]. The C auchy m ethod yields a surprisingly
accurate m atch o f the com puted points, betw een and even exterior to the sam pled points with
the exact solution.
However, m ost o f the papers published on applications o f C auchy’s m ethod deal with one­
dim ensional interpolation, nam ely frequency response interpolation. Here, we extend the one­
dim ensional interpolation for frequency response interpolation to m ulti-dim ensional Cauchy in­
terpolation with respect to both frequency and geom etric dim ensions. Two different approaches
are suggested to achieve a m ulti-dim ensional approach: A recursive one-dim ensional application
o f the standard Cauchy m ethod and m ulti-dim ensional rational function expansion.
We start the chapter w ith a short review o f interpolation schemes. In the second part o f this
chapter, w e propose the new m ulti-dim ensional interpolation scheme, called multi-dim ensional
37
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 4. M U LTI-D IM EN SIO N AL C A U C H Y M ETHO D
38
Cauchy method. We explain why this schem e is superior to linear, quadratic, and polynom ial
interpolation. A t the end o f this chapter, som e exam ples o f applications o f m ulti-dim ensional
Cauchy m ethod are given.
4.1
Multi-Dimensional Interpolation
In general, the interpolation problem consists o f determ ining a function y o f p from
7
given pairs
o f num bers (p1, y 1), (p2, y 2) , . . . , (p7 , y 1 ). These pairs are here called sam ple p oint pairs. The
set o f all available sam ple points is denoted by TLThe num ber p m ay be real, com plex, or— for m ulti-dim ensional interpolation- vector-valued (p).
The function value y is here considered as one specific response of the circuit, e.g. return loss,
w hich is typically complex.
The linear m ulti-dim ensional interpolation defines a linear function in the n-dim ensional param ­
eter s p a c e p = ( p i , p 2, ■■• ,p n ) by
y(Pl , P2, ■■■i P n ) =
Ao
+
A 1P 1
+
A2P 2
-I
(4.1)
A„ P n
w here the A’s are unknown coefficients. The A’s can be derived from n + 1 sam ple points from a
system o f linear equations. We can then interpolate any point in the param eter space using eqn.
(4.1).
The algorithm can be expanded to higher-order approxim ations o f the function. In order to
include quadratic term s, the function in eqn. (4.1) can be rewritten as
y ( P l , P 2 , ■■■,Pn) = A0 +
A lP i
+ A2p \ +
A 3 P 2 + A4 P 2 + A 5 P 1P 2
b
M nPn
(4-2)
This technique requires 2 n + l response samples. The sam e algorithm can obviously be expanded
to higher order approxim ations.
4.1.1
Neural Networks
Recently, neural netw orks have been introduced into the m icrowave engineering com m unity
[125, 126, 130] as a fast and flexible tool for m icrowave circuit response interpolation. The
input and output param eters o f the circuit are related by a netw ork of activation functions usually
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 4. M U LTI-D IM EN SIO N AL C A U C H Y M ETHO D
39
arranged in several layers. A typical three-layer feed-forw ard neural netw ork is shown in figure
4.1. F or m ore inform ation about neural netw orks refer to [119]-[134].
N e u ro n s
W e ig h te d
L in k s
input l.a \
-ik k ten I .aver
( Juiput L ayer
Figure 4.1: Neural N etw ork M odel o f M icrow ave Circuit
Follow ing our definitions, the input and output layer o f the netw ork correspond to the vector p
and the vector y, respectively. The input layer and output layer are connected by a hidden layer.
In each layer, w e find neurons that are activated depending on the input from the layer on the
left. The activation level is com puted from the so-called activation function. A typical activation
function is the logistic function
1
ah
1 + e x p (— £ " = i aiWih - @h )
(4.3)
w here Wih are the adjustable weights between ith and hth neuron, and ©h is a threshold value.
The netw ork has to be trained by a num ber o f sam ple input-output vector pairs (p, y ) in order
to determ ine the weight factors
and thresholds 0 ^ . These data have to be com puted by an
expensive EM simulation. The sam ple p-vectors are chosen randomly, and y is com puted by an
EM -sim ulation.
During the training process, the neural net autom atically adjusts its weights w iy and threshold
values Oh such that the error E betw een predicted and sam pled outputs is m inim ised.
The num ber o f hidden neurons m ust be determ ined carefully. Too few neurons m odel the system
Reproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 4. M U LTI-D IM ENSIO NAL C A U C H Y M ETHO D
40
imprecisely. Too m any neurons tend to over-train the network, such that the training data are
m atched nicely, but subsequent testing for unknown points fails.
The perform ance o f the netw ork is tested by a second set o f sample vector pairs. These sam ples
are not included in the training data set. If these unknown sam ple pairs are m odelled correctly,
the netw ork is likely to represent a valid m odel for the m odelled system.
A fter training and testing, the netw ork is ready to be used as a substitute to the EM -sim ulator.
The netw ork can predict the output y for any given input param eters p within the trained region.
The calculation is very fast since only a few basic algebraic operations are required.
The m odel once trained can be reused in the design process many tim es w ithout the cost of
additional EM -sim ulations.
4.1.1.1
Neural Network Filter Model
This exam ple dem onstrates a neural netw ork m odel o f a narrow band three pole filter as shown
in Figure 4.2. The m odelling is difficult due to extrem ely non-linear dependence o f the circuit’s
response.
P3
h
Pb
Figure 4.2: Planar Three-Pole Filter
The netw ork consists o f 5 input neurons, 20 hidden neuron and 4 output neurons and is m od­
elled using the SNNS [119] neural netw ork generator. The m odel im plem ents five input circuit
param eters, including the frequency. The output neurons represent the com plex S-param eters.
T he netw ork is trained on 5000 random ly taken sam ples and tested on 500 additional samples.
T he com putation o f the sam ples took 80 h. Training o f the network required 20 h.
Figures 4.3 and 4.4 show the correlation o f S 'a for a random ly generated layout geom etry and
for an optim ised layout, respectively.
4.1.1.2 Limitations of Neural Network Applications
We can see that the neural netw ork m odels the approxim ate behaviour o f the circuit’s response
correctly. However, an exact m odelling, particularly for values o f the S-param eters close to zero,
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 4. M U LTI-D IM EN SIO N AL C A U C H Y M ETH O D
41
-10
-15
-20
N eural N e t
E M -S im .
-25
3.86
3.9
3.92
3.94
3.96
3.98
f [GH z]
Figure 4.3: Neural N etw ork Approxim ation and Exact EM Sim ulation for R andom Input Vector
-105
p—i i <
N e u ra l N e t P re d ic tio n
§
“
-2 0
E M -S im u la tio n
-30
-35
-40
-45
3.7
3.75
3.8
3.85
3.9
3.95
4
4.05
4.1
4.15
4.2
f [GHz]
Figure 4.4: N eural N etw ork A pproxim ation and Exact EM Sim ulation for an O ptim ised Circuit
Layout
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 4. M U LTI-D IM EN SIO N AL C A U C H Y M ETHO D
42
is not achieved. The accuracy o f the neural net can be im proved further by adjusting the netw ork
size and activation function, and by taking m ore training data points. The latter com es with the
penalty o f m ore EM simulations.
The requirem ent for m any sam ples originates from the fact that the standard neural netw ork
is a black box. Theoretically, it can m odel anything, from m edical diagnosis to stock m arket
fluctuations. This apparent advantage is actually a draw back when it com es to m icrowave cir­
cuit modelling. The netw ork requires a vast num ber o f response sam ples to m odel the circuit
response accurately, since it does not have any background knowledge about microwave circuits.
In the last years, several researchers focused on the use o f neural networks for m icrowave cir­
cuit m odelling and synthesis. The general idea is to include knowledge about the general cir­
cuit behaviour into the netw ork structure. W a t s o n , G u p t a , and M a h a ja n [127] as well as
Z h a n g et al. [131]-[134] use special neurons, so called knowledge-based neurons. Each of
these neurons includes a coarse m odel o f the circuit’s response. The netw ork itself is only used
for m inor adjustm ents o f the coarse m odel. A nother approach by BURRASCANO and MONGIArdo
[128, 129] applies Kohonen self-organising maps. The netw ork is split into tw o sections,
one for m odelling the general behaviour o f the circuit and another one for adjusting the general
behaviour according to a specific response with respect to the input param eter vector.
T he new knowledge-based neural networks are an exciting new paradigm for circuit modelling.
Including knowledge seem s to be the key towards an efficient application o f neural networks.
One drawback o f this idea is, however, that we have to develop a specific network for each type
o f circuit.
In the following section, we discuss the Cauchy method. The Cauchy m ethod allows a general
description o f the response o f passive microwave circuits. N o specific knowledge o f the circuit
is required.
4.2
Cauchy Method
Linear, quadratic, and spline interpolation schemes lack in ability to m odel poles for finite argu­
ments of a function property. Due to their nature, there is no possibility for the function value
y (p ) to approach infinity for a finite value o f p.
The transfer function o f passive m icrowave circuits, however, often includes such poles for finite
values o f the circuit variables. An obvious exam ple is the input im pedance with respect to the
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 4. M U LTI-D IM EN SIO N AL C A U C H Y M ETHO D
43
frequency u> o f an ideal parallel m icrowave resonator as shown in Figure 4.5. A t resonance
frequency, the circuit im pedance approaches infinity.
Zresi^res) =
1™ ( j
U l^ -L O r e s
—
:C
Z D
x ) ----- > OO
^
(4.4)
-jj-
gL
Figure 4.5: Parallel R esonator and Input Im pedance
R esponses, like the one shown above, can be represented by rational functions. In A ppendix C
we show, that any response from a m icrowave circuit can be approxim ated by a rational function.
Consequently, rational polynom ials as interpolation functions yield a m uch closer representation
o f the system ’s response than other schemes, e.g. splines. Below, w e develop the m athem at­
ical fram ew ork for a one-dim ensional rational function interpolation. This algorithm , known
as Cauchy method or Pade approximation, is a frequently applied m ethod for interpolation of
the frequency response. M any published papers show the alm ost m agical pow er o f frequency
response interpolation. Today, the technique is incorporated into many com m ercial microwave
sim ulation software packages as so-called f a s t f requency sweep.
4.2.1
One-Dimensional Cauchy Method
The system response S ( f ) is described by a rational polynom ial function o f num erator order N
and denom inator order D . F or one param eter, m ost often the frequency / , the function has the
form
n(
a0 +
°1
/ +
a 2f
2H
Qp + H n = l an f n
^
l +6
1
/ +6
2
/2+... - 1+££=
!&
*/"'
In order to determ ine a function in the form o f eqn.(4.5) we need k = N + D + 1 arbitrary
sam pling points. W hen using k = N + D + 1 sam ple points w e can set up a system o f linear
equations whose solution yields the values for the coefficients a o , . . . , ajv and b \ , . . . , bnIn order to avoid m atrix inversion problem s o f the likely ill-conditioned m atrix, another faster
and m ore stable algorithm was found by S to er and B u r l is c h . This algorithm o f the Neville
type perform s the interpolation on tabulated data in a recurrent manner. The B urlisch-Stoer
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 4. M U LTI-D IM EN SIO N AL C A U C H Y M ETHO D
44
algorithm does not require the inversion o f a matrix. The algorithm is outlined in the following
paragraph.
L et R \ be the value at f \ o f the unique rational function o f degree zero (i.e. a constant) pass­
ing through the point ( f i , S ( f i ) ) . Likew ise, define R 2 , R 3 , . . . , Rk- Now let R n be the ra­
tional polynom ial o f degree one passing through both ( / i , 5 ( / i ) ) and (J 2 , S ( f 2 ))- Likewise
-^ 2 3 ) -^ 2 3 ) • • •, R ( k - i ) ( k ) ) ■ Similarly, we proceed for higher order polynom ials up to R m . . . k The polynom ial R. m.. .k is the value o f the unique interpolating polynom ial through all k points,
i.e. the desired answer. The various R ’s form a tableau w ith ancestors on the left leading to a
single descendant at the extrem e right. F or example, with k — 4 we have
S im u la ted
S im u la ted
Sim u la ted
S im u la ted
The B urlisch-Stoer algorithm is a recurrent way o f filling in the num bers in the tableau a colum n
at a time. It is based on the relationship betw een a daughter and her parents. B y [81], we have
R(i+l)...(i+m)
R i . . . ( i + m —1 )
(4.6)
This technique produces the so-called diagonal rational function, where the degrees o f the nu­
m erator and denom inator are equal (if k is odd) or where the degree o f the denom inator is larger
by one (if k is even). For the derivation o f the algorithm , refer to Stoer and B urlisch [80],
To dem onstrate the efficiency o f the one-dim ensional Cauchy m ethod, the results o f different
interpolation schem es (linear, spline, and Cauchy) are shown in Figure 4.6. The response o f
a four-pole narrow-band filter is sam pled at 20 frequency points only. The linear interpolation
does not yield m uch inform ation about the behaviour o f the circuit’s response. The spline inter­
polation provides better correlation, due to the sm ooth curve fitting; however, the curve does not
m atch with the exact solution. The Cauchy m ethod, on the other hand, shows the exact response,
even though only a few sam ple points are used.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 4. M U LTI-D IM EN SIO N AL C A U C H Y M ETH O D
5
S ll
45
Sll
[dB]
0
■5
-10
-10
-15
-15
-20
-20
-25
-3 0
3 .7
-25
® E xact Solu tion
X Sam p le P oints
— L inear Interpolation
3.8
3.9
4
-3 0
f [G H z]
4.1
4 .2
4 .3
4 .4
3.7
* E xact Solution
X Sam ple P oin ts
Spline Interpolation
3.8
3.9
4
4.1
f [GHz]
4 .2
4 .3
4 .4
S ll
A
•»»>»>»»«•»>»>»
-10
-15
-20
-25
-3 0
• E xact Solu tion
X Sam p le P oints
— C auchy Interpolation
f [GHz]
3.7
3 .8
3.9
4
4.1
4 .2
4 .3
4 .4
Figure 4.6: Interpolation o f Four-Pole F ilter Response, a) Linear, b) Cubic Spline, c) Cauchy
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 4. M U LTI-D IM EN SIO N AL C A U C H Y M ETHO D
4.3
46
Adaptive Sampling
As seen from Figure 4.6, m any frequency sam ples are taken at frequency bands where the re­
sponse is totally linear. There is not m uch inform ation gained from these sam ples; rather, we
should concentrate the sam pling in bands o f high non-linearity. D haene et al. show in [82] that
an adaptive sam pling scheme can be applied in order to focus the sampling in those areas.
The technique is illustrated by the follow ing example. First, a few samples, let’s say five, are
taken. Using these samples, tw o different rational polynom ial approxim ations are com puted.
These two m odels are then scanned for the frequency with biggest mismatch. A t this point,
the next sam ple is taken, and the procedure is repeated until both m odels agree. The procedure
is illustrated by an exam ple in Figure 4.7. The Cauchy m ethod starts with 5 samples. From
these samples, the response is interpolated for a rational function o f num erator order
2
and
denom inator order 2. A nother response is interpolated using a rational function o f num erator
order 1 and denom inator order 3. The curves are com pared, and the next sam ple is taken at the
frequency point with the biggest m ism atch, here / = 9.6 GHz. The com parison is repeated,
and m ore sam ples are taken, as long as necessary. A t the end, w e observe, that the sam ples are
concentrated in areas w here the response is highly non-linear. Only a few sam ples are taken in
intervals o f the response with very linear behaviour.
4.4
Multi-dimensional Cauchy Method
We now extend the one-dim ensional interpolation for frequency response interpolation to a
m ulti-dim ensional Cauchy interpolation with respect to both frequency and geom etric dim en­
sions. Two different approaches are suggested to achieve a m ulti-dim ensional approach: a recur­
sive one-dim ensional application o f the standard C auchy m ethod, and m ulti-dim ensional rational
function expansion.
4.4.1
Recursive Cauchy Method
The recursive m ethod solves the m ulti-dim ensional interpolation problem by a recursive algo­
rithm. The algorithm itself perform s a one-dim ensional Cauchy interpolation as described in
Section 4.2. From a given set H o f
7
sam ple points and an arbitrary point p*, the algorithm C
calculates the interpolated function value S*(p*).
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 4. M U LTI-D IM EN SIO N AL C A U C H Y M ETHO D
Sll
5 points
[d B )
-10
t y-v
-1 5
•
•••'*
-20
order(num) = 2
' order(denom)= 2
o rder(num ) = 1
o rd e r(d e n o m )s 3
-2 5
20
-10
-1 5
•
o rder(num ) = 3
' o rd er(d e n o m )= 2
■,
o rder(num ) = 2
*• o rd er(d e n o m )= 3
6 points
-20
-2 5
f [G H z ]
-3 0
20
-10
-15
7 points
-20
«
y-\ o rder(num ) = 3
' order(d e n o m )= 3
;
.**•. o rder(num ) = 2
’• o rd er(d e n o m )= 4
-2 5
-3 0
-35
flG H z ]
-4 0
20
10
0
-10
-20
points
-3 0
»
'•S
-4 0
•,
y-\ o rder(num ) = 4
' order(d e n o m )= 3
o rder(num ) = 3
*• o rd er(d e n o m )= 4
-5 0
f IG H z ]
-6 0
10
4
14
20
18
-10
-1 5
-20
-2 5
-3 0
-3 5
-4 0
-4 5
9 points
i
C'
y-\ o rder(num ) = 4
\ o rd er(d e n o m )= 4
o rder(num ) = 3
’• o rd er(d e n o m )= 5
I IG H z ]
-5 0
20
Figure 4.7: Adaptive Sam pling Procedure
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 4. M U LTI-D IM ENSIO NAL C A U C H Y M ETHO D
48
The set TL is put together by the pairs o f sam pling points p ' to p 1 . The algorithm C can be defined
as the function C (p*,TL) w hich yields the interpolated response S* for p* using the sam ples TL.
c (p*,n) =
(4.7)
(p", W , S ' ) , ( p " , S " ) , ( p ' " , S ' " ) , . . . , (p“ , S ” )} ) — > S '
Using these definitions, the algorithm can now be extended to a m ulti-dim ensional interpolation.
F or this purpose, the set o f sam ple points m ust be extended from the one-dim ensional sample
point set TL to a m ulti-dim ensional sam ple point array.
4.4.1.1
Choice of Sample Points in Parameter Space
The sam ple points in an n-dim ensional param eter space are now represented by vectors p =
( p i , P 2 , . . . , p n )- For the recursive algorithm , the set o f sam ple point pairs (p, S ) m ust fall on
a com pletely filled grid o f points. The grid does not have to be equidistant. Typically sample
locations in the param eter space are shown in Figure 4.8 for tw o and three param eters.
P2 ’
Pi
pl’ pl” pl
pl
Figure 4.8: Sam ple Locations for Two-dim ensional and Three D im ensional P aram eter Space
E ach dim ension has its own subset o f param eter values, as seen from Figure 4.8, w hich do not
change for different values o f the other param eter dim ensions. In the given exam ple the set of
sam ples in the p \-dim ension has the four param eter values p[ to p'{".
4.4.1.2 Algorithm Implementation
The goal is to interpolate the function value S* o f an arbitrarily located point p* = {p\ ,p*2, . . . , p n* )
The algorithm can be divided into three steps:
Step 1: The root process starts by interpolating the point p* with constant p 2 =
Ps = P% etc.
parallel to the p \ -axis, shown as a dashed line in Figure 4.9(a). This is a one-dim ensional
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 4. M U LTI-D IM EN SIO N AL C A U C H Y M ETHO D
49
interpolation, so the algorithm C defined in eqn. (4.7) can be used. This step yields the
desired interpolated point
S* = C ( \ p l , p * 2, . . . ] , A )
(4.8)
w here A is the set o f sam pling points for p[ to p'{" and p 2 — p,\, • • .=const. These are
the points m arked with ©
in Figure 4.9. They m ay not fall on the grid o f known sam ple
points. If that is the case, the algorithm proceeds to step 2 in order to determ ine the points
© . O therw ise (the points are known), the algorithm proceeds with Step 3.
Step 2: The algorithm calls itself for each o f the unknown points © . In the exam ple, the
algorithm starts four new child processes
C ((p i,^ )^ t)
,C(( p '",P2),B3) ,C((p?",rf),B4)
(4.9)
as shown in Figure 4.9(b). The B ’s are sets o f sam ple points with a fixed value for p\
as seen from Figure 4.9. The interpolation is now perform ed along the p 2 -axis. The
routine called is exactly the sam e routine as already used in Step 1. The algorithm is thus
recursive. Again, each subprocess checks if the set B is from known samples. If not, the
algorithm starts another instance o f subprocesses in order to interpolate the points included
in B using the next higher dim ension. In the example, this would be p^.
Step 3: In case the subprocess determ ines that all sam ple points are known, it calculates the
interpolated point using C and hands it back to the parent process w hich requested that
point.
Finally, the answ er for the root process in eqn. (4.8) will be found.
T he iterations o f the algorithm can be represented in an inverted tree diagram (Figure 4.10).
Each process C(TL,p*) requests the unknown sam ple points in its dim ension from a num ber of
child processes and hands the result o f its interpolation to its parent. The tree term inates with
“leaves” -processes that interpolate from know n sam ple points.
Adaptive sam pling can be applied to the m ulti-dim ensional Cauchy m ethod. U sing the recursive
m ethod, the sam ples m ust lie on a— not necessarily uniform — grid and adaptive sam pling can
only be used in one dim ension w ithout constraints. However, when com bining the approaches,
the adaptive sam pling can be applied to m ore dimensions.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 4. M U LTI-D IM EN SIO N AL C A U C H Y M ETHO D
CO
N3
O’
<D
.
50
—a— -fc-a-Q — a
o.
a.
c
o
io
X X
»-l
£
T3
IS
£
s
'o
0-
3
cr
o
«3 °
-*
x •
X X
X X
a
X
a
-0 --X -X ---X
0)
1 -4-*
Ul
es
--© --X -X -
(b) Step 2
X X --X -
o)
Da-
X
a
a
X X
<N
P .
X X
o.
a
Figure 4.9: Steps in Recursive C auchy M ethod in Two Dim ensions
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 4. M U LTI-D IM EN SIO N AL C A U C H Y M ETHO D
Figure 4.10: Inverted Tree Representation for Recursive C auchy M ethod
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
51
CHAPTER 4. M U LTI-D IM ENSIO NAL C A U C H Y M ETHO D
52
W hen using the m ulti-dim ensional polynom ial expansion, the choice o f sam ple points is uncon­
strained in the com plete param eter space, and the sam ples can be picked by the sam e m ethods
as in the one-dim ensional case. This m eans, w e can use adaptive sam pling in the sam e way as it
was used in the one-dim ensional case.
4.4.2
Multi-Dimensional Cauchy Expansion
The m ulti-dim ensional m ethod can also be im plem ented by a single m ulti-dim ensional rational
polynom ial o f the form
\
S (Pi,P 2,P3,--0 =
where P nu m { P i , P 2 , P 3 ,
■ ■ •)
T>n u m { P l i P 2 , P 3 i ■■■)
s — ------------------------ r
P d e n ( P l , P 2 , P 3 , - ■■)
and Pden i ‘P \ >P 2 ■,P s ,
■•
,. l m
(4 -1 0 )
■) are arbitrary polynom ials o f the param eters
p i to p n in the num erator and denom inator, respectively. U sing this approach, the coefficients of
the function (4.10) are determ ined directly. S a k a ta [ 8 3 ] showed the extension into tw o dim en­
sions and the general scheme is briefly discussed here.
Eqn. 4.10 can be w ritten as
a(
,
QQ + E j i l d j P j ( p i , P2, • • • , Pn)
1 + £?=1 b j P j ( p i , P 2 , ■■■i P n )
w here P j ( p i , p 2, • ■■, p n ) are the m ixed term s o f the existing param eters. The general form of
the m ixed term elem ents is
Pj = t l P i U)
i= 0
( 44 2 )
where u ( j ) is an integer function o f j .
M uch attention m ust be paid to the choice o f the functions u ( j ) . The m ethod used here is to
get all m ixed term s o f all param eters p \ , . . . , p n up to a specified m axim um pow er and then sort
by power-sum s u sum = Y a =\ u i- Starting w ith the lowest pow er sum, the polynom ial is built.
U sing this scheme, the start o f the polynom ial for tw o dim ensions is
P { p i , P 2 ) = a 0 + a i p i + a 2p 2 + a 3p i p 2 + a4p l + a 5 p | + a 6p \ p 2 + . . .
(4.13)
This approach yields a linear equation system. Solving this system determ ines directly the co­
efficients o f eqn.(4.11) and, hence, a closed-form and differentiable equation o f the system ’s
response S ( p 1, p 2, p 3, •. ■)■
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 4. M U LTI-D IM EN SIO N AL C A U C H Y M ETHO D
53
As there are no restrictions whatsoever, a m ulti-dim ensional adaptive sam pling o f the param eter
space can be applied in analogy to the one-dim ensional case.
It should be m entioned that for large dim ensional problem s, it would be m ore efficient to split
the problem into several problem s o f low er order, which are then solved recursively as shown in
Section 4.4.1.
4.5
Numerical Results
Three exam ples are given here to dem onstrate the interpolation using m ulti-dim ensional Cauchy
method. All EM -sim ulations are perform ed using the Sonnet em planar electrom agnetic solver.
A ll CPU tim es given refer to com putations on a H ew lett Packard K-class machine.
4.5.1
Microstrip Line Impedance
The m ulti-dim ensional rational function expansion is dem onstrated by m odelling the line im pedance
o f a m icrostrip line with respect to the line’s width-to-height-ratio f and the relative dielectric
constant er o f the substrate.
The m odelling algorithm described in Section 4.4.2 returns the follow ing closed-form rational
function:
Z ( p x = er , p 2 = *jr) =
5 4 9 .6 2 3 2 + 1 1 7 . 0 0 7 8 p x + 9 7 9 9 . 3 8 3 9 p 2 1 .0 + 0 . 4 5 0 2 1 p ! + 3 6 .5 7 0 9 p 2 -
(4 -14)
7 . 4 3 6 7 ^ + 1 5 3 6 .0 6 5 2 p | + 2 9 9 .3 8 4 8 /> ip 2 + 0 .1 6 1 2 3 p f -
0 .0 1 9 4 6 p 2 + 1 2 . 5 5 1 9 p 2 + 3 6 . 3 2 6 3 0 p i p 2 + 7 . 9 0 6 E
-
5p f -
1 7 .9 0 0 3 2 5 p | + 3 6 3 .8 9 4 7 2 p fp 2
0 .1 1 6 0 5 p 3 + 1 7 .7 3 3 8 7 p 2 p 2 _ 4 . 6 6 4 5 p i p |
The sam ples are determ ined by an adaptive sam pling technique sim ilar to the m ethod used in
[82]. Figure 4.11 shows the m odel approxim ation and the 19 adaptively taken sam ples. The
m odel provides an accuracy within
4.5.2
0 .1 %
error.
Recessed Line Fed Microstrip Antenna
Heuristic m odels o f m icrowave circuits are restricted to very basic topologies such as lines,
gaps, step etc.. The m ulti-dim ensional Cauchy m ethod allows the creation o f fast and accurate
m odels for any kind o f topology, as long as a param eterised EM -sim ulation is feasible. Here,
the building o f a generic m odel for an recessed line fed m icrostrip antenna in C -band on 10
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 4. M U LTI-D IM EN SIO N AL C A U C H Y M ETHO D
Z[Q]
54
200
150
100
0 .9
0.8
0/7
S
0.6
0 .5
0.3 ^ w / h
Samples points
0.2
20
25
Figure 4.11: Im pedance o f M icrostrip L ine as a Function o f er and w / h using M ulti-D im ensional
Rational Function Expansion
m il-Rogers RT/duroid™ 5870 substrate (Figure 4.12) is discussed. An accurate m odel for such
elem ents does not exist, and up to now, the design is based on simplified cavity m odels [85] or
an expensive search by a vast num ber o f E M -sim ulations, which m ust be repeated whenever a
new slightly m odified design is requested.
Figure 4.12: M icrostrip A ntenna with Recessed line feed
The m ulti-dim ensional Cauchy M ethod overcom es these problems. A ll expensive EM -sim ulations
for the m icrostrip antenna are com puted at once. The problem has four dim ensions, the param ­
eters p i to pz and the frequency / . Five param eter values per dim ension are required. H ence,
the responses o f 5 4 = 625 sam ples have to be com puted for a com pletely filled grid o f sample
points. This data acquisition phase requires approxim ately 10 h o f C PU time. Even though this
initial effort is quite high, it pays off, as the established m odel can be used for a large variety o f
antenna layouts w ithout any additional expensive com putations.
The sim ulated geom etries fall on the follow ing grid:
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 4. M U LTI-D IM EN SIO N AL C A U C H Y M ETH O D
55
pi
=
17.5 m m ,
20.3 m m ,
P2
-
0.0 m m ,
2.8 m m ,
5.6 m m ,
8.4 mm,
11.2 m m
pz
=
2.8 m m ,
5.6 m m ,
8.4 m m ,
11.2 mm,
14 mm
/
=
3.4 GHz
3.7 GHz
4.0 GHz
4.3 GHz
4.6 GHz
2 3 .1 m m ,
25.9 mm,
28.7 m m
The frequency response for the m odel is verified at an unsam pled arbitrary geom etry with p \ =
24.5 mm, p 2 = 1.4 m m , pz= 4.2 mm. The response com puted by full E M sim ulation and by
the m odel is shown in Figure 4.13. As seen, the m odel is in very good agreem ent with the EM
sim ulated result.
a
o
-J
E M -Sim ulation
— M ulti-D Cauchy
E
3
(U
Ot
3 .6
3 .7
3 .8
3 .9
4
4 .1
4 .2
f [GHZ]
Figure 4.13: R eturn Loss o f A ntenna G eom etry p \ = 24.5 m m , p 2 = 1.4 m m , pz= 4.2 m m C om ­
puted by M ulti-D C auchy M odel and Full EM -Sim ulation
4.5.2.1
Parameter Plot of Feed Location
Using the m ulti-dim ensional Cauchy m ethod m odel, the antenna is optim ised for m inim al return
loss at 4 GHz. The effects o f the feed location on the return loss is shown by a param eter plot of
the return loss in Figure 4.14.
The param eter plots contain the inform ation o f 2000 frequency points. To obtain the sam e in­
form ation using EM -sim ulators, one has to perform a sim ulation that takes 33 C PU hours. O ur
Cauchy m odel com putes the sam e inform ation in less than 4 CPU seconds. It can clearly be
seen, that the established m odel enables the designer to investigate the circuit’s perform ance
with respect to m odifications w ithout any additional expensive calculation.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
56
CHAPTER 4. M U LTI-D IM EN SIO N AL C A U C H Y M ETHO D
O'
-5
..__= 1 8 .0 6 m m
p 3= 5 .1 8 m m
-10
-1 0
W -1 5
2.
3?-20
P2 =
\
o
c -2 5
1 2 .6 m m
S -3 0
11 .2 m m
^
/
/
9 ‘ 15
“C/5 -2 0
/
-40
-45
3 .9
p 2= 7 .0 m m
4 .2 m m
4 .3 4 m m
4 .4 8 m m
4 .6 2 m m
4 .7 6 m m
4 .9 m m
5 .0 4 m m
V u X ///
" 1.4 m m
e ’25
§ -3 0
9 .8 m m - '''" ”" ^ W —— —2 .8 m m
“ -35
" 4 .2 m m
-4 0
OS
-35
Pj = 18.06 m
8.4 m m
(M— "
- 5 .6 m m
7 .0 m m 3 .9 2 3 .9 4 3 .9 6 3 .9 8
4 .0 2 4 .0 4 4 .0 6 4 .0 8
5 .4 6 m m
5 .3 2 m m
5 .1 8 m m
-45
1
4
5 .6 m m
4.1
3 .9
3 .9 2 3 .9 4 3 .9 6 3 .9 8
4
4 .0 2 4 .0 4 4 .0 6 4 .0 8
4.1
f [G H z]
f [G H z]
Figure 4.14: E ffect o f Inset location on A ntenna’s R eturn Loss
4.5.3
Narrow-Band 3-Pole Filter
In this example, the S-param eters o f the response o f a planar superconductive m icrostrip filter as
shown in Figure 4.15 are interpolated. The five param eters are the four geom etrical param eters,
nam ely the gap length and resonator lengths, and the frequency. A sam ple grid with 5 sample
points per dim ension for the geom etrical param eter is used.
Figure 4.15: L ayout o f Three-pole Filter Showing Param eters p l to p4
D ue to the underlying lim itation o f the EM -sim ulation software, the param eter values are forced
to lie on a 1.75 mils grid. The frequency dependency is determ ined by adaptive sam pling, as the
last level o f the recursive algorithm . The frequency is chosen because it shows by far the largest
variation o f the S-param eter response.
Shown in Figure 4.16 is the interpolated response for the param eter values
G ap length p l
= 6.125
G ap length p2
= 16.625
Line length p3
= 23 8.875
Line length p 4
= 238.875
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 4. M U LTI-D IM EN SIO N AL C A U C H Y M ETHO D
51
In com parison w ith the exact solution— obtained by a finer m eshing and finer frequency stepping—
one notices the very good agreem ent o f the interpolated response with the exact solution.
In Figure 4.17, the response is shown when the m ulti-dim ensional Cauchy m ethod is not applied.
D ue to the restrictions that all geom etrical values have to fall on the 1.75 mils grid, the param eter
values m ust be snapped to the next grid-point. Figure 4.17 shows the response for both rounded
up and rounded down values for the param eters p \ to p^. In addition, the frequency resolution
is lost, as only the sam pled frequencies can be shown. A s a result, the tw o-ripple filter response
degrades to a m eaningless polygon.
4.5.3.1
Optimisation, Monte-Carlo Analysis
U sing the m odel, the filter is optim ised with respect to the specifications:
S n < -20 dB for 3.9 GHz < / < 3.95 GHz
S 21 < -20 dB for / = 3.85 GHz and / = 4.00 GHz
The optim isation by gradient search requires 19 iterations, which corresponds to 1140 function
calls. A full EM -sim ulation would require a C PU tim e o f at least several days, as the circuit’s re­
sponse m ust be repeatedly com puted on a very fine grid. The m ulti-dim ensional Cauchy M ethod
solves the problem within 3 CPU-m in.
Having the m odel on hand, the developer can check the sensitivity o f the m odel by a M onteC arlo analysis. Again, the analysis— known to be expensive or not feasible w hen EM -sim ulation
is used— can be applied with little effort using the predefined model. In Figure 4.19, a random
variation o f ±
0 .1
m ils is added to all geom etrical param eters.
N either optim isation nor M onte-C arlo analysis can be perform ed using direct EM -sim ulation
within reasonable com putation time. The num ber o f required points exceeds even the capabil­
ities of m odem m ulti-processor com puters. M oreover, the planar electrom agnetic solver is not
capable o f handling geom etries w ith sm all variations, as all vertices m ust fall on a relatively
rough grid. The m ulti-dim ensional Cauchy m ethod m odel overcom es these problem s. A fter
one expensive generation o f an on-grid database, all follow ing com putations o f arbitrary circuit
variations can be obtained by the inexpensive recursive algorithm.
4.6
Conclusion
In the past, m any publications have shown the rem arkable reduction o f com putational cost when
Cauchy m ethod and adaptive sam pling are applied to the frequency response interpolation.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 4. M U LTI-D IM ENSIO NAL C A U C H Y M ETHOD
58
5
S ll
[dB ]
°
■5
-10
-1 5
-20
-2 5
-3 0
M u ltid im en sion al C auchy
° E xact S olu tion
-3 5
-4 0
3 .7
3 .7 5
3 .8
3 .8 5
3 .9
3.9 5
4
4 .0 5
4.1
4 .1 5
4 .2
f [G H z]
Figure 4.16: Interpolated 5 n Param eter U sing M ulti-D im ensional Cauchy M ethod
S ll
[d B ]
-10
-15
S am p le b e lo w
-20
S am p le ab ove
-25
-30
-35
3.7
3 .7 5
3 .8
3 .8 5
3 .9
3 .9 5
4
4 .0 5
4.1
4 .1 5
4 .2
f [G H z]
F igure 4.17: S u Param eter W ithout U sing M ulti-D im ensional Cauchy M ethod Show ing the
Sam pled Points Above and Below
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 4. M U LTI-D IM EN SIO N AL C A U C H Y M ETHO D
59
-10
-35
*■ E M -Sim ulation
-40
— C auchy-M ethod
-45
3.85
3.9
3.95
4 .05
f [G H z]
Figure 4.18: R esponse o f O ptim ised Three-Pole Filter
This C hapter has shown that the m ethods can be extended and applied to m ulti-dim ensional
problem s. This can be either done by recursive application o f the Cauchy m ethod or by a m ulti­
dim ensional rational polynom ial expansion.
In doing so, sim ilar savings o f com putational expenses for m ulti-dim ensional problem s can be
achieved as for the one-dim ensional case. In the exam ples given, we showed that a com plete
and accurate num erical m odel o f a three-pole filter w ith 4 geom etrical param eter plus frequency
dependency can be obtained.
A s shown, we can apply the derived m odel to com plex operations, such as optim isation or
M onte-C arlo analysis. The operations are perform ed at very little com putational expense.
The m odelling o f m icrowave circuits using the m ulti-dim ensional C auchy m ethod is restricted
to a few (~ 5 ) param eters, though, because the num ber o f simulations increases exponentially
with the num ber o f param eters. Optim isation o f large circuits is not feasible with the m ethod
proposed in this chapter.
In the next chapter, we develop an optim isation schem e for large circuits. The algorithm uses
com plete and decom posed circuit sim ulations to determ ine the optim ised layout fast and accu­
rately. F or an extrem ely fast calculation o f the decom posed circuit response, the decom posed
sim ulation uses block models based on the m ulti-dim ensional Cauchy method.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 4. M U LTI-D IM EN SIO N AL C A U C H Y M ETHO D
-10
-3 0
-35
-4 0
-45
L
-5 0
3.8
3 .85
3.9
3.95
f [G H z]
4
4 .05
4.1
Figure 4.19: M onte-C arlo Analysis o f O ptim ised Three-Pole filter
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
Chapter 5
Hybrid Optimisation Using Coupling
Matrix Alteration
"There are only tw o w ays to live y o u r life. One is as though
nothing is a m iracle.
The o th er is as though everything is a
m iracle. ", Albert Einstein
The quality o f an iterative optim isation o f m icrowave circuits is determ ined by the underlying
circuit analysis m ethod. Full electrom agnetic sim ulation is the m ost desirable analysis m ethod
in the optim isation process as it attains the highest possible accuracy. Full E M sim ulation, how ­
ever, com es w ith very high com putational costs. It is w ell known that these costs can becom e
prohibitive for large circuits. Consequently, direct full EM optim isation o f large circuit is not
feasible with the current tools.
Large circuits can be usually optim ised by a sim plified circuit analysis only. The simplification
m ay decrease accuracy, but it m ay t be the only choice if optim isation is necessary.
Using
simplified circuit m odels for optim isation often yields an inaccurate but sufficient solution o f the
problem. The use o f the exact m odel can be restricted to the final verification o f the solution.
If w e require both the high accuracy o f the full E M m odel and the speed o f the sim plified m odel,
w e can com bine them. We call this a hybrid optim isation technique. In this chapter, w e discuss
existing hybrid optim isation techniques and their limitations. We also propose a new m ethod
especially tailored to an efficient m icrowave filter design. O ur technique is based on the coupled
resonator representation o f the filter.
61
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 5. H Y B R ID O PTIM ISATIO N USING COUPLING M A T R IX A LT E R A T IO N
5.1
62
Background
Any hybrid optim isation relies on tw o sim ulation models, a fine model and a coarse m odel. The
fine m odel is accurate but com putationally expensive, w hereas the coarse m odel is fast but less
accurate. The coarse m odel is used to optim ise the circuit, and the fine m odel is used to verify
o f the results and to correct the m odel based on the fast calculations. B oth m odels are essential
in the hybrid optim isation process. For planar circuit optim isation, as discussed here, the coarse
m odel is usually represented by a decom posed E M analysis or circuit theory analysis. The fine
m odel is norm ally represented by a com plete E M analysis. A detailed outline o f the tw o m ethods
is given in C hapter 2.
H ybrid techniques generally w ork in tw o phases: a training phase and an application phase.
D uring the training phase, w e collect data about the coarse and the fine m odel for a few specific
circuit geom etries. B ased on this data, we derive an algorithm which adjusts the coarse model.
O ur goal in the training phase is to derive an adjusted coarse m odel w hich m im ics the fine m odel
very closely. In the application phase, we use this corrected coarse m odel to inexpensively
com pute new, unknown circuit geometries.
In the next section w e discuss some hybrid optim isation techniques and their lim itations. Later
in Section 5.2 we will introduce a new hybrid optim isation scheme. O ur technique is particularly
tailored towards filter optimisation. The proposed approach uses a coupled resonator filter m odel
for correcting the coarse m odel.
We could apply the correction o f the coarse m odel directly to the m odel’s output param eter, i.e.,
the S-param eter response. This approach, however, is not successful since the corrected m odels
do not obey the law o f conservation o f com plex power. In this section, w e outline why.
L et us define the S-param eter response o f the coarse m odel as the m atrix [S]c . T he response of
the fine m odel is defined as the m atrix [S]f. B oth m atrices are functions o f the circuit param eter
vector p. The vector p represents all param eterised layout features, e.g. line width and substrate
thickness. Now, w e define a difference S-m atrix [A S ^ p ) by
[A S ](p) = [S]f (p) - [S]c(p).
(5.1)
T his m atrix describes the differences betw een the fine m odel’s response and coarse m odel’s
response. Since w e investigate frequency responses, the difference m atrix is actually an array of
m atrices corresponding to a list o f frequency points, i.e.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 5. H Y B R ID OPTIM ISATIO N USING COUPLING M A T R IX A L T E R A T IO N
([A S]) = ( [ A S K /i), [A S K / 2 ), [A S ]( /3) , . . . )
63
(5.2)
For convenience, w e refer to this array sim ply as the m atrix [A S], even though it is really a list
o f matrices.
Now, we perform one coarse m odel sim ulation and one fine m odel sim ulation each with the
arbitrary param eter start vector po. Then, w e calculate [A S ](jJq). Next, w e w ant to optim ise
[S]f with respect to p. Since the calculation o f [S}j(p) is expensive, the optim iser uses the
coarse m odel. To im prove the coarse m odel’s accuracy, w e adjust its response [S]c(p) to obtain
a new S-m atrix
[S]*(p) = [5]c(p) + [ A S ] ( f t ) .
(5.3)
In other words, the difference determ ined at po is added to all subsequent calculations o f [S]c.
W e can divide hybrid optim isation algorithm s into tw o phases: a training phase and an appli­
cation phase. D uring the training phase we determ ine a [AS] from a coarse and fine model
simulation. D uring the application phase, w e use the coarse m odel sim ulation exclusively and
adjust its response by [AS]. The process is sum m arised is in Figure 5.1.
Training:
Application:
F in e M o d el
S im u la tio n
l.\S|
Simulation
[AS]
Figure 5.1: Training and Application Phase s
In the case o f p = po, i.e., the layout used for training, the corrected m atrix [S']* is identical to the
fine m odel’s m atrix [S]j by definition. However, if p A Po the adjusted m atrix [S]* represents
an approxim ation o f [ S] f only. The corrected m atrix [S]*.(-p) and the accurate response [S'] / (p)
are equal only if the difference m atrix [AS] is not a function o f p.
F or real problem s, [AS] is always a function o f p. The question is w hether it is a slowly changing
function. If yes the the approxim ation is likely to approxim ate the fine m odel well. If no the
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 5. H Y B R ID OPTIM ISATIO N USING COUPLING M A T R IX A L T E R A T IO N
64
approxim ation is not valid. The answ er to this question depends on the problem. In general,
however, [AS] changes rapidly with p. A typical exam ple is outlined next.
L et us calculate [AS] for the tw o-pole filter given in Figure 2.3. The coarse m odel is given by
the circuit theory sim ulation, and the fine m odel is given by the full-em simulation.
The difference m atrix [AS] is calculated for tw o slightly different geom etries. The first layout,
w ith the param eter vector p \ , is the layout discussed in Figure 2.3. The second layout, with
the param eter vector p 2 , is identical except for a slightly longer resonator line. As a result, the
response S ( p 2 ) is shifted downwards in frequency w ith respect to S (p \). The tw o responses in
logarithm ic scale are shown in Figure 5.2.
f r eq
freq
f r eq
freq
T racel-dBtSU .tJ)
Tricel-dBCSC1,1])
Trace2»dB(FULLEM_DESIGN_AFS. . SC I, 13)
Tr 8 ce 2 *d B( DE SI GN _s hi f te d _f ul lent AFS. . S C I , 1])
Figure 5.2: R esponse o f C ircuit L ayout p \ and j72■
Now, let us take a look at the difference m atrix [A S]. The real and im aginary parts o f A S n
for both layouts are shown in Figure 5.3. It can be seen, that A 5 n (pi) and A S u (]>>) are quite
different.
Consequently, a correction as proposed by eqn. (5.3) fails because [AS] changes dram atically
when we go from p \ to p ‘2 - Figure 5.4 shows [S*}(p 2 ) when we use [AS] ( p \ ) for correction, i.e.
[S*}(P2) = [Sc](p2) + [ A S M ) •
(5-4)
The corrected coarse m odel response S * u (p 2 ) is obviously not a good approxim ation o f the
a ccu ra te fin e m o d e l r e s p o n s e [S]^(p 2 )- E v e n w o r s e , S * 11( p 2 ) e x c e e d s 0 d B in d ic a tin g n e g a tiv e
resistance, hence, violating the unitary condition.
In conclusion, the fixed S-param eter correction does not w ork properly, even for a sim ple exam ­
ple as shown above. The application o f this algorithm to the Y-, Z-, and A B C D -m atrix, generally
results in the sam e failure.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 5. H Y B R ID O PTIM ISATIO N USING COUPLING M A T R IX A LT E R A TIO N
R eal-P art of Delta-511 of
L ay o ut LI and L a y o ut L2
65
Imag-Part of D e l t a - S l l of
Layouy LI and L ay o ut L2
TraceB
Trace?
2.0
2.0
« BQ
.
*
A
’
j
/
-2.0
-2.0
V_
fti
t
■\
GHz
freq
freq
3.9
3.9
Trace5-Real(OESIGN_orig..SCI,m-Real(DESiGN_orig..SC3,3])
TraceG»real(design_shifted..SC I , U)-Real(design_shifted..SC3,33)
GHz
GHz
freq
freq
4.1
4.1
GHzB
GHz A
Trace8*Ii*ag(DESIGN_orig..SCI,13)-Inag(DESlGN_orig..S[3,31)
Trace7*Imag(deiign_shifted..S[l,13)-Iinag(design_shifted..SC3,3])
Figure 5.3: R eal Part o f A 5 n ( f ) for L ayout p\ and L ayout p 2 (Left), and Im aginary Part o f
A S ' n ( f ) for L ayout p i and L ayout p 2 (Right)
freq
diff=design_orig.siml.SP.S[3,3]-des
adj_ shifte ds design_shifte6.siml.SP.!
Trace3=dB(adj_shifted)
Figure 5.4: A djusted S-Param eter S * n ( f ) for L ayout p 2 U sing
[ A 5 ] (pi)
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 5. H Y B R ID O PTIM ISATIO N USING COUPLING M A T R IX A L T E R A T IO N
5.1.1
66
Space Mapping
From the responses shown in Figure 5.2 we observe that the responses are very similar, except
for a slight frequency shift. It m ay be possible, to adjust the response o f the coarse model
by simply rescaling the frequency axis. L et us assum e the frequency shift betw een [Sjc and
[S]f is constant or nearly constant for different geom etry param eters. Then, we can correct the
coarse m odel response by rescaling the frequency variable. This approach changes the correction
focus from the output (S-param eter) to the input (frequency). We can show that the adjustm ents
can also be applied to other input variables, nam ely the param eter vector p. The system atic
application o f this m ethod, called space m apping, was first suggested by
B a n d le r
[72]. An
outline o f the space m apping algorithm is given in A ppendix A. The training and application
phase of the space m apping algorithm is discussed in Figure 5.5.
Training:
Application:
C o arse M odel
S im u latio n
Figure 5.5: Training and Application U sing Space M apping
Space m apping always yields results com pliant to the unitary condition. Since we change the
input parameter, the derived response m ust conform with the unitary condition as long as the
param eter values are physical m eaningful, e.g. non-negative line-width.
Stability and robustness o f the space m apping algorithm is questionable . A global m apping as
described in A ppendix A.3 cannot be found in all cases. The param eter correction is a m athe­
m atical construct only, with little or no physical m eaning. The m apping is always local in the
param eter space. Thus, strong derivation betw een the fine and coarse m odel can yield to a m ap­
ping legitim ate at the m apped point 'p*c but not at the desired optimal point j7?.pt. This m apping
is said to be local. These problem s can often be circum vented with m ore iterations and sophis­
ticated algorithm s for deriving the m apping function as suggested in [73] w ith the thrust-region
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CH APTER 5. H Y B R ID O PTIM ISATIO N USING COUPLING M A T R IX A L T E R A T IO N
67
algorithm .
M ore im portant is the fact that the param eter adjustm ents— as done by space m apping— have
fundam ental limitations. The control o f the coarse m odel response is restricted to m anipulations
w hich can be controlled by the m odel input param eters. Certain electrom agnetic phenom ena de­
tected in the fine m odel only— and thus not accessible in the coarse m odel— cannot be integrated
in a param eter space m apping. F or instance, m any coarse m odels neglect radiation effects and
stray coupling etc. in their derivation. These effects cannot be adjusted or included by tw eaking
the accessible param eter vector p.
A n exam ple o f this effect is given by a typical three-pole filter described in Figure 5.6. We
sim ulate the filter as a com plete circuit (fine m odel) and by decom position (coarse m odel). F ig­
ure 5.7 shows the sim ulation results. It can be clearly seen, that the filter response is quite
different betw een com plete and cascaded sim ulation. M ost notable is the transm ission zero at
/ = 3.96 GHz. W e see it as a dip-down in the logarithm ic S 2 1 response. Finite transm ission
zeros are only possible for elliptic filter transfer functions. Thus, the zero m ust be caused by a
stray coupling betw een the first and third resonator. This coupling is sketched in Figure 5.6.
3
Figure 5.6: Three-Pole Filter, a) C ascaded (Decom posed) Layout, b) C om plete Layout
The decom posed m odel does not allow param eter adjustm ents which can excite this stray cou­
pling, because there is no physical connection betw een the first and last resonator in the cascaded
m odel. Consequently, space m apping is not feasible for resonator coupled structures o f this kind.
5.2
M-Parameter Correction
In order to include all couplings into the coarse m odel, we switch to the coupled resonator filter
representation. This representation describes a resonant-coupled filter com pletely including all
cross-couplings.
Cross couplings are defined as couplings between non-adjacent resonators.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 5. H Y B R ID O PTIM ISATIO N USING COUPLING M A T R IX A L T E R A T IO N
68
0
-10
-10
-20
c2o
CM
-20
-3 0
-4 0
-3 0
-5 0
<n
-4 0
-6 0
-7 0
-5 0
-8 0
-6 0
-9 0
3 .9 3 .9 2 3 .9 4 3 .9 6 3 .9 8 4 4 .0 2 4 .0 4 4 .0 6 4 .0 8 4.1
f [GHz]
3 .9 3 .9 2 3 .9 4 3 .9 6 3 .9 8 4 4 .0 2 4 .0 4 4 .0 6 4 .0 8 4.1
f [GHz]
Figure 5.7: C ascaded and C om plete Sim ulation R esults o f Three-Pole Filter. Clearly Seen in
the Com plete Sim ulation is a Transm ission Zero at f=3.96 GHz Due to Stray C oupling B etween
Resonator O ne and Three.
Figure 5.8 illustrates the description o f a filter using direct coupled resonators. The theory of
direct coupled resonant structures is developed in detail in [30], [31], and [32].
Resonator 1
Input
Coupling
Resonator 2
Coupling
Resonator 3
Output
Coupling
Coupling
Resonator 4
Resonator 5
Coupling
lie
Figure 5.8: D irect C oupled Resonators
A brief review o f the coupled resonator filters is given in the next section. In Section 5.2.3, we
will show how we can extend the theory to a new hybrid optimisation scheme.
5.2.1
Coupling Matrix Representation
Figure 5.9 shows a m ultiple inductive coupled iV-resonator filter structure. The resonators are
assum ed identical, i.e. all resonances are at the sam e frequency
m alised im pedance Z
ujq
—
and with the nor­
— 112. The inductive coupling between the ith and j t h resonator
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 5. H Y B R ID O PTIM ISATIO N USING COUPLING M A T R IX A L T E R A TIO N
is denoted
69
— w hich are real num bers1— and assum ed to be frequency independent2. The
resonators can include loss denoted by an ohm ic resistor R n . Further we denote the source and
load im pedance as i ^ a n d R j y respectively.
M
M
12
M
23
M
34
N -2 .N -1
M
n
_ i >n
Figure 5.9: Inductive C oupled Resonators
The loop equations can be w ritten as
h (R i + j{uL —j i \ M \ 2 +
*2
- j i 2u M i 2
( R 2 + j(uL —
jiNMiN
=
e%
— j i2wM23 — ■■■ — j i N M 2N
—
0
=
0
=
0
:
:
:
—j h M N i — j i 2LoMN2 — - ' - + j i N M i N
w here M mn = M nm for all couplings.
U sing the above norm alisations for
ujq
=
and Z q =
= 1 we can use the following
abbreviation
j { u) L - - ^ ) + R — j Z o
w here A = Z q
cu
u)o
U)o
U)
+ Rn — j A + Rn ,
(5.5)
. N o w , o u r lo o p e q u a tio n s ca n b e w r itte n in m a trix n o ta tio n
[[R\ + y(A[7] - u[M})} i = [Z] i = e
'Appendix B discusses w hy w e can use pure real coupling values
2Frequency dependence can be included by a dispersion factor
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
(5.6)
CHAPTER 5. H Y B R ID O PTIM ISATIO N USING COUPLING M A T R IX A LT E R A TIO N
70
where
( Ri
0
0
R2
0
0
•••
■■■
0
0
0
0
\
(5.7)
[ R ]
V o
Rn
o
Rn
J
and
( Mu
M 2\
M\2
M\n
22
M 2n
M 2n
M nn
M nN
m
■
M
in
[M} =
(5.8)
M ni
\ M
n
i
M
M
n2
■■■ M
nti
■■■ M N n }
and [/] is the identity m atrix and e is the voltage excitation vector, i.e.
( ei \
e =
(5.9)
V o
T h e main diagonal elem ents M nn are zero in the tuned case, i.e. all resonators have the sam e
resonant frequency
loq.
The coupling m atrix com pletely describes the electrical behaviour of
the filter. Thus, it is possible to represent any resonant band-pass lum ped elem ent filter by its
coupling m atrix. Since distributed elem ent filters can be approxim ated with lum ped elem ent
resonators using R ichard’s transform (see[27]), we can also represent those filters by their cou­
pling matrix. R ichard’s transform , however, is only valid for relatively small band-w idths (10%
to 20%). Thus, the coupling m atrix representation is, strictly speaking, only valid for lum ped
elem ent band-pass filters and narrow -band distributed elem ent band-pass filter.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 5. H Y B R ID O PTIM ISATIO N USING COUPLING M A T R IX A L T E R A T IO N
71
We can com plem ent the coupling m atrix representation with a dual circuit. A detailed outline of
this duality is discussed in A ppendix B .l.
5.2.2
Coupling Matrix o f Three-Pole Filter
We can use the coupling m atrix representation to describe the three-pole filters from the last
section in Figure 5.6. The coupling m atrix o f the cascaded filter is
/
[M\c =
V
.0 0
.82
.0 0
.82
.0 0
.82
.0 0
.82
.0 0
\
(5.10)
The self couplings are zero, i.e. the resonators are tuned exactly to the frequency luq. The m utual
coupling M
\2
and M 2 3 are different from zero, i.e. adjacent resonators are coupled. There is no
coupling betw een resonator one and three, as seen from M
= M 31 =
13
0
.
The coupling m atrix o f the com plete full EM model
-.2 0
.83
.83
1
\
.0 0
(5.11)
0
.83
[M]f
-.2 0
to
.83
-.2 0
is different. We observe now a coupling betw een resonator one and three, corresponding to the
stray coupling. D ue to the m utual inductance
M
13
the self coupling o f M \ \ and
M
33
changes as
well. D etails about the reasons are given in A ppendix B.
M ost notable is the difference
[A M ] = [ M \ f - [M ]c =
V
.20
- .0 1
.20
- .0 1
.00
-.0 1
.20
- .0 1
(5.12)
.20 /
which contains all stray couplings not considered by the coarse m odel, but included in the fine
model. H ence, we can use this difference m atrix [A M ] for correcting a coarse m odel in a hybrid
optimisation. The detailed algorithm is presented in the next section.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 5. H Y B R ID O PTIM ISATIO N USING COUPLING M A T R IX A L T E R A T IO N
5.2.3
72
M-Parameter Corrected Optimisation
A s illustrated by the exam ple above, the difference coupling m atrix representation can m odel
additional couplings introduced by upgrading from a cascaded analysis towards a full-em anal­
ysis. Consequently, we can use the difference m atrix [A M ] for correction o f a cascaded circuit
analysis. The exact procedure is outlined below.
L et us define the response o f the cascaded analysis with the layout param eter vector p as [S}c (p)
and the response o f the Full-EM analysis as [S]f(p). From the response w e can derive the
respective coupling m atrices [M ] c(p) and [ M] f (p ). The difference m atrix o f M is then defined
as
[A M ] ( p ) = [M]f (p) - [M]c(p)
(5.13)
The m atrix [A M ] (p) contains all additional couplings not included in the cascaded analysis plus
corrections o f all other couplings. We can use [A M ] (p) to adjust the coarse m odel response at
any param eter vector p x such that the adjusted response
[M]*c(Px) = [M}c{p) + [A M ] ( p )
(5.14)
m odels the real (full-EM ) circuit m uch closer. Since [A M ] ( p ) is constant for small changes o f
the geometry, the same [A M ] (p) can be used to correct the coarse m odel response for any layout
containing sm all geom etrical changes. In the exam ple below, we will see that [A M ] is constant
with respect to p.
Sim ilar to the other hybrid m ethods, our m ethod requires a training phase to acquire [A M ]. The
subsequent application phase em ploys the difference m atrix for correction o f the coarse model.
The procedure is outlined by Figure 5.10.
Training:
Application:
Fine
— |S |P—
» | M ] , ------
;r
-
—
|S ] C
1M] ,— < 3— (M i; — IS) •
M |c ------
Figure 5.10: Training and Application o f Coupling M atrix Correction
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 5. H Y B R ID O PTIM ISATIO N USING COUPLING M A T R IX A L T E R A T IO N
73
T he proposed technique can be incorporated into a hybrid optimisation scheme. R ather than
calculating the accurate response in the loop using direct EM simulation (the fine m odel), we
derive the response from the adjusted coarse m odel. The m atrix [A M ], required for adjustm ent,
is derived in the training phase from just one response o f one full EM sim ulation and the cor­
responding coarse m odel analysis. This techniques offers a high accuracy in the optim isation
w ithout the costs associated with direct EM simulation.
F or the coarse m odel w e usually em ploy a decom posed circuit analysis. F or even m ore effi­
ciency, w e m odel here the sub-circuits o f the decom posed circuit using our m ulti-dim ensional
Cauchy m ethod presented in C hapter 4. In doing so, w e take all expensive calculations out o f the
optim isation loop. T he circuit analysis within the optim isation becomes ju st a m atter o f database
requests and m inim al algebraic com putations. The flow o f the proposed optim isation schem e is
outlined in Figure 5.11.
( Start ^
Parameter
M odifications
C ascaded Circuit
A n alysis
by M -C auchy
Interpolation
C ascaded Circuit
A n alysis
Full-EM Circuit
A n alysis
A djustm ents
using [AM]
R equirem ents
S atisfied ? /
,yes
F ull-EM Circuit
A n alysis
R equirem ents
. S atisfied ? /
-yes
F igure 5.11: Optim isation Including Correction o f Cascaded R esponse
Note, that at the end o f the optim isation process the perform ance o f the optim ised structure is
checked by another single Full-EM simulation. If the perform ance passes the specification the
synthesis is com plete. If the check fails, the correction [A M ] is not valid for the new layout,
i.e. [A M ] is a function o f p. In that case, a new correction [A M ] is calculated using the already
known responses o f the full EM and cascaded sim ulations. The optim isation process is repeated
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 5. H Y B R ID O PTIM ISATIO N USING COUPLING M A T R IX A LT E R A T IO N
74
using the new [A M ] .
5.3
Numerical Example
O ur novel optim isation technique is dem onstrated in the design of a
6
-pole narrow -band filter.
T he filter— part o f a superconductive satellite com m unication system [74]— is realised as a pla­
nar thin-film T B C C O 3 m icrostrip circuit on a L aA 1 0 3 -substrate. Specifications for the filter are
given in Table 5.1.
C entre Frequency
4.5 GHz
3dB -B andw idth
Pass band return loss
Stop bands
Stop band insertion loss
100 M H z
<-17 dB
f<4.44 GHz and f>4.56 GHz
<-30 dB
Table 5.1: Specifications for Narrow B and HTS Filter
The initial filter layout is shown in Figure 5.12. We em ploy a com plete full EM sim ulation using
Sonnet em [150] as the fine m odel. A decom posed circuit analysis as shown in Figure 5.13 is
used for the coarse model. We m odel each sub-section o f the circuit using the m ulti-dim ensional
Cauchy method.
Figure 5.12: Layout o f Six-pole Filter
The optimisation o f the filter breaks down into
6
steps:
1. Param eterisation and sim ulation o f the sub-circuit
2. Optim isation o f the coarse m odel
3 Thallium-Barium-Calcium-Copper-Oxide,
a high-temperature superconductor, see [139][140]
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 5. H Y B R ID O PTIM ISATIO N USING COUPLING M A T R IX A L T E R A T IO N
75
Figure 5.13: Cascaded Six-pole Filter
3. Com parison o f the coarse m odel’s response with the fine m odel’s response
4. Deriving the correction m atrix [A M ]
5. O ptim isation w ith corrected coarse m odel using [A M ]
6
. Verify validity o f optim ised circuit.
Step 1: Parameterisation and Simulation of the Sub-Circuit
Each resonator in the filter is described by one sub-circuit. The param eters o f each resonator are
the resonator length and the width o f the tw o coupling gaps as seen in Figure 5.14. We build a
database o f the responses o f a sub-circuit for the three geom etrical param eters plus the frequency
dependence as described in Section 4. The param eter values given here are always relative to the
initial layout. A draw ing o f the initial layout can be found in A ppendix D.
The database covers a sufficiently large range o f all param eter values using 5 sam ples per di­
mension. The com putation o f the database takes 2.4 hours o f C PU tim e on an H P-700 Series
workstation. Once the database is built the C auchy m ethod developed in C hapter 4 is then ex­
clusively used to m odel the circuit.
Figure 5.14: Definition o f Variable Param eters in Single R esonator
Step 2: Optimisation of Cascaded Circuit
The com plete
6
-pole filter is com posed o f six sub-circuits. Since the filter is sym m etric about
the y-axis, we deal with three different resonators with three param eters each. In total, the
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CH APTER 5. H Y B R ID O PTIM ISATIO N USING COUPLING M A T R IX A L T E R A T IO N
76
optim isation o f the filter requires the optim isation o f nine param eters. The raw filter layout
w ithout any optim isation, i.e. param eter vector p geoi = ( 0 , 0 , 0 , 0 , 0 , 0 , 0 , 0 , 0 ), has a response
as shown in Figure 5.15.
[dB]
-10
-1 5
-20
-2 5
-3 5
-4 0
' 5 ^ .3 8
4 .4
4 .4 2
4 .4 4
f [GHz]
Figure 5.15: R esponse o f Initial Filter Layout
The goal o f the optim isation is to design the filter such that it satisfies the specification as defined
in Table 5.1. The filter is optim ised using Genetic algorithm search and m inim ax gradient-search
com bined. F or now, optim isation is only perform ed on the coarse m odel, i.e. the cascaded
circuit. The optim ised filter response is shown in Figure 5.16. We define the param eter vector of
the layout that generates the optim ised response shown in Figure 5.16 as p geo2 . The num erical
value is
Pgeo2 = (2.62,1.02, - 5 . 4 , - 3 .9 8 , - 4 .8 3 , - 2 .9 9 , - 1 2 .3 , - 8 .2 9 , - 2 .1 )
(5.15)
A ll num bers are in mils. The optim ised layout is not exact, however, as we used the coarse
m odel for the optimisation. In the next step, we will com pare the coarse m odel’s response with
a response calculated by the fine model.
Step 3: Comparison with Full-EM
In order to com pare the coarse m odel’s results w ith fine m odel’s results we have to sim ulate the
sam e layout using both models. D ue to lim itations o f the full E M simulator, however, we cannot
com pare arbitrary layouts. The reason behind this is the rigid layout grid used by the full EM
simulator. A ll vertices o f the layout m ust fall on a grid o f fixed pitch as explained in Section
2.3.2. Hence, w e cannot check the layout p g e o 2 derived above.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 5. H Y B R ID O PTIM ISATIO N USING COUPLING M A T R IX A L T E R A T IO N
11
[dB]
-10
-15
-20
-25
-30
-35
-40
-45
-50,
4 .3 8
4.4
4 .6
4.5
f [GHz]
Figure 5.16: R esponse o f O ptim ised Filter
Instead, the circuit’s vertices m ust be snapped to the nearest grid point. The sim ulation is per­
form ed on this snapped-to-grid layout. In doing so, w e can perform the sim ulation for the coarse
m odel as w ell as the fine m odel. The snapped layout is defined as p geo3 . Since the snap grid is
0.2875 m ils the param eter vector becom es
pgeo-i = (2 .5 8 7 5 ,1 .1 5 , - 5 .4 6 2 5 , - 4 .0 2 5 , - 4 .8 8 7 5 , - 2 .8 7 5 , - 1 2 .3 6 2 5 , - 8 .3 3 7 5 , - 2 .0 1 2 5 )
(5.16)
The responses [S ] c ( p geo3) and [S]r(P[iro:i) are shown in Figure 5.18. N ot surprising, the re­
sponses, for the cascaded and full-EM analysis respectively, are different, because the full-EM
analysis includes additional couplings. The additional couplings are shown in Figure 5.17.
The conclusion is, that the circuit needs further optim isation using a m ore accurate m odel in
order to m eet the specifications. For an accurate and rapid optim isation we now switch to our
hybrid technique.
Step 4: Finding the Correction [AM]
O ur hybrid optim isation technique calls for the difference coupling m atrix derived [A M ]. This
m atrix is derived from the coupling m atrices o f the coarse m odel’s response and the fine m odel’s
response. In Step 3 w e already derived a response o f the coarse and the fine m odel, that is the
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 5. H Y B R ID O PTIM ISATIO N USING COUPLING M A T R IX A LT E R A T IO N
78
M ,
Figure 5.17: A dditional Couplings in Full-EM Sim ulation
0
[dB]
-10
-20
-30
-40
-50
I____________t ____________I____________ I____________I____________ I____________ I____________ I____________ I____________I____________ I____________ I
4 .3 8
4 .4
4 .4 2
4 .4 4
4 .4 6
4 .4 8
4 .5
4 .5 2
4 .5 4
4 .5 6
4 .5 8
4 .6
f [GHz]
Figure 5.18: R esponse o f (Grid-Snapped) F ilter L ayout p g e o 3 using Fine M odel and C oarse
M odel
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 5. H Y B R ID O PTIM ISATIO N USING COUPLING M A T R IX A LT E R A T IO N
79
response o f layout p geo3 - From the responses [S]c(pgeo3 ) and [ S] j(pge„?,) w e can calculate the
respective coupling m atrices [M]c(pgeo3 ) and [ M ] f ( p geo3 ). The two m atrices are
/
- .0 0 9 6 4
.78917
.00000
.78917
.06138
.58990
.00000
.00000
.00000
.00000
.00000
.00000
.58990
.04509
.56839
.00000
.00000
.00000
.00000
.00000
.00000
.56839
.04756
.59129
.00000
.00000
.59129
.05796
.79165
.00000
.79165
- .0 1 1 3 8 /
V
.00000
.00000
.00000
.00000
\
and
/
[M
] f ( p geo3
) —
V
\
.13423
.84662
- .0 0 7 8 1
.00000
.00054
.00000
.84662
.25104
.65190
- .0 2 0 9 0
- .0 0 7 0 7
.00002
- .0 0 7 8 1
.65190
.25251
.63468
- .0 2 1 7 2
.00000
.00000
- .0 2 0 9 0
.63468
.26261
.65203
-.0 2 0 0 1
.00054
- .0 0 7 0 7
- .0 2 1 7 2
.65203
.26221
.85066
.00000
.00002
.00000
-.02001
.85066
.17549 /
( 5 . 18)
The difference o f the above coupling m atrices is
/
[ / \ M ] ( p geo
3) —
V
.14387
.05745
- .0 0 7 8 1
.00000
.00054
.05745
.18966
.06199
- .0 2 0 9 0
- .0 0 7 0 7
- .0 0 7 8 1
.06199
.20742
.06629
- .0 2 1 7 2
.00000
- .0 2 0 9 0
.06629
.21505
.06074
.00000
.00002
.00000
-.02001
.00054
- .0 0 7 0 7
- .0 2 1 7 2
.06074
.20425
.05901
.00000
.00002
.00000
-.02001
.05901
.18687 /
\
( 5 . 19 )
This m atrix represents all additional couplings detected by the full-EM sim ulation, but not con­
sidered in the cascaded sim ulation. W e note non-zero couplings of non-adjacent resonators due
to stray coupling. Further, the values o f direct couplings (i.e. M ig+\ ) and self couplings change
significantly. These changes are the result o f m any effects, e.g. stray coupling, different de­
em bedding, and changes in the housing wall locations. We do not have to know the exact cause
o f these extra couplings. It is im portant only, that these additional couplings are accounted for
in the matrix [A M ],
We use the m atrix [ A M ] ( p ge03 ) to include the additional couplings into the coarse m odel sim u­
lation. The corrected coarse m odel m atrix becom es
[M]*c = [M]c + [A M ] (pgeoZ)
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
(5.20)
CHAPTER 5. H Y B R ID O PTIM ISATIO N USING COUPLING M A T R IX A LT E R A T IO N
80
O ur first application o f the corrected m odel is the verification o f layout p geo2 - This layout is the
optim ised layout found by the coarse m odel optim isation w hich was not snapped to the grid.
The response o f the corrected m odel in com parison with the uncorrected m odel, is shown in
Figure 5.19. Clearly seen, the response w ith the correction— assum ed to be accurate— violates
the specification. H ence, the circuit needs further optimisation.
[dB]
-10
-1 5
-20
-2 5
-3 0
-3 5
-4 0
C ascad ed resp o n se
C orrected resp o n se
-4 5
-5 0
4 .3 8
4 .4
4 .4 2
4 .4 4
4 .4 6
4 .4 8
4 .5
4 .5 2
4 .5 4
4 .5 6
4 .5 8
4 .6
f[GHz]
Figure 5.19: R esponse o f Filter L ayout p g e o 2 using C orrected Coarse M odel and U ncorrected
Coarse M odel. M arked in Gray is the Forbidden A rea o f the Specifications.
Step 5: Repeat Optimisation using Cascaded Corrected Response
The optim isation is repeated on the cascaded response including the correction algorithm . The
layout o f the optim ised circuit using the corrected coarse m odel is called p geo4 . The response o f
this geom etry is shown in Figure 5.20. This response is a very good approxim ation o f the real
response, since we included all stray couplings in our m odel. Still, w e have to verify that the
response calculated by the corrected coarse m odel is accurate.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 5. H Y B R ID O PTIM ISATIO N USING COUPLING M A T R IX A LT E R A T IO N
81
0
[dB]
-5
-10
-15
-20
-25
-30
-35
-40
-45
'5 ^ .3 8
4 .4
4 .4 2
4 .4 4
4 .4 6
4 .4 8
4 .5
4 .5 2
4 .5 4
4 .5 6
4 .5 8
4 .6
f[GHz]
Figure 5.20: R esponse o f F ilter W ith L ayout p ge0A O ptim ised U sing C orrected C oarse m odel
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 5. H Y B R ID O PTIM ISATIO N USING COUPLING M A T R IX A L T E R A T IO N
82
Step 6: Verify Validity of Optimised Circuit
The quality o f the corrected coarse m odel calculations for filter layout o f p g e o 4 is validated by a
full-EM sim ulation. Again, due to the nature o f the discrete layout grid, the geom etry vertices
have to be snapped to the next grid point, yielding layout p geo5 - The response o f the corrected
coarse m odel using layout p g e o 5 is com pared with the full-EM counterpart. The com parison is
shown in Figure 5.21.
[dB]
-1 0
-20
-25
-30
-35
F u ll-E M
-4 0
Cascaded + Corrected
-45
-50,
4 .3 8
4 .5
f[GHz]
F igure 5.21: Com parison: R esponse o f Filter by C ascaded Corrected Sim ulation and Full-EM
Sim ulation at N ext G rid-Snap C alled L ayout p g e o 5
The graph shows an excellent agreem ent o f the corrected coarse m odel’s response with the full
E M m odel’s response. W hen w e assum e the full E M sim ulation is exact, the filter layout p g e o 4
m ust be very closely to the exact solution. The optim ised filter is verified by hardw are m easure­
ments. The reader m ay refer to C hapter 7 for the results o f the m easurem ent.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 5. H Y B R ID O PTIM ISATIO N USING COUPLING M A T R IX A L T E R A T IO N
5.4
83
Comparison of Computation Time
T his section gives an overview on the com putation tim es required using different optim isation
approaches. The benchm arks are perform ed on an H P-U X 733 workstation. All tim es given are
CPU-tim es.
A com plete electrom agnetic sim ulation (frequency sweep) o f the layout of Figure 5.12 requires
22,080 C PU -seconds
(6
h,
8
min). The sim ulation o f a single resonator (Figure 5.14), on the
other hand, requires only 70 s for a com plete frequency sweeps.
The database for the sub-circuit response includes five values for each param eter dim ension.
H ence for building the database, we require 5 3 sim ulations o f the resonator layouts. This deter­
m ines the database generation tim e to be 8750 s (2.4 h). A sim ulation o f the com plete cascaded
circuit using the database and Cauchy m ethod takes 3 s.
The optim isation o f the circuit requires 15 iterations using first-derivative gradient search. The
tim e required for the com plete standard gradient optim isation process can be sum m arised in the
following formula,
I = tsingle ' (1 3“ ftp a r ) ’ f^freq ' ^ i t e
(5.21)
where t Singie is the sim ulation tim e for a single circuit at a single point, n par is the num ber o f pa­
rameters, n f req is the num ber o f frequency sam ples required, and n u e is the num ber o f iterations
necessary to find an optim al solution. In this form ula, n lte is the only variable undeterm ined
prior to the optim isation. Experience shows that an optim um result can be achieved with 10 to
20 iterations for problem s w ith around 5 to 10 variables.
W hen using the com plete full EM sim ulation for analysis the optimisation w ould last for 920 h
or 39 days. The exact tim e could not be determ ined for obvious reasons. The response o f the
optim ised circuit w ould be close to the desired response but lim ited by the discrete grid-snap.
Using the cascaded circuit and rough grid m eshing for analysis, the optim isation takes 31500 s
or 8.75 h. The result is not satisfying. Firstly, the calculations are based on a very rough grid
and, secondly, errors introduced by decom position are present. The accuracy can be enhanced
by finer m eshing. However, the tim e penalty is severe.
Im plem enting the Cauchy m ethod, the total optim isation tim e is reduced to 2.5 h. This includes
the tim e required for generating the database. The final optim ised response is closer to the de­
sired response because w e interpolate off-the-grid points. The simulation, however, still suffers
from inaccuracies introduced by the decom position.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 5. H Y B R ID O PTIM ISATIO N USING COUPLING M A T R IX A LT E R A T IO N
84
W hen applying our hybrid M -m atrix optim isation as suggested in this chapter, the optim isation
tim e is 12 h 16 min. The tim e includes database generation, tw o com plete EM sim ulations and
25 optim isation iterations. The num ber o f iterations is increased because we have to perform
two optim isation runs as shown in Section 5.11. The error o f the m ethod is very sm all as seen
from the com parison o f Figure 5.18.
The benchm ark results are sum m arised in Table 5.2.
Conventional C om plete Circuit Optim isation
Conventional C ascaded Circuit Optim isation
Case. Circuit Opti. from Database-i- Cauchy
O ur H ybrid O ptim isation
C PU Time
E rror
920 h = 39 days
8 h 45 min
2 h 33 min
1 2 h 16 min
defined as exact
very large
large
very small
Table 5.2: C PU Tim e Com parison for D ifferent O ptim isation Strategies
5.5
Conclusion
H ybrid optim isation com bines the advantages o f coarse and fine m odelling in the analysis. On
the one hand, we exploit the coarse m odel’s com putational speed in the optim isation loop for a
rapid— but inaccurate— calculation o f the optim al circuit layout. On the other hand, we exploit
the fine m odel’s accuracy for verifying and adjusting the coarse model.
We discussed in this chapter three hybrid optim isation techniques: the m odel output correction,
the space m apping algorithm and our new coupling m atrix correction. M odel output correction
adjusts the output param eters (e.g. S-param eters) according to the difference o f the port param ­
eters of one coarse m odel and fine m odel simulation. A s shown above, this technique is not
successful. The correction o f the coarse m odel often yields non-physical responses.
Space m apping corrects the input param eters— rather than the output param eters— o f the coarse
model. There is no problem with non-physical behaviour, because the corrected m odel output is
always derived from a real physical m odel. However, space m apping is limited. The m apping
can suppress only faults correctable by the input param eters. Space m apping cannot correct
physical phenom ena w hich are not m odelled by the coarse model.
These restrictions prevent us from applying space m apping to bandpass filter optim isation. For
bandpass filter design we often use cascaded circuits as the coarse m odel and com plete EM
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 5. H Y B R ID O PTIM ISATIO N USING COUPLING M A T R IX A LT E R A T IO N
85
sim ulations as the fine m odel. The stray coupling betw een circuits is m odelled by the fine model
but not by the coarse m odel. As shown, w e cannot find a m apping o f the param eter spaces.
Stray coupling is considered in the coupling m atrix representation o f a filter. H ence, we pro­
posed a correction scheme based on the coupling m atrix representation. The coupling m atrix is
calculated for both the fine and the coarse m odel’s response. Then we derive the difference m a­
trix by subtracting the m atrices o f the tw o m odels. The difference matrix represents all couplings
not included by the coarse m odel, but m odelled by the fine m odel.
Subsequent sim ulations o f the coarse m odel are adjusted using the difference matrix.
This
m ethod derives a response which includes all stray couplings usually only m odelled by the fine
model. O ur corrected coarse m odel, however, has the com putational costs o f the coarse model,
only.
The new optim isation technique is dem onstrated on the design of a planar six-pole filter. The
m ethod incorporates a cascaded EM analysis m odel based on the m ulti-dim ensional Cauchy
m ethod (see C hapter 4) for the coarse m odel, and a com plete EM simulation for the fine model.
W ith only tw o com plete EM sim ulations, the filter could be very accurately sim ulated and opti­
mised. W hen com pared with conventional approaches, our technique is the only one providing
a very accurate solution within reasonable com putational expenses.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
Chapter 6
Optimisation by Time-Domain
Reflectometry
"Time is natures w ay o f keeping everything from happening at
the sam e tim e .", Unknown
This C hapter focuses on the optim isation o f wide-band and low-pass filters. For these filters
w e cannot apply the hybrid optimisation technique proposed in Chapter 5 because our hybrid
technique is restricted to narrow -band band-pass filters.
W ide-band and low-pass filters are often realised as non-uniform transm ission line filters. The
design of such filters is described in [1 0 3 ,1 1 6 ,1 1 8 ,9 9 ,1 1 1 ,1 1 0 ,1 1 2 ]. A very interesting sub-set
o f non-uniform transm ission line filters are the transversal filters. Transversal filters, also known
as Kallm ann filters, are finite im pulse response (FIR) filters. These filters can— in theory— be
synthesised for any type o f causal response. D etailed design instm ctions on transversal filters
are given in [109, 108].
Sim ilar to the form al filter synthesis o f coupled resonator filters, the synthesis o f non-uniform
transm ission line filters assum es ideal circuit elem ents. Since real transm ission lines and line
discontinuities (steps) are not ideal the response o f the synthesised layout differs from the de­
sired response. D irect E M optim isation o f the com plete circuit is very difficult, as these filters
typically include a large num ber o f optim isation param eters. Cascading is not feasible either, as
the decom position generates significant errors. This chapter suggests a new m ethod o f perform ­
ing an accurate optim isation o f transm ission line filters using full EM sim ulations even for a vast
num ber o f design param eters.
86
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 6. O P T IM ISA T IO N B Y TIM E -D O M A IN R E F LE C TO M E TR Y
87
O ur new technique perform s the optim isation within the tim e-domain. The advantage o f using
the tim e-dom ain is that each section o f the tim e-dom ain response corresponds to a geom etrical
section in the circuit. H ence, we can optim ise the circuit layout section by section watching
the respective response section exclusively. This approach simplifies the optim isation process
significantly.
O ur m ethod is adapted from tim e-dom ain reflectometry. Tim e-dom ain reflectom etry (TD R) is a
well known m easurem ent technique for detecting discontinuities in both R F and optical trans­
m ission paths. TD R detects type and m agnitude of the discontinuity and— m ost interesting— the
location o f the discontinuities. The characterisation o f discontinuities by tim e-dom ain reflec­
tom etry is the topic o f m any publications, e.g. [100, 105, 107, 106, 115].
O ur m ethod uses the tim e-dom ain response from an ideal circuit as a tem plate and com pares
it with the actual response calculated by full electrom agnetic sim ulation or m easurem ent. We
m ake adjustm ents based on the differences in the tim e-dom ain response o f the desired and actual
circuit.
In the next section, w e will introduce the theory o f our new tim e-dom ain optim isation algorithm
for non-uniform transm ission lines. A pplications o f the algorithm and num erical exam ples are
given in Section 6.3 and 6.4.
6.1
Theory
Figure 6.1 describes an W -section transm ission line, w here each section i is characterised by
its characteristic im pedance Zi and propagation delay r* . The length o f the ith line section is
li. D elay and length o f the fine section are associated via the ith section’s wave propagation
constant /% by
h —
( 6 . 1)
For an incident wave a(t) at port one, the reflected wave b(t) at port one is a sum m ation o f
internal single or m ultiple reflections caused by the line’s discontinuities[106]
_ | I m,m+ 1 n L o (:Tk,k+iTk+i,k) a(t-£™ = 17*)+m ult.Reflections
(6.2)
where
(6.3)
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 6. O PTIM ISATIO N B Y TIM E -D O M AIN REF LEC TO M ETRY
C)
a(t)
Z,
b(t)
T i
Z2
Zj
Zn
T i
T n
T 2 ..........
L1
L2
L
n
Figure 6.1: N on-U niform Transm ission Line
T he m ultiple reflections m entioned in eqn. (6.2) are discussed in detail in [107]. The m ulti­
ple reflections occur— following causality— always later in tim e than the single reflections at a
discontinuity and do not have influence on the first reflections as defined above.
One notices from eqn. (6.2), that contributions in the response with respect to the line section
i, do not occur earlier in tim e than E L i 2-7*-
other words, the tim e-dom ain response up to
ti=Yj\=i 2rfc *s not altered by the line sections * + 1 to N . We can now associate a tim e-interval
T i o f the response b(t) with a length interval Hj o f the circuit using the relation
Y i = [ t i , t i + i]
H = [ L i , L i + i]
(6.4)
w here Li = E L t h ■
The intervals are related by t j = 4f to each other follow ing eqn.( 6.1).
Pi
Reflections observed in tim e interval Yj are caused by discontinuities in the length interval S i to
Hj. Consequently, when we observe tim e interval Yj only, we can concentrate our optim isation
on the param eter adjustm ent o f the discontinuities in the geom etrical interval Hi to Hj only.
W hen we have optim ised all param eters o f the interval Hi to H j_ i, w e can fully concentrate our
optim isation on the param eters in interval Hj .The next section discusses the algorithm in m ore
detail.
6.1.1
Ideal Filter Synthesis and Real Filter Performance
The circuit theory design o f non-uniform transm ission line filters is described in [103, 116,
118, 99, 104, 111, 110, 112, 109]. All o f the suggested design techniques, however, assum e
ideal circuit elem ents. Analogous to the problem s described for the synthesis in the frequencydom ain, the response o f the ideal circuit m ay differ significantly from the actual realised circuit.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 6. O PTIM ISATIO N B Y TIM E -D O M AIN REF LEC TO M ETRY
89
A n exam ple o f a sim ple three line section filter is given in Figure 6.2 (a). The desired response
in frequency dom ain is a notch type filter response that resonates at /o = 3 .7 5 GHz. The circuit
is synthesised using transm ission line theory [27], The m iddle line section needs an electrical
length of
w here A is the wavelength at / = 3.75 GHz along the m iddle line section.
T he ideal TEM -line section is converted into a m icrostrip line section as shown in Figure 6.2 (b).
D ue to the im perfect m icrostrip m odel the actual frequency response— com puted using full elec­
trom agnetic analysis— is different from the ideal response. The ideal (desired) and actual re­
sponse is depicted in Figure 6.3.
[m m ] 0 4
0
-0 .4
-
0.8
-1 0
0
10
20
30
40
50
60
70
80
90
[mm]
F igure 6.2: Circuit L ayout U sing Circuit Theory and R ealisation in M icrostrip Topology and
T heir Respective R esponse
The goal o f our new optim isation algorithm is to adjust the actual circuit such that its response
m atches the response o f the ideal circuit, w hich is our desired response. Thus, w e define the
response o f the ideal circuit as the desired response S des( f ) and the response o f the real circuit
as the actual response S act( f ) .
The technique requires both responses to be available in the time-dom ain. W hen using a timedom ain sim ulation tool, the tim e-dom ain responses are known. W hen using a frequency-dom ain
sim ulation tool, the response has to be transform ed into tim e-dom ain using Fourier transform .
Fourier transform ation o f frequency-dom ain signals (obtained by EM sim ulators) is known to be
com putationally expensive. In the next section, w e will show that this problem can be overcom e
using the Cauchy m ethod applied to the frequency sweep.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 6. O PTIM ISATIO N B Y TIM E -D O M AIN REF LEC TO M ETRY
90
-20
-25
-30
-35
-4 0
-45
d e s ire d
a c tu al
-5 0
-55
3 .5 8
3 .6
3.6 2
3 .6 4
3 .6 6
3 .6 8
3 .7
3 .7 2
3 .7 4
3 .7 6
3 .7 8
3.8
f [G H z]
Figure 6.3: Frequency R esponse o f Ideal and Real Circuit
6.1.2
Transformation into Time-Domain
The response has to be transform ed into the tim e-dom ain when a frequency-dom ain sim ula­
tor is used.
The tim e-dom ain response is com m only attained using the inverse fast Fourier
Transform (FFT). However, applying the inverse F FT [98] to a frequency response requires an
abundant num ber o f frequency sam ple points. The calculation o f these sam ples is com putation­
ally expensive, thus, m aking the overall process ineffective. With the im plem entation o f the
Cauchy m ethod [76] and adaptive sam pling [82], as described in Section 4.1, these drawbacks
are overcome. The response o f the circuit is sam pled at a few salient frequency sam ples, and the
in-between points are interpolated by rational function interpolation.
The procedure o f obtaining the tim e-dom ain response is sum m arised in Figure 6.4. The m ini­
m um required num ber o f sam pling points is calculated by the E M -sim ulator im plem enting adap­
tive sam pling, as proposed in C hapter 4. U sing the exact samples, an abundant set o f sam ples
is interpolated using C auchy’s method. Since the frequency data is band-lim ited, the response
S ( f ) has to be m ultiplied by a window -function A ( f ) to avoid G ibb’s phenom enon[113]. The
function S ( f ) ■A ( f ) represents the response o f the system excited by the pulse A ( f ) in the
frequency-dom ain. From the product, the inverse fast Fourier transform (IFFT) com putes the
tim e-dom ain response b(t). The output b(t) is the convolution of the tim e-dom ain system re-
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 6. O PTIM ISATIO N B Y TIM E -D O M AIN REF LEC TO M ETRY
S(f)
EM
S im u la to r
P u lse
fu n c tio n
abundant
s e t of
s a m p le s
s a lie n t
s a m p le s
A(f)
S(f)
C a u c h y ’s
M e th o d
91
S(f)A(f)
s(t) ® a(t)
FFT
A d a p tiv e S a m p lin g
Figure 6.4: C om putation o f Tim e-dom ain From Frequency-D om ain Electrom agnetic Sim ulation
sponse w ith the selected pulse function a ( t ):
6
(f) = s(f) <8 >a(t) d=
f
s(r)a(t —r ) d r
(6.5)
Jo
D epending on the application four different functions for a(t) are used, nam ely a Gaussian
pulse, a differential Gaussian pulse, a step function and a sine-m odulated Gaussian pulse (see
Figure 6.5).
1
I
1
1
0
0
0
0
-2
0
2
4
6
3
•1
10
-2
0
2
4
6
3
10
-2
0
2
4
6
8
10
•1
Figure 6.5: Excitation Pulses in Tim e-D om ain a) Gaussian b) Differential Gaussian c) Gaussian
Step d) Sine M odulated Gaussian Pulse
O ur algorithm uses the B ase-2 FFT algorithm. The frequency response consisting o f 2" sample
points is transform ed into a 2"-point tim e-dom ain response. A m ore detailed discussion o f the
FFT, the convolution theorem , and the w indowing functions is given in [98].
6.1.3
Pinpointing of Fault Location
L e t’s recall the problem o f adjusting the resonator circuit such that the actual circuit response
m atches the desired response. Before w e start the actual optim isation o f the real circuit we
pinpoint the fault location. We define fault locations as sections in the filter w here differences
betw een the desired and actual responses originate.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 6. O PTIM ISATIO N B Y TIM E -D O M AIN R E F LEC TO M ETRY
92
Com paring the actual and desired frequency-dom ain responses o f our exam ple line filter, we
cannot obtain any inform ation about the fault location from the frequency response (see Figure
6.3). R ecalling Section 6 .1, however, we know that any discontinuity in the line can be connected
to a certain interval in the tim e-dom ain response. The tim e-dom ain responses s f f s (t) and s f f 4 (f),
using a Gaussian pulse excitation, are depicted in Figure
6 .6
.
0.4
0.3
desired
actual
0.2
0.1
- 0.1
- 0.2
- 0.3
100
300
200
500
400
Ups]
Figure
6 .6
: D esired R esponse s f l s (t) and A ctual R esponse S i f ( i ) in Tim e-Dom ain
W e observe that the first reflections in tim e are identical. Consequently, the length interval S i,
corresponding to the tim e interval Y i = [0 n s , 300 ns], behaves identically in both the ideal
circuit and the actual c irc u it. The section S* o f the filter does not need any further adjustment.
Reflections after t = 300 ns are different due to differences betw een the ideal circuit and the
actual real circuit. By taking the difference of the tim e-dom ain responses, i.e.
A s ( t ) = s act(t) - s des( t ),
( 6 .6 )
all echos which originate from identical discontinuities are elim inated. Only echos o f additional
or different discontinuities remain. The difference pulse is zero up to the tim e when the first o f
these rem aining echos reaches the port. This can be seen in Figure 6.7. W e include the excitation
pulse into the figure, so that w e can read off the tim e-delay t \ betw een exciting the circuit and
getting the echo.
From the elapsed tim e t \ betw een the excitation o f port one and getting the echo pulse (see
Figure 6.7) w e can derive the distance l\ o f the discontinuity from port one by
h =
,
(6.7)
w here /3 is the velocity o f the waves in the m icrostrip line. This m eans that the fault location can
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 6. O PTIM ISATIO N B Y TIM E -D O M AIN REF LEC TO M ETRY
93
0.4
S„
0.3
xciting G aussian pulse
0.2
- 0.1
-
0.2
-0.3
-0.4
100
200
300
400
500
t lp s l
Figure 6.7: Exciting Gaussian Pulse and Difference o f D esired and A ctual R esponse; t \ Denotes
the Time Elapsed B etw een Excitation and Getting the Echo o f the Fault Location
be pinpointed at l\ units away from port one.
T he pinpointing o f the fault location am ong other discontinuities, w hich are identical in both
structures, is analogous to clutter 1 elim ination in R A D A R systems. N on-coherent clutter elim i­
nation, as explained in [101], extracts targets from the clutter in the sam e manner: First a radar
echo is recorded w ith the fixed echos (due to clutter) only. Then this recorded radar response is
subtracted from the actual response so that all echos o f fixed targets disappear. The target echos
rem ain and stand out.
In our exam ple, the tim e t.\ is found to be 325 ps from Figure 6.7. For simplicity, w e used a
free-space sim ulation o f the circuit(i.e. (3 — v = 3 ■10 8 m/s). U sing equation (6.1) we finally get
l\ = 48m m
( 6 .8 )
Looking at Figure 6.2, we pinpoint the fault location at the second discontinuity. In the next
section, w e use this inform ation to adjust the actual circuit’s output to the desired output.
6.1.4
Elimination of Faults
To elim inate the fault w e m odify the circuit section causing the fault. H ere com es the advantage
o f tim e-dom ain optimisation: The circuit is optim ised in one specific area only. We deal with
one or two optim isation variables.
'clutter: undesired echos in the radar response due to fixed ground targets, weather etc.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 6. O PTIM ISATIO N B Y TIM E -D O M AIN REF LEC TO M ETRY
94
We located the fault at the second step around I = 48 mm. Hence, we concentrate our optim i­
sation on adjustm ents within this area. L et us call the length interval around this area E p . We
m odify the circuit by varying the m etallisation at discontinuity in the interval
Figure
6 .8
As seen from
, varying the param eters p \ and p 2 m odify the m etallisation at this point.
Pi
48mm
Figure
6 .8
: Optim isation Param eters at Fault Location
T he optim isation process adjusts these param eters until the actual tim e-dom ain response m atches
the desired response in the corresponding tim e interval T p. Form ally speaking, the optim iser
m inim ises the least square error
(6‘9>
( W )2 - 6*^)2) *
E tf = L
o f the time interval T p starting at t l ='^2'k7=\ 2Tk an<l ending at t2=^2 'k=12 Tk using the optim isation
variables p \ and p 2 This optim isation process is very robust, as the param eter p i and p? effect the error function
alm ost proportionally within the small param eter range. The optim isation m ay take three or four
steps. A possible outcom e o f the optim isation is described in Figure 6.9.
a) Modification 1
-1 0
0
10
20
30
40
50
60
70
80
90
[m m ]
Figure 6.9: O ptim ised Circuit Layout
A fter the optim isation, the actual response is m atched with the desired response as shown in
Figure 6.10. This m eans that w e can m anufacture the circuit with the desired response, as we
assum ed the EM -sim ulation to be correct.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 6. O PTIM ISATIO N B Y TIM E -D O M A IN REF LEC TO M ETRY
95
-10
[dB ]
-20
-3 0
-4 0
-5 0
-6 0
a c tu a l
d e s ir e d
-7 0
3 .5 8
3 .6
3 .6 2
3 .6 4
3 .6 6
3 .6 8
3 .7
3 .7 2
3 .7 4
3 .7 6
3 .7 8
3 .8
f [G H z ]
Figure 6.10: D esired Frequency R esponse and Frequency R esponse o f M odified Circuit
6.1.5
Non-Uniqueness o f Fault Elimination
Evidently, there exists m ore than one m odification leading to the desired response. Three possi­
ble m odifications, bringing out the desired response, are shown in Figure 6.11. The final circuit
layout is said to be non-unique. The actual m odification used is left up to the user.
Note that the circuit could even be fixed at a different location (see M odification 3) in case
the fault location is not accessible. An increase o f capacitance in a line at one location— as
done here— is equivalent to an increase o f inductance j away from that location. H ence, we
can increase (thinning the line) the inductance in the m iddle o f the resonator alternatively to
increasing the capacitance at the end o f the resonator.
This dem onstrates that non-uniqueness can w ork to our advantage. We do not have to find the
m odification, since many different solutions provides us with the desired fix. The optim isation
becom es even m ore efficient, as w e m ay have to optim ise only one param eter rather than two
param eters.
6.1.6
Eliminating Multiple Faults
So far we dem onstrated our algorithm on the elim ination o f one fault only. In practice, however, a
circuit needs m odifications at m ultiple locations. Below w e will outline the iterative elim ination
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 6. O PTIM ISATIO N B Y TIM E -D O M AIN REF LEC TO M ETRY
96
a) Modification 1
[m m ] a 4
0
-0 .4
-
0.8
-1 0
0
10
20
30
40
50
60
70
80
90
[m m j
0.8
[mm] Q 4
b) Modification 2
0
-0 .4
-
0.8
-1 0
0
10
20
30
40
50
60
70
80
90
[m m ]
c) Modification 3
[m m ] a 4
0
-0 .4
-
0.8
-1 0
0
10
20
30
40
50
60
70
80
90
[m m ]
Figure 6.11: Three D ifferent M odifications o f the L ayout Yielding the D esired Response: a)
Extending the w idth at Fault; b) Shortening the Line with Im pedance Zo ; c) A djusting Circuit
Away F rom Fault
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 6. O PTIM ISATIO N B Y TIM E -D O M AIN R E F LEC TO M ETRY
97
o f faults.
L et us recall the line filter o f Figure 6.1. The algorithm works by detecting and optim ising
discontinuities sequentially. The iterative process starts w ith interval one with the goal o f m in­
im ising the error E \ in the first tim e interval by optim ising Z \ and w \ . Then, section tw o is
m odified by adjusting Z 2 and w 2 and so on. We repeat this process for all intervals up to interval
N . U sing this technique the high-dim ensional optim isation problem o f 2N variables is split into
N tw o-dim ensional optim isation problem s w hich can be solved m ore easily.
We apply the sequential tim e-dom ain optim isation m ethod on a another exam ple in Section 6.4.
6.1.7 Pulse Tracking
In the above exam ples, we assum ed that the propagation constant k and, consequently, the prop­
agation speed j3 inside the circuit is known. Furtherm ore, w e assum ed to know how the pulse
is travelling through the circuit so w e could use the sim ple equation (6.7) to pinpoint the fault
location. However, in som e circuit structures the exact propagation speed and propagation path
o f the pulse is not obvious.
Thus, we im plem ented a pulse tracking routine in our program . The routine generates a timesequence o f the pulse travelling through the structure. The output o f the routine provides us the
m ost wanted inform ation, i.e. location w ith respect to the tim e t, o f the pulse propagation inside
the circuit.
The tim e-sequence representation can be used to obtain the fault location from the echo tim e
11 — instead o f equation 6.7— even if the propagation speed and exact path is unknown.
The propagation o f the Gaussian pulse in the first exam ple is shown in Figure 6.12.
Sum m ing up, we synthesised a circuit w ith a desired response by a com bination o f a circuitbased program and EM -sim ulation. The circuit-based program is used for a fast synthesis o f the
circuit. Faults in the derived layout are detected by an accurate EM -sim ulation. A pinpointing
o f the faults in tim e-dom ain and an optim isation aim ed at the fault location fixes the faults. In
doing so, the use o f the slow optim isation in a high-dim ensional param eter space is avoided.
6.2
Tuning of Filter Hardware
Som e of the challenge in tuning filter hardware is in identifying which circuit section needs to
be tuned and determ ining how to tune the filter to obtain the desired response. O ur proposed
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 6. O P TIM ISATIO N B Y T IM E -D O M AIN REF LEC TO M ETRY
98
t= ,014 ns
t= ,027 ns
t= ,057 ns
t= ,071ns
t= ,085 ns
t= ,114ns
t= ,127 ns
t= ,143 ns
t= ,157 ns
t= ,171ns
t= ,185 ns
t= ,200 ns
t= ,214 ns
Figure 6.12: Tim e Sequence o f Pulse Travelling Through a Circuit
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
t= ,043 ns
CHAPTER 6. O PTIM ISATIO N B Y TIM E -D O M AIN R E F LEC TO M ETRY
99
tim e-dom ain technique can help the person at the test bench to tune the filter hardware.
In that case, the m easured response becom es the actual response and the theoretical response
becom es the desired response. The fault locations can be pinpointed in analogy to the m ethod
described above. However, adjusting the fault location is different. W e cannot change the circuit
shape as easily as in the com puter simulation. On planar circuits we can add m etallized patches
or cut away m etallisation, as done e.g. with laser beam s. Alternatively, w e can m ove dielectric
blocks on the substrate surface. In waveguide filters, we com m only introduce tuning screws.
6.3
Optimisation and Tuning of Narrow-Band Resonant Filter Cir­
cuits
A direct application o f the tim e-dom ain optim isation technique to narrow-band resonant filter
structures is not practical. The im pulse response o f such filters stretches out over a very long
tim e period. The energy is reflected in very sm all portions by the filter discontinuities. Accurate
adjustm ents o f the discontinuities from these m inim al reflections is not feasible.
The problem can be circum vented using a band-pass weighting function in the frequency do­
m ain. The frequency-dom ain response is m ultiplied by the weighting function, and the timedom ain transform ation is applied to this weighted frequency-dom ain response. A typical w eight­
ing function (also known as window function) is shown in Figure 6.13. The introduction o f a
band-pass weighting function m eans physically, that the system is excited by a pulse function
m ixed with a carrier frequency at /o . The excitation pulse can be interpreted as an am plitude
m odulated signal. H ere w e im plem ent the sine m odulated Gaussian pulse as depicted in Figure
6.5 (d).
Response
Band of Interest
32730836^6
Window Function
Figure 6.13: Band-pass W indow Function
W hen excited with the sine m odulated Gaussian pulse, a tuned band-pass filter’s response shows
characteristic dips. The m agnitude envelope (in log-scale) o f the tim e-dom ain response o f a
tuned 5-pole filter is shown in Figure in 6.14.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 6. OPTIM ISATIO N B Y TIM E -D O M AIN REF LEC TO M ETRY
100
-10 dB
-20 dB
-30 dB
-40 dB
-40 ns
0 ns
260 ns
F igure 6.14: Tim e-D om ain R esponse o f Tuned N arrow -Band Filter
B eginning with the first dip near t=0 ns, the next N dips represent the response o f the resonators
1 through N in the filter. This relationship was recently identified in [117] by a research group
from Hew lett-Packard Test & M easurem ent. T he tim e span betw een the dips is the group delay
from one resonator to the next. W hen one o f the resonators is de-tuned, i.e. it does not resonate
at the desired frequency, the dip disappears. Figure 6.15 shows the response with resonator three
de-tuned.
-10 dB
-20 dB
-30 dB
-40 dB
-40 ns
0 ns
2 6 0 ns
Figure 6.15: Tim e-Dom ain R esponse o f N arrow -Band F ilter W ith R esonator three Detuned
The geom etry o f this de-tuned resonator can now be changed such that the characteristic dip
appears. M inim ising the tim e-dom ain m agnitude at the dip locations results in a properly tuned
filter as outlined in [117]. In analogy to the unw eighted tim e-dom ain optim isation, w e fix one
fault location— here a resonator— after another, starting from the excitation port.
The same group o f people at H -P expanded the technique to allow tuning o f the filter couplings.
Observing the peaks betw een the dips in the response, they can also tune the resonator couplings.
The m ethod relies, exactly as our m ethod, on the existence o f an ideally tuned filter. The cou­
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 6. O PTIM ISATIO N B Y T IM E -D O M AIN REF LEC TO M ETRY
101
plings are adjusted with the goal to m atch the actual with the ideal response. Once all peaks are
m atched the filter is tuned.
6.4
Numerical Results
This num erical exam ple illustrates the functionality o f the tim e-dom ain optim isation technique,
on the designed o f a non-uniform line low-pass filter. O ne o f the filter specifications is a return
loss o f less than 25 dB in the passband below / = 1.5 GHz. From filter and transm ission line
theory [27] the initial layout as shown in Figure 6.16 is synthesised.
L1 L2
L3
L4
L5
Figure 6.16: Planar N onuniform Transm ission Line Low-Pass Filter
The ideal response o f the initial layout in frequency dom ain is shown in Figure 6.17 m arked as
circuit theory.
We also com pute the response o f the initial layout using the full EM sim ulation H P-M om entum
[149]. This response is the actual response. A s seen in Figure 6.21, the actual response does not
yield the desired 25 dB rejection in the passband. We convert the response into the tim e-dom ain
by inverse FFT. The inverse FFT is perform ed on 64 frequency samples in a 5 GHz frequencyband. Using C auchy m ethod the 64 sam ples are com puted from 9 original frequency samples.
Only these nine sam ples are actually derived using full EM -sim ulation.
Not surprisingly, the tim e-dom ain response (Figure 6.18) shows a significant deviation between
circuit theory and EM -sim ulation. The filter requires a full-EM optim isation in order to m eet the
specifications.
The filter is optim ised to the specifications using our new tim e-dom ain optim isation m ethod.
First we divide the filter into five geom etrical intervals. Each interval covers one discontinuity,
i.e. im pedance step. W hen changing the tw o param eters o f the respective im pedance step (line
length and width), the tim e-dom ain response is only m odified in the corresponding and later
time-intervals.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 6. O PTIM ISATIO N B Y TIM E -D O M AIN REF LEC TO M ETRY
102
S ll
[dB]
-10
-15
Full-EM
-20
-25
-30
Circuit-Theory
-35
-40
-45
f [GHz]
-50
0.5
3.5
2.5
F igure 6.17: Frequency-D om ain response S n From Circuit Theory (D esired Response), FullE M Sim ulation (Actual R esponse)
— T ra n sien tjS ll}
Circuit Theory
Full-EM
0
1
2
4
3
t [ns]
5
- Transient] S 2 2 } !
Circuit Theory
Full-EM
1
0
1
2
3
4
t [ns]
Figure 6.18: Tim e-D om ain R esponse s n ( f ) From Circuit Theory and raw Full-EM
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
5
CHAPTER 6. O PTIM ISATIO N B Y TIM E -D O M AIN R E F LEC TO M ETRY
103
F or example, Figure 6.19 illustrates the effects o f layout variations on both the frequency and
tim e-dom ain response. Varying w 4 the frequency-dom ain response is affected over the com plete
frequency band, w hereas in tim e-dom ain the effects are lim ited in the interval [0.25ns,
00
]. The
interval[0ns, 0.25ns] rem ains untouched by w ,\, as predicted in Section 6.1.4.
0
T im e-D o m ain R esp o n se
f
F re q u e n c y -D o m a in
6 m il R esp o n se
8 mil
10 mil
12 mil
14 m il
16 m il
-10
-20
-30
-40
-50
-60
-70
-80
-90
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
. 0 .5
0
0.5
1
1.5
2
Figure 6.19: Variation o f Frequency and Tim e-D om ain R esponse by Changing W4
An optim isation o f the response using the m ethod described in this C hapter m atches the EM sim ulation response w ith the theoretical response as shown in Figure 6.20. It can be seen from
F igure 6.21, that the adjusted frequency-response o f the optim ised layout m eets the specifica­
tions.
6.4.1
Comparison of Computation Times
The synthesis o f the low-pass filter is a ten-dim ensional optimisation problem . T he full EM
sim ulation o f the circuits takes 110 C PU seconds. Hence, a frequency-dom ain optim isation
requires about 300 C PU min. The tim e-dom ain technique optim ises tw o param eters in parallel.
It needs five optim isation runs for the five intervals. In total the optim iser perform s 35 full EM
sim ulations, w hich is equivalent to about 64 C PU min. The results are sum m arised in Table 6.1.
As seen, the required C PU tim e is reduced by a factor o f five. Savings in com putation tim e are
even bigger for larger circuits.
Conventional direct full E M optim isation
O ur tim e-dom ain optim isation m ethod
C PU Time
E rror
300 min
64 min
very small
very small
Table 6 .1: Com parison o f C PU Tim es for Conventional and Tim e-dom ain Optim isation
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 6. O PTIM ISATIO N B Y TIM E -D O M AIN REF LEC TO M ETRY
104
_ T ransient{S ll}
Circuit Theory
Full-EM
0
1
3
2
4
t [ns]
5
-T ransient[S22}
Circuit Theory
Full-EM
0
■1
1
3
2
4
t [ns]
5
Figure 6.20: Tim e-D om ain R esponse o f s n (£ ) from Circuit Theory and C orrected Circuit Using
Full EM and TD R Optim isation
0
Sl l
■5
[dB]
-10
-15
-20
Full-EM
-25
L Full-EM tim e-dom ain adjusted
-30
Circuit-Theory
-35
-40
-45
-50
f [G H z]
0.5
2.5
3.5
Figure 6.21: Frequency-D om ain R esponse 5 n From C ircuit Theory, Full-EM , C orrected FullE M Simulation
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 6. OPTIM ISATIO N B Y TIM E -D O M AIN REF LEC TO M ETRY
6.5
105
Conclusion
In this chapter w e introduce a new optim isation technique for non-uniform transm ission line
filters. O ur technique is based on the theory o f tim e-dom ain reflectometry. We perform the
optim isation in tw o steps: First the fault locations are located by com paring the tim e-dom ain
response o f the circuit, i.e. the actual response, with the desired response. Then, the faults are
rem oved using an optim isation aim ed at the specific fault location.
The large set o f optim isation variables is divided into small sets o f one or tw o variables, which are
m uch easier to handle than the high-dim ensional problem . Hence, w hen dealing with m edium
or large size circuits, the proposed optim isation technique is superior to the optim isation o f the
frequency-response. We showed that for a circuit with ten degrees of freedom the com putational
costs are already reduced by 80%. The m ethod has in principle no upper bound o f the num ber
o f param eters. A pplication o f the m ethod lies in the com puter aided design and com puter aided
tuning of filters.
A lso in this chapter, w e suggest an algorithm that overcom es the problem s with the Fourier
transform o f a very sm all set o f frequency-response samples. B y applying Cauchy m ethod and
adaptive frequency sam pling, the num ber o f required sam ples is reduced to a m inim um .
It has been shown that w e can apply the proposed technique to narrow-band resonator filters as
well. For optim um perform ance, w e have to excite the filter ports with a sine-m odulated pulse.
This technique reduces the sam pling acquisition for the FFT to the frequency band o f interest.
The response o f the filter exhibits characteristic dips and peaks in the dem odulated tim e-dom ain
response. This characteristic can guide the user in the optim isation or tuning o f the filter.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
Chapter 7
Hardware Verification
"The only theorem o f superconductivity that can be p ro v e d is
that any theory o f superconductivity is refutable", F elix Bloch
The design tools proposed in this thesis are pointless, if w e cannot use them for the developm ent
o f m icrowave device hardware. Consequently, w e verified the usability o f these techniques on
the design o f functional hardware. In this chapter, w e present m easurem ents for som e o f the
hardware designed with the help o f our new synthesis techniques.
The hardware is developed at C O M D EV Space Group. C O M D EV is a specialised supplier
for m icrowave satellite equipm ent. All devices appearing here are part o f experim ental satellite
payloads. The em phasis is on applications o f high-tem perature superconductors (H TS) in satel­
lite m icrowave systems. The superconducting com ponents presented here are part o f the current
DARPA High-Tem perature Superconductor Technology D evelopm ent project. Inform ation on
this project can be found in [144]. Form er H TS space experim ents projects are discussed in e.g.
[142, 143]. The project is a joint C anadian/U S program directed by the N ational A eronautics
and Space A dm inistration (NASA), and funded by DARPA, the Canadian space Agency (CSA)
and the D epartm ent o f National D efence (DND).
7.1
Ka-Band Rejection Filter
The rejection filter— designed and optim ised using our enhanced genetic algorithm s— is inte­
grated into a K a-B and relay satellite. The typical block diagram of a relay satellite is shown in
Figure 7.1. The purpose o f the rejection filter is to suppress any feedback from the down-link
antenna into the up-link antenna. This coupling, if not suppressed, can cause oscillations.
106
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 7. H A R D W A R E VERIFICATION
107
L ocal
I O sc .
L o w N o is e
Pow er
A m p lif ie r
A m p lif ie r
R e je c tio n
D ow n
F ilte r
C o n v e r te r
P a ra s itic F e e d b a c k
Figure 7.1: B lock D iagram o f a Typical R elays Satellite
The optim ised design for the rejection filter shown in Figure 3.7 is verified by m easurem ent. The
m easured results com pared to the sim ulated results are shown in Figure 7.2.
The m easurem ents show excellent agreem ent with the sim ulation of the optim ised circuit.
7.2
3-dB Power Divider
For a high-tem perature superconducting (HTS) L-B and beam form ing netw ork (shown in the
next section) we design a 90°-branch-line coupler. This device is com m only known as a hybrid.
Superconducting planar circuits achieve the sam e perform ance while using less volum e and
w eight com pared to conventional coaxial hybrids. Figure 7.3 shows a the size com parison shown
for an conventional versus HTS 90°-H ybrid pow er divider a L-band. A num erical com parison is
given in Table 7.1.
Loss
Volume
W eight
W aveguide
Co-axial
HTS
HTS drop-in
0.3 dB
'1 0 in3
'3 0 0 g
0.3 dB
0.75 in3
100 g
0.3 dB
0.75 in3
40 g
0.3 dB
0.75 in3
5g
Table 7.1: Com parison o f B ranch-C oupler Technologies
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 7. H A R D W A R E VERIFICATION
108
1 9 .5 -2 0 .5 GHz
l
2 9 .0 - 3 0 .0 GHz
o
-5
-10
-1 5
-20
-25
-3 0
-35
~4 0 5
10
15
20
25
1 9 .5 -^2 0 .5 G H z
0
30
f[G Hz]
35
2 9 . 0 -J30.0 G H z
-10
-20
-30
-40
-50
-60
'7°
10
15
20
25
30
f[G Hz]
35
Figure 7.2: Sim ulated (Thin Line) and M easured R esponse (Fat Line) o f Stripline F ilter .At the
Top S'nfdB J and at the B ottom S'i 2 [dB]
Conventional
0 .7 5
HTS
0 .5 " (
1.5"
HTS
drop-in
Figure 7.3: C om parison o f L-B and Hybrids
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
109
CHAPTER 7. H A R D W A R E VERIFICATIO N
It can be seen from the table, that the volum e and w eight can be significantly reduced by replac­
ing coaxial hybrids with HTS technology. Especially when several hybrids can be integrated on
one substrate, the w eight reduction can be enormous.
Conventional pow er dividers are typically realised using a ring or square shaped layout. H ow ­
ever, such layouts occupy a large area. The area o f the hybrid can be reduced by about 60%
w hen using a folded structure as shown in Figure 7.4.
1 - |---------------1 - 2
1—
i
4-1
i—
'
------U 3
Figure 7.4: F ayout o f Hybrid
We m odel the hybrid using the m ulti-dim ensional Cauchy m ethod in conjunction with a full EM
simulator. U sing this approach, w e obtain a fast and accurate m odel o f the circuit. We optim ise
the circuit’s perform ance (using a gradient-search algorithm ) for best return loss and equal pow er
distribution at the outputs with a 90° phase shift. The optim isation uses four param eters: length
and width o f the through lines and branch lines.
A fter optimisation, the hybrid exhibits a return loss o f less than 25 dB and an isolation o f 30 dB
over a bandw idth o f 300 M Hz. Pow er split is very close to the desired 3dB/3dB, and the phase
shift is at the desired 90°. T he perform ance o f the structure is not degraded by folding. The
response o f the optim ised hybrid is shown in Figure 7.6. A photograph o f the hybrid is shown in
Figure 7.5.
The perform ance o f the filter is verified by m easurem ents. The m easured results are shown in
Figure 7.7. As seen, the m easured response is in good agreem ent with the simulation.
7.3
Butler-Matrix Beamformer
On satellites it is som etim es desirable to feed an antenna array through a feed system that will
excite a num ber o f beams. This netw ork is called a beam form ing network or simply beamform er.
Each input port o f the beam form er excites the array so that the it produces one particular beam.
D ifferent input ports generate beam s pointing in a different direction. The concept is illustrated
in Figure 7.8. The beam s can be used sim ultaneously w ithout cross-talk in the beam form ing
networks.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 7. H A R D W A R E VERIFICATIO N
110
Figure 7.5: Photograph o f the Folded H ybrid in C om parison W ith O ne C ent Coin
200
-10
150
-15
100
lap ]
-20
-2 5
-3 0
-50
-35
-100
-4 0
1.5
1.52
-45
1.54
1.56
frequency [GHz]
1.58
1.6
1.56
1.6
-5 0
1.5
1.52
1.5 4
1.5 6
1 .5 8
1.6
frequency [GHz]
Figure 7.6: Full EM Sim ulated R esponse o f Optim ised H ybrid
o
200
-s
150
•10
100
•15
50
-20
0
-25
•50
•30
-100
-35
•150
•40
-45
1.5
1.52
1.54
1 56
f [GHz]
.56
1.6
-200
.5
1.52
f [GHz]
Figure 7.7: M easured R esponse o f O ptim ised Hybrid
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 7. H A R D W A R E VERIFICATION
P a tte r n
g e n e ra ted
111
P a tte rn
g en erated
P a tte r n
g en e ra ted
B eam form er
B eam form er
B eam form er
P a tte rn
g en e ra ted
B eam form er
♦
4
4
4
F eed
F eed
F eed
Feed
Figure 7.8: B eam Generation by M ultiple Fixed B eam form ing N etw ork
B eam form ing networks are n -to -m netw orks and their purpose is to interconnect the n input
ports to the individual m antenna ports w ith the required am plitude and phase. They can be
either passive or active networks.
Probably the m ost widely known m ultiple-beam m atrix feed is the B utler m atrix. It is well
docum ented in the literature [38], Figure 7.9 is a schematic representation o f an 8x8 butler
matrix.
Antenna Ports
x
45°
<t>
IX
67.:
X
+5°
-2 0 °
+10°
-1 5 °
Beam Ports
F ig u r e 7 .9 : S c h e m a t ic B u tle r M a tr ix B e a m fo r m e r , T h e X ’s R e p r e se n t 9 0 ° H y b r id s W ith a P o w e r
Split Ratio o f 3dB/3dB.
In this section, we describe the layout and operation o f an 8x8 B utler m atrix. Twelve o f the
hybrids designed in the last section are necessary to build the 8x8 beam form ing network. The
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CH APTER 7. H A R D W A R E VERIFICATIO N
112
hybrids are connected by delay lines with a phase delay as seen in Figure 7.9.
T he beam form er is optim ised using gradient optim isation in connection with a cascaded elec­
trom agnetic simulation. The cascaded sim ulation yields satisfying results because we deal with
relatively w ide-band structures. The interested reader m ay be referred to [7] and [8] for detailed
inform ation on the beam form er design.
T he m atrix is built using T B C C O films supplied by DuPont. A detailed layout o f the m atrix is
shown in Figure 7.10. The overall size o f the m atrix is only 3.0 x 2.0 x 0.2 in3. The beam ports
(inputs for transm it m ode) are num bered respective to the beam generated. That is, 1R m eans
the first beam generated to the right. The antenna ports are num bered according to the order
o f the connected antenna elem ents. A photograph o f the here discussed 8x8 m atrix is shown in
F igure 7.11.
1
5
2
6 3 7 4 8
Plug-In
GPO-Connectors
M ulti-Layer
Cross-Over
D elay Lines
Folded 90°-Hybrids
1R 4 L 3 R 2L 2 R 3 L 4 R 1L
— - - -------
2.046
-------------------------------------------------------
Figure 7.10: L ayout o f 8x8 M atrix in Housing
The m easured and sim ulated results are sum m arised in Figures 7.12 to 7.15. Figure 7.12 shows
the return loss o f the beam ports 1R to 4R. Figure 7.13 shows the insertion loss for the path from
one beam port (shown here port 1R only) to the various antenna ports 1 to 8. The beam form er
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 7. H A R D W A R E VERIFICATION
113
is designed for uniform am plitude taper, i.e., all ports should receive | o f the input pow er (9 dB on a logarithm ic scale). T he sim ulated and m easured phase tapers o f all ports for the
centre frequency 1.55 G H z are sum m arised in Figure 7.15. As seen, all phases are within a 10°
accuracy. The total device insertion loss o f the m atrix is less than 0.2 dB.
Figure 7.11: Photograph o f 8x8 B utler M atrix B eam form er
o
•s
•10
•15
-20
-25
-30
•40
-45
-50
f req
1.5
1.52
1.54
f(QHz)
.56
1.56
1.6
Figure 7.12: R eturn Loss o f 8x8 M atrix for B eam Ports 1R, 2R, 3R, 4R Sim ulated (left), and
M easured (right)
The m easured am plitude and phase taper are used to calculate a beam -pattem generated by an
im aginary linear antenna array. The generated beam s are shown in Figure 7.16. It can be clearly
seen that the beam form er perform s its intended function.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
114
CHAPTER 7. H A R D W A R E VERIFICATION
1.6
GHz E
Figure 7.13: Insertion Loss o f 8x8 M atrix W ith B eam -Port Input 1R, sim ulated (left), and m ea­
sured (right)
200
150
90
cn
0
•50
•200
1.5
GHz
1.6
f
GH z H
1.5
1.52
f(GHz)
1.56
1.55
1.6
Figure 7.14: Phase Taper o f 8x8 M atrix Sim ulated and M easured
1500
1000
500
0
-500
•1000
•1500
0
2
3
4
5
Antenna Port
6
7
8
9
Figure 7.15: Phase Distribution o f 8x8 m atrix (Sim ulated and M easured).
Each Line corresponds to O ne Input Port. The D otted Lines M ark the Theoretical Phase Taper,
W hereas The D ots M ark The Sim ulated and M easured Taper, Respectively.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
115
CHAPTER 7. H A R D W A R E VERIFICATIO N
2L
2R
3R
4R
4L
F igure 7.16: Radiation Pattern (Sim ulated from M easured Data)
7.4
Cryogenic Switches
The pow er am plifiers aboard a satellites m ay be switched to provide redundancy as illustrated in
Figure 7.17. The scheme shown is term ed a four-to-two redundancy, m eaning that four channels
are provided with two redundant amplifiers. Using the m atrix shown in the diagram , any channel
can use three different amplifiers.
Ch. A
Ch. A
Switch
Switch
Ch. B
Ch. B
Switch
Switch
Ch. C
Ch. C
Switch
Switch
Ch. D
Ch. D
Switch
Position I:
&
Switch
Position II:
X /
&
Figure 7.17: Four-For-Two Redundancy Sw itch A rrangem ent o f A N IK D (see [141])
F or the incorporation into a cryogenic payload, we develop a new class o f cryogenic switches.
The switches use planar thin-film HTS technology with integrated sem iconductor devices.
7.4.1
HTS C-Switch
In this section we describe the design o f an planar integrated HTS C-switch. The switch has
been designed and optim ised using the m ulti-dim ensional Cauchy method.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 7. H A R D W A R E VERIFICATIO N
116
A C -switch is a four port network. The tw o operation m odes are defined as shown in Figure
7.18. We choose the layout o f the switch as shown in Figure 7.19. U sing a special gap coupling
o f the stub line with the circuit, w e attain D.C. de-coupling betw een the switch branches.
S w itc h
M o d e I:
S w itc h
M o d e II:
Figure 7.18: C -sw itch Operation M odes
P IN D io d e s /
F E T s/
H T S S w itc h L in e
R F C hoke
In d u c to rs
O n e L a y e r C irc u it
( n o D .C . C o u p lin g )
K -C o n n e c to r
D C -B ia s F e e d
F igure 7.19: L ayout o f C -Sw itch . Inset Shows One o f the Eight D.C. D ecoupled Stub Lines.
For an efficient sim ulation and optim isation w e decom pose the circuit into four sub-sections,
which are identical. The subsection is shown in Figure 7.21. We m odel this subsection using
m ulti-dim ensional Cauchy method. T he response analysis o f the switch is perform ed on the
cascaded circuit o f four identical subsection. W e optim ise the switch for best perform ance in the
frequency-band from 4.5 GHz to 5.5 GHz. B est perform ance m eans m inim ised return loss and
insertion loss with m axim ised isolation.
The optim isation results are sum m arised in Figure 7.22(left). M easurem ents on the device (Fig­
ure 7.22 right) show a good agreem ent o f the simulation.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 7. H A R D W A R E VERIFICATIO N
117
F igure 7.20: Photograph o f Conventional C o-A xial C -Switch, HTS C -Sw itch and O ne D ollar
Coin
P o rt 3
p4
P o rt 1
P o rt 4
p3
-► 4 pl
P3
P4
■P o rt 2
-► 4 p2
pi
Figure 7.21: Subsection o f C -Sw itch Showing the four O ptim isation Param eters
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 7. H A R D W A R E VERIFICATIO N
118
Transmission (on-state)
<L
O
Transmission (on-state)
o
•10
Return Lass
Return Loss
-15
-20
(u0
ra
CD
Transmission (off-state)
L
-30
Trans mi 5S10I1
ff-stite)
O
/ Isolation
-50
freq
4
4.2
4.4
4.6
4.8
5
5.2
5.4
56
5.8
6
f [GHz]
Isolation
Figure 7.22: Sim ulated (Left) and M easured (Right) Perform ance o f C -sw itch
7.4.2
HTS High Isolation Switch
The C -switch assum ed an ideal switching elem ent, i.e. ideal FET or Diode. In reality these
elem ents are non-ideal, i.e. they are lossy in the on state and not perfectly open in the off-state.
T he m easurem ents o f Section 7.4.1 are obtained by opening and closing m etal strips at the end
o f the line. U sing sem iconductors does not yield the desired response due to the losses in the
device.
A new design o f the switch is optim ised for best perform ance when operating with P IN diodes.
W e obtain the exact m odel o f the PIN diode from de-em bedded m easurem ents o f the diode at
cryogenic tem peratures.
T he assem bly o f the first tested switch using PIN diodes is shown in Figure 7.23.
M easurem ents show a very good agreem ent betw een our optim ised sim ulated response and the
m easured response. The com parison is shown in Figure 7.25.
7.5
Six-Pole Narrow-Band Filter
H ere presented are the m easurem ents o f the six-pole filter designed in C hapter 5. This type of
narrow-band filter is used in satellite filter assem blies for channel splitting. The detailed design
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
119
CHAPTER 7. H A R D W A R E VERIFICATIO N
XT
Side-View:
•DC-Feed
Spiral
Pi n
lndoctorDi ode
Ribbon
Bond
Wire
Bond
Figure 7.23: L ayout o f Four PIN D iode Sw itch and B low -U p o f the D iodes Including D.C. Bias
Netw ork and R F Chokes
Figure 7.24: Photograph o f High-Isolation Sw itch W ith Lid
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 7. H A R D W A R E VERIFICATION
[dB]
120
S21
-10
-15
-20
Sll
-25
-30
On State
-35
-40
o Measured
-45
-50
— Simulated
3.5
4.5
f [GHz]
5.5
0
[dB]
Sll
-10
-20
-30
-40
-50
Off State
S21
-60
-70
» Measured
-80
-90
— Simulated
3.5
4.5
f [GHz]
5.5
F igure 7.25: Sim ulated and M easured R esponse o f The High-Isolation Integrated H TS/PIN
D iode Switch
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 7. H A R D W A R E VERIFICATION
121
techniques is described in C hapter 5. The com plete assem bled circuit in a housing with Kconnectors is shown in Figure 7.26. A photograph o f the filter without lid can be seen in Figure
7.27.
1,337
Figure 7.26: L ayout o f Six-Pole Filter in Housing
Figure 7.27: Photograph o f the six-Pole Filter
The results o f the m easurem ents are shown in Figure 7.28. The m easured response is in excellent
agreem ent w ith the theoretical response, obtained by our hybrid full E M and cascaded sim ula­
tion. The m easured response shows a slightly higher centre frequency and slightly w ider band­
width. We can attribute these deviations to im perfect substrates and m anufacturing tolerances.
Indeed, past m easurem ents on other HTS devices com m only exhibit m uch stronger m ism atch b e­
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 7. H A R D W A R E VERIFICATION
122
tw een sim ulation and m easurem ent. This indicates, that we alm ost com pletely elim inated errors
from the num erical sim ulation and optim isation when applying our new developed techniques.
7.6
Conclusions
O ur new techniques have been tested on the developm ent o f various m icrowave devices. In this
chapter we have presented the design and m easured results for an L-band hybrid, an 8x8 B utler
m atrix beam form er, tw o C -Band switches, and a narrow -band six-pole filter. The devices are all
developed with the aid o f the here proposed synthesis techniques. W hen using our m ethods, we
reduced developm ent tim e significantly w ithout neglecting an accurate design.
The rem aining differences betw een the theoretical responses and m easured responses o f the
devices can be attributed to im perfect m aterials and m anufacturing processes. These problem s
can norm ally be fixed with circuit tuning.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 7. H A R D W A R E VERIFICATIO N
123
-10
-15
-20
-25
-30
-35
o M easured
-40
—
Sim ulated
-45
-50 I—
4 .3 5
4 .4
4 .4 5
4 .5
4 .5 5
4 .6
4 .6 5
-10
S21 [dB]
-20
-30
-40
-50
-60
-70
<po
0 M easured
—
Sim ulated
-80
-90 I—
4.35
4.4
4.45
4.5
4.55
4.6
4.65
f [GHZ]
Figure 7.28: Sim ulated and M easured Frequency R esponse o f Six-Pole F ilter W ithout Tuning
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
Chapter 8
Summary and Future Work
"If som ething is so com plicated that you c a n ’t explain it in 10
seconds, then i t ’s p robably not w orth knowing anyway", Calvin
in Calvin and H obbes
This thesis is m otivated by the lack o f fast, accurate, and reliable synthesis techniques for m i­
crowave filter design. As discussed, filter synthesis is currently perform ed by form al synthesis
and iterative optim isation in conjunction w ith an EM simulator. The trend is towards iterative
optim isation since it is m ore accurate and not lim ited to regularly shaped circuit layouts. H ow ­
ever, this approach still has shortcomings. W e have identified the main problem s o f the available
techniques as follows:
• The search for the optim al solution is unreliable That is the algorithm s are non-robust.
• The com putational costs are high.
• W hen sim plifications are used the accuracy o f the sim ulation is low.
• The spatial grid for the circuit vertices is discrete.
To resolve these problem s, we have proposed four new techniques, that can be com bined or used
separately. These techniques are
• Genetic algorithm s in circuit design
• M ulti-dim ensional Cauchy m ethod
124
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 8. S U M M A R Y A N D FU TURE W O R K
125
• Hybrid optim isation
• Tim e-dom ain based optim isation
We have dem onstrated that genetic algorithm s can increase the robustness o f m icrowave circuit
optimisation. Genetic algorithm s do not get trapped in sub-optim al solutions. Rather, they seek
the optimal solution by applying the law o f the survival o f the fittest. Further, genetic algorithm s
do not need derivative inform ation. H ence, the optim isation can be as fast or even faster than
a gradient guided search. In several exam ples, w e have shown the excellent perform ance of
GA’s com pared to gradient search. Genetic algorithm s are not m eant to be a replacem ent for
gradient-based optim isers, but a supplement.
For a faster analysis step, w e have developed a new m ulti-dim ensional Cauchy m ethod as an al­
ternative to neural networks. O ur technique can predict the response o f unknown circuit layouts
based on the knowledge o f several sam ple layouts and responses. This approach takes the expen­
sive calculations out o f the optim isation loop. The optim isation can be perform ed exclusively
on the very inexpensive m ulti-dim ensional Cauchy m odel. As shown, the m ethod can speed
up the optim isation by several orders o f m agnitude. The m ulti-dim ensional Cauchy m ethod is
now a vital part o f the num erical design package used for the synthesis o f high-tem perature
superconductive circuits at C O M D E V ’s corporate R & D departm ent.
O ur enhanced Cauchy m ethod is in practice lim ited by the m axim um num ber o f param eters.
Large circuits, still m ust be decom posed. As w e have shown, decom position neglects stray
coupling betw een the circuit sub-sections. To include this stray coupling, we have developed
here a new hybrid technique. O ur m ethod is based on the coupled resonator representation of
band-pass filters. W e have dem onstrated that our techniques reduce the optim isation tim e by a
factor o f several m agnitudes w ithout sacrificing accuracy. W ith the new technique, we accurately
designed large circuits, that previously seem ed im possible to optim ise with today’s com puters.
This new optim isation paradigm is now also part o f C O M D E V ’s filter design package.
A s m entioned above, the coupled resonator representation is lim ited to band-pass structures. To
tackle the optim isation o f low-pass and non-resonant filters, w e propose an optim isation based
on tim e-dom ain reflectometry. We have shown that we can split one iV-dim ensional optim isa­
tion problem into N one-dim ensional optim isation problem s. The latter arrangem ent is m uch
easier to handle than the former. H ence, w e gain significantly in optim isation perform ance. We
can exploit tim e-dom ain optim isation in the com puter sim ulated design process as well as in
hardw are m easurem ents.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 8. S U M M A R Y A N D FU TURE W O R K
126
T he new techniques have been verified by num erous hardware m easurem ents. The m easured
hardware is designed for upcom ing superconductive satellite microwave systems. The m easure­
m ents have m et or even exceeded our expectations. Slight differences in the responses between
theory and practice are largely due to m anufacturing tolerances.
8.1
Recommendations
Table 8.1 sum m arises the recom m ended design techniques for various filter types. Sm all filter
circuits with up to 5 optim isation variables can best be designed by our new m ulti-dim ensional
Cauchy method. The actual optim isation algorithm could either use the genetic algorithm or the
gradient search technique. It is often recom m ended to start the optim isation with the GA, and
then refine the solution w ith a gradient search.
Filter Type
Param eter
Recom m ended
D esign Techniques
Recom m ended
Opt. A lgorithm
Arbitrary
Filters
R esonant
Filters
Non-U niform
Transm ission Line filters
< 5
M ulti-Dim.
Cauchy M ethod
H ybrid optim isation with
C oupling M atrix
Tim e-Dom ain
Optim isation
GA and/or
G radient Search
GA and/or
G radient Search
G radient Search
5-40
5-100
Table 8.1: R ecom m ended Optim isation M ethods for D ifferent F ilter Types
Large resonant filters should be designed with our hybrid optim isation technique using the cou­
pling adjustments. Again, w e favour searching the param eter space with both G A and gradient
search.
F or the design o f non-uniform transm ission line filters o f large order, we advocate our timedom ain based optim isation. We recom m end using gradient-search for the param eter adjustm ents
only. GA’s are less favourable in this case. Since the param eter adjustm ents problem is very
linear, gradient search is m ore effective.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 8. S U M M A R Y A N D FU TU RE W O RK
8.2
127
Final Remarks
One conclusion o f our work is that a universal design technique for all filter types does not exist.
Rather, the synthesis is m ost efficient when geared to a specific filter type. The perform ance of
the design technique is the better the m ore inform ation about the particular filter type is presented
to the design routine. On the other hand, w e could also show that the choice o f the optim isation
algorithm (e.g. GA or gradient search) itself does not have a large im pact on the perform ance
o f the design process. We can state that not a specific optim isation algorithm but additional
inform ation can im prove the design process. This statem ent is known as the N o Free Lunch
Theorem o f Optimisation. A proof can be found in [64].
Consequently, w e m ust focus our attention not on the optim isation algorithm itself, but on how
to incorporate the m ost inform ation into the design process. A good exam ple is the circuit design
using neural networks. Standard “black-box” netw orks perform poorly for very specific optim i­
sation problem s. As shown in e.g. [131, 132, 133, 134], neural networks perform m uch better
when they include some fundam ental current-voltage relations relevant to the device. Including
this knowledge reduces the required data acquisition from several hundred to a few samples. The
sam e principle applies to space m apping, adjoint-netw ork m atrix method, and so on.
The four techniques introduced in this thesis follow this path. W hen we include background
knowledge in the genetic algorithm , it out-perform s gradient-based m ethods. O ur m ulti-dim en­
sional Cauchy m ethod includes know ledge by im plying that the circuit’s response is a rational
polynom ial.
H ybrid optim isation provides additional knowledge from the fine m odel to the
coarse model. O ur tim e-dom ain based optim isation m akes use of inform ation in the transient
response. K nowledge about the relationship betw een the response interval and the geom etrical
interval helps us guide the optimiser.
Incorporation o f knowledge about the circuit structure is the key to an efficient m icrowave cir­
cuit design. Throughout this thesis, w e have dem onstrated increased perform ance when using
know ledge-based algorithm s.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
CHAPTER 8. S U M M A R Y A N D FU TURE W O RK
8.3
128
Future Work
Filter synthesis is a very broad field o f research. In this thesis, we could only concentrate on a
few filter topologies and design techniques. The follow ing research topics can be considered for
future work:
• Som e applications require filter topologies that are not covered in this thesis. A ctive filters
are an example. Future research m ight concentrate on developing efficient techniques for
those filter topologies. In particular, w e w ould like to consider algorithm s that incorporate
know ledge about these filter topologies.
• As shown in C hapter 7, the theoretical results do not m atch the m easurem ents perfectly.
This is m ainly due to lim itations in the m anufacturing process and applied m aterials. We
will im prove this situation in tw o ways. First, we will concentrate on filter layouts de­
signed for easier m anufacturing. Second, w e w ill try to im prove the m anufacturing process
itself and the material.
• The software used here is still in the developm ent stage. At this time, the use o f tools
requires thorough understanding o f com puter program m ing and filter theory. It is desirable
to im prove the program s to the point o f integration into an actual engineering environment.
It is planned to incorporate these algorithm s into user-friendly C A D packages as part of
C O M DEV-proprietary m icrowave sim ulation tools.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
Appendix A
Theory of Space Mapping
Space m apping, first suggested by B a n d l e r , establishes a relationship betw een the coarse
m odel param eter vector and the fine m odel param eter. W e refer to the vector o f the param e­
ters for the fine and coarse m odel as p p and p c , respectively. The aim o f the space m apping
algorithm is to establish a m apping p — { p c '— >7 p f} (see Figure 2.7) such that
IIR
f
{p
f
)
— R c { p c )\\ < £
(A .l)
w here Ftp and R c is the response o f the fine and coarse m odel, respectively, and || || is a suitable
norm.
L et
p
(p
f
) be
defined as a linear com bination o f some predefined fundam ental fixed functions
flip),
M P),
fn(p)
(A. 2)
in the form
n
(A 3 )
A m apping is established when w e know the coefficients a. Finding the a ’s for the m apping
function
p (p f )
is an iterative process. In the first step o f the space m apping algorithm the
circuit is optim ised for an optim al response using the coarse model. The param eter vector o f the
optim ised layout is defined as p*c . The circuit is now sim ulated with the geom etric param eters
P q using the fine m odel, i.e. R(p*F = P c ) is calculated. This step is illustrated in Figure (a).
129
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
APP EN D IX A . T H E O R Y OF SPACE M APPING
-►*
•*C
(a)
coarse model space
fine model space
(b)
coarse model space
Pi
fine model space
-►*
(c)
coarse model space
Pi
fine model space
Pi
(d)
coarse model space
Pi
fine model space
Pi
Figure A. 1: Illustration o f Space M apping
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
131
APP EN D IX A . TH E O R Y OF SPACE M APPING
T he response R f (Pf ) plus som e additional responses o f points in the vincity o f p lF is now
com pared to corresponding coarse m odel responses. By param eter extraction a corresponding
P c and surrounding points are derived such that the error
\\Rc(pb)
-
R
f
( P f )\\
(A.4)
is m inim ised. From the tw o sets o f corresponding coarse and fine m odel points (i.e. their re­
sponses are alm ost identical in their respective spaces) a m apping as used in eqn.( A. 3) is derived.
_ii)
T he derivation o f the correct m apping function p is an iterative process. L et p p 'b e the /'th
iteration to approach the solution. The next iterate is found by a quasi-N ewton iteration
w here M ^is obtained from
B - h ® = -E ( j$ )
(A . 6 )
w here B_ is an approxim ation to the Jacobian o f the vector E with respect to p at the ith iteration.
The m atrix B is updated at each iteration using B royden’s update [75]. The error function E is
then evaluated from
p
(p
f
),
w hich can be found by a param eter extraction as described in [72]
and [73].
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
Appendix B
Coupling Matrix Algebra
B .l
Coupling Capacitor Representations for Band-Stop Filters
This appendix outlines a coupled resonator m odel that is dual to the inductive coupled resonator
m odel described in Section 5.2.1.
L et us assum e a netw ork with N resonators. The coupling between the resonators is achieved
using shunt capacitors as described in Figure B .l.
2
Figure B .l: Capacitive C oupled Resonators Z
There are
+ ^ u n k n o w n currents in the netw ork as shown. Using K irchhoff’s rules we can
find the N nodes equations:
132
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
APP EN D IX B. COUPLING M A T R IX A L G E B R A
*1
+
*21
* 12
+
*2
133
+ ■ ■• + *ijV
= iE
(B .l)
+ ' ' ’ + *2 iV = 0
(B.2)
:
= 0
(B.3)
I n i + *1V2 + • ■■+ *n
—0
(B.4)
w here we define i ki = - i ki and C mn = C nrn.
A nd also there will be ( ^
— y ) loop equations:
Z m im,
“
Z n in — 0
*mn
(B.5)
J iO L /m n
w ith all com binations o f m = 1 . . . N and n = ( m + 1 ) . . . N .
Eqn. B.5 can be rearranged to
imn =
j ^ C m nZ n i n .
(B . 6 )
Now, we can substitute all i mnterm s in eqn.( B.4)
*1
+ j u C w Z i h — iw (7 i2 ^ 2 * 2 4------- + j u C i N Z \ i i — J u C i n Z n i n
=
juCwZih —jujCnZ2i2 + *2 4----- + iwCiiv^2*2 —J+’CWZjyijv =
•■■ + ••■ + *jv
=
*£
(B.7)
0
(B . 8 )
0
(B.9)
0
(B.10)
using O hm ’s law i k Z k = Vk we get the adm ittance m atrix
(
rn
y
12
Y21
Y22
y 1n
• ■■
\
Y2N
/
V\
\
( iE ^
V2
0
(B .l 1)
0
\ Yn 1
Yiv2
• ■■ Y n n )
\ V N )
\
0
/
w here the m atrix elem ents are
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
APPEN D IX B. COUPLING M A T R IX A L G E B R A
Lmn
—
134
jw C n
N
Ymm — En "F ^ ]
jwCmk
k= l,k^n
for bandpass filter structures we define Ymas a resonator structure as shown in Figure and nor­
m alise such that
Figure B.2: Lossy Parallel R esonator Circuit
co L
ui
— + R — Zo
uC
.Wq
U>0
+ R = A
co
(B.12)
and then the full m atrix is
( x-jwE^iCik
-j w C i 2
-jw C x i
\
( )
- j w C 2N
~ jw C 21
V
-jw C iN
-jw C m
\
( i E )
v2
0
vs
0
VN )
10
)
(B.13)
B.2
Duality of Representations
The Z-representation o f the left circuit is:
The Y-representation o f the right circuit is
vi = (R +
+ j w ( L + M )) i i + M i 2
*! = (G ' + JUU + M C ' + £>')) «i + D ' V2
v 2 = M i \ + ( R + -Gjq + ju j(L + M ) ) i2
^ = (G ' + ^ U + M C ' + D' ) ) v x + D ' v 2
M ore general for a filter o f n th order the coupling m atrix and capacitor m atrix is
The Z-representation o f the left circuit is:
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
APPEN D IX B. COUPLING M A T R IX A L G E B R A
C
HI—
M
> H h
=L+M
135
R
czh
?L+M
Y -P re s e n ta tio n
Z - P re s e n ta tio n
F igure B.3: Duality o f C oupling Netw orks, On the left Coupling by Lum ped Elem ent Inductive
T-Network, on the R ight C oupling by Capacitive 7r-N etw ork
using A = R +
+ juL
( A - jui Y ,k =
-ju M
\
1
-ju> M i2
-J w M ijv
A-ju;E£LiM2*
—j u ) M 2N
M lk
21
—j u M j y i
—j w M N2
A -
\
J 2 k = i M ? ik J
(
h
(
\
VE ^
*2
0
*3
0
10
V lN )
)
(B.14)
The Y-representation o f the right circuit is
using A = G ' + j
+ ju C '
( A' - ju j Y ,k = i D i k
- ju jD - n
\
-ju )D Ni
-jw D n
—j u D i N
A' - j u Y ,k= i D 2k
—j u D 2N
—ju )D N 2
A ' - E f = i
D Nk
\
J
^
(
\
( IE ^
v2
0
Vi
0
VN )
10
)
(B.15)
This shows that the Z-presentation m odel is a dual to the Y-presentation m odel. In the dual
m odel with have to swap all param eters by
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
136
APP EN D IX B. COUPLING M A T R IX A L G E B R A
1
< ^ = f>
V
R
B.3
V
i
4 = f-
G'
L
C'
C
L'
M
D'
Coupling Values are Pure Real Numbers
W hen looking at arbitrary inductive or capacitive coupled resonators, w e m ay think the cou­
pling values m ust be com plex in order to reflect phase delays betw een the coupled resonators.
Nevertheless, com plex valued M -m atrices are not necessary. The reasoning is outlined here.
L et us recall the arrangem ent o f Figure B.3. The coupling network is an inductive T-network.
First, we assum e the elem ent values to be positive real. Thus, the T-network can create trans­
m ission phase shifts o f 0° to 90° in the com plex plane depending on the values o f M an L . As
seen from the equivalent circuit, the inductance L is absorbed in the resonator inductance and
will show up in the m atrix o f eqn. (B.14) as a non-zero value in the m ain diagonal. H ence, a real
valued positive coupling m atrix can represent a 0° to 90° phase shift in the coupling structure.
Further, w e are not lim ited to positive real values. N othing prevents us from choosing L and M
negative real. Then, we can create any phase shift with the proper choice o f L and M , while still
m aintaining a real coupling m atrix. Som e M -values m ay be negative though. Negative coupling
values can be observed in the num erical coupling m atrices throughout C hapter 5.
The theory is lim ited to narrow -band signal, because the frequency dispersion o f the T-networks
is not identical to the dispersion o f the real coupling networks, i.e. delay lines. In band-lim ited
signals (bandw idth<10% ), however, T-networks are a very good approxim ation o f delay lines.
The errors are usually less than 1%.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
Appendix C
Validity of Rational Function
Representation
The port im pedances o f any lum ped linear passive elem ent m icrowave circuit can be represented
as a rational function o f its elem ent values Inductance L , capacitance C , and resistance R , as
well as its harm onic operating frequency
uj.
T he reason for this is as follows: Any lum ped passive elem ents network can be represented by
a com position o f T-networks or 7r-networks as shown in Figure C .l. The decom position and
com position rules are described in detail in [39].
H
hJLH b
z3
T ---------- 0
Figure C .l: T-Network and 7r-Network
The Y-parameters (Y-matrix) as defined in [39] for a T-network is
PI
1
Z2 + Z3
Z \Z 2 + Z ‘i Z,\ + Z \ Z$
—Z 3
—Z
3
Z \ + z%
A nd for the 7r-network
137
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
(C .l)
APPEN D IX C. V A L ID IT Y OF R A T IO N A L FU N C TIO N REPRESENTATIO N
138
w here Y and Z o f lum ped linear passive elem ents are generally proportional or inverse propor­
tional to the frequency to and their respective elem ent values C , L , and R , i.e.
Z oc ju)
or
Z oc —
or
Z oc
j u j
C
or
Z oc L
or
Z oc R
(C.3)
Hence, w e can describe the m atrix elem ents o f [Y] or [Z] analytically by rational polynom ial
functions. T he polynom ial variables can be either the frequency or the elem ent values R , C , and
L.
Com posing the com plete circuit from cascaded T-networks and 7r-networks includes only the
basic algebraic operations o f addition, subtraction, m ultiplication, and division. Consequently,
the resulting transfer function rem ains a rational polynom ial function.
This explains, that all netw orks o f lum ped passive elem ents can be represented as rational poly­
nomials.
C.l
Distributed Elements
W e can use the rational function approxim ation also for a band-lim ited response representation
o f distributed elem ents. In that case the response o f the distributed elem ent m ust be approxi­
m ated using a Pade approxim ation. U sing this m ethod the response o f a band-lim ited signal can
be arbitrarily closely approxim ated. For exam ple, w e can approxim ate ta n (x ) by the rational
polynom ial
ta n (x ) «
1
x + ^-x 3
r, - 21—=— - .
+ 3r * 2 +
(C.4)
This Pade approxim ation o f ta n (:r) is shown in Figure C.2. Similarly, w e can approxim ate m ost
distributed circuit elem ent param eter, e.g. transm ission lines, stubs etc. The S-param eters of
such elem ents are functions o f the line lengths and line im pedance proportional to trigonom etric
functions, representable by Pade approxim ation.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
A P P E N D IX C. V A L ID IT Y O F R A T IO N A L F U N C T IO N R E P R E S E N T A T IO N
30
20
10
0
-10
-20
-30
•3
•2
-1
0
x
1
2
3
Figure C.2: Tangents o f x and Pade Approxim ation
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
Appendix D
Circuit Layouts
140
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
APP EN D IX D. CIRC U IT L A Y O U T S
D cfUUO
h
0.3/40
304 /
O.bBP/
3 3/b 4
3383
0. 8161
? 8130
Figure D .l: L ayout o f LC Low -Pass Circuit
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
APP EN D IX D. CIRC U IT LA Y O U T S
142
1.1127
0.0069
0,0069-
J.0069
1.00f?3 - 1
•- 0 0092
- 0 0 7 8 2 ------
0,0644
1 .0 0 69
0.0069 1. 1334
Figure D.2: L ayout o f sub-section o f six-pole filter
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
APPEN D IX D. C IRC U IT L A Y O U T S
~r
T -~ ^ — r
F igure D.3: L ayout o f Six-Pole filter
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
Bibliography
[1] S.F. Peik, Y.L. Chow, F ilter Synthesis with Genetic A lgorithm s and Constraints, Proceedings
o f the P IERS Conference 1997, Vol I, p. 229, 1997
[2] S.F. Peik, Y.L. Chow, Genetic Algorithm s A pplied to M icrowave Circuit Optim ization, Pro­
ceedings o f the A sian Pacific M icrow ave Conference 1997 , pp. 857 861, 1997
[3] S.F. Peik, R.R. M ansour, Y.L. Chow, M ultidim ensional Cauchy M ethod and Adaptive Sam ­
pling for an A ccurate M icrow ave Circuit Modeling, IE E E M TT-S Sym posium D igest 1998
Baltimore, pp. 21-24, June 1998
[4] S.F. Peik, R.R. M ansour, Y.L. Chow, M ultidim ensional Cauchy M ethod and Adaptive Sam ­
pling for an Accurate M icrowave Circuit M odeling, IE E E Transactions on M icrow ave Theory
and Techniques Volume 46 N um ber 12, pp. 2364—2372, D ecem ber 1998
[5] R.R. M ansour, S. Ye, S.F. Peik, B. Jolley, V. Dokas, T. Rom ano, G. Thom son, HTS Filter
Technology for Space A pplications, IEEE M TT-S Workshop on Comparative Filter Technolo­
gies fo r C ommunication System s, June 1998
[6 ] R.R. M ansour, K.L. W u, S.Ye, S.F. Peik, A. Sivadas, M odem Filter Synthesis, IE E E M TT-S
Workshop on C om puter-Aided D esign fo r Filter Synthesis, June 1998
[7] S.F. Peik, R.R. M ansour, H igh-Tem perature Superconductive Butler M atrix B eam form er for
Satellite Application, Proceedings o f the A N T E M Conference 1998, Ottawa, Canada, August
1998
[8 ] S.F. Peik, B. Jolley, R.R. M ansour, Super-Conductive Butler-M atrix B eam form er for Satel­
lite Applications, Proceedings o f the M TT-S Sym posium 1999, Anahein, CA, July 1999
[9] S.F. Peik, G. Coutts, R.R. M ansour, A pplication o f N eural Netw orks in M icrow ave Circuit
M odelling, Canadian Conference on Electrical and C om puter Engineering C C E C E 1998,
Waterloo, Ontario, M ay 1998
144
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
B IB L IO G R A P H Y
145
[10] S.F. Peik, Y.L. Chow, Full-EM Synthesis o f N onuniform M icrowave Transm ission Lines by
Tim e-Dom ain Reflectometry, Canadian Conference on Electrical and C om puter Engineering
C CECE 1998, W aterloo, O ntario, M ay 1998
[11] G.C. Temes, D. A. Calahan, C om puter A ided N etw ork O ptim ization the State o f The Art,
Proceedings IEEE, Vol. 55, pp 1832-1863, 1967
[12] G. M atthaei, L. Young, E.M .T. Jones, M icrowave Filters, Im pedance-M atching Networks,
A n d Coupling Structures, A rtech House, 1980
[13] G. G uglielm i, Sim ple CAD procedure for microwave filters and m ultiplexers, IEEE Trans­
actions on M icrowave Theory and Techniques, Volume: 42 7 1-2 , pp. 1347 -1352
[14] R. Levy, A Generalized Technique for Practical Distributed Reciprocal L adder Networks,
IEE E Transactions on M icrow ave Theory and Techniques, Vol MTT-21, pp. 519 525, Aug
1973
[15] C. Kudsia, R. Cam eron, W.C. Tang, Innovations on M icrowave Filters and M ultiplexing
Netw orks for Com m unications Satellite System s, IEEE Transactions on M icrowave Theory
and Techniques, Vol 40, pp 1133 1149, June 1992
[16] R. Levy, Filters with single Transm ission Zeros at R eal or Im aginary Frequencies, IEEE
Transactions on M icrowave Theory and Techniques, Vol. 24, pp. 172 181, April 1976
[17] R.R. Bonetti, A.E. W illiam s, New D esign Techniques for Coupled-line filters with trans­
m ission zeros, pp. 240-243, Proceedings o f the 23rd European M icrow ave Conference,
M adrid, 1993
[18] A. Zverev, H andbook o f Filter Synthesis, John Wiley & Sons, New York, pp 168-289, 1967
[19] F. Huang, Quasi-Transversal Synthesis o f M icrow ave Chirped filters, Electronic Letters,
Vol. 28, No. 11, M ay 1992
[20] H.C. H. Cheung, F. Huang, M. J. Lancaster, Im provem ents in Superconducting Linear
Phase M icrow ave D elay Line B andpass Filters, IEEE Transactions on A pplied Superconduc­
tivity, Vol. 5, No. 2, June 1995
[21] N. M arcuvitz, Waveguide Handbook, M IT Radiation Labratory Series, Vol. 10, M cGraw
Hill, 1951
[22] R. W. Rhea, H F Filter D esign and C om puter Simulation, M cGray Hill, 1994
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
B IB L IO G R A P H Y
146
[23] R. Saal, E. U lbrich, On the D esign o f Filters by Synthesis, IRE Transactions on Circuit
Theory, D ec 1958
[24] Van Valkenburg, M odern N etw ork Synthesis, John W iley and Sons, 1960
[25] E.A. G ullem in, Synthesis o f Passive Networks, John W iley & Sons, 1957
[26] M. Saito, Synthesis Procedure for general C oupled Resonator Transm ission N etw orks,
Electronics and Comm unications in Japan, Vol. 53-A, No. 5, pp. 2633, 1970
[27] D.M . Pozar, M icrow ave Engineering, Artech H ouse, 1993
[28] J.A.G. M alherbe, M icrow ave Transmission Line Filters, Artech House, D edham , M ass.,
1980
[29] R.E. Collin, Foundations fo r M icrowave Engineering, M cGray Hill, 1992
[30] S.B. Cohn, D irect-C oupled R esonator Filter, Proceedings o f IRE, Feb 1957, pp. 187 196
[31] R. Levy, Theory o f Direct-Coupled-Cavity Filters, IE E E Transactions on M icrow ave The­
ory and Techniques, Vol M TT-15, N o 6 , June 1967, pp 340 348
[32] L. Young, D irect-C oupled Cavity Filters for W ide and Narrow B andw idth, IE E E Transac­
tions on M icrow ave Theory and Techniques, M ay 1963, pp 162 178
[33] R.F. Harrington, Field Computation by M om ent M ethods, K rieger Publishing Company,
1989
[34] A. Wexler, Solution o f W aveguide D iscontinuities by M odal Analysis, IE E E Transaction
on M icrowave Theory Techniques, Vol. 15, pp 508-517, Sept. 1967
[35] T. Okoshi, P lanar Circuits fo r M icrowaves A n d Lightwaves, Springer Series in E lectro­
physics, Springer Verlag 1985
[36] T. Itoh, N um erical Techniques For M icrowave A n d M illimeter-W ave Passive Structures,
John Wiley, 1989
[37] J. Wang, Generalised M om ent M ethod in Electromagnetics, John W iley \& \Sons, 1991
[38] J.L. Butler, Digital, M atrix, and Frequency Scanning, in Hansen(ed.): M icrow ave Scanning
Arrays, A cadem ic Press, New York, 1966, C hapter 3
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
B IB L IO G R A P H Y
147
[39] K.C. Gupta, R. Garg, R. Chadha, C om puter A ided D esign o f M icrow ave Circuits, A rtech
House, 1981
[40] J.W. Bandler, Optim ization M ethods for Com puter-A ided D esign, IE E E Transactions on
M icrowave Theory and Techniques, pp. 533 552, Aug 1969
[41] A. W eisshaar, M. M ongiardo, A. Tripathi, C A D -oriented equivalent circuit m odels for rig­
orous full-wave analysis and design o f waveguide com ponents and circuits, IEEE M TT-S
Sym posium D igest 1996, Vol 3 , pp. 1455 -1458, 1996
[42] F. Arndt, F.; R. Beyer, J.M . Reiter, T. Sieverding; T. Wolf, A utom ated design o f waveguide
com ponents using hybrid m ode-m atching/num erical EM building-blocks in optim izationoriented CAD fram ew orks-state o f the art and recent advances, IE E E Transactions on M i­
crowave Theory and Techniques, Vol 45 5-2 , pp. 747 -760
[43] J.W. Bandler, S. Ye, R.M . B iem acki, S.H. Chen, D.G.Sw anson, M inim ax M icrostrip Filter
D esign U sing D irect E M Field Sim ulation, IEEE M TT-S Sym posium 1993, p. 889-892
[44] J. A. D obrow olski, Introduction to C om puter M ethods fo r M icrowave Circuit Analysis and
D esign, A rtech House, 1991
[45] S. Ye, R. M ansour, A n innovative CAD Technique for M icrostrip Filter D esign, IE E E
Transactions on M icrow ave Theory and Techniques, Vol 45, pp 780 - 786, M ay 1997
[46] RS. Barretto, Theory, C.A.D. optim ization and R F & M W circuits for space use, M i­
crowave and Optoelectronics Conference 1997, Linking to the N ext Century. Proceedings.,
1997 SBM O /IEEE MTT-S International Volume: 2 , pp 501 -508 vol.2
[47] N. Jain, P. Onno, M ethods o f using C om m ercial Electrom agnetic Sim ulators for M icrowave
and M illim eter Wave Circuit D esign and O ptim ization, IEEE Transaction on M icrowave The­
ory and Techniques, Vol 45, No. 5, pp. 724746, M ay 1995
[48] F. Allesandri, M. M ongiardo, R. Sorrentino, New Efficient fullwave Optim ization o f M i­
crowave C ircuits by A djoint N etw ork M ethods, IEEE M icrowave G uided Wave Letters, Vol.
3, pp 414416, Nov 1993
[49] F. Allesandri, M. Dionigi, R. Sorrentino, M. M ongiardo, A Fullwave CAD tool o f W aveg­
uide C om ponents U sing H igh-Speed D irect Optim izer., IE E E M TT-S Int. Sym posium Digest,
pp 15391542., M ay 1994
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
B IB LIO G R A P H Y
148
[50] P. M isra, K. N aishadham , Order-recursive Gaussian elim ination (O RG E) and efficient CAD
o f m icrowave circuits, IEEE Transactions on M icrow ave Theory and Techniques, pp. 2166 2173, Vol. 44 12 1 , D ec 1996
[51] J. Simkin. C. Trow bridge, O ptim izing Electrom agnetic Devices C om bining D irect Search
M ethods w ith Sim ulated A nnealing. IE E E Transactions on M agnetics, Vol. 28, pp. 15451548, M ar 1992
[52] D.A. Lowther, K now ledge-B ased and N um erical Optim isation Techniques for the D esign
of Electrom agnetic D evices, Int. Journal on N um erical M odelling, Vol. 9, pp 35-44, Jan-April
1996
[53] T.Sieverding, F. Arndt, C om bined circuit-/field theory C A D procedure for m anifold m ul­
tiplexers with circular cavities, 24th European M icrow ave Conference Proceedings, Cannes,
France, 1994, pp. 437-442
[54] P.C. Sharm a, K.C. Gupta; C.M . Tsai; J.D. Bruce; R. Presnell, Two-dim ensional field anal­
ysis for CAD o f R otm an-type beam -form ing lenses, International Journal o f M icrow ave and
M illim eter Wave C om puter A ided Engineering, vol.2, no.2; April 1992; p.90-7
[55] J. Purviance, M. M eehan, CAD for statistical analysis and design o f m icrowave circuits, In­
ternational Journal o f M icrow ave and M illim eter Wave Com puter A ided Engineering, v o l.l,
no. 1; Jan. 1991; p.59-76, 1991
[56] ____ , M odem perspectives in supercom puter-aided microwave circuit design, Interna­
tional Journal o f M icrow ave and M illim eter Wave C om puter A ided Engineering, v o l.l, no.2;
April 1991; p.201-23
[57] P. Pram anick, P.A. B hartia, A New Efficient CAD Technique for Inductive Reactance C ou­
pled W aveguide B andpass Filters, International Journal o f M icrowave and M illim eter Wave
Com puter A ided Engineering, vol.3, no.2; April 1993; p. 134-42, 1993
[58] X.J. Zheng, Three-dim ensional electrom agnetic sim ulation o f electronic circuits o f general
shape, International Journal o f M icrow ave and M illim eter Wave C om puter A ided Engineer­
ing, vol.4, no.4; Oct. 1994; p.384-95, 1994
[59] L.D. Khalaf, A.F. Peterson, Perform ance o f the sim ulated annealing and genetic algorithm s
for the design o f periodic devices, International Journal o f M icrowave and M illim eter Wave
Computer A ided Engineering, vol.7, n o .l; Jan. 1997; p .108-16, 1997
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
B IB LIO G R A P H Y
149
[60] Qi.J. Zhang, W. Fang, M. N akhla, Optim ization o f high-speed V LSI interconnects, a re­
view, International Journal o f M icrow ave and M illim eter Wave Computer A ided Engineering,
vol.7, n o .l; Jan. 1997; p.83-107, 1997
[61] Optim isation, J. Kowalik, M .R. Osborne, M ethods fo r U nconstrained O ptimization Prob­
lems, New York, Elsevier, 1968
[62] IM SL Library, Problem Solving Softw are System s, User Handbook, IM SL Houston, Texas
[63] J.W. Bandler, R.M . B iem acki, C ourse N otes Com puter Engineering 3KB3, Optim ization,
Vol. 1, January 1997, M cM aster University, D undas Ontario
[64] D.H. W olpert, W.G. M acready, N o Free Lunch Theorem s for Optim ization, IEEE Transac­
tions on Evolutionary Computation, V ol.l, No. 1, April 1997
[65] J.A. Nelder, R. M ead, A Sim plex M ethod for Function M inim ization, C om puter Journal,
Vol. 7, pp. 308313, Jan 1965
[6 6 ] R. Fletcher, M.J. Powell, A rapidly Convergent D escent M ethod for M inim ization, C om ­
puter Journal, Vol. 6 , pp. 163168, June 1963
[67] R. Achar, M. Kolbehdari, M. M akhla, A U nified A pproach for M ixed E M and Circuit S im ­
ulation U sing M odel-R eduction Techniques, IEEE M T T Sym posium D igest 1997, pp 1017—
1020, 1997
[6 8 ] M. Goldfarb and A. Platzker, The effects o f electrom agnetic coupling on M M IC design,
Int. Journal M icrowave M M -W ave C om puter-Aided Eng., vol. 1, no. 1, pp. 38-47, 1991.
[69] W. P. Harokopus, L. P. B. K atehi, E M coupling and radiation consideration in m icrostrip
(M )M IC design, IEEE Trans. M icrow ave Theory and Techniques, vol. 39, pp. 413-421, Mar.
1991.
[70] K. C. G upta and M. D. Abouzahra, Segm entation and desegm entation techniques, in A nal­
ysis and Design o f P lanar M icrow ave Components, Piscataway, NJ: IEEE Press, 1994, ch. 3,
pp. 75-86.
[71] S.P. Liu, C.K. Tzuang, Full-w ave segm entation analysis o f arbitrary shaped planar circuits,
IEEE Trans, on M icrow ave Theory and Techniques, vol. 45, pp. 1554-1562, Sept. 1997.
[72] J. W. Bandler, B iem acki, S.H. Chen, R.A. Grobelny, R.H. Hem m ers, Space M apping Tech­
nique for Electrom agnetic Optim ization, IE E E Trans, on M icrowave Theory and Techniques,
Vol 42, pp 2536 - 2544, Dec. 1994
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
B IB L IO G R A P H Y
150
[73] M .H . Bakr, J.W. Bandler, R. B iem acki, S.H. Chen, K. M adsen, A Trust Region Aggressive
Space M apping Algorithm for E M Optim ization, IEEE Transactions on M icrow ave Theory
and Techniques, Vol. 46, N o 12, D ecem ber 1998
[74] R.R. M ansour S. Ye, V. Dokas, B. Jolley, G. Thom son, W.C. Tang, C.M . Kudsia, D esign
C onsiderations o f Superconductive Input M ultiplexers for Satellite Applications, IE E E Trans­
actions on M icrow ave Theory and Techniques, July 1996, Vol 44, Part II, pp. 1213-1228
[75] C.G. Broyden, A class o f M ethods for Solving nonlinear Sim ultaneous Equations, M athmatical Computation vol. 19, pp 577-593
[76] R. S. Adve, T.K. Sarkar, S.M . Rao, E.K. M iller,D.R. Pflug, Application o f the Cauchy
M ethod for Extrapolating/Interpolating N arrow -Band System R esponses, Transactions on
M icrowave Theory and Techniques, Vol 45, N o 5, pp 837 - 845, M ay 1997
[77] K. Kottopalli, T. Sarkar, Y. Hua, E. M iller,G. J. Burke, Accurate Com putation o f W ideband
R esponse o f Electrom agnetic System s U tilising N arrow -Band Inform ation, IE E E Transac­
tions on M icrow ave Theory and Techniques, Vol 39, N o 4, pp 6 8 2 - 687, A pril 1991
[78] J.W. Bandler, Electrom agnetic Optim isation o f 3-D Stm ctures, IE E E Transactions on M i­
crowave Theory and Techniques, Vol. 45, No. 5, pp. 770777, M ay 1997
[79] C. Brezinski, Pade-type Approxim ation and General Orthogonal Polynomials , Basel,
Switzerland, B irkhauser Verlag,1980
[80] J. Stoer, R. Bulirsch, Introduction to N um erical Analysis, Springer-Verlag, New York, 1980
[81] W.T. Vetterling, N um erical Recipes, The A rt o f Scientific Computing, C am bridge Press,
1989
[82] T. D haene, J. Ureel, N. Fache, D. D e Zutter, Adaptive Frequency Sam pling A lgorithm
for Fast and Accurate S-Param eter M odelling o f G eneral Planar Stm ctures, IE E E MTTSym posium Digest, pp. 1427-1431, 1995
[83] S. Sakata, Partial Realization o f 2D discrete linear system and 2D Pade A pproxim ation and
Reduction o f 2D Rational Transfer Functions, Proceedings IEEE, Vol. 78, pp 604-613, April
1990
[84] J. Ureel, N. Fache, D. D e Zutter,P. Lagasse, Adaptive Frequency Sam pling A lgorithm of
Scattering Param eters Obtained by Electrom agnetic Sim ulation, IE E E A ntennas and Propa­
gation Sym posium , pp 1162 - 1167, 1994
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
B IB L IO G R A P H Y
151
[85] W.F. Richards, Y.T. Lo, D.D. Harrison, An Im proved Theory o f M icrostrip A ntennas with
A pplications, IEEE Trans, on Antennas and Propagation, Vol.29, N o .l, pp 38—46, January
1981
[8 6 ] T. Dean, J.A llan, Y. Aloim onos, Artificial Intelligence, Theory and Practice, B enjam inCum m ings Publishing Company, 1995
[87] I. Rechenberg, Evolutionsstrategie, Optimierung Technischer System e N ach Prinzipien der
Biologischen Evolution, Stuttgart-Bad Cannstadt, Germany, From m berg Holzboog, 1973
[8 8 ] C olom i A., M. D origo, V. M aniezzo. Genetic Algorithm s: A New A pproach to the Tim eTable Problem , NATO AS1 Series, Vol.F 82, C om binatorial Optimization, (Ed. M .A kguel and
others),Springer- Verlag, 1990
[89] J. Schaffer, L. Eshelm an, D esigning M ultiplierless D igital Filter Using Genetic A lgo­
rithm s, Proceedings o f the Fifth International Conference on Genetic algorithm s, University
o f Illinois,U rbana-C ham paign, Editor Stephanie Forster, 1993
[90] D.Pow ell, M. Skolnik, Using Genetic A lgorithm s in Engineering D esign O ptim ization with
N on-Linear C onstraints, Proceedings o f the Fifth International Conference on G enetic algo­
rithms, University o f Illinois, U rbana-Cham paign, Editor Stephanie Forster, 1993
[91] Y. Tanaka, A. Ishiguro, Y. Uchikawa, Genetic A lgorithm s A pplication to Inverse Problem s
in Electro-m agnetics, Proceedings o f the Fifth International Conference on G enetic A lgo­
rithms, University o f Illinois, U rbana-Cham paign, E ditor Stephanie Forster, 1993
[92] J.H. Holland, Genetic A lgorithm s and the O ptim al Allocations o f Trials, SIAM Journal o f
C omputing 2, p.
88
- 105, 1965
[93] .H. Holland, Adaption in N atural and Artificial Systems, A nn Arbor, University o f M ichi­
gan Press, 1968
[94] Z. M ichalew icz, Genetic Algorithm s + D ata Structures = Evolution Programs, Springer
Yerlag, 1992
[95] Goldberg J., Genetic Algorithm s in Search, Optimization and M achine Learning, AddisonWesley, 1989
[96] Krishnakum ar, K., M icro-G enetic Algorithm s fo r Stationary and N on-Stationary Function
O ptimization, SPIE: Intelligent C ontrol and Adaptive System s, Vol. 1196, Philadelphia, PA,
1989.
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
B IB L IO G R A P H Y
152
[97] D.L. Carroll, F O R T R A N Genetic A lgorithm (G A) D river ,Internet W eb page URL:
http://uxh.cso.uiuc.edu:80/~carroll/ga.htm l, D epartm ent o f A eronautical and A stronautical
Engineering , University o f Illinois at U rbana-Cham paign, Urbana, Illinois 61801
[98] E.O. Brigham , The Fast Fourier Transformation and Its Application, Prentice H all, New
Jersey, 1988
[99] P.P. R oberts, G.E. T o w n ,D esign o f m icrowave filters by inverse scattering M icrow ave T he­
ory and Techniques, IEEE Transactions on M icrow ave Theory and Techniques, Volume: 43 4
1-2 , Page(s): 739 -743
[100] K. Lu, T. Brazil, A system atic A nalysis o f H P8510 Tim e-Dom ain Gating Techniques with
Experim ental Verification, IE E E M T T Sym posium Digest, 1993, p. 1259 - 1262
[101] I. Eaves. J.L. Reedy, K. Edw ard, Principles o f M odern Radar, Van N ostrand Reinhold,
1987
[102] W.J.R. Hoefer, Tim e dom ain Electrom agnetic Sim ulation for M icrow ave CAD A pplica­
tions, IEEE Transactions on M icrow ave Theory and Techniques, Vol. 40, No. 7,July 199 pp
1517-27
[103] M. Forest, W.J.R. Hoefer, A Novel Synthesis Technique for Conducting Scatterers Using
T LM Tim e Reversal, IE E E Transactions on M icrow ave Theory and Techniques, Vol. 43, No.
6
, June 1995, pp 1371-78
[104] R Sorrentino, L. R oselli, P. M ezzanotte; Tim e Reversal in Finite D ifference Tim e D om ain
M ethod, IE E E M icrow ave and G uided Wave Letters, Vol. 3, No. 11, Nov 1993, pp 402-4
[105] P. Som lo, D. Hollway, M icrow ave Locating Reflectometer, Electronic Letters, Vol. 5, pp
468-^169, Oct. 1969
[106] Q. Gu, J. Kong, Transient A nalysis o f Single and C oupled Lines with C apacitively-loaded
Junctions, IE E E Transactions on M icrow ave Theory and Techniques, Vol34, pp 952 —964,
Sept. 1986
[107] C. Hsue, T. Pan, R econstruction o f N on-U niform Transm ission Lines from Tim e-Dom ain
Reflectometry, IE E E Transactions on M icrow ave Theory and Techniques, Vol.45, N o 1, Jan­
uary 1997
[108] F. Thom son, Transversal and Elliptic Function Filters on M icrostrip Utelizing HighTemperature Superconductive Thin Films, M aster Thesis, M cM aster University
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
B IB LIO G R A P H Y
153
[109] H. Kallm ann, Transversal Filters, Proceedings o f the I.R.E., July 1940
[110] K.M. Fidanboylu, S.M . Riad, A. Elshabini, A Novel Tim e Dom ain Synthesis Technique,
Proceedings o f M TT-S Sym posium , pp. 641-644
[111] K.L. Su, Time-Domain Synthesis o f Linear N etw orks, Prentice-H all, New Jersey, 1971
[112] L.Hayden, V.Tripathi, Characterisation and M odeling o f M ultiple Line Interconnections
from Time D om ain M easurem ents, IE E E Transactions on M icrow ave Theory and Technique,
Vol. 42, No. 9, Septem ber 1994
[113] H. Blinchikoff, A. Zverev, Filtering in Time and Frequency Domain, John W iley \& Sons,
New York, 1976
[114] N. Sifi, G. Angenieux, P. Ferrari, A C om bined A lgorithm fo r Optimizing M icrowave C om ­
ponents M odel in Time D om ain, IE E E Trans, on M agnetics, Vol 31, No 3, pp 1980-1983,
M ay 1995
[115] S. Sorey, A. Yang, Interconnect Characterization U sing Tim e-Dom ain Reflectometry,
IEE E Transactions on M icrow ave Theory and Techniques, Vol. 43,No. 9, pp 2151-2156,
Sept. 1995
[116] S. Burkhart, R. W ilcox, A rbitrary Shape Synthesis via Nonuniform Transm ission Lines,
IE E E Transactions on M icrow ave Theory and Techniques, Vol. 38, No. 10, pp 1514-1518,
Oct. 1990
[117] H -P Test & M easurem ent 8510-8720 N ew s, Vol. 9, No. 2, pp 6 -7 , O ctober 1998
[118] T.W. Pan C.W. H sue, M odified Transm ission and Reflection Coefficients o f N onuniform
Transm ission Lines and Their Application, IE E E Transactions on M icrowave Theory and
Techniques, Vol. 46, No. 12, D ecem ber 1998
[119] SN NS User M anual, Version 4.1, University o f Stuttgart,Inst, for Parallel and D istributed
H igh Perform ance System s (IPVR)
[120] S. Haykin, Neural Networks, Prentice Hall, 1994
[121] P.I. Somlo, D.J. Hunter, M icrow ave Im pedance M easurement, Peter Peregrinus Ltd., L on­
don, UK, 1985
[122] S.I. Gallant, N eural N etw ork Learning A n d Expert System s, M IT Press, 1993
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
B IB L IO G R A P H Y
154
[123] E. Yamashita, A nalysis M ethods fo r Electromagnetic Wave Problems, A rtech House, 1990
[124] A. Cichocki, R. U nbehauen, N eural N etw orks fo r Optimization and Signal Processing,
John Wiley, 1993
[125] ___ , Applications o f N eural N etw orks to M icrowave Design, IEEE M TT_S Sym posium
W orkshop, Denver, Colorado, June 1997
[126] P.M. W atson, K.C. Gupta, D esign and Optim ization o f CPW C ircuits Using EM -A N N
M odels for CPW Components, IE E E Transactions on M icrowave Theory and Texchniques,
Vol. 45, No. 12, Dec 1997, pp 2515 - 2523
[127] P. W atson, K. Gupta, R.L. M ahajan, D evelopm ent o f K nowledge B ased Artificial Neural
N etw ork M odels for M icrowave C om ponents, IE E E M TT-Symposium D igest 1998, pp 9-12
[128] P. B urrascano, M. Dionigi, C. Fancelli, M. M ongiardo, A Neural N etw ork M odel for CAD
and optim isation o f M icrowave Filters, IE E E M TT-Sym posium D igest 1998, pp 13-16
[129] P. B urrascano, M. M ongiardo, A Review o f Artificial Neural N etw ork Applications in
M icrow ave C om puter A ided D esign, International Journal o f R F and M icrow ave ComputerA ided Engineering, Special Issue on Applications o f Artificial Neural N etow rks to R F and
M icrowave D esign
[130] A. Zaabab, Q.J. Zhang, M. Nakla, A pplication o f Neural Netw orks in C ircuit Analysis,
IE E E Conference on Neural N etw orks, 1995, pp 423 - 426
[131] F. W ang, Q.J. Zhang, K nowledge B ased Neural M odels for M icrow ave D esign, IEEE
Transactions on M icrowave Theory and Techniques'!, VOL. 45, No 12, pp D ecem ber 1997
[132] A. Zaabab, Q.J. Zhang, M. Nakla, A nalysis and O ptim ization o f M icrow ave Circuits
& D evices U sing N eural M odels, IE E E M T T Sym posium D igest 1994, pp 393-396 1994
[133] F. W ang, Q.J. Zhang, Incorporating Functional Knowledge into N eural N etw orks, IEEE
International Conference on N eural N etw orks, Houston, TX, pp 266—270, June 1997
[134] F. W ang, Q.J. Zhang, Know ledge B ased Neural M odels for M icrow ave D esign, IEEE
M T T Sym posium D igest 1997, pp 627-630,1997
[135] A. Patniak, R. M ishra, G. Patra, S. D ash, An Artificial Neural N etw ork M odel for Effec­
tive D ielectric Constant o f M icrostrip Line, IE E E Trans, on A ntennas and Propagation, Vol
45, N o l l , p 1697 N ovem ber 1997
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
B IB L IO G R A P H Y
155
[136] S. Wu, M. Vai, S. Prasad, Reverse M odeling o f M icrow ave Devices using a N eural N et­
work approach, Asian Pacific M icrow ave Conference 1994
[137] K. Yale, Preparing the Right D ata D iet for Training N eural N etw orks, IE E E Spectrum
M agazine, pp 64—66, M arch 1997
[138] S. R osloniec, Algorithm s For C om puter-Aided D esign o f Linear M icrow ave Circuits,
Artech House, 1990
[139] T. P. Sheahen, Introduction to High-Temperature Superconductivity, Plenum Press, 1994
[140] Z.Y. Shen, High-tem perature Superconducting Circuits, A rtech House, 1994
[141] D. Roddy, Satellite Communications, Prentice H all, New Jersey, 1989
[142] R.R. M ansour S. Ye, V. Dokas, B. Jolley, G. Thom son, W.C. Tang, C.M . Kudsia, D e­
sign Considerations o f Superconductive Input M ultiplexers for Satellite A pplications, IEEETransactions M icrowave Theory and Techniques, July 1996, Vol 44, Part II, pp. 1213-1228
[143] ___ , IE E E Transactions on MTT, Special Issue on the M icrowave and M illim eter Wave
A pplications o f High-Temperature Superconductivity, Vol. 44, No. 7, July 1996
[144] R.R. M ansour, Targets o f US/Canadian HTS Space Program s, in w orkshop “Supercon­
ducting M icrow ave Applications to M obile C om m unications” , 1998 Asia-Pacific M icrowave
Conference Proceedings- A PM C 98, p. 340, D ec 1998
[145] Touchstone, H P-EEsof, Hew lett-Packard Co. Test & M easurem ent B usiness C enter 9780
S. M eridian Blvd. Englewood, CO 80112
[146] H P-8510 User M anual, H ew lett-Packard Company, 3000 Hanover St., Palo Alto, CA
94303
[147] OSA90/hope, Version 3.1, Optim isation Engine, O SA Inc. , Sept. 1995, P.O. Box 8083,
D undas, Ontario, Canada, L 9H 5E7
[148] Hewlett-Packard M icrowave D evelopm ent System (M DS), Version B .07.00, HewlettPackard Company, 3000 H anover St., Palo Alto, C A 94303
[149] Hewlett-Packard M om entum , Electrom agnetic Planar Simulator, H ew lett-Packard C om ­
pany, 3000 H anover St., Palo Alto, CA 94303
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
B IB L IO G R A P H Y
156
[150] Sonnet em, 2.5-D M om ent M ethod Solver, R elease 4.0, Sept. 1996, Sonnet Softw are Inc.
, 1020 Seventh N orth Street, Suite 210, Liverpools NY, 13088
[151] IE-3D, 2.5-D M om ent M ethod solver, R elease 4, Seland Software Inc. , 39210 A rgonaur
way, Suite 499, Frem ont, CA, 94538
[152] EM -Pipe, Optim isation link to Sonnet em and A nsoft H FSS, OSA, P.O. Box 8083, Dundas, Ontario, Canada, L9H 5E7
[153] EM -M U X, W aveguide Electrom agnetic Sim ulator, M ode M atching, proprietary to C O M
D EY
[154] Citius, Electrom agnetic Planar Sim ulator, M om ent M ethod, developed at the University
o f W aterloo
[155] A nsoft Em inence HFSS, Electrom agnetic A rbitrary Shape Sim ulator, Finite Elem ent
M ethod,
[156] M AFIA, integral equation solver, C ST Gm bH, Lautschlaeger Str. 38, D arm stadt, Germ any
[157] M icrowave Studio, integral equation solver, CST Gm bH, Lautschlaeger Str. 38, D arm ­
stadt, Germ any
R eproduced with perm ission of the copyright owner. Further reproduction prohibited without permission.
Документ
Категория
Без категории
Просмотров
0
Размер файла
7 951 Кб
Теги
sdewsdweddes
1/--страниц
Пожаловаться на содержимое документа