rights / license: research collection in copyright - non ...23266/eth-23266-02.pdfmethodsin optical...
TRANSCRIPT
Research Collection
Doctoral Thesis
Analysis, optimization, and synthesis of planar integratedlightwave circuits for WDM applications
Author(s): Spühler, Michael-Martin
Publication Date: 2000
Permanent Link: https://doi.org/10.3929/ethz-a-003856734
Rights / License: In Copyright - Non-Commercial Use Permitted
This page was generated automatically upon download from the ETH Zurich Research Collection. For moreinformation please consult the Terms of use.
ETH Library
Diss. ETH No. 13438
Analysis, Optimization, and Synthesis of
Planar Integrated Lightwave Circuits
forWDM Applications
A dissertation submitted to the
Swiss Federal Institute of Technology Zurich
for the degree of
Doctor of Technical Sciences
presented by
Michael-Martin Spühler
Dipl. El.-Ing. ETH/HTL
born April 30, 1969
citizen of Zürich and Wasterkingen ZH, Switzerland
accepted on the recommendation of
Prof. Dr. Werner Bächtold. examiner
Prof. Dr. René Dändliker. co-examiner
Dr. Gian-Luca Bona, co-examiner
2000
Contents
Zusammenfassung 1
Abstract 3
1. Introduction 5
1.1. Motivation 6
1.2. Analysis of Optical Devices 6
1.3. Optimization With Evolutionary Algorithms 6
1.4. Analysis and Optimization of Planar Lightwave Circuits.. 7
2. Analysis of Planar Waveguide Structures With the
Beam Propagation Method 9
2.1. Introduction 9
2.2. Theory of the Finite Difference Beam PropagationMethod 10
2.3. Mode Solving Using the Imaginary-Distance Beam
Propagation Method 12
2.3.1. Overview 12
2.3.2. Theory 13
2.3.3. The Optimal Parameters 16
2.3.4. Examples 19
2.4. Discussion 23
3. Design and Implementation of a Short Spot-Size
Converter on S1ON/S1O2 25
3.1. Introduction to Spot-Size Converters 25
3.2. Design 27
3.3. Implementation 31
3.4. Measurements 32
3.5. Summary 38
4. A Study of the Optimization Behavior of Evolutionary
Algorithms 39
4.1. Introduction ..39
4.2. The Spot-Size Converter 40
4.3. Structure Representation 42
4.4. The Evolutionary Algorithm 43
4.5. Results 47
4.6. Post Processing of the Evolution 50
4.7. Discussion 55
Contents
5. A Design and Simulation Concept for Planar
Integrated Lightwave Circuits 57
5.1. Introduction 57
5.2. The Concept of the Forward Solver 59
5.3. Details of the Forward Solver 61
5.3.1. The Waveguide Description 61
5.3.2. The Geometry Description 62
5.3.3. The Semantic Analyzer and the Netlist 63
5.3.4. The Scattering Matrix Compilation 66
5.3.5. The Waveguide Library and the WaveguideDatabase 68
5.3.6. The Eigenmode Solver 69
5.3.7. The Calculation of the Coupling Coefficients 69
5.3.8. Elementary Scattering Matrices 71
5.4. Examples 73
5.4.1. Resonant Couplers 73
5.4.2. Add—Drop Filters Using Ring Resonator Devices 76
5.5. Discussion ..79
6. An Optimization Concept for Planar Integrated
Optics ...,81
6.1. Introduction .81
6.2. The Optimizer Strategy 82
6.3. The Fitness Definition 83
6.4. Mutation Operators 84
6.5. Example: A Resonant Coupler Add-Drop Device 85
6.6. Discussion 89
7. Conclusions and Outlook 91
References 93
Acknowledgments 107
Curriculum Vitae 109
Publications ...Ill
Zusammenfassung
In Kommunikationssystemen werden immer häufiger
integriert-optische Elemente eingesetzt. Diverse Wellenleiter¬
technologien und -Herstellungsverfahren wurden in den letzten
Jahren entwickelt. Dieser Trend wird bestimmt noch einige Zeit
anhalten.
Die Entwicklung von optisch integrierten Schaltungs¬elementen ist eine anspruchsvolle Aufgabe. Der Computernimmt auch in diesem Gebiet eine immer wichtigere Rolle ein.
Dies nicht nur für die notwendigen Simulationen, sondern
insbesondere auch für die Optimierung neuer Strukturen. Da die
Konkurrenz gross ist, muss der Entwickler über Analyse- und
Optimierungswerkzeuge verfügen, welche es ihm erlauben,kurze Entwicklungszeiten zu erreichen.
In der vorliegenden Arbeit wurde gezeigt, dass die Anwendungunkonventioneller Entwicklung^- und Optimierungsprozedurenes erlaubt, Lösungen zu finden, welche vorher unbekannt waren.
Evolutionäre Algorithmen stellten sich als sehr effizientes
Werkzeug heraus.
Als Grundlage für die Simulation von optischen Strukturen
wurde eine verbesserte Methode zur Berechnung von Grund¬
moden und auch Moden höherer Ordnung von beliebigen Wellen¬
leiterkonfigurationen entwickelt. Diese Technik ist dann sehr
sinnvoll, wenn BPM (Beam Propagation Method) zur Simulation
von optischen Wellenleiterstrukturen benutzt wird. Sie erweist
sich als sehr genau und reduziert gleichzeitig die Rechenzeit in
den meisten Fällen.
Um die Verluste beim Übergang von der Glasfaser auf den
Chip zu reduzieren, wurde ein neuartiger Modenkonverter
entwickelt. Er besteht aus einem nicht-periodisch segmentiertenWellenleiter mit unregelmässigem Breitenverlauf. Die Struktur
ist nur 100 bis 140 um lang und reduziert die Verluste um mehr
als 2 dB pro Übergang. Zudem hat die Struktur sehr geringeReflexionen. Die Synthese des Modenkonverters wurde mit der
Hilfe eines effizienten genetischen Algorithmus gemacht.
Das interne Verhalten des für den Modenkonverter benutzten
Optimierungsalgorithmus wurde daraufhin analysiert. Es war
1
2 Zusammenfassung
damit möglich, den impliziten Parallelismus des Algorithmus zu
zeigen, indem gezielt die Evolution von Mustern, die
Subpopulationen bildeten, beobachtet wurden. Die Erkenntnisse
die daraus entstanden wurden dazu verwendet, ein Qualitäts-mass zu definieren, welches es erlaubt, die Optimierung zu
beobachten und zu beurteilen. Es ist damit möglich, die
Evolution laufend auf deren Effizienz zu prüfen.
Im letzten Teil dieser Arbeit wurde ein neues Analyse- und
Optimierungskonzept für planare integrierte Optik entwickelt.
Es basiert auf der Streumatrix Methode und beinhaltet drei
Repräsentationsebenen für die Strukturen. Der daraus
entstandene Vorwärtslöser ist sehr genau und schnell und bildet
die Grundlage für einen Optimierungsalgorithmus. Auch hier
wurde eine evolutionäre Technik verwendet, welche allerdings,wegen der geometrischen Natur der einzelnen Strukturen, einigeRestriktionen beinhaltet.
Als Produkt dieser Arbeit entstand eine benutzerfreundliche
Microsoft Windows® Anwendung. Sie enthält alle in dieser
Dissertation entwickelten Komponenten. Das Programm erlaubt
BPM Simulationen und Modenberechnungen und integriert auch
die entwickelten CAD Funktionen für planare optisch-integrierte
Schaltungen.
Abstract
The domain of integrated optics for communication systems is
a field that is growing rapidly. Many waveguide technologies and
fabrication techniques have been developed in the last few years.
This development will probably continue in the future.
Designing integrated lightwave circuits in this context is a
very demanding task. The computer becomes more and more
important in this field. Not only for the necessary simulations
but also for the optimization of new structures. Since the
competition is very intense, the designer must utilize analysisand optimization tools to shorten the design cycle.
In this work it was shown that using non-conventional designand optimization procedures, it is possible to find solutions that
were not known before. Evolutionary optimization procedures
proved to be very efficient for this task.
As a base element for the simulation of integrated opticalstructures an improved method to extract fundamental and
higher-order eigenmodes of arbitrarily shaped waveguides was
developed. This technique is very useful, when using BPM (Beam
Propagation Method) for the simulation of optical waveguidestructures. The results are very accurate, while the calculation
time is reduced in most cases.
For the reduction of the coupling losses at the interface
between optical fibers and integrated lightwave circuits, a novel
spot-size converter was designed. It has the form of a non-
periodic segmented waveguide with irregular tapering. This
structure is only 100 to 140 urn long, reduces the coupling losses
more than 2 dB per interface and has very low reflections. The
synthesis of the spot-size converter was completed with the helpof a breeder genetic algorithm.
The intrinsic behavior of the algorithm used for the
optimization of the spot-size converter was analyzed. It was
possible to demonstrate the implicit parallelism of the algorithm
by observing the evolution of patterns included in the optimized
population. Using this insight a new evolution quality figure was
proposed. This figure was used to observe and judge the
evolution process with respect to its efficiency.
3
4 Abstract
In the last part of this thesis, a new concept for the analysisand optimization of planar integrated lightwave circuits was
developed. It is based on a scattering matrix approach and
includes three representation levels of the structures. The
resulting forward solver is very accurate and fast. Based on this
forward solver an optimizer was built. It is based on an
evolutionary procedure with some restrictions because of the
geometrical nature of the individuals.
As a result of this work a user friendly Microsoft Windows®
based application was developed. All components that were
developed and used during this thesis are included. The
application allows BPM simulations as well as eigenmodecalculations, and integrates the developed CAD functionality for
planar integrated lightwave circuits.
1. Introduction
The rapid development of the information technology, the
global activity of the companies as well as the growing privatedesire for information require an increasing data transmission
capacity of the communication networks. With the invention of
the semiconductor laser and the introduction of optical fibers as
transmission medium for large distances, new doors were
opened. Only optical fibers have the necessary bandwidth for
modern wide-area networks (WAN). The available bandwidth of
the optical fibers, is orders of magnitude higher than the
capacity of today's electronic systems. In order to take advantageof the whole capacity of an optical fiber, one wavelength is not
sufficient and therefore multiple-wavelength systems (calledwavelength division multiplexed (WDM) systems) have been
proposed [1J. In WDM-systems the output of several lasers are
multiplexed onto one single-mode fiber [2]. The parallel channels
can be transmitted over the fiber without any interference.
Because of the possibility to amplify optical signals without
having to convert them to electronic signals, the idea of an ail-
optical network becomes reachable. Thus, the data is processedoptically wherever possible. The routing and multiplexing of the
data has to be solved in the optical domain, too. The data rate at
each electrical must be reduced, thus the higher rates would be
processed optically. To achieve this goal of optical processing and
routing, very selective optical filters are needed to separate the
individual channels in a WDM system. Today, filters are
implemented as integrated lightwave circuits. A crucial filter
device is the add-drop device [3] which performs the extraction
and insertion of one single wavelength without perturbing the
others, or demultiplexers which separate all channels at once [4,5, 6]. The advantage of add-drop filters is that a system needs
only to extract the channel which contains the required data
while the others are not affected. In this way the multiplexernode needs not to detect and decode all data that is transmitted
through the optical fiber.
o
6 Introduction
1.1. Motivation
Many problems are yet unsolved in the domain of opticaldevices. Actual optical engineering tools lack the efficiency or the
generality that would be necessary to design the optical
integrated devices. In this work several aspects are treated
within the context of WDM communication. Especially the
optimization problem is treated in some detail to show new
methods in optical engineering.
1.2. Analysis of Optical Devices
For the analysis of integrated optical devices, one of the
standard methods used is the beam propagation method (BPM)[7]. This method is able to calculate the propagation of the
optical field through a dielectric waveguide. In most implemen¬tations it is assumed that the beam propagates very close to the
optical axis and that radiation occurs only with small angles.Additionally only small differences in the refractive index may
occur in the propagation direction. These assumptions
theoretically restrict the applicability of the method. In practiceit can be shown that very accurate results may be obtained even
if the above conditions are not well met. In Chapter 2 an
introduction to BPM is given together with the introduction of a
mode solver using BPM [8, 9], and in Chapter 3 it is shown that
BPM is an accurate tool, even for complex structures. As an
example a novel spot-size converter [10] is demonstrated.
1.3. Optimization With Evolutionary Algorithms
Evolutionary algorithms [11, 12, 13, 14] are optimizationmethods which mimick the natural evolution process. When
applying evolutionary algorithms a thorough understanding of
the optimization behavior is required. Only if this process is well
understood, can the procedure be successfully applied to real
problems [10, 15, 16, 17]. It is necessary to have enoughinformation about the problem to optimize, to find an
appropriate representation scheme to be used by the
optimization algorithm. However, a precise idea of the possiblesolutions is not required. This knowledge can be acquired while
the optimizer is pursuing its way.
Introduction 7
In Chapter 3 the solution of the mentioned example, the mode
converter [10], is presented. The resulting structures from the
optimizer are very robust against fabrication tolerances and
other factors such as the cleaving position, and have better
performance than other structures to date.
In Chapter 4 an analysis of the optimization behavior [13] is
presented using again the example of the spot-size converter
introduced in Chapter 3. The insight gained in this study can be
used in many optimization problems in future.
The goals of any optimizing strategy are, first, a small overall
optimizing time, and second, to find a solution that is as close as
possible to the global optimum. Evolutionary algorithms help us
to approach both objectives.
It is not always favorable to find the real global optimum. This
solution might not be very robust against fabrication tolerances
and can possibly include a number of calculation inaccuracies. To
judge if a solution is robust against small variations of the
parameters, we can observe the population during the evolution.
If the population contains many individuals which have nearlythe same fitness and are close to each other in the parameter
space, then the optimization has very likely found a robust
solution.
1.4. Analysis and Optimization of Planar
Lightwave Circuits
In the real engineering world, almost every design procedureends up with an optimization task. In electronics many CAD
systems exist to help engineers to simulate and optimize the
circuits. This is partially true for optical lightwave circuits [18,
19, 20], but the optimization part is not yet well established.
Only very few optimizations are possible and those are usually
parameter optimizations. A topology optimization does not exist
(this is also the case in the field of electronic network analysis
tools). The Chapters 5 and 6 treat a simulation and optimization
strategy that is a big step towards real topology optimization. As
a special feature this system includes an inverse problem solver
for optical integrated lightwave circuits.
8 Introduction
Every inverse problem solver consists of two parts: the forward
solver and a strategy to optimize the result. In Chapter 5 the
forward solver for optical integrated lightwave circuits is
described. It is based on a scattering matrix approach. The
particularity of this implementation of the forward solver is that
it operates on three representation levels. Each of them having a
different level of abstraction. This enables an optimizer to use
information that would not be available in a standard implemen¬tation. The resulting forward solver proves to be very accurate
and robust. It can be used as a stand-alone simulation tool for
many applications.
In Chapter 6 the optimizer part of the system, which is based
on an evolutionary approach, is presented. It is able to transform
the topology of the lightwave circuits in order to obtain possiblebetter structures. Because the system can access information on
different levels of abstraction, it is also possible to optimize on
these different levels. The applicability of this concept is
demonstrated using two examples.
2. Analysis of Planar Waveguide Structures
With the Beam Propagation Method
The beam propagation method (BPM) has evolved to a
standard numerical analysis tool of integrated optics. The
method has an intuitive approach to the problem of propagatingelectromagnetic fields in dielectric waveguide structures because
it uses a scheme where the transversal scattering problem and
the longitudinal field propagation are solved separately (this is
called a split-step method). Therefore it is easy to implement the
method on any computer. This is the reason why it is so popularand many research groups are interested in developing improvedBPM-codes.
Nevertheless separate mode solvers have been required to
obtain eigenmodes that can be used as starting fields for the
BPM simulations. These eigenmodes obtained from sophisticatedsolvers may be very precise but they don't match the simulation
context (discretization of the waveguide structure and
propagation mechanism) of BPM. It would be useful to obtain
mode profiles that behave as 'real' eigenmodes in a BPM
analysis. In the second part of this chapter a technique is
presented to obtain eigenmodes (fundamental and higher order)using a special parameter set in the BPM analysis. The
calculated eigenmodes really behave as eigenmodes within the
BPM simulations when the same grid is used. They are
propagated along the real axis through the waveguide structure.
2.1. Introduction
The beam propagation method (BPM) is presently one of the
most widely used tools for the investigation of complex photonicstructures such as non-uniform waveguides, optical junctions or
directional couplers. BPM was introduced in the field of
underwater acoustics [21] and seismology in 1973 as the
parabolic equation (PE) method [22] before it was adapted to
optical problems by Feit and Fleck [7] in 1978. Since then a great
variety of different BPM solvers have been presented, includingFFT-BPM (fast Fourier transform BPM) [7, 23]. FE-BPM (finiteelement BPM) [24, 25] and FD-BPM (finite difference BPM) [26,
9
10 Analysis of planar waveguide structures
27, 28, 29, 30]. Both unidirectional and bidirectional [31, 32, 33]BPM codes as well as full vectorial codes were developed [27, 34,
35, 36, 37, 38, 39].
When using the beam propagation method, it is usuallyassumed that the light propagates along an optical axis and that
scattering occurs within small angles only. In this case the
computational cost can be reduced. When dealing with complexstructures like Y-junctions or bends several techniques exist to
overcome the problem of non-straight propagation. These include
conformai mapping [40, 41, 42, 43] and other grid transformation
techniques [44, 45, 46].
It is possible to use BPM as mode solver. Using the imaginary-distance BPM eigenmodes and effective indices of complex
waveguide structures can be determined. The corresponding
theory is presented in Section 2.3.
2.2. Theory of the Finite Difference Beam
Propagation Method
In this section an introduction to the finite difference BPM is
given. The propagation of the electromagnetic waves in an
inhomogeneous medium is governed by the vector wave equationfor the electric field
AE(x,y,z)-n2(x,yiz)k2E(x,y,z) = 0, (2.1)
where k = m-sfs^jUQ and n(x,y,z) is the refractive index of the
medium. In the paraxial limit, the Helmholtz equation can be
reduced to the paraxial wave equation [47] which is
-
., cE(x,y) û'Ei.x.y) à2E(x,y). r . .o ,ir. . /n oX
2jknQ —V^~ = j^- + ±^- + k n(x,y,zy -
n0 E(x,y) , (2.2)œ ex oy
L J
where w0 = ßfk is the reference index and ß is the propagation
constant of the fundamental mode of the waveguide. For
multimode waveguide systems the reference index can be chosen
as the mean value of the propagation constants of the modes.
The advantage of the paraxial approximation is the reduced
numerical effort because of the simplified operators. On the
other hand, the assumption is made that the field propagates
Analysis of planar waveguide structures 11
close to the optical axis. Several wide-angle codes exist [48, 49] to
overcome this problem where it is absolutely necessary.
The equation (2.2) can be rewritten as
.
dEj — = HE,ôz
(2.3)
where
H = -
2n0k L
•Q. "io o
+
ox" cy
+ (n2 -n2)kJ (2.4)
In FD-BPM (finite difference BPM), the partial differential
equation is replaced by the finite difference approximation,which yields
â2 E E(x - Ax,y,z) - 2E(x,y,z) + E(x + Ax, y,z)
ck" Ax~
c^E E(x,y - Ay,z) - 2E(x,y,z) + E(x,y + Ay,z)(2.5)
0' Ay2
Using these equations it is possible to calculate the
propagation of the electric or magnetic field through an arbitrary
waveguide structure. For this purpose the equation (2.3) is
discretized in z -direction. The equation may then be written in
an explicit, implicit or mixed form. The general equation is
E - E" » n-
= aTIEAz
n+\ (1 -a)HEn, (2.6)
where En represents the known field at any position z and En+l
the yet unknown field at the position z + Az. a represents the
implicitness parameter where a = 0 means explicit, a-\ means
implicit and a = 0.5 represents the Crank-Nicolson case.
The resulting sparse matrix can be inverted using different
standard methods [50, 51].
Because of the limited computational domain, appropriate
boundary conditions are to be used to avoid reflections and other
effects at the boundaries. Several efficient formulas have been
developed in past [52, 53, 54. 55].
12 Analysis of planar waveguide structures
2.3. Mode Solving Using the Imaginary-DistanceBeam Propagation Method
When replacing the propagation step by an imaginary value, it
is possible to extract eigenmodes for arbitrary shaped
waveguides. Higher-order modes can be computed when applyingthe sequential mode extraction [8, 56, 57]. This technique has
been used for several applications already [58, 59]. In this
section we present an extension to the standard method of
eigenmode-extraction using the imaginary-distance beam
propagation method. We show that higher-order propagationmodes of arbitrary shaped waveguide structures can directly be
extracted by propagating the field along the imaginary axis when
the parameters are chosen in an appropriate manner. This
method requires a starting guess of the propagation constant of
the eigenmodes. In many cases this value can be determined
using fast approximate techniques such as the effective index
method [60]. Additionally, the approximate mode shape may be
introduced as starting condition and can further accelerate the
extraction of the eigenmode. The overall number of propagation
steps needed to extract multiple eigenmodes is then significantlysmaller than in the case when extracting the modes sequentiallywith the method described in [57].
2.3.1. Overview
The imaginary-distance beam propagation method in its
original implementation has been successfully used to extract
fundamental modes [8, 34, 56, 58, 61] and higher-order modes
[57, 59, 62] of arbitrary shaped waveguide structures. The modes
are extracted sequentially starting with the lowest mode by
propagating the field along the imaginary axis, and for higher-order modes all previously calculated lower modes are subtracted
periodically during the propagation. This means that to compute
a higher mode, the complete set of lower modes is required. As a
consequence, the precision and stability of the solution are
dependent on the accuracy of all previous eigenmodes.
Additionally, when the interval between the effective indices
becomes small, it turns out to be very difficult and cumbersome
to extract further modes. Consequently many propagation stepsare required.
Analysis of planar waveguide structures 13
The new technique presented here is able to directly extract a
higher-order propagation mode by choosing an appropriate set of
parameters for the propagation. This has several advantages:
1. It allows the computation of a particular mode directly,without much knowledge of all other modes.
2. A higher selectivity is reached between two modes with
similar effective indices. Therefore, even the fundamental
mode is obtained more rapidly with a smaller number of
propagation steps.3. When extracting a set of modes, the standard procedure of
calculating the modes sequentially is further accelerated byusing the set of parameters introduced in this chapter.
In the following section we first present the theory behind this
technique. In Section 2.3.3 we show how the optimal parameterscan be found. Finally, we present some examples of waveguidestructures and eigenmodes calculated with the method.
2.3.2. Theory
The following development is done for the electric field
component E. The operators and results would be exactly the
same for the magnetic field component H. We start with the
scalar approximation of the field propagation equation (2.3) (as it
was developed in section 2.2)
j^- = GE (2.7)dz
where E = E(x,y,z) is the electric field distribution, z the
propagation axis and G the operator given in (2.8).
G= —J-^~+^T + k2(n(x,yy -w02)L (2.8)2/?0,
where w0 represents the reference refractive index, k is the
vacuum wave number, «(v,y) is the index distribution of the
waveguide. In the case of standard beam propagation z is a real
value. For the imaginary-distance beam propagation method, z
is substituted by jz, which yields the new equation
14 Analysis of planar waveguide structures
^= GE. (2.9)
dz
The general solution of (2.9) is a simple exponential function of
the operator G
E(x,y,z) = E(x,y,0)e6-', (2.10)
where E(x,y,0) now represents a starting field profile before any
propagation. This starting field can be freely chosen and can
always be represented by a weighted sum of eigenmodes of the
index distribution n(x,y)
E(x,y,0) = jrV'^V,}'), (2.11)
where E{,)(x,y) represents the eigenmode /' and c{,) is the
corresponding weighting factor. In general, the weighting factors
c{'] are complex numbers but this theory deals with real values
only, thus it is not dealing with evanescent field contributions.
We can now introduce equation (2.11) into equation (2.10) and
obtain
£(x,y,z)= IV'^%^^ (2.12)
where G{,) is now the operator for only one mode /', which may be
simplified by removing the derivatives in x and y, and where
the index distribution n(x,y) reduces to a single value «^
representing the effective index of the mode /
ôf"^(«)2~4 (2'l3)
Applying a generalized Crank-Nicolson scheme to the equation(2.9) we obtain
Ëjizï ~.A= a qE)i i+(\-a) GEn, (2.14)
Az
where En and Einl are the field distributions at the steps n and
n + l respectively, and a is the implicitness parameter as it was
introduced in section 2.2 (a = l means fully implicit). Solving(2.14) for EmX yields
Analysis of planar waveguide structures 15
En ii
l + (l-q)GAz
1-aGAzE,„. (2.15)
Using (2.12) it is possible to write
'l + (l-^)G('W«i-i
~~ 2-ii=\
l~aG(,)AzE]'] (2.16)
where E^ is the portion of the eigenmode i included in the field
En, and E^ =c{')E(,). Inspecting (2.16) we can interpret the term
within brackets as an amplification factor A(,) for each
eigenmode / propagating through the waveguide
1 + (1-ûOG(,)AzA -
l-aG{*Az(2.17)
a) b)
Figure 2.1. Graphs of the amplification factor A containing a pole and a
zero. In a) the parameters are chosen such that the fundamentalmode is near the pole. All other modes have a smaller
amplification. In b) another pai am eter set sets the mode 2 near
the pole. Higher as well as loiver modes have smaller
amplification in this case. Theiefore, the mode 2 can be
extracted. In both graphs the vertical lines represent the
amplifications of different modes alien the parameters a, nQ
and Az remain constant.
The function (2.17) is shown in Figure 2.1 and has the three
parameters a, n() and Az.These parameters can be chosen to
16 Analysis of planar waveguide structures
place one individual eigenvalue n^i into (or near) the pole of
(2.17). In this case, the corresponding eigenmode is much more
amplified compared to all others.
2.3.3. The Optimal Parameters
As stated in the previous section, the amplification factor
(2.17) has three parameters a, n0 and Az.In this section we will
develop the optimal parameter set to directly obtain any mode of
a given waveguide structure. To obtain a particular mode i it is
necessary to place the corresponding effective index at or near
the pole of the amplification factor A{!). Thus, the denominator of
A{,) of equation (2.17) is required to become zero which leads to
the equation
aG(,)Az = l. (2.18)
If we assume that the step size Az has to be positive, then, as
a consequence, G(0 (see (2.13)) must also be positive. The
condition
"o<"# (2-19)
results.
The best performance can be expected if both neighboringmodes (modes adjacent to mode /) have an equal (relativelysmall) amplification
Av-T)
In this case it is guaranteed that no other mode has a strongeramplification than the desired one. Solving equation (2.20) for a
positive Az yields
te-ZAzJlzm, (2.21)4/
with
f = a(l-a)G{'-"G(l l) and g-(la- 1)(g(M) + G('+1)). (2.22)
A\nl)A (,-])
= -Al (2.20)
Analysis of planar waveguide structures 17
As we will further show, the parameter a is restricted to the
range ]0.5,1[. From equation (2.17) we can see that if we choose
a to be 0 no pole is present. If a~\ then Az would become
infinite according to (2.21). To get a positive value for Az, a
must be larger than 0.5. In general we can assume that a is not
really a free parameter. It must be chosen such that the solver
provides the most stable results. When a is near 0.5 the step size
Az is already quite large (e.g., Az > 103) which usually produces
poor results. When a is near 1 (mostly implicit) the best results
have been observed but in this case the step size Az increases to
even larger values. A good tradeoff between the accuracy of the
solver and the mode selectivity is to use a value for a between
0.99 and 0.999.
From (2.13) and (2.17) it is easy to see that if n^ =n0, the
amplification A{m) =1. For simplicity, to get the mode i, we can
set n^-n^ + An (where An is an offset) and then use (2.21) to
obtain Az. Experience shows that An = (1[---y)(n{e'i)g -n^) is a good
range. The desired result is obtained when the eigenmode is not
directly on the pole and not too far away either.
Only these parameters ensure that the desired mode has a
higher amplification than all others and that the two adjacentmodes have the same smaller amplification.
To obtain Az out of (2.21), n^ must be replaced by an
estimate of the effective index of the mode m. The accuracy of
the value should be better than 1/3 of the effective index
difference between the wanted mode i and the nearest adjacentmode. If this condition is not met, either the convergence will be
very slow or a wrong eigenmode will come up. If, due to an
inaccurate estimation of the effective indices, two eigenmodesencounter nearly the same maximal amplification, no
convergence may occur at all.
After every propagation step Az the resulting field should be
normalized in power. The effective index of the field distribution
can be calculated as follows [63]
18 Analysis of planar waveguide structures
e'sk
^k2n2(x,y)\E('\x,y)c?E(n(x,y)
âx
âE{'\x,y)
âydx dy
§E('\x,yi\dxdy. (2.23)
The effective index given by equation (2.23) is only correct, if
the field distribution corresponds to an eigenmode of the
waveguide structure. Therefore, an accurate value of the
effective index is only available once the field solution has
converged.
The discretization of the waveguide structure has a majorimpact on the effective index value obtained with this method.
Especially structures with very thin layers less accurate values
may result when using a uniform grid. Usually sophisticatedinterface- or boundary conditions (e.g., Hadley's transparent
boundary conditions [52]) fail for imaginary (nonphysical)propagation. Therefore only a boundary condition like E = 0 may
be used, while the calculation window has to be chosen largeenough to prevent mode guiding due to the boundary conditions.
The starting field must contain a certain amount of the mode
to be calculated. In non-symmetrical waveguides this condition is
not very critical. On the other hand, in symmetrical waveguidestructures it is important that the start field contains parts with
the same symmetry as the searched mode. When the fraction of
the mode to be calculated is very small in the start field, the
method may take longer to converge. If an estimate of the mode
profile is available, the convergence can be very quick.
Additionally to the parameter set defined above it is possibleto subtract previously found waveguide modes (as in the
sequential mode extraction) and even get much faster
convergence.
The method presented here has the advantage that all a prioriknowledge available can be introduced to further accelerate the
mode calculation. Therefore, if all modes are needed, the
sequential extraction is still applicable and the calculation is
much faster when using the parameter set (a , ;?0 and Az ) that
was developed in this section.
Analysis of planar waveguide structures 19
2.3.4. Examples
As a demonstration and verification example we show the TM
mode extraction of a 5 jim by 6 um rectangular bar waveguidewith a core and cladding refractive index of 1.5 and 1.45
respectively. For this waveguide, approximate effective indices to
be used as starting values may easily be calculated with the
effective index method. In order to demonstrate the behavior of
the mode extraction we choose as starting field a superposition of
all propagating modes of the structure. For each mode the
optimal parameters are determined and used for the calculation.
Figure 2.2 shows the fraction of each mode through the
propagation. It is clearly visible that the desired eigenmode is
rapidly dominant over all the others which finally become
negligible. This works even for very high order modes near
cutoff. In Figure 2.3 the corresponding field distributions are
shown.
As a more complex benchmark example we chose the same
geometry that was used in [57]. There a directional coupler with
asymmetric waveguides was calculated with the sequential
technique. In Figure 2.4 the fundamental and second lowest
mode are shown, as it was calculated with the method presentedin [57]. After less than 10 steps both eigenmodes have already
converged, comparing to the more than 100 steps required in
[57]. The fundamental mode was not subtracted after each
propagation step and the same start field - a sum of two
gaussians of equal amplitude centered at both waveguides - was
used.
20 Analysis of pianar waveguide structures
sa
a)M
hhhhhhhh s7 !
0
o —' O N " N
CS O IN T-H
1 0
09
08
07
06
0.5
04
.03
0?
01
00
30
&25
fi 20
o
S 15Mca
P, 10ou
© O ^o ^ o
O <NCI O
H H H H HS...H H H
d)
S-3
o o ^o ^ o
O (N(N O
H
W4 >•—i *ÇH i^ ^ *^
H H H H H H
(N r-H
H
.3
O O <-*
O T-H O© tS <~-<
CS O CS
s-s
10
09
08
07
06
05
04
03
02
0.1
00
30
f*5CO
Ö20.2"§15M03
g,10(H
30 ;
25,
20
151 ;
10
5! I.
o o <-+ —< o es *-<
O r-i O -h (N O CS
hhhhhhhh-s;
(N ^2 o oo
o M
NON(N w
U, U [h [H L., L< £—< L, r^
ä;g
Figure 2.2. Using the parameters of Table 2.1 all propagating modes were
extracted. Each graph a) - h) shows vertically the fraction \c \
(see eqs. (2.11) and (2.12)) of every mode through the
propagation. The last column (residuals) shows the part of any
spurious field that has no correlation with any guided mode. On
a horizontal line the sum is always 1.0. In general the desired
mode becomes dominant after a few steps. From h) it can be
concluded that the mode TM 12 is not guided. The mode does
not converge at all and the residuals rise due to non guided,contributions. Therefore, the highest propagating mode is
TM21.
Analysis of planar waveguide structures jù X
Table 2 1 Parameters used for the mode extraction
Mode netf
a nn Az Figs 2.2 and 2 3
TMOO 1.4919 0.99 1.4851 3429 4 a)TM10 1.4818 0.99 1.4796 10784.1 b)TMOl 1.4787 0.99 1.4719 2340.6 c)TM11 1.4686 0.99 1.4667 10337 6 d)TM20 1.4658 0.99 1.4611 3446 8 e)TM02 1.4588 0.99 1.4549 5365 9 f)TM21 1.4532 0.99 1.4510 10155.6 g)TM12 1.4500 0.99 1.4486 15434 6 h) (Fig. 2.2 only)
Bi
5
-5
-5 0 5 Y |>m]
a)
TM 10 JE?:
5
-5
0 5 \ [um]
b)
TM20
/-"-N ^
! flti^ ^ -,r^
\
: t i;
\"
l1*O^
I
5
5 X[nm] -5 0 5 X[ixm]
e)
Ei
5
TM21
/""~^ ^~^\i y Ä '^
,f N
« S V
^
•« f^
u_ i \
-5 0 5 Y[pm]
g)
TMOl
-5 0 5 X[\im]
C)
TM02
aSÉsÊ?^ë^m
y
5 0 5 Xjurn)
f)
Figure 2 3 Propagating hvbnd EH modes as they were found with the
imaginary-distance BPM (7s(
- field of the TM approximation)
22 Anahsis of planar waveguide structures
b)
Figure 2.4. a) Normalized fundamental mode (E - field), b) normalizedsecond lowest order mode of a directional coupler structure Thisstructure is the same as in [5f. For both eases the effective indexand Die field distribution comerged after only 7 propagationsteps. The start field was the sum of two gaussians of equalamplitude centered at the centers of the two waveguides for bothcases. Inset plots shou the fast cornergence of the effective index.
Analysis of planar waveguide structures 23
2.4. Discussion
BPM is an excellent method to simulate the propagation of
electromagnetic fields through dielectric waveguide structures.
The limitations and the applicability are to be checked for every
case [64]. Together with the imaginary-distance propagation it is
possible to build a complete simulation platform around one sole
numerical method, however, one must be aware of the possiblenumerical instabilities of the mode solving method presented in
this chapter. Especially if the starting field contains parts of
evanescent modes (with imaginary propagation constants) the
mode solver may have a poor convergence due to the evanescent
parts which cause oscillatory contributions to the solution.
In many cases the presented eigenmode calculation method is
very fast, i.e., requiring little simulation time to converge. For
nearly symmetric directional couplers or poor starting field
conditions the convergence may be slower than with other
methods. In the second example presented in this chapter, the
simulation time was reduced by about a factor 10 in comparisonto the standard sequential technique even though the calculation
time required for one imaginary propagation step may be higher.
A further step would be to automate the mode calculation by
generating a guess of the eigenmodes and by calculating an
approximated effective index. Determining the eigenmodes of
arbitrary shaped waveguide structures would then be possiblewithout any user interaction.
"UTS' »i* ^ -*u*f ! v
*
3. Design and Implementation of a Short
Spot-Size Converter on SiON/Si02
In WDM systems [2, 4], various integrated optical devices have
to be developed and optimized towards a minimal loss. Such
devices are, e.g., couplers [65, 66], splitters [67, 68], filters (add-
drop-devices) [69. 70], multiplexers and demultiplexers [5, 71],and spot-size converters [10].
In this chapter transitions from fibers to integrated waveguidestructures are addressed. A spot-size converter [10], which is
optimized using a genetic algorithm is proposed [13]. The
implementation is described in this chapter. Measurements are
also shown.
In Chapter 4 the optimization algorithm used for the spot-sizeconverter is analyzed in more detail. The analysis gives insightto such algorithms and their intrinsic behavior. Only a full
understanding of these optimization techniques ensures
successful applications of the genetic algorithm.
3.1. Introduction to Spot-Size Converters
At the interface of integrated optical waveguide structures and
fibers, the problem of fiber-to-chip coupling arises because of the
mismatch between the spot-sizes. Similar losses are encountered
when the light is coupled onto the chip and when coupled from
the chip into the fiber. Planar optical lightwave circuits on silica
[4] with waveguides matched to single-mode fibers show
negligible loss for butt fiber-to-chip coupling, for certain
conditions. When using small effective refractive index contrasts
[2] or implementing oversized single-mode waveguides [72, 73]losses are small. However, the low refractive-index contrast
associated with such designs leads to bent waveguide structures
having radii of more than 15 mm. This considerably hinders a
further device miniaturization in devices such as phased array
waveguide multiplexers or ring resonators. Increasing the index
contrast allows us to reduce the bonding radius, but at the
expense of a reduced butt fiber-to-chip coupling efficiency. Byproperly designing and integrating spot-size or mode converters,
25
26 Short spot-size converters
this drawback can be overcome. Spot-size converters are
implemented to reduce the coupling losses when butt coupling
single-mode fibers to an integrated waveguide structure or to a
laser diode [74. 75, 76, 77, 78, 79, 80, 81, 82, 83]. The advantage
of device miniaturization or optical integration would be lost by
using long converters. Thus, compact spot-size converters are
required from a system point of view.
Several approaches to spot-size converters have been reportedin literature. Most of them implement combined laterally and
vertically tapered waveguide structures. Such structures have
been implemented on InP/InGaAsP [76, 77, 80], InGaAsP [79],
InP [81, 83], and InGaALWInAlAs [75], as well as on SiOa/SiON
[78]. They are designed to convert the field shape almost
perfectly into the fiber mode such structures. However, the
waveguide structures are difficult to integrate and requireadditional fabrication steps. Another technique uses a taperedfiber to reduce the coupling loss [82]. In industrial applicationssuch structures are not very practical. To integrate the spot-size
converters with the whole waveguide topology in only one step,
planar structures are required. Planar spot-size converters were
presented using periodic or quasiperiodic segmented waveguideseither with [84, 85] or without [86, 87] lateral tapering. Recently,another type of converter using a grating structure [88] was
reported. Such designs transform the mode shape to obtain the
fiber mode quite well, but the structures become very long (fromseveral hundred micrometers up to several millimeters).
In this chapter another approach to planar spot-size converters
is presented where the structures are very short, using a general
nonperiodic segmented waveguide approach with irregularlateral tapering [12]. Since these structures are short, small
losses result. Negligible reflections occur due to the
nonperiodicity and the low Fresncl reflections. The converters
were designed and implemented in a SiCVSiON materials
system [89, 90], as used for the fabrication of passive optical
add/drop wavelength division multiplexing filters [1]. For this
materials system, simple manufacturing procedures are
available.
Short spot-size converters 27
3.2. Design
The spot-size converter structure was designed using an
evolutionary optimization procedure (see also Chapter 4). Such
evolutionary algorithms prove to be very efficient especially in
the case of demanding combinatorial optimization problems, e.g.,
when having a large set of discrete parameters. The algorithmitself relies on a collective learning process within a population of
potential solutions comparable to the process of natural
evolution. The process of evolution is mimicked by using genetic
reproduction operators such as selection, crossover and
mutation. In this optimization scheme, better individuals
inherently have higher reproduction probabilities and will
therefore survive longer. Because evolutionary algorithms do not
necessarily require a well prepared starting point, they are well
suited for true synthesis tasks finding novel, unexpectedsolutions.
Segment S} Segment SN
L
14 14 118 8 0 6 0 0 0 0 5 5 5 5 03 0 0 0 0 0 0 4 4 0 0
»is.
Figure 3.1. The representation of the spot-size converters for the optimizer.The converter is divided into segments S{ to SK, which can
have any discrete value (within given limits) for the width. The
length of the segments is Ls = 3 :um.
For the purpose of the evolutionary optimization procedure the
structure is divided into a number of segments of equal length.Each segment represents a short waveguide with a certain width
that is coded by an integer between 0 and 21, where, for
example, 4 means the width of the original waveguide of 3 am.
Therefore a finite number of discrete values, in steps of 0.75 am.
can be created by the optimizer. A segmentation is obtained
28 Short spot-size converters
when the width is zero. Figure 3.1 shows this representation for
spot-size converters.
The representation was chosen to allow the total length of the
converter generated by the optimizer to be variable [12]. A
comprehensive description and analysis of the evolutionary
optimization procedure used here has been presented in [13]. A
possible resulting converter structure is shown in Figure 3.2.
Figure 3.2. A possible implementation of the spot-size converter. For
visualization purposes the upper cladding is omitted. The lower-
cladding is at Y < 0 (see also Figure 3.3 for more details on the
waveguide structure). A mode traveling in the + Z-direction is
widened.
The initial waveguide structure is a buried ridge waveguide as
shown in Figure 3.3. The waveguide structure was optimized for
small losses, reduced bending radii, and to be single-mode at
1550 nm wavelength. Upper and lower claddings are made of
Si02 with a refractive index of 1.45. The core consists of SiON
with a refractive index of 1.50. At 1550 nm, this results in a highlateral effective refractive index step of 0.02. When the
waveguide is segmented, the residual base layer - with the same
refractive index as the central core - remains. That layer
considerably hinders a full vertical expansion of the mode.
Short spot-size converters 29
Figure 3.3. Schematic representation of the waveguide geometry and
refractive indices at X = 1550 nm.
The field simulations were performed with a semi-vectorial 3D
finite difference beam propagation method (3D FD-BPM). The
coupling losses were calculated using an overlap-integralbetween the field at the output of the spot-size converter and the
fiber mode, including the radiation losses with an appropriatenormalization [12]. The evolutionary algorithm proved to be very
efficient in solving such problems. Good results were obtained
(less than 1.5 dB loss) after only about 1000 optimization steps.The best result of the optimizer (after 1132 optimization steps)was then optimized with respect to the segment length Ls.
An optimum for the coupling losses was found at a segment
length Ls of 2.7 am. The resulting structure is slightly shorter
than the original one and therefore has smaller radiation losses.
The final structure is shown in Figure 3.4, together with the
evolution of the coupling efficiency through the converter. A
minimal theoretical loss of 1.3 dB can be achieved with this
structure. Inspecting the optimized converter structure shown in
Figure 3.4, two different converter sections may be identified: A
first section (between 0 am and around 60 am) without
segmentation, i.e., only with changes in the waveguide width and
a second consecutive section, where true segmentation takes
place.
30 Short spot-size convert eis
20 40 60 80 100 120 140 160 180 200
Propagation distance / [jim]
/
go-Tjffo-o >"£ aoapoil
40 60 80 100 120
Propagation distance Z [|im"|
140 160 180 200
40 60 80 100 120 140
Propagation distance Z [fim]
180 200
Figure 3.4. Structure resulting from the evolutionary algorithm with
contour plots of the propagating field (top and center) The
converter structure starts at the position 0/urn The theoretical
mode mismatch (overlap between the propagating mode and the
fundamental fiber mode) is reduced from 3 3 dB to 1 3 dB
(bottom) The best conversion is achieved at the distance of140 urn, marked with a vertical line Ou. ing to the smooth curve
the structure may be cut at any position between 120jam and
150 jiim without a significant degradation of performance
Looking at the actual field pattern within the structure, this
classification is additionally underpinned by a different mode
treatment considering the two converter sections. The first
converter part performs a significant field expansion m the
horizontal direction whereas the segmentation m the second part
provides a vertical enlargement of the mode profile This sort of
functional partitioning is characteristic for all well performingcoupler topologies. The last periodic part of the converter was
Short spot-size converters 31
lengthened to show that the further guidance of the mode
provides only small changes in coupling efficiency. Thus, the
structure may be cut with a large tolerance of about 30 u.m
without degrading the performance more than 0.1 dB (Figure
3.4).
Reflections are estimated to be very low because of the small
difference in the effective refractive index between segmentedand non-segmented waveguides of about 0.02.
The theoretical direct butt coupling loss was calculated to be
3.3 dB. The performance with a spot-size converter is a couplingloss of 1.3 dB, which results in a coupling improvement of 2 dB
as shown in Figure 3.4.
3.3. Implementation
The SiON waveguide structures are deposited by PECVD
(plasma enhanced chemical vapor deposition) onto thermallyoxidized silicon wafers. Channel waveguides are formed byreactive ion etching. The single mode waveguide design is based
on the demand to obtain low-loss bends for radii as small as
1 mm. Small bending radii allow optical components, e.g.,
resonant couplers, to be designed on a smaller chip size and
complete optical building blocks to be integrated on one chip. To
meet this requirement the lateral effective index contrast of the
waveguide was chosen to be 0.02, resulting in a width of 3 am for
single-mode operation [89].
For a precise characterization of the modeled spot-sizeconverter, straight waveguides of 1.5 cm length were fabricated.
Every second waveguide ended with a converter structure. The
chip was diced using a wafer saw at the end of the converter and
the end facet polished to optical quality.
A direct comparison between a spot-size converter and an
adjacent conventional butt-coupled straight waveguide was
performed, thereby avoiding the influence of fabrication
inhomogeneities across the chip.
32 Short spot-size converters
>l-IMlOMNHHHh-
140 um
Figure 3.5. Optical microscope image of a 140-pm-long spot-size converter
structure as developed in photoresist. The rounded comers
visible in the zoomed section of the structure are due to the
lithographic process. The segment length is 2.7p.m.
In the zoomed section of the optical microscope image shown in
Figure 3.5, the waveguide segments are rounded due to the
lithographic process. This effect can even be advantageousbecause small, sharp corners are difficult to overgrow
conformally in PECVD. Experimentally, we found no negative
impact on the performance of the converter. To estimate the
influence of geometrical errors different versions of converter
designs were implemented, ranging from simple tapered
segments to those having sharp or rounded corners. All these
variations showed similar results. A theoretical analysis of
waveguide roughness and small-scale inhomogeneities [91, 92,
93] shows that such effects have little influence on the
eigenmode of the waveguide.
3.4. Measurements
Figure 3.6 shows a block diagram of the measurement setup.The light emitted from a 1550-nm laser source is guided throughan isolator to prevent the light from returning into the laser. The
single-mode fiber is butt-coupled directly into a straightwaveguide with a converter on the other end of every other
Short spot-size converters 33
waveguide. After coupling into a fiber, the light strikes a
detector. The signal is measured with a power meter.
Laser Source ->- Isolator
^)r
Converter
HHfMlSingle-mode Fibers
V —1É1ÉÉI
Power Detector ->- Power Meter1111111
Figure 3.6. Schematic of the measurement setup. The waveguide chip is
mounted on a translation stage which, permits the measurement
of adjacent waveguides. For reference purposes waveguides with
and without couplers arc implemented alternately.
The results of the measurements are shown as a comparisonbetween waveguides with and without a spot-size converter. As
shown in Figure 3.6, converters are placed on only one side of the
chip. The improvement of the coupling efficiency is greater than
2 dB per interface (see Figure 3.7).
Only small variations between consecutive waveguides were
measured. The devices have no significant polarization
dependence, which was to be expected because of the very small
difference in TE and TM mode shapes in addition to the small
converter length. No measurable difference between TE and TM
field expansion was observed. A very low wavelength dependencyof less than 0.1 dB in the range from 1.50 to 1.60 jim is expectedfrom simulations. Below the wavelength of 1.50 urn the originalwaveguide is no longer guaranteed to be single-mode.
34 Short spot-size converters
-1
O
Ph
ë-4
öCO
•5 -
PÜ
-6 -
-7
1 1 1 !— —
,• •
* ** ^
~M
*^ *
*
-
with coupler
+
.
-_t_- . -
+.-
+_ i
4- +
-
[
-L 1 _
without coupler
f , , ,
12 3 4 5 6 7 8
Sample
Figure 3.7. Comparison, between the coupling losses with and without spot-size converter. The improvement with converter is greater than
2 dB per interface, which agr-ees well luith theoretical
predictions. The vertical axis shows only a, relative difference.No absolute calibration was made. The solid and the dashed
horizontal line represent the mean value of the four best
structures with and without coupler respectively.
Figure o.<
1
0.9, n
CO
•3 0.8
4 0.7
-0.6
PÜ 0.5
a
P.0.4
o
Ü0.3
13^ 0.2
0.1h
-15
—with coupler
- without coupler
-5 0 5
X Offset [pm]
10 15 20
Measured horizontal fiber alignment sensitivity with and
without, spot-size converter. Using a, converter the horizontal
(and also the vertical) alignment is slightly less critical, because
of the widened field shape. In both cases the fiber is butt-coupledto the chip.
Short spot-size converters 35
One important factor is the horizontal and vertical alignmentsensitivity when coupling a fiber to the chip. The alignment is
slightly less sensitive with the converter (see Figure 3.8) as is
expected because of the decrease of the numerical aperture when
using a spot-size converter. The coupled power as a function of
the gap size between the fiber and the chip shows interference
effects in the first 20 urn with only small variations in the order
of 5% of the coupling efficiency. The Z -alignment for butt-
coupling is less sensitive than for coupling to a lensed fiber.
A further improvement of the coupling efficiency may be
achieved when using an index-matching oil. The difference was
measured to be 0.19 dB (the theoretical value is 0.18 dB) when
an index-matching oil of n = 1.5 is used. Therefore, when usingindex-matching oil, the improvement of the coupling efficiency is
further increased.
Single-mode Fiber
Figure 3.9. Setup for near-field measurements. At the output of the
converter a lens images the field intensity into a CCD camera.
The chip used for this measurement contained converter
structures shifted along the uavcginde with respect to each
other.
The propagating field intensity is compared to the simulations
with near-field measurements. For this purpose a series of
identical converters displaced along the waveguide at intervals
of several microns are used (see Figure 3.9). This allows us to
determine the field shape at several positions in the mode
converter on a single chip. To measure the near field, a
microscope lens was used to image the field intensity into an
36 Shoit spot-size convert eis
infrared CCD camera as shown m Figure 3.9. All elements are
specified for a wavelength of 1 55 (am
Simulated Measuied
^3 0
-4-Z= 15 um
-5 0 5 -5 0 5X[umJ a) X[|um]
-5 0 e) -5 0 S
X^m] b) X$m\
"5 X[?ml 5
c)"5 xSm] 5
X[um] d) X[um]
-5 0 5 -5 0 5X[um] e)
v
X[um]
Figure 3 10. The calculated intensit\ distribution (left) is compared with the
measured near field (right) through the converter structure The
intensities agree lerv uell with the simulations The fields are
shown for Z = 15, 65, 95, 135 urn and at the end of the
converter from a) to e), respecta ely
As shown m Figure 3 10 the measured near field intensity
compares very well with the calculated field profiles through the
converter.
Short spot-size converters 37
The reflections due to the spot-size converter were estimated
to be very low. A reflection measurement using a position-sensitive mterferometric tool (HP 8504B) with a resolution of
40 am confirms this estimate. Figure 3.11 shows a reflection
measurement performed on a double converter structure. The
maximum reflection level induced by a converter structure is
found to be about -40 dB, which is more than 20 dB smaller than
the Fresnel reflection that would occur if a lensed fiber were
used to couple to the waveguide. This measurement shows that
BPM is an appropriate tool to simulate such structures. This
assumption proves to be valid because the reflections are so
small.
[dB]
-10
-20
-30
-40
-50
-60
-70
-80
-90
START 73 5 mm 1300 nm n=1 5 STOP 74 167 mm
Figure 3.11. Reflection measurement on a double converter structure at
1300 nm This uavelength was used because of the higher
precision The second best optimized structure was used for this
measurement The maximum reflection of one structure is about
-40 dB A picture of the structure is overlaid to show the
positions of the reflections Following the inset reflection scheme,the slight asymmetry in the reflection signal is caused by an
asymmetrical optical excitation of the structure
When taking into account a loss of 0.5 dB/cm of the
waveguides, the absolute butt-coupling loss without spot-sizeconverter is 3.7 dB, whereas the coupling loss with spot-sizeconverter is only 1.6 dB, both using index-matching oil. These
eai
\
I I
"
Incoming light
<— «_D «D 4J> «pDistnbutc drcfle rtion
A
(\
\ f\f\A*A,
AAi 1/\»
yv Y *j \j1/vA/V
38 Short spot-size converters
values compare well with theoretical calculated couplingefficiencies of 3.3 dB and 1.3 dB, respectively. The main source of
discrepancy may come from a residual mode mismatch of the
calculated and real waveguide mode with respect to the optical
single-mode fiber. Nevertheless the spot-size converter results in
an improvement of 2.1 dB per interface.
3.5. Summary
A compact spot-size converter was optimized with an
evolutionary algorithm and implemented on high refractive-
index contrast Si02/SiON material as nonperiodically segmented
planar waveguide structure. Planar structures on the chosen
material are very cost-effective and may be fabricated using a
simple single-step lithography process. The evolutionary
optimization results in a converter length of less than 140 \xm,
which is very short compared to earlier designs. The fabrication
tolerances are larger than in classical designs. Measured
coupling losses and field shapes show very good agreement
compared to BPM simulations. A measured improvement of the
coupling efficiency of more than 2 dB per interface compared to
direct-butt coupling was obtained.
4. A Study of the Optimization Behavior of
Evolutionary Algorithms
In the Chapter 3, an evolutionary algorithm was successfullyapplied to the synthesis of an integrated spot-size converter.
Evolutionary algorithms proved to be well suited for the solution
of very complex problems having strongly nonlinear cost
functions defined over the solution space. They in most cases
work faster than other optimization techniques such as random
search or the Monte-Carlo Method because of their parallelsearch mechanisms, also referred to as intrinsic or implicit
parallelism [94. 95].
To obtain an understanding of the intrinsic behavior of the
optimization, the optimization of the spot-size converter that is
implemented as a non-periodic segmented waveguide structure
is further analyzed.
In this chapter, after a analysis of the intrinsic behavior of the
optimizer, an observation method is proposed, introducing an
evolution quality figure. This figure is used to visualize and to
qualify the evolution of the algorithm. Based on this figure a
termination condition is suggested.
4.1. Introduction
Evolutionary algorithms have been applied to solve very
complex problems [96, 97, 98]. Especially for problems includinga large number of discrete variables without any predefinedneighborhood relationship and an associated nonlinear, unsteadycost function [12], these methods show an efficient behavior and
do not stick in local optima as certain other optimizationtechniques. However, special care has to be taken when choosingof an appropriate representation and the optimizationparameters. A general overview for genetic algorithms used in
electromagnetics is given in [99]. In the previous chapter, an
evolutionary algorithm based on a breeder genetic algorithmscheme was applied to the design of an optical spot-sizeconverter. An approach using non-monotonic lateral taperingand non-periodic segmentation was used. In order to obtain all
39
40 A study of evolutionary algorithms
possible solutions, the chosen representation is kept as generalas possible. The corresponding analysis method is able to
evaluate the performance of all possible structures that can arise
during the optimization process.
If the evaluation of the 'cost' or 'fitness' value takes a
considerable amount of computation time one would be
interested to see at an early stage, whether a good repre¬
sentation and appropriate parameters have been chosen.
Statistically available information concerning a final state of a
population's evolution does not correctly represent the
optimizer's potential for a further improvement. What is missingso far is a tool that enables the possibility of optimizersupervision. Therefore, based on the data obtained during an
optimization run a post processing' procedure is applied. There
an evolution quality figure is defined which is used to indicate
the progress and the termination point of the optimizer. Such a
technique is useful, when the supervision time is negligiblecompared to the total computation time of the optimization.
4.2. The Spot-Size Converter
Spot-size converters are used to reduce the coupling loss when
connecting a single-mode fiber to the integrated waveguidestructure. Different approaches to mode converters have been
reported. The horizontal expansion of the field is achieved with a
lateral tapering of the integrated waveguide. Vertical expansioncan either be obtained with a vertical tapering of the waveguide[81] or by implementing a periodic [84, 85, 87] or non-periodic[12] segmentation. x\ simplified explanation of the effect of a
periodic segmentation is an averaging of the effective indices in
the segmented and not segmented waveguide parts that gives a
broader fundamental mode profile.
To enable simple manufacturing vertical tapering is not
included in this design because of the required additional
fabrication steps which would result in higher costs and more
elaborate equipment.
The structure is integrated on silicon [89], with a waveguide of
3am width and 1.94 um height (see Figure 3.3). The core
material consists of SiON with a refractive index of 1.50. The
upper and lower claddings are made of Si02 with a refractive
A study of evolutionary algorithms 41
index of 1.45. This results in a refractive index contrast of about
0.02. A waveguide structure with a relatively high index contrast
has the advantage of allowing small bending radii. This
facilitates
devices.
the further miniaturization of optical integrated
15
10
a
-10
Air
-15
Core
Upper Cladding /^gjj^X £„
Lower Cladding
Silicon Substrate
¥,
15 10 0
X[iim]
a)
0Y [Urn]
b)
Figure 4.1. a) The fundamental mode o\ the integrated waveguide. The
structure of the waveguide is also shown h) The Gaussian
shaped eigenmode of the fiber together with the view of a large
integrated waveguide. Because of the field parts overlapping
into the air or silicon an optimal mode conversion cannot be
obtained.
The narrow eigenmode intensity profile of the integrated
waveguide has to be adiabatically converted into a field
distribution as close as possible to the eigenmode of the single-mode fiber (see Figure 4.1). A residual base layer which has a
height of 0.64 um (see Figure 3.3) and the same refractive index
as the central core exists. Such a layer considerably hinders the
vertical expansion of the propagating field. This layer is also
present where the waveguide is segmented. Therefore it is not
possible to obtain an optimal eigenmode of the fiber using such a
design. Furthermore the total vertical extent of the waveguidestructure is not large enough to allow an adiabatic conversion of
the field into the eigenmode of the fiber. As shown in Figure 4.1
b), parts of the fiber mode overlap into the air and into the
silicon, even if we allow verv large waveguides.
42 A study of evolutionary algorithms
The eigenmode of the integrated waveguide was calculated
using the imaginary distance beam propagation method [8]. An
example of such a planar spot-size converter structure was
shown in Figure 3.2.Ltoi
4.3. Structure Representation
The representation of a general converter structure is chosen
to allow different values for the total length of the converter and
arbitrary variations of the segmentation of the rib. Therefore the
length and the width of the structure are discrotized. Now a
converter can be represented as a fixed length string of multi¬
valued bits {£r..S^}. Each bit S, indicating the property of one
segment. Every segment can have a value from -1 up to Nw,
where Nw represents the code for the largest possible segment
width. A zero value means that the ridge of the waveguide is
omitted. A special value (—1) is used as don't care representing a
segment of zero length and zero width, e.g., segments with this
value are non-existent in the final converter and are used to
allow structures with a variable length when using fixed lengthmulti-valued bit strings (see Figure 4.2). The width Wt of the
segment number / is then calculated as
don't care, S, =-1
^ = io, S, =0 , (4.1)
(S,-\)AW+W0> S,>0
where W() is the smallest width of the waveguide and AW
represents the step size. The length of the segment Ll of the
segment / can be determined according to
L,=\'
,
4.2)'
[Ls, S,>0'
where Ls stands for the length of a segment. A converter can
therefore have the maximal length of A\t Ls.
A study of evolutionary algorithms 43
,BitS, Bit Sf
1 0 0 4 4 0 2 0-1-10 2 -1 2 2 0 8 8 -i -i o o|TJT|o 0 6
Lv — 3 urn«—-
8 8 64 4 0 2 0 0 2 2 2 01 0 0 0 0 1 1 0 0
Figure 4.2. Representation example of a converter structure (top). The
physical structure is obtained by removing the don't care's prior
to the simulation (bottom).
4.4. The Evolutionary Algorithm
An evolutionary algorithm uses a population where each
individual represents one point m the search space. Each point of
the search space may be reachable. No physically invalid
parameter combination may be generated by the optimizer. To
ensure this, problem related constraints and rules are included.
For the case of the mode converter, the values of W0 and Ls must
be chosen correctly in order to guarantee that all possiblestructures are realizable. These two values do not have any
impact on the general representation of a converter structure for
the evolutionary algorithm.
44 A study of evolutionary algorithms
/ : maximal number of iterations;
Initialize Population P(0);
Evaluate Population P(0);
i=0;
While (i<l and termination condition == False)
{
Select two different individuals indi, inch, out of'P(i);
Crossovetfindi, ind2) => offspring!, offspring.?
Mutate(offspringi);
Mutate(offsprings) ;
if (Valid(offspringi)) Evaluate fitness ofoffspringi;
elsefitness ofoffspringi = 0.0;
if (Valid(qffspring2)) Evaluatefitness ofoffspring?;
elsefitness ofoffspring?. = 0.0;
Sort(Population P(i) u { offsprings offsprings };
Reject worst two => P(i+1);
i=i+l;
Figure 4.3. Pseudo-code of the breeder genetic algorithm. The functionValidate checks if the genotype of the individual (as an
argument of the function) is already contained in the
population.
The evolutionary algorithm used here is based on a breeder
genetic algorithm scheme [11] and is shown in Figure 4.3.
A heuristic is used to initialize the population generatingindividuals containing concatenated blocks with LB segments of
constant waveguide widths. Block lengths LB are set to values
between LBmm and LBmx.. The corresponding value of the width is
chosen according to given probabilities Ps :
Pn
P>o
= the segment is set to don't care,
= the segment is set to zero width, and
= the segment is set to any width from 1 to N.
After generating and evaluating all Np initial individuals, they
are sorted according to their fitness.
A study of evolutionary algorithms 45
Two bit strings are selected from the population, the
probability of selection being an increasing function of fitness.
Here the so called 'roulette wheel selection' [100] is used.
Two offsprings are generated by exchanging a part of the bit
string between the two parent structures. The part to be
exchanged is determined by selecting two positions on the stringwith a uniform probability. After this 'crossover' operation, the
two offsprings are mutated by reinitializing the bit string at
every location with a very small probability.
Figure 4.4. Simulated field propagation through a converter. Left of the
dashed line the width of the original waveguide is shown. The
propagation steps are 0.3pm and the discretization in X and
Y are 0.25pm. A horizontal slice of the waveguide is
superposed. The expansion of the propagating field is clearlyvisible.
The propagating field through the converter structure is
calculated using a 3D FD-BPM (finite difference beam
propagation method) program. An example of a propagating field
through a converter structure is shown in Figure 4.4. The fitness
of a structure is calculated after each propagation step of 0.3 urn
using an overlap-integral between the propagating field and the
fundamental mode of the fiber. The integral has the followingform:
46 A study of evolutionary algorithms
F(z) =
\fra(xiy)z)%{x,y)dA
JJi % (x,y) f dA |J| Tr (x,y) f dA(4.3)
where X¥I is the fundamental mode of the waveguide, T^ the
fundamental mode of the fiber and vVa the propagating field
through the converter (see Figure 4.1). By using the double
integral J |jxr,/(x,y)|"c//l as normalization factor, the radiation
A
losses are taken into account. The initial field x¥l for the BPM
simulation is calculated once prior to the optimization. The
highest value obtained along the structure is returned as the
overall fitness. A possible fitness function F(z) through such a
converter is shown in Figure 4.5. The final converter structure is
cut at the location achieving the highest fitness value. Therefore,
a physically fabricated converter may be significantly shorter
than the structure defined by its multi-valued bit string (verticalline in Figure 4.5).
180
Propagation Distance [jam]
HhH11" \\ "I
Figure 4.5. The fitness evolution through the converter is shown here. The
real structure will be cut at the position where the highest
fitness is obtained. Therefore the implemented converter is
usually considerably shorter than the total structure. The fitnessis calculated after each BPM propagation step. The best fitness
ever encountered (here at about 110 urn, shown by the vertical
line) is retained as the overall fitness of the converter.
A study of evolutionary algorithms 47
The algorithm differs from a conventional genetic algorithm in
the following elements:
Every new individual is checked whether it is alreadycontained in the population. Allowing no duplicates avoids
premature convergence and is advantageous for the algorithms
finishing behavior because the population never fully converges.
In [101] the authors also reported a large improvement of the
effectiveness allowing only one copy of any bit-string at any time.
Only better individuals than the worst contained in the
population are inserted, e.g., a strict breeding is done. New
individuals are immediately available for the next recombination
step. This strategy is therefore a kind of multi-point hill climbingwhere only better individuals are introduced into the population.The best points found are kept and the search is kept as diverse
as possible.
The use of the combined crossover/mutation operator takes
into account that in a population with a high diversity, the main
effect of disrupting the structures will be due to the use of
crossover. The more the population converges to a narrow fitness
distribution the more a local search will be carried out where the
main effect will be due to the mutation operation which
reintroduces new information at randomly chosen locations in
order to achieve a fine tuning of the converged population.
Offsprings are immediately available for recombination. The
algorithm therefore has the opportunity to exploit a promisingindividual as soon as it is created.
This algorithm is described in more detail in [97] and has also
been applied to other problems.
4.5. Results
For the runs of the algorithm the population consists of 100
structures. The following parameters were taken for the
optimization: N = 70 segments of A\ = 3 urn length, a minimum
segment width of W0=1.0 urn and width steps of AW =0.5 um.
This representation reflects the manufacturing conditions for
such structures. The block length range for the initialization is
defined by LBmm =3 and LB = 7. The width-coding range for the
48 A study of evolutionary algorithms
S, is [-1. 40], giving a solution space of 42 =424-10in possible
converter structures Some of the real converters may be
identical - due to the don't cares - and therefore the physicalsolution space is slightly smaller but still exceptionally largeThe propagating step size for the BPM simulations was chosen to
be 0.3 urn and a grid with Ax = 0.25 am and Ay = 0.25 urn was
used.
Simulated overall coupling losses of about 1 3 dB were
obtained which represents a theoretical improvement of the
coupling efficiency of 2 dB This was achieved by evaluating10350 individuals requiring only one structure out of 10107 to be
calculated. Other optimization methods require a much highernumber of evaluations to achieve a similar result
Figure 4 6 Microscope photograph of a spot-size conter ter The resist is
shown here The corners are clearly rounded due to the
photolithography process and in part due to the restricted
resolution of the optical microscope Indeed, this effect has no
impact on the perform an ce of the converter [10, 15]
Two resulting structures have been implemented on a silicon
substrate and were optically characterized An example of an
implemented spot-size converter is shown in Figure 4 6 In
general, the genetic algorithm will not find the absolute globaloptimum in the search space This would even not necessarily be
a good achievement, because the very optimum might not be
very robust against fabrication tolerances and could be achieved
A study of evolutionary algorithms 49
due to a number of numerical side effects. Solutions found by the
algorithm turned out to be very robust.
i l ' ' l l
0 2000 4000 6000 8000 10000 12000 14000
Evolution Steps
Figure 4.7. The évolution of the population is shown. The top curve shows
the best fitness as a function of the evolution steps. The lower
line represents the lower fitness boundary of the population. For
every evaluated individual a dot is plotted. When a structure
has a lower performance than the lower fitness boundary it is
not inserted into the population
The evolution of the optimization is shown in Figure 4.7. The
upper line represents the fitness of the best individual and the
lower line is the fitness of the worst converter contained in the
population. For each evaluation a dot is plotted. As shown in
Figure 4.7, bad individuals may be created at every stage of the
optimization.
Compared to the length of a conventionally designed converter
(usually 800 urn up to several 1000 urn) these optimizedstructures tend to be very short (generally less than 150 urn).
Thus the requirements are met with the additional benefit that
the radiation and absorption losses are lower in shorter
structures. The error introduced into the calculation by the fact
that BPM does not take reflections into account is very small.
The reflections were measured to be m the order of -40 dB. This
small value can be explained by the non-periodicity of the
structure and the relatively small refractive index difference of
0.02.
50 A study of evolutionary algorithms
4.6. Post Processing of the Evolution
Evolutionary optimization procedures also provide overall
information about a possible solution strategy. One of the main
differences between classical optimization procedures such as
Monte Carlo or simple hill-climbing methods and evolutionary
optimization procedures is their parallel search mechanism. As
is demonstrated later, any successful converter contains
substructures that may be important for optimal performance. In
our procedure it is possible to keep track of such substructures
during the evolution. To obtain the corresponding data of the
traces, substructures of 10 segments length were compared. If no
more than 3 segments of that substructure differ from one
individual to an other the individuals are considered to be part of
one trace. The position index and the fitness of all individuals
taking part of a trace are stored. Three different types of traces
can be distinguished:
1. Traces from the initial population: Are substructures of the
initial population still contained in the population later in the
evolution process?2. Backward, traces from fitness steps: What is the history of the
substructures contained in the best individual of the
population?3. Backward traces from the final population: Are there
substructures that still remain in evolution but have not yet
contributed to the best individual?
These three types of traces will be discussed in detail.
(1) Traces from the initial population: During evolution,
substructures reproduce in order to generate new individuals
with a possibly higher fitness levels than the previous ones
that contain the same substructures. If an individual has a
lower fitness than the worst one in the population, it will be
eliminated immediately. Once a substructure is no longercontained in any individual of the population that
substructure dies out. The probability of reintroduction of the
same substructure by crossover and mutation is very low. The
first type of traces may show how long the initial patternssurvive. In the example shown in Figure 4.8 a major part of
A study of evolutionary algorithms 51
the initial substructures die out m the first 1000-2000
evolution steps. Only two such examples (number 16 or 300)are shown. Some of them remain longer (172, 100 and 60) and
only one initial substructure is really successful and takes
part in several fitness steps (228).
0.75
0.7
0.65
S 06
0 55
0
"'% 1000 2000 3000 4000 5000 6000 7000 8000
Evolution Steps
Figure 4.8. Destiny of the initial population. The stars mark the creation
time of a substructure and the dots show, where the
substructures fall out of the population.
(2) Backward traces from fitness steps: Most of the substructures
will be mixed by crossover during evolution to form other
arrangements that never existed before. They can be traced
back, when fitness steps occur during the evolution, or at the
end of the optimization. This second type of trace is capable of
showing the history of a substructure that has produced an
increase of the fitness. They indicate where a substructure
first occurred, and if multiple substructures coexist in the
population and evolve parallel. A certain competition between
these structures may therefore be revealed. Some
substructures will temporarily be at the top of the
population's fitness ranking, while other substructures are
successful at another time. In Figure 4.9 two backward traces
from fitness steps are shown. The structure according to
Figure 4.9 a) causes a singular fitness increment whereas the
structure of Figure 4.9 b) has been successful at an earlier
stage.
52 A study of evolutionary algorithms
» «)!- 1
s-
6
fi CD "f^*'
.*
LO, /
o
/"2000 4000
Evolution Steps
6000 2000 4000
Evolution Steps
6000
Figure 4.9. Examples of backward traces from fitness steps. Several good,
substructures are created during the optimization, see a) as an
example. They are represented in sub-populations that evolve in
parallel. One or the other will contribute to a, fitness step at a,
certain stage. There is a, competition between the different sub-
populations.
(3) Backward traces from the final population: It is possible and
desirable that new substructures are generated during the
evolution process. Such structures may survive a certain time
in the evolution process, but they are seldom at the top of the
population's fitness ranking. By tracing back substructures
from the final (or intermediate) population it is possible to see
i « » - -
ji.i 4 ——H- -l
d—•
^
ggppw^""*^"'i * * i -f
Fitness 0.6
0.7-
imess 0.6r / /
J-l
\ f ; ./m f Lfi /o
JO j
(
2000 4000 6000
Evolution Steps
a)
2000 4000 6000
Evolution Steps
b)
Figure 4. 10. To observe if there are still different sub-populations in the
actual or final, population, a trace back to earlier stages of the
population's evolution may be created. By doing so, it is possibleto observe how the evolution of sub-populations takes place.
Therefore the parallelism in the evolution is clearly visible. For
these examples, the backward traces are shown for a populationat 7300 evolution steps.
A study of evolutionary algorithms 53
if this competition still takes place. Figure 4.10 shows that
some substructures may be included in the population but
have never produced a fitness improvement so far (Figure4.10 a)). Other substructures participated in fitness steps
earlier (e.g., Figure 4.10 b)) and may become successful againat a later stage of the evolution process.
The three types of traces show that a parallel optimization of
different structures takes place in the evolutionary process as
implemented. Each substructure may be interpreted as a part of
a sub-population containing this unique substructure. When
crossover takes place, the combination of substructures may
produce very successful new substructures as a possible basis for
a new sub-population. From the trace type (3) we can see
whether or not there are still latent substructures that may
become the fittest later. When the whole population is mainlyconstructed with only one or a few significant substructures, the
parallel search mechanism is mostly lost and the evolution
process is no longer efficient. Only a very high mutation rate
could produce other substructures. Thus, the process converges
very slowly much like a random search.
When applying evolutionary algorithms to complex and
lengthy optimization processes, a criterion to decide whether the
optimizer is efficient or not and at what time it may be
terminated would be useful. To get an idea of the progress of the
optimizer the user may observe or supervise certain variables
during the optimization. In small problems such supervising maynot be necessary, because the result is obtained rapidly.
An evolution quality figure used to represent the capability of
the population to produce further more successful structures
when continuing the optimization would be useful. Such a figure
may be defined using the number of sub-populations togetherwith the fitness of the best representing of each sub-population,normalized using the temporal maximum fitness in the whole
population.
c'w=t4?^w- {iA)
where Cp(n) is the evolution quality figure after n optimization
steps, F(n) is the temporal maximum fitness. Nsl\n) the number
54 A study of evolutionary algorithms
of sub-populations and F,SF(n) the fitness of the most successful
représentant of the z'th sub-population. Figure 4.11 shows the
value of that evolution figure during the optimization. To get the
sub-populations, structure parts of 10 segments length were
compared. If any pattern of 10 segments in length is exactlyincluded in more than one individual, these individuals form a
sub-population.
30
Initialization Phase
o
I Evolution Phase
Terminal Phase
11LWTLL X.
Termination
0 2000 4000 6000 8000 10000 12000 14000
Evolution Steps
Figure 4.11. Value of the evolution figure during the optimization. Four
phases may be distinguished.
When observing this evolution quality figure, four phases of
the evolution process as a possible interpretation are
distinguishable. (1) Initialization phase: during this phase the
optimizer rapidly eliminates structures that are not resistant
enough to survive the evolution. The number of sub-populationsis more or less steadily decreasing during this phase. This phasemust not be confused with the generation of the initial
population. (2) Evolution phase: during this phase as many new
sub-populations are generated than are eliminated. This is the
most productive phase of the optimization. (3) Terminal phase:
during this phase all but a few sub-populations are eliminated.
The best sub-population begins to dominate. At the end of the
terminal phase the corresponding sup-population includes over
80% of the total population. (4) Termination of the evolution:
A study of evolutionary algorithms 55
When the previous phase has ended, the evolution process may
be terminated. The termination condition has been reached.
When continuing the evolution process, because of the physical
background of this problem type, we can assume that no
significant increase of the fitness will be obtained.
Traditional methods for obtaining information about the
optimizer's state usually include values such as the fitness
evolution, fitness spreading in the population or mean values of
obtained fitnesses during the optimization. Such values do not
represent the diversity of the individuals and therefore the
optimizer's potential for further improvement. An evolution
quality figure as described here may help gain a view of the
inside state of the population.
4.7. Discussion
An evolutionary optimization procedure has been applied to a
problem in integrated optics, namely a non periodically
segmented waveguide structure for spot-size conversion. This
example shows that optimization of a very large and complex
problem is possible within a reasonable number of evaluated
structures.
The analysis of the evolution process demonstrated the
parallel search mechanism of our procedure. A supervisionmethod was presented by defining an evolution quality figure.Most of the statistically available information concerning a final
state of a population's evolution (e.g., the decreasing spread of
fitness values) usually does not accurately represent the
optimizer's potential for a further improvement. From the
viewpoint of the few but still competing patterns this soberingprospect may be reassessed into a promising one.
The evolution process shown in Figure 4.11 represents a
successful optimization. In less successful cases the evolution
phase may be shorter or even missing. Several reasons may
result in such a behavior: a poor choice of the initial population,a non-appropriate definition of the genotype, which hinders the
definition of substructures by taking neighboring segments, or
other effects, such as a bad mutation rate. It is then more
difficult to determine the terminal phase. In any case, if one
substructure starts to dominate the population, the terminal
56 A study of evolutionary algorithms
phase has started. This condition is very easy to implement in an
optimization, but it is of course not sufficient for a correct
judgment of the optimization process.
5. A Design and Simulation Concept for
Planar Integrated Lightwave Circuits
When designing integrated optical filters there are two main
criteria to fulfill. First, the filter characteristics have to be met.
And second, the design and fabrication costs should be held low.
While the filter characteristics have to be met without
compromises, the factor of cost is always to be minimized.
The overall cost of a product is a combination of the designcosts and the fabrication costs. Depending on the production
volume, the weighting of the two factors varies in a wide range.
The following two chapters present a new approach of a designand optimization platform which can potentially reduce both cost
contributions. The design costs are reduced because the system
is able to operate autonomously with minimal user interaction,
and find solutions that are as cost-effective as possible.
The whole system is composed of a forward solver partdescribed in this chapter and a optimizer part detailed in
Chapter 6.
5.1. Introduction
Several implementations of optical filters with well known
design methods exist. Three of the most convenient structures
are waveguide grating filters [71], resonant coupler (cascadedMach-Zehnder) filters [70], and cascaded ring-resonators [102].For all three types, straight forward filter design methods exist
and can be applied to obtain the required design parameters
[103, 104]. In Figure 5.1 these three types of filters are shown
together with some design information.
The problem of all these structures is the chip space they
require. For all types, the required chip space is linearlydependent on the filter selectivity or the free spectral range that
is to be achieved.
57
58 Design and simulation concept for integrated lightwave circuits
a) Star Coupler
Waveguide Grating Filter
[71, 105,106]
Star Coupler T^esi§n Parameters:
- Number of waveguides- Waveguide spacing- Phase shifts
b)Add-
In /Directional Delay Lines
Coupler
Resonant Coupler Structure
[3, 69, 70, 107]
P Design Parameters:
Through- Number of stages- Coupling coefficients
- Delay line lengths
Coupled King Resonators
Directional Couplers [102J
c)In -
VKfa1 rvOut Design Parameters:
Drop -—A
s. ,«*M UA —Add
-- Number of rings- Coupling coefficients
Delay Lines - Ring radii (resonance
wavelengths)
Figure 5.1. Three types of optical filters for which well known designmethods exist, a) waveguide grating filter, b) resonant couplerstructure (cascaded Mach-Zehnder interferometers, FIR Filter),c) coupled ring resonator array (IIR Filter).
Approaches using ring resonators [108, 109], should be more
suited for compact designs (see examples in Figure 5.2). For such
structures, a general design method does not exist. For every
new topology, the calculation method has to be developed almost
from scratch. The work presented in this chapter directly allows
the filters to be calculated and optimized.
Design and simulation concept for integrated lightwave circuits 59
a) b)
Figure 5.2. Two examples of filters composed of ring resonators which are
more compact than standard designs. The couplers are
encircled, a) Triple-coupler ring-based waveguide resonator
[108], b) compound triple-coupler ring resonator [109].
A system, where the user can just enter the required filter
characteristics and then the system would design the most
compact filter that meets the given requirements would be very
powerful. Such a system would solve the inverse problem for the
optical filter design. It has to be composed of a forward solver
and a general optimizer to solve the inverse problem.
5.2. The Concept of the Forward Solver
The general forward solver relies on several software
components and several design representation schemes. There
are three representation formats: 1) the geometry description, 2)the semantic or functional description and 3) the netlist. Figure5.3 shows the architecture of the forward solver.
The reason to include the three representation levels into the
forward solver is to enable an optimizer to act on geometrical
structures, the simplest representation. Users of the system can
enter any of the three representations. To make the transition
between the three abstraction levels, there are two functions: 1)the semantic analyzer and 2) the netlist generator. The
implementation of these components is described in separate
sections.
Design and simulation concept for integrated lightwave circuits
InputForward Solver
Output
IncreasingAbstraction
Level
Figure 5.3. Architecture of the forward solver including three levels of
representation. They can be seen as different levels ofabstraction, of the same structure containing more and more
information about its functionality.
Waveguide
DescriptionGeometryDescription
Functional
Description
Netlist Generator
Netlist
Description
Scattering-Matrix Compilation
WaveguideDatabase
Scattering-MatrixDescription (internal)
Scattering-Matrix Analysis
Results
Figure 5.4. Flow diagram of the forward solver. Sec text for a description ofthe different steps.
Design and simulation concept for integrated lightwave circuits 61
To obtain a transfer function of a given structure, the
geometry is first transformed into a functional description, then
into a netlist. Together with the waveguide description, it is now
possible to obtain a scattering matrix description of the filter.
Using this scattering matrix the transfer function may be
computed. Any data that can be generally used across different
structures is stored in a waveguide database to allow rapidaccess for subsequent calculations. Such information includes
effective indices, eigenmodes and coupling coefficients. The flow
diagram of this procedure is depicted in Figure 5.4.
5.3. Details of the Forward Solver
5.3.1. The Waveguide Description
Any symmetric waveguide may be described as a superpositionof layers and laterally limited bricks, which constitute the
waveguide's cross section. Figure 5.5 shows an example of a
waveguide structure composed of the two elements. It is
assumed that the main guiding layer or brick has the highestrefractive index. From the waveguide description the eigenvaluesand eigenmodes as well as coupling coefficients can be directlycomputed.
BrickLayer
Figure 5.5. Sample of a waveguide structure composed of layers and bricks.
In the waveguide description, a standard waveguide width is
defined. To obtain other widths, the width of all bricks is
changed accordingly. For this reason, the defined waveguidewidth does not have to correspond to the width of any specific
62 Design and simulation concept for integrated lightwave circuits
brick allowing the user to introduce fabrication dependent
changes.
5.3.2. The Geometry Description
Almost any practical filter topology can be represented as a
concatenation of straight and bent waveguides. Additionallytapers can adapt for different waveguide widths. This means
that with only three generic elements (see Table 5.1) it is
possible to construct virtually any planar lightwave circuit.
To maintain generality, it is important that the base
description consists of only few, very simple building blocks.
In the geometry definition, absolute coordinates are given for
all elements. No connectivity check is done at all. The user has
the sole responsibility of defining geometric structures that are
correctly convertible into a functional description.
Table 5.1. Definition of the geometry elements used to define any filter
topology. The general variables X]; Yx, X7, and Y2 define the
start and end points of the element.
Element Description
Straight (A7,, F,,Ay, Y7, W)
L
W
BendLY,, F,,X2, Y7,W.R,a)
If R^O then the sign of a
defines the bending direction.
Taper (X, Y,. X, F„ Wv W?)
w,
Figure 5.6 shows an example of a simple geometric definition
composed of straights, bends and tapers.
Design and simulation concept for integrated lightwave circuits 63
Figure 5.6. Sample of a waveguide geometry composed of straights, bends
and tapers.
5.3.3. The Semantic Analyzer and the Netlist
The geometric description does not contain any information
about functionality of the structure. A human user could
probably easily distinguish between elements such as couplersand Y-branches by simply looking at the picture. But to enable a
program to find the function of any topology is a difficult task. A
semantic analyzer, which searches the geometry for different
functional elements is required.
Therefore it is necessary to define a set of functional elements
with which any planar lightwave circuit can be constructed. In
Table 5.2 the different functional elements and their
corresponding descriptions are shown. Constraints of each
functional element define the limits of applicability. These
constraints are stored in a waveguide library which is linked to
the waveguide definition. The corresponding values can be
defined by the user according to experimental results or
experience.
64 Design and simulation concept for integrated lightwave circuits
Table 5.2. Definition of the functional elements used by the semantic
analyzer to extract the function of any filter topology. The
general variables X{, Yx, X-,, and F, define the start and end
points of the elements, ini and outl represent the interface
nodes. The description syntax is the one used by the system.
Element Description
W
outA
L
StraightGuide (/«,) -> {oiltx) (L, W)
StraightGuide (A7,, 7,, X2, Y2, W)
a
W \
VétvX
BentGuide (z;?,) -> (oü/j) (R,CX, W)
BentGuide (A" Y,, X„ 7„ R,a, W)
.mt (yivCT"
D [m w^
my j rOUt'y\.
StraightCoupler {inv in7) ->
(outv out,) (D, L, Wi,W2)
StraightCoupler (Xla, Yta, X2a, Y2a, Wv
^x\b> 1IA> 'v2/i' l2b> "2>
BendStraightCoupler (/«,, in-,) ~>
{outv out2) {Y R, Dmm, Dmax, Wv
IF,, a )
BendStraightCoup 1 er ( 1, Xla, Yla, X2a,
*Ice ''I'^li' '
W ^2A> *2b> **> "2>
D.
BendStraightCoupler (in{, in0) ~>
{out,, out,) (2, R,Dmm)Dm^ Wv
W2,a)
\ p) BendStraightCoupler (2, Xx , YXa,Xla,
Y2l,Wx.X]b,Yxl,XmY2b,R,W2)
Design and simulation concept for integrated lightwave circuits 65
Elementi
Description
BendBendCoupler [inx, in2) ->
{outv out2) ( R^, Rz ,Dmm, DmaK,Wl}
W2, a )
BendBendCoupler (XXa, YXa, X2a, Y2o, R
y¥x,AXB. lXy A-,b, J2b'
*2' "2'
BendlnBendCoupler ( /«,, m2) ->
(outv ouf) (1,7?,, R2, Dmm, Dm^, Wv
W2,a)
BendlnBendCoupler ( 1, XXa, YUl, X2a, Y2d)R W X Y X Y R W )
>' 1' MA' 1"' 2b> -/2A'iv2> rr2>
BendlnBendCoupler {mv m2) ->
{outv ouf) (2,RX,R2, Dmm, Dmax, Wv
W2,a)
BendlnBendCoupler (2, Xia, Y]a, X2a, Y2fi,R W Y Y X Y R W )
Table 5.2. Continued.
The functional description still contains all coordinates, but it
now contains information about related elements. The geometry
elements are cut accordingly to obtain the functional elements.
Almost every filter design can be constructed out of directional
couplers and connecting waveguides.
From the functional description created by the semantic
analyzer the generation of a netlist is relatively straightforward.It defines the connectivity between the functional elements. The
netlist definition contains a minimal set of parameters necessary
to completely describe each element. Additionally the input and
output ports are numbered to define how the elements are
connected. The netlist generator automatically detects any global
input and output ports. If the user does not define and number
the ports, the program numbers them.
A.
a
ID
66 Design and simulation concept for integrated lightwave circuits
The program can then transform the netlist description back
into a functional description and into a geometry. This will be
necessary for the operation of any optimizer. It must be
guaranteed that the transformation back and forth does not
modify the geometry of the structure.
If any illegal constellation of geometric elements is detected bythe semantic analyzer, an internal error code is generated. Thus,the optimizer can eliminate illegal structures.
5.3.4. The Scattering Matrix Compilation
The netlist represents a number of individual elements
connected to each other by ideal links. Every element can be
represented by an individual scattering matrix. Assuming that
the whole system is single mode, only one port is required for
each interface node. Figure 5.7 shows an example of connected
scattering matrices.
B
ial nodes (k)
al nodes (p)
C
Figure 5.7. Filter structures are composed of several elements. These can be
combined such that an overall scattering matrix can be definedto make the direct connection with the external ports.
Using the connection scattering matrix method [110, 111] it is
possible to combine the whole system into one scattering matrix
representing the characteristics of the overall system with
respect to the global system inputs and outputs. To do this, all
scattering matrices of the individual elements are first placed on
the diagonal of a matrix.
Design and simulation concept for integrated lightwave circuits 67
[4.
[4 [°] M [o][o] [4 [o] ... [0][o] [o] [4 [0] (5.1
I»! M [«I - [4,
where \S] is the scattering matrix of the element / and M is the
total number of elements. The resulting matrix is in general very
sparse.
The rows and columns of \S] are then rearranged to place theL hot ö ^
external nodes in the uppermost lines and in the leftmost
columns. With this operation \S] is divided into four distinct^ L lioi
VmYS[S]kk,[S]pp,[S}pk and [4,.
K
KFml 1
Mkk
mPk
5] kP
s\
a,
a,
ct..
Ü..
a.
a.
(5.2)
In (5.2) at and bt represent the incoming and outgoing waves
respectively of the port i. \S\k now depends on the external
ports only and [S]?depends on the internal ports only. The
other two ([S]A and [S]k7) refer to external and internal ports.
To make the connections between the internal ports a
connection matrix [C] is needed. The matrix [C] has the same
dimensions as [S]^. The value 1 is inserted wherever two ports
are connected.
68 Design and simulation concept for integrated lightwave circuits
M
C C
Mi L12
C C
r c
Ci\
G2\
\\
(5.3)
where
0, port k and port / not connected
1, port k and port / connected(5.4)
are the elements and X is the number of internal ports. Onlypairs of nodes can be connected together.
The external scattering matrix LSI can then be calculated as
follows:
(5.5)[^L-^L-^LfL-M lsL-
Since the individual scattering parameters are wavelengthdependent, it is more efficient to evaluate the expression
analytically before sweeping over the wavelength. In the system
implementation the matrix inversion is therefore solved
analytically with an LR-decomposition [50]. For realistic devices
consisting of 50 to 100 elements, this evaluation takes less than
one second of computation time, but requires a considerable peakamount of memory (up to several megabytes). It is then possibleto rapidly calculate the external parameters for different
wavelengths. Any calculated values that do not change for one
wavelength are cached. Therefore for one wavelength different
scattering matrix parameters can be calculated with virtually no
additional computation effort.
5.3.5. The Waveguide Library and the WaveguideDatabase
The semantic analyzer is coupled with a waveguide librarywhich includes information such as constraints (e.g., minimum
radius for bends or minimum and maximum waveguide widths),and other conditions used for the detection of couplers. This
information is mainly provided by the user and is based on
experience concerning the fabrication process, and on
measurements.
Design and simulation concept for integrated lightwave circuits 69
In order to accelerate the calculations the forward solver
makes use of a waveguide database. Eigenmodes, couplingcoefficients (as a function of the wavelength and waveguidewidths) and other information used by the forward solver are
stored in this database. The solver can then rapidly access this
information without having to recalculate it every time. As soon
as an information is missing in the waveguide database, the
corresponding solver (see the following sections) is launched and
the missing data is calculated and introduced into the waveguidedatabase. The data can then be accessed, and the forward solver
can continue its calculation.
5.3.6. The Eigenmode Solver
The eigenmodes and the effective indices of the individual
waveguides are computed using the imaginary-distance beam
propagation method, as described in chapter 2. The eigenmodesolver can either be directly called from a menu, or invoked
automatically when the forward solver detects a lack of
information in the waveguide database.
5.3.7. The Calculation of the Coupling Coefficients
The most important parameters to describe the physicalbehavior of directional couplers are the coupling length Lc and
the coupling coefficients kX2 and k2X ,which can be calculated
from the propagation constants/?, and ß0 of the even and the odd
mode of the coupled waveguide structure.
Lc represents the propagation distance that is required to
couple the maximum power from one into the adjacentwaveguide and can be determined as follows:
Therefore the difference between the even and the odd
propagation constants ßi — ßn is an essential quantity when
calculating waveguide couplers.
Exact coupling coefficients could be calculated if the even and
odd modes in a coupled waveguide structure are known
70 Design and simulation concept for integrated lightwave circuits
analytically or numerically with a high precision. But accuratelycalculating these modes in a general fashion is a difficult and
time consuming task. It would be much easier to operate with
the eigenmodes ElX and El2 of the individual waveguides only.
This is possible using the following approximation for ße—ß0[112]:
ße-ßo =
iklK]2K2X+(ßx- ß2y ß2(i~- N^y(5.7)
ß-(\-N-)
where k is the vacuum wave number, ßx and ß2 are the
propagation constants of the individual (uncoupled) waveguides,and ß represents the average propagation constant
ß = M+ßo) = HA+A)> (5-8)
and N is the overlap integral between the two separate
eigenmodes
N= IJEn-El2dxdy. (5.9)CO
Further the coefficients K2l smclKx2 are defined as follows:
K2X = \$[n;(x>y)-n:(x<yJ\(E,i EI2-NE„ • ElX)dxdy, (5.10)
and
Kn = jj nl(x,y) - n](x,y) (ElX • El2 - XEt2 El2)dxdy, (5.11)
where nx and n1 represent the index distributions with only the
core waveguide 1 or waveguide 2 respectively, and if is the index
distribution with no waveguide core at all. No waveguide core
means that the width of the corresponding waveguide is assumed
to be zero.
The coupling coefficients k]2 and k2] can then be determined as
follows:
^12 ~
T BN + yJl-B-(l-N2) (A ~ßo) (5.12)
and
'21BX + ^l-B"(i^X-) (£-#V (5.13)
Design and simulation concept for integrated lightwave circuits 71
with
B=ßylA_ (5.14)ßc-ßo
This theory gives very good results for symmetrical directional
couplers. A comparison of different methods is given in [113].
It can easily be observed from (5.12) and (5.13) that the two
coupling coefficients /c12 and /c21 are always different, unless the
two waveguides are identical (B = 0) or they are uncoupled
(A = 0).
Commonly two other quantities 5 and 5,called detuning and
effective detuning, are defined as follows
S = j(ß-ß2), (5.15)
where ßx and ß2 are the propagation constants of the two
individual waveguides, and
%« =Hßc-ßo) = ^M::W:^^' (5-16)
where Kn and k2] are the two coupling coefficients between the
two waveguides which are defined in equations (5.12) and (5.13).
The problem of asymmetrical directional couplers has been
extensively treated in literature [112, 114, 115, 116, 117], but no
satisfactory result has yet been reported. The simplificationsresult in a violation of the energy conservation law. The only wayto bypass the problem is to use rigorous methods without
simplifications which require precise knowledge of the even and
odd modes of the coupler. Then, the propagation constants of
these modes (ße and ßo) are exactly known and the coupling
coefficients result directly from equations (5.9) and (5.14), which
are introduced into equations (5.12) and (5.13).
5.3.8. Elementary Scattering Matrices
The most widely used scattering matrices are those of straightand bent waveguides, and directional couplers. In this work it is
assumed that the bending radii are large enough such that no
significant radiation losses occur. Therefore the scatteringmatrices of straight and bent waveguides have the same form.
72 Design and simulation concept for integrated lightwave circuits
V=
S=
^st/aiehl ^beiui
( 0 e
(iß a)Lq
(;/}-a)L\
(5.17)
where ß is the propagation constant, a represents the losses,
and L is the length of the waveguide. For bent waveguides, the
length L is approximated by their neutral path, which may be
adjusted, because the field propagates slightly on the outer side
of the bend.
For directional couplers, the scattering matrix depends on the
coupling coefficients /cp and /c7, as well as on the detuning factors
8 and 8cff .See Figure 5.8 for the definition of the nodes.
S.31
i3i n— cos(8cff L) - j
_
S$,m{8,, L)
o_
c_
°42~'
J24~~
V
f(
Vcos(8Lf,L)+ j
,<5sm(8 L)
8
532 = S2i - ~-[jKn8$ii\(8elj L))e
if
-ccL
J
\
J
-al
-aL
(5.18)
^41 =- S\4 = ~(/K12Ss'm(3tifL))e11=
k 12^
21~ *^22 = ^Vi = ^U
~
^4") "^ ^44 =
This scattering matrix is valid for directional couplers with
two parallel waveguides. If the coupler consists of two bent
waveguides or any other configuration, where the distance
Figure 5.8. Example of a directional coupler with varying gap between the
waveguides. To calculate the overall scattering matrix it may be
subdnided into several segments of length LI with constant
gaps d].
Design and simulation concept for integrated lightwave circuits 73
between the waveguides changes over the length of the coupler,the scattering matrix can be determined by cascading small
couplers with constant distances [118] (see Figure 5.8).
The overall scattering matrix can be calculated as the cascade
of the partial scattering matrices S^'Jsegment
V_
CO)A
C(-) A...AÇ(') A-.-A Ç(A-1) A <s(A') f5 1 Q">u
couplerusegment
' N usegment
' v ' x usegment
/x ' yusegment ' v usegment ' \u.a.*j;
evaluated from left to right, where the cascade operator a is
defined in the following manner
n7 çib , eu ob
Sa a Sb =
u o31o31 + o41o32 o3]o41-t-04,0,42
b . p« c*b eta c<b 1 pa oA0 0 s^ + s:;2s;2 $°2s°x + s:;2s.42
c7 c^ j„ cc/ c^1 c7 c^ -i_ 00 c^ a n^31 31 41^32 ^32^31 ' *-*42i->32
V03,o41-t-o41o42 o32o41-i-o42o42 U 0
(5.20)
After calculating the individual scattering matrices of all
separate elements, the overall scattering matrix of the lightwavecircuit can be determined using the connection scattering matrix
method that was described in the previous section.
5.4. Examples
5.4.1. Resonant Couplers
As primary examples a number of resonant couplers are
calculated. Such resonant couplers can be used as add-dropfilters in WDM systems. A number (1 to 9) of Mach-Zehnder
stages are cascaded to show the growing free spectral range (seeFigure 5.9 for one and two stages). In this example all stageshave the same parameters. The goal is to analyze the
performance of an add-drop filter with various stages and not to
construct an optimal filter.
74 Design and simulation concept for integrated lightwave circuits
Drop Add-
-Through In
Figure 5.9. One and two Mach-Zehnder interferometer stages. The outputsare interchanged for even and odd numbers of stages.
The first parameter to determine is the length difference of the
delay lines. To obtain this value, it is first necessary to calculate
the delay time difference Ar [107]
Ar = 4"-' (5-21)/o
where f0 is the desired spectral period, chosen to be 800 GHz for
this example. The corresponding wavelength period of 6.4 nm
gives Ar = 1.25-1 (T12 s and a delay line length difference of
253.4 urn for an effective index of 1.48. The coupling coefficients
and coupler lengths are chosen to couple 50% of the energy into
the adjacent waveguide at about 1.56 urn.
Table 5.3 shows the results of the simulation of the devices
from 1 to 9 stages. A strong wavelength dependency of the
devices can be observed mainly due to the wavelengthdependency of the directional couplers. Nevertheless this
dependency may be neglected over one period of 800 GHz.
The selectivity is increased as can be seen in the cases with
one, five and nine stages in Table 5.3. But without optimizationthe side lobe suppression is very poor, even for 9 stages. By justadding more identical stages this problem cannot be solved. It
can only be overcome with a non-uniform distribution of the
coupling coefficients and the delay line lengths [3, 69]. A flatter
passband and a smaller crosstalk can then be achieved. Usually,to meet these requirements, an optimization is necessary. Such
an optimization will be demonstrated in the next chapter.
Design and simulation concept for integrated lightwave circuits 75
Table 5.3. Filter characteristic of resonant couplers with 1 up to 9 Mach-
Zehnder interferometer stages. The drop output is shown on the
left, and the through, output on the right. The results were
obtained, considering a, waveguide loss of 0.1 d,B/cm.
Number
of
Stages
Add-
In-
Drop
• Drop/Through
"Through/Drop
Through
1.55 1.57 1 1.53 1.55 1.57 X
g0.5
H
1.57 X 1.55 1.57 X
30.5
H
&3 1.55 1.57 X
4
H
0 5|t
h y
( A
1.53 1.55
05
H
A A 1
W
odiLiiiiiïiiiii,
1.53 1.55 1.57 A 1.55 1.57 X
76 Design and simulation concept for integrated lightwave circuits
Number
of
Stages
Add-
In -
Drop
'Drop/Through
"Through/Drop
Through
6
57 X 1 57 X
57 X 1 53 1 55 1 57 X
8
.57 X 1 57 X
9
57 X f57 X
Table 5.3. Continued.
5.4.2. Add-Drop Filters Using Ring Resonator Devices
Many filter topologies using ring resonators and corresponding-
analysis methods were reported in literature [102, 103, 108, 109.
119, 120]. In this section two types of ring resonator devices are
calculated. Both types are compact add-drop devices that use
Vernier operation to obtain both a high wavelength selectivity as
well as a large free spectral range (FSR).
Design and simulation concept for integrated lightwave circuits 77
Add _
In —,
Figure 5.10. Compound triple coupler ring resonator (CTCRR) add-drop
filter ]109]'.
The first filter structure treated in this section is a compound
triple coupler ring resonator (CTCRR). The schematic diagram of
the CTCRR (see Figure 5.10) has two rings with radii r, and r2 as
well as three different types of directional couplers with
amplitude coupling ratios Kv K2 and Kv The two waveguides
external to the rings have a radius of curvature at least equal to
the smallest between the radii of curvature of the rings. The
same propagation constant ß is assumed for all waveguides of
the structure. More details about this configuration may be
obtained from [106].
To enable a direct comparison with literature, the following
parameters were chosen: i\ =5.7 mm, /*, = 6.5 mm, and a
waveguide loss of a = 0.5 dB/cm. The result of the calculation is
shown in Figure 5.11. Comparing this with the result in [106], a
slight difference is visible and comes from the wavelength
dependency of the directional couplers that is also taken into
account in the present calculation.
78 Design and simulation concept for integrated lightwave circuits
I I
ÖO
CO
0
-10
20
-30
-40
z
2 -io
TTYTlP/rHTH0: i
[1061
12 16 20 24
ßl/K
1.5495 1.55 1.5505
Figure 5.11. Transfer cliaracteristics (Out) of the compound triple couplerring resonator structure shown in Figure 5.10.
A second filter structure to be treated in this section is a
double-ring resonator (DRR). It consists of two ring resonators
with different radii r, and r2 located between the input and
output waveguides (see Figure 5.12). The waveguides are
coupled by three directional couplers with amplitude couplingratios of Kx, K2 and A"3. Each ring has a different resonance
wavelength. By correctly choosing the ratio between the two
radii r, and r2 it is possible to obtain a free spectral range which
is related to the least common multiple of the two resonance
wavelengths. This can be referred to as Vernier operation.
In
Add
^Out
Drop
Figure 5.12. Double-ring resonator (DRR) filter structure [121].
To enable a direct comparison with literature the structure has
been calculated with the parameters r, = 5.7 mm, ;\ = 6.5 mm, and
Design and simulation concept for integrated lightwave circuits 79
a waveguide loss of a - 0.2 dB/cm. The calculated transmission
characteristics (see Figure 5.13) is in very good agreement with
the measurements presented in [121].
CQ
«3
O
80
60
40
(YfWV\
'121]
37.2 GHz
0 10 11 3D »0 !0 $0
RELATIVE FREQUENCY (GHz)
201.549 1.55
X
1.551
Figure 5.13. Characteristics of the dropping output of the double-ringresonator structure, shown in Figure 5.12.
5.5. Discussion
The forward solver approach described in this chapter proves
to be very accurate and fast for practical filter structures.
Because the solver is based on a geometric description and does
not require the specification of any functionality by hand, it is
very suitable to be used by a general optimizer. It is possible to
freely switch between the three representation levels and
therefore to extract any information needed. The semantic
analysis is quite complex but still very fast.
Since each representation level is available as an ASCII text,the user can easily supervise the steps and also interact at
different levels. Editors and viewers are available at every stageof the forward solver.
The use of scattering matrices for the analysis of integratedwaveguide structures reduces the complexity of the calculations
while only degrading the obtained results very slightly. In this
implementation all waveguides are assumed to be single-mode.
80 Design and simulation concept for integrated lightwave circuits
For a multimode analysis (this is also possible for single mode
waveguides), all partial scattering matrices would have an order
that is proportional to the number of guided and radiation modes
considered. It would then imply that the coupling coefficients
between the individual modes have to be determined as well. The
overall complexity would grow to an extent that the analyticdetermination of the overall scattering matrix is no longer a
realistic approach. It would then be easier to use a numerical
analysis method for the calculation of the scattering matrix
elements. Then, they are combined into a global scatteringmatrix numerically. The result of such calculations may
potentially be more accurate than with a strict single-modeapproach, but the increase in calculation time would be too largefor the gain in precision.
6. An Optimization Concept for Planar
Integrated Optics
In an industrial context almost every design procedure results
in an optimization process. The optimization of integrated
lightwave circuits has not yet been completely solved. Onlyparameter optimizations have been presented so far. In this
chapter a new optimization strategy is developed based on the
forward solver described in Chapter 5. Since information is
available at different levels of abstraction, the optimizer is not
only able to optimize any filter characteristic, but it is also
possible to optimize structures with respect to other parameterssuch as chip size, aspect ratio, complexity, etc.
The resulting inverse problem solver is based on an
evolutionary optimization procedure. Modifications of the
structures are done by special mutation operators, acting directlyon the geometry definition. Because it would be very difficult to
define a crossover operator between two (possibly completelydifferent) lightwave circuits, a crossover operator has not yetbeen implemented.
The following implementation of the optimizer represents a
mandatory step in the direction towards a system capable of
"inventing" new lightwave circuits.
6.1. Introduction
To build a successful inverse problem solver with an
evolutionary optimization procedure, the following elements are
necessary:
1) A general format that can represent every generated structure
(this topic has been covered in Chapter 5).2) A robust forward solver that gives useful results for realistic
structures (this has also been described m Chapter 5).3) A fitness definition which allows a correct qualification of the
individuals with respect to the given specifications.4) Mutation operators to transform the structures.
81
82 Optimization concept for planar integrated optics
5) Crossover operators or other statistical procedures to
integrate information about several individuals into newly
generated ones.
6) A supervision method delivering an abortion criterion (such a
method has been proposed in Chapter 3).
First, the overall optimizing strategy is described, then the
points 3-5 are explained.
6.2. The Optimizer Strategy
The optimizer strategy is depicted in Figure 6.1. The forward
solver is not a black box for the optimizer, therefore it is possiblefor it to use all information about the different stages and levels
of representation. This enables a hybrid structural and
parameter optimization.
Inverse Problem Solver
Figure 6.1. General architecture of the inverse problem solver for planar
integrated lightwave circuits.
The optimization procedure itself resembles the breeder
genetic algorithm used in Chapters 3 and 4 to design the spot-
size converter. The only major difference is that no crossover
operator has been defined in this implementation.
It is possible to introduce several restrictions for the optimizer.Such restrictions are useful when the overall topoLogy should not
change, and optimization of the remaining parameters is desired.
For example, the user can forbid the separation of directional
Optimization concept for planar integrated optics 83
couplers, or let rings move in horizontal or vertical directions
only. Many restrictions act on the functional description. A
design rule check ensures the correct functionality of the
structures. If the optimizer produces a geometrical structure that
results from illegal operations, the structure is rejected, and a
new structure will be generated. The restrictions are mostly used
in cases where the designer knows that a structure is near the
topological optimum.
6.3. The Fitness Definition
The definition of the desired filter characteristics is a
piecewise linear function in the linear scale (see Figure 6.4 as an
example). Each part of the definition consists of either an upper
or a lower limit and a weighting factor that defines the
importance of that limit. For each segment a selection can be
made, which scattering parameter is concerned, and if the
corresponding power, group delay or the dispersion should be
considered for the limit.
The difference between the calculated filter response and the
constraint is integrated over the ranges where the constraints
are not met. This results in the following fitness definition:
F =
1 JX \[max(0Sl(X)^Eu(X))
y
7= 1 ?Af/l1î!ll
dX +
max(ÖJ'L(X)-~S'L(X)) dX
, (6.1)
-i'/ m,,/
where the symbols are defined as follows:
N0-, XL : number of upper and lower limits,
w'LJ, w'L : weighting factors for the limits,
^mm^l/nm'^mm^/max: wavelength ranges for the limits,
ELI (X) : definition of the upper limit /,
EL(X): definition of the lower limit /,
S'b(X): value of the transfer function, group delay or
dispersion that has to be compared with the
upper limit /,
84 Optimization concept for planar integrated optics
S'L(X): value of the transfer function, group delay or
dispersion that has to be compared with the lower
limit /, and
p: defines the p-norm of the distance.
This fitness definition has a lower limit of zero and an upper
limit of one. The weighting factors Wv and w'L, can implement
any special behavior without modifying the fitness calculation
procedure.
Additionally the filter function may be shifted along the
wavelength axis. This makes perfect sense because a wavelengthshift of the final design can easily be achieved by scaling the
geometry. Tuning elements can also be implemented to achieve a
fine tuning of the filter response [3, 122].
6.4. Mutation Operators
The geometry is modified with different mutation operators.
Each mutation operator has a different probability that can be
set individually. Two of the most important are shown in Figure
6.2, shifting of any node and rotation of the node axis. In both
cases a number of statistically selected nodes is locked prior to
the mutation to limit the geometrical extent of the mutation.
H-..
ts<U
^ a\
a) b)
Figure 6.2. Two mutation operators, a) node shift, b) node rotation. The
black dots show fix-points that are not moved by the mutation
operator. Depending on these fix-points the operation may have
different results.
Optimization concept for planar integrated optics 85
Other mutation operators are, e.g.. the addition and deletion of
rings as well as the scaling and displacement of ring elements.
For all mutation types the connectivity must be maintained and
constraints like the minimum bending radius or the minimum
distance between waveguides have always to be fulfilled.
Since the optimizer has access to functional information about
the current structure, constraints such as not to separate
couplers, not to modify certain types of elements, etc. can be used
to optimize the structure without modifying the network
topology.
6.5. Example: A Resonant Coupler Add-DropDevice
In this section an optimization is shown. It consists of a 5-
stage resonant coupler add-drop filter. As input, five identical
stages are introduced resulting in a poor dropping performanceand a large crosstalk. Figure 6.3 shows the device together with
its filter characteristic.
A number of constraint levels were defined for the
optimization. Figure 6.4 shows the constraints and the
corresponding weighting factors that were used (see the
definition of the fitness (6.1) for a explanation of the different
factors). A rectangular filter shape was given with several levels
of constraints. This prevents the optimizer from producingstructures that fulfill the constraints in a large wavelengthrange and have very poor performance in the remaining part.The constraints are much more demanding than the structure
can ever fulfill. The optimization will converge to a structure
that gives the best fitness. This structure best approximates the
given filter characteristics.
Optimization concept for planar integrated optics
-Drop
-Through
a)
1.56
1.56
3. a) Schematics of the 5-stage resonant coupler consisting of 6
directional couplers, b) and c) the initial filter characteristics
(drop-channel) with a linear and logarithmic scale respectively.This starting structure has large losses in the passband, and a
very poor performance in the stop band.
Optimization concept for planar integrated optics 87
w[ =3000 ,--4W
w2L = 9000
>X4(^)
L\;(X)\
\\
4(X)\
= 600 \
4(x) 4(a.)
\ w* = 60° /\
i K = 200 l i w?,=200 ,
0.95
0.85
0.1
0.05
1.544
>A
un > 'W/r
1.55
1.5496 1.5504
1.556
Xl ^
!\\ ^3 "l
1,2 -Vt^t/nux ' ^C/max
AT/mm> 'Vmm'
T.1 ">2
Figure 6.4. The optimization, constraints are shown in this figure. 6
different constraints are introduced, each, of which, having a,
different weighting factor, shoivn beneath the lines.
For this optimization example, a population size of 20 was
chosen. Experience shows that for topological optimizations the
population size should be kept small. Figure 6.5 shows the
fitness evolution during the optimization.
0.7
0.6
0.5
0.4
^ 0.3
0.2
0.1
a
1
r-ffîy.^^'yfry^y-,
500 1000 1500
Optimization Steps2000
Figure 6.5. This graph, shows the evolution of the fitness during the
optimization. The top line represents the fitness of the best
individual, the line below the least fitness which is still included
in the population. Dots show the fitness of every evaluation.
88 Optimization concept for planar integrated optics
The best performing structure was found after about 2000
optimization steps and has a much lower crosstalk than the
starting structure (see Figure 6.6). The losses in the passbandare also low. Depending on the fitness definition, some regions
may have a higher importance. In this example every part of the
filter characteristics had the same importance. The equiripple in
the stop band could be improved by introducing additional
constraints.
1.56
1.56
Figure 6.6. Filter characteristics of the optimized, structure, a) linear scale,
b) logarithmic scale. A large improvement is visible when
comparing this with, the characteristics shown in Figure 6.3.
For this particular example more efficient optimizationmethods exist. However, we show that this optimizer is able to
treat known problems. The advantage of the present optimizingapproach is that it is potentially able to treat problems where no
theory or design method exists. The user will not have to develop
Optimization concept for planar integrated optics 89
such a theory, he will just have to enter the new topology and
optimize.
Figure 6.7 shows the coupling ratios and the difference of
delay line lengths for the optimized structure.
249 I —, ,
r——|
| 248 8 r—,
= 248 6-
S 248 4 •
g 248 2 -
123456 12345
Coupler Stage
a) b)
Figure 6.7. Distribution of a) the amplitude coupling ratios and b) the delaylines for the optimized filter structure. All values are valid at a
wavelength of 1.55 /urn.
6.6. Discussion
A procedure was implemented for the topological optimizationof planar lightwave circuits. It is very flexible and allows the
optimization of many aspects of the structure. Especially it is
possible to optimize the filter characteristics and the
corresponding dispersion at the same time.
Since the phase conditions are very critical to filter
characteristics, two almost identical structures may have
completely different performance. Therefore, finding a goodtopology by just randomly putting together basic waveguideelements is very unlikely. Given a topology that corresponds to
the desired structure, the optimizer is then able to find an
optimal configuration. A given topology defines the number and
types of directional couplers, as well as the feed-backs. If an
optimal structure cannot be found using the given topology the
optimizer is able to modify the functional structure to obtain a
new topology.
Additional work has to be done to enable the system to
autonomously discover new filter topologies. Such a functionality
30%
erf
U 20%
o
U
-i no/
90 Optimization concept for planar integrated optics
could be implemented by acquiring statistical information duringthe optimization and using a database of well performingstructures. The statistical information could partially replace the
crossover operator. A structure database may be used to compose
filters out of more complex functional blocks. It is nevertheless
not yet clear, how the optimization would function in future. The
rapid development of the computer performance will certainlyallow "inventor" software to become a reality.
7. Conclusions and Outlook
In this thesis a number of topics were addressed in the domain
of optical lightwave circuit design. In a first part, the simulation
of waveguide structures with the beam propagation method was
investigated. In this context, an improved method for eigenmodeextraction has been developed. This technique is based on the
imaginary-distance beam propagation method. Using a
mathematical trick in the context of a finite difference solver it is
possible to directly extract higher order propagation modes of
arbitrary shaped waveguide structures. This method is very fast
and accurate, and the results are especially well adapted for
BPM simulations.
The beam propagation method, despite of its limitations, is
very suitable for complex calculations. This was shown by the
design of a very short spot-size converter of typically less than
140 pm length. The structure of the converter is a non-
periodically segmented waveguide with irregular tapering. Usingan evolutionary optimization procedure a solution that improvesthe coupling losses from an optical fiber to the chip from 3.5 dB
down to about 1.3 dB was found. Thus, evolutionary algorithmsare well suited for such optimizations. Using these algorithmswe have found that non conventional designs potentiallyoutperform straightforward implementations.
A thorough analysis of the genetic algorithm applied for the
optimization of the spot-size converter showed the internal
behavior of the evolution. The implicit parallelism was
demonstrated using the notation of traces of sub-populations.These sub-populations were also the basis for the definition of a
evolution quality figure. As a first application, this figure can be
used as an abortion criterion of the optimization. It is also
possible to qualify the efficiency of the evolution when analyzingthe temporal behavior of the evolution quality figure. This is a
promising method of qualifying an evolutionary algorithm.
In the following, a new optimizer architecture for planarintegrated lightwave circuits was designed and implemented.The forward solver part is based on three representationschemes (geometry, functional description and netlist). Each
level contains an increasing amount of information about the
91
92 Conclusions and outlook
functionality of the structure. A scattering matrix solver is then
used to calculate the output spectra of the structures. Because
information is available at every stage of the forward solver, an
optimizer can act on different levels of abstraction. In particularit can modify the geometry by using different mutation
operators, but it can also use the functional description to test
constraints on a higher level of abstraction. The actual
implementation of the optimizer is based on an evolutionaryprocedure. The presented implementation of the optimizer is a
very important step towards an autonomous program which is
able to construct new filter topologies based on different
previously designed structures.
During this work, a Microsoft Windows® based application was
written, combining all elements of this thesis into a user friendly
program. Editors and viewers are available for every step of the
design. An interpreted high level language (the syntax is similar
to that of C) allows a very flexible use of the BPM solver.
Further investigations are required to improve the "inventor"
part of the optimizer. Statistical information about the
optimization and the structures may be useful to guide the
optimizer towards the right topology that can implement a
desired filter characteristics. A database of designs could be used
as a template library to obtain different successful substructure.
These could then be combined into a new topology. Scientific
research should go on in this direction.
The program needs some further development if it should be
made commercially available. In particular, graphical editors
should be implemented for a simpler drawing and modification of
the optical structures. It would be very desirable to have such a
tool for BPM simulations as well as for the CAD of planarintegrated lightwave circuits.
References
[1] Denzel, W.E., and B. Meekers, "Photonics in the backbone
of corporate networks - the ACTS COBNET project" in
Photonic Networks, Optical Technology and Infrastructure,
NOC'97, D.W. Faulkner and A.L. Harmer, eds., IOS Press,
1997, pp. 25-32.
[2] Kawachi, M., "Silica wavegiddes on silicon and their
application to integrated-optic components" Opt. and
Quantum Electron., vol. 22, no. 5, September 1990,
pp. 391-416.
[3] Offrein, B.J., R. Germann, G.L. Bona, F. Horst, and
H.W.M. Salemink, "Tunable optical add/drop componentsin silicon-oxynitride waveguide structures," Proc. ECOC'98,
September 1998, pp. 325-326.
[4] Li, Y.P., and C.H. Henry, "Silica-based optical integratedcircuits" IEE Proc.-Optoelectron., vol. 143, no. 5, October
1996, pp. 263-280.
[5] He, J.-J., B. Lamontagne, A. Delâge, L. Erickson, M. Davis,and E.S. Koteles, "Monolithic integrated wavelengthdemulitplexer based on a waveguide rowland circle gratingin InGaAsP/InP" J. Lightwave Technol., vol. 16, no. 4,
April 1998, pp. 631-638.
[6] Gini, E., W. Hunziker, and H. Melchior, "Polarization
independent InP WDM multiplexer/demultiplexer module"
J. Lightwave Technol., vol. 16, no. 4, April 1998, pp. 625-
630.
[7] Feit, M.D., and J.A. Fleck, Jr., "Light propagation in
graded-index optical fibers," Appl. Opt., vol. 17, no. 24,December 1978, pp. 3990-3998.
[8] Jüngling, S., and J.C. Chen, "A study and optimization ofeigenmode calculations using the imaginary-distance beam-
propagation method," IEEE J. Quantum Electron., vol. 30,no. 9, September 1994, pp. 2098-2105.
93
94 References
[9] Spühler, M.M., D. Wiesmann, P. Freuler, and M. Diergardt,"Direct computation of higher-order propagation modes
using the imaginary-distance beam propagation method,"Opt. and Quantum Electron., vol. 31. no. 9, vol. 9/10,October 1999, pp. 751-761.
[10] Spuhler, M.M, B.J. Offrein, G.-L. Bona, R. Germann, 1.
Massarek, and D. Erni, "A very short planar silica spot-sizeconverter using a nonperiodic segmented waveguide," J.
Lightwave Technol., vol. 16, no. 9, September 1998,
pp.1680-1685.
[11] Mühlenbein, H., and D. Schlierkamp-Voosen, urpae science
of breeding and its application to the breeder genetic
algorithm (BGA)f' Evolutionary Computation, vol. 1, no. 4,Winter 1993. pp. 335-360.
[12] Erni, D., M. Spuhler, and J. Fröhlich, "A generalizedevolutionary optimization procedure applied to waveguidemode treatment in non-periodic optical structures" Proc.
ECIO'97, April 1997, pp. 218-221.
[13) Spuhler, M.M., D. Erni, and J. Fröhlich, "An evolutionaryoptimization procedure applied to the synthesis of integratedspot-size converters" Opt. and Quantum Electron., vol. 30,no. 5/6, May 1998, pp. 305-321.
[14] Erni, D., M.M. Spühler, and J. Fröhlich, "Evolutionaryoptimization of non-periodic coupled-cavity semiconductor
laser diodes," Opt. and Quantum Electron., vol. 30, no. 5/6,May 1998, pp. 287-303.
[15] Spuhler, M.M, B.J. Offrein, G.-L. Bona, R. Germann, and
D. Erni, "Compact spot-size converters using non-periodic
segments for high refractive index contrast planarwaveguides," Proc. CLEO/EUROPF/98, September 1998,
p. 234.
[16] Spuhler, M.M., B.J. Offrein. G.L. Bona, D. Erni, and I.
Massarek, "Design and implementation of short opticalspot-size converters" Bulletin SPG/SSP, vol. 15, no. 1,
February 1998, p. 94.
[17] Paul, H., and J. Tmdle, "Passive optical netivork planningin local access networks - an optimisation approachutilising genetic algoiithms" BT Technol. J., vol. 2, no. 2,
April 1996, pp. 110-115.
References 95
[18] Koh, S., and L. Ye, "Modeling, simulation, and verification
of photonic integrated circuits using VHDL." Proc. SPIE,vol. 2659, January 1996, pp. 234—211.
[19] Jakubczyk, Z., M. Florjanczyk, and V. Tzolov, "Modelingtools in waveguide optics," Proc. SPIE, vol. 2994, February1997, pp. 310-318.
[20] Glingener, C, J.-P. Elbers, and E. Voges, "CAE for photonic
waveguide devices," AEU Int. J. Electron. Commun.,vol. 51, no. 1, January 1997, pp. 20-28.
[21] Spofford, C.W., "A symposis of the AESD workshop on
acoustic propagation modelling by non-ray-tracing
techniques," Rpt. TN-73-05, Office of Naval Research,
Washington D.O., 1973.
[22] Tappert, F.D.. "The parabolic equation approximation," in
J.B. Keller and J.S. Papadakis, eds., Wave propagation and
underwater acoustics, Lecture Notes in Physics, vol. 70,
Springer, Berlin, 1977, pp. 224-287.
[23] Yevick D., and B. Hermansson, "Efficient beam propagationtechniques," IEEE J. Quantum Electron., vol. 26, no. 1,
January 1990, pp. 109-112.
[24] Tsuji Y., and M. Koshiba, "A finite element beam
propagation method for strongly guiding and longitudinally
varying optical waveguides," J. Lightwave Technol., vol. 14,no. 2, February 1996, pp. 217-222.
[25] Tsuji Y., M. Koshiba, and T. Shiraishi, "Finite element beam
propagation method for three-dimensional opticalwaveguide structures," J. Lightwave Technol., vol. 15, no. 9,
September 1997, pp. 1728-1734.
[26] Chung. Y., and N. Dagh, "An assessment of finite differencebeam propagation method" IEEE J. Quantum Electron.,vol. 26, no. 8, August 1990, pp. 1335-1339.
[27] Huang, W., C. Xu, S.-T. Chu, and S.K. Chaudhun, "The
finite-difference vector beam propagation method: analysisand assessment," J. Lightwave Technol., vol. 10, no. 3,March 1992, pp. 295-305.
[28] Hoekstra. H.J.W.M., G.J.M. Krijnen. and P.V. Lambeck,"New formulation of the beam propagation method based on
the slowly varying envelope approximation,'' Opt. Comm.,vol. 97, no. 5/6, April 1993, pp. 301-303.
96 References
[29] Lui, W.W., W.-P. Huang, and K. Yokoyama, "Precision
enhancement for optical waveguide analysis," J. LightwaveTechnol., vol. 15, no. 2, February 1997, pp. 391-396.
[30] Yamauchi J., J. Shibayama, O. Saito, 0. Uchiyama, and H.
Nakano, "Improved finite-difference beam-propagationmethod based on the generalized Douglas scheme and its
application to semivectorial analysis," J. LightwaveTechnol., vol. 14, no. 10, October 1996, pp. 2401-2406.
[31] Rao, H.. R. Scarmozzmo, and R.M. Osgood, "A bidirectional
beam propagation method for multiple dielectric interfaces,"IEEE Photon. Technol. Lett., vol. 11, no. 7, July 1999,
pp. 830-832.
[32] Hayashi, K., M. Koshiba, Y. Tsuji, S. Yoneta, and R. Kaji,"Combination of beam propagation method and mode
expansion propagation method for bidirectional opticalbeam propagation analysis," J. Lightwave Technol., vol. 16,
no. 11, November 1998, pp. 2040-2045.
[33] Mitomi, O., and K. Kasaya, "An improved semivectorial
beam propagation method using a finite-element scheme,"IEEE Photon. Technol. Lett., vol. 10. no. 12, December
1998, pp. 1754-1756.
[34] Xu, C.L., W.P. Huang, and S.K. Chaudhuri, "Efficient and
accurate vector mode calculations by beam propagationmethod" J. Lightwave Technol., vol. 11, no. 7, July 1993,
pp.1209-1215.
[35] Fogli, F., G. Bellanca, P. Bassi, I. Madden, and W.
Johnstone, "Highly efficient full-vectorial 3-D BPM
modeling of fiber to planar waveguide couplers," J.
Lightwave Technol., vol. 17, no. 1, January 1999, pp. 136-
143.
[36] Mansour I., A.-D. Capobianco. C. Rosa, "Noniterative
vectorial beam propagation method with a smoothingdigital filter" J. Lightwave Technol., vol. 14, no. 5, May1996, pp. 908-913.
[37] Xu, C.L., W.P. Huang, J. Chrostowski, and S.K. Chaudhuri,"A full-vectorial beam propagation method for anisotropicwaveguides," J. Lightwave Technol., vol. 12, no. 11,November 1994. pp. 1926-1931.
References 97
[38] Kriezis, E.E., and A.G. Papagiannakis, "A joint finite-difference and FFTfull vectorial beam propagation scheme"
J. Lightwave Technol., vol. 13, no. 4, April 1995, pp. 692-
700.
[39] Huang, W.P., C.L. Xu, S.T. Chu, and S.K. Chaudhuri, "A
vector beam propagation method for guided-wave optics,"IEEE Photon. Technol. Lett., vol. 3, no. 10, October 1991,
pp. 910-913.
[40] Heiblum, M., and J.H. Harris, "Analysis of curved optical
waveguides by conformai transformation," IEEE J.
Quantum Electron., vol. 11, no. 2, February 1975, pp. 75-
83.
[41] Lee C.-T., M.-L. Wu, L.-G. Sheu, P.-L. Fan, and J.-M. Hsu,
"Design and analysis of completely adiabatic taperedwaveguides by conformai mapping" J. Lightwave Technol.,vol. 15, no. 2,'February 1997, pp. 403-410.
[42] Wu, M.-L., P.-L. Fan. J.-M. Hsu, and C.-T. Lee, "Design ofideal structures for lossless bends in optical waveguides byconformai mapping" J. Lightwave Technol., vol. 14, no. 11,November 1996, pp. 2604-2614.
[43] Fan, P.-L., M.-L. Wu, and C.-T. Lee, "Analysis of abruptbent waveguides by the beam propagation method and the
conformai mapping method" J. Lightwave Technol., vol. 15,no. 6, June 1997, pp. 1026-1031.
[44] Rivera, M.. "A finite difference BPM analysis of bent
dielectric waveguides," J. Lightwave Technol., vol. 13, no. 2,
February 1995, pp. 233-238.
[45] Sewell, P., T.M. Benson, T. Anada, and P.C. Kendall, "Bi-
oblique propagation analysis of symmetric and asymmetricY-junctions" J. Lightwave Technol., vol. 15, no. 4, April1997, pp. 688-696.
[46] Lee, C.-T., M.-L. WYi, and J.-M. Hsu. "Beam propagationanalysis for tapered ivaveguides: taking account of the
curved phase-front effect in paraxial approximation," J.
Lightwave Technol., vol. 15, no. 11, November 1997,
pp.2183-2189.
[47] Haus, H.A., "Waves and fields in optoelectronics," Prentice-
Hall, Englewood Cliffs. NJ. 198 1, 402 p.
98 References
[48] Lee, P.-C, and E. Voges, "Three-dimensional semi-vectorial
wide-angle beam propagation method" J. LightwaveTechnol., vol. 12, no. 2, February 1994, pp. 215-225.
[49] Yamauchi, J., J. Shibayama, M. Sekiguchi, and H. Nakano,
"Improved multistep method for wide-angle beam
propagation"
IEEE Photon Technol. Lett., vol. 8, no. 10,October 1996, pp. 1361-1363.
[50] Press, W.H., S.A. Teukolsky, W.T. Vetterling, and B.P.
Flannery, "Numerical recipes in C, the art of scientific
computing" 2nd ed., Cambridge University Press, 1992,994 p.
[51] Vinsome, P.K.W., "Orthomin, an iterative method forsolving sparse sets of simultaneous linear equations," Proc.
Fourth SPE Symp. on Numerical Simulation of Reservoir
Performance, Soc. Pet. Eng. of AIME, February 1976,
pp. 149-159.
[52] Hadley, G.R., "Transparent boundary condition for the
beam propagation method" IEEE J. Quantum Electron.,vol. 28, no. 1, January 1992, pp. 363-370.
[53] Vassallo, C. and J.M. van der Keur, "Highly efficienttransparent boundary conditions for finite difference beam
propagation method at order four," J. Lightwave Technol.,vol. 15, no. 10, October 1997, pp. 1958-1965.
[54] Vassallo, C, and F. Collino. "Highly efficient absorbingboundary conditions for the beam propagation method," J.
Lightwave Technol., vol. 14, no. 6, June 1996, pp. 1570-
1577.
[55] Hoekstra, H.J.W.M., G.J.M. Krijnen, and P.V. Lambeck,
"Efficient interface conditions for the finite difference beam
propagation method" J. Lightwave Technol., vol. 10, no. 10,October 1992, pp. 1352-1355.
[56] Yevick, D., and B. Hermansson, "New formulations of the
matrix beam propagation method: application to rib
waveguides," IEEE J. Quantum Electron., vol. 25, no. 2,
February 1989. pp. 221-229.
[57] Chen, J.C, and S. Jüngling, "Computation of higher-orderwaveguide modes by imaginary-distance beam propagationmethod," Opt. and Quantum Electron., vol. 26, no. 3. March
1994, pp. S199-S205.
References 99
[58] Yamauchi, J., M. Sekiguchi, 0. Uchiyama, J. Shibayama,and H. Nakano, "Modified, finite-difference formula for the
analysis of semivectorial modes in step-index opticalwaveguides," IEEE Photon. Technol. Lett., vol. 9, no. 7,
July 1997, pp. 961-963.
[59] Wijnands, F., H.J.W.M. Hoekstra, G.J.M. Krijnen, and
R.M. de Ridder, "Numerical solution method of nonlinear
guided modes with a finite difference complex axis beam
propagation method," IEEE J. Quantum Electron., vol. 31,
no. 5, May 1995, pp. 782-790.
[60] Lee, J-S., and S.-Y. Shin, "On the validity of the effective-index method for rectangular dielectric waveguides," J.
Lightwave Technol, vol. 11, no. 8, August 1993, pp. 1320-
1324.
[61] Yamauchi, J., G. Takahashi, and H. Nakano, "Modifiedfini te-difference formula for semivectorial H-field solutions
of optical waveguides," IEEE] Photon. Technol. Lett.,vol. 10, no. 8, August 1998, pp. 1127-1129.
[62] Wijnands, F., H.J.W.M. Hoekstra, G.J.M. Krijnen, and
R.M. de Ridder, "Modal fields calculation using the finitedifference beam propagation method," J. LightwaveTechnol, vol. 12, no. 12, December 1994, pp. 2066-2072.
[63] Huang, WT., and H.A. Haus, "A simple variational approachto optical rib ivaveguides," J. Lightwave Technol., vol. 9,no. 1, January 1991, pp. 56-61.
[64] Thylén, L., "The beam propagation method: an analysis ofits applicability" Opt. and Quantum Electron., vol. 15,no. 5, September 1983, pp. 433-439.
[65] Takagi. A.. K. Jinguji, and M. Kawachi, "Silica-based
waveguide-type wavelength -insensitive couplers (WINC's)with series-tapered coupling structure," L. LightwaveTechnol., vol. 10, no. 12, December 1992, pp. 1814-1824.
[66] Ramadan, T.A., R. Scarmozzmo, and R.M. Osgood, Jr.,
"Adiabatic couplers: design rules and optimization" J.
Lightwave Technol., vol. 16, no. 2, February 1998, pp. 277-
283.
100 References
[67] Veldhuis, G.J., J.H. Berends, and P.V. Lambeck, "Designand characterization of a mode-splitting lF-junction" J.
Lightwave Technol., vol. 14, no. 7, July 1996, pp. 1746-
1752.
[68] Klekamp, A., P. Kersten, and W. Rehm, "An improvedsingle-mode Y-branch design for cascaded 1:2 splitters," J.
Lightwave Technol., vol. 14, no. 12, December 1996,
pp. 2684-2686.
[69] Offrein, B.J., G.L. Bona, F. Horst. H.W.M. Salemink, R.
Beyeler, and R. Germann, "Wavelength tunable optical add-
after-drop filter with flat passband for WDM networks,"IEEE Photon. Technol. Lett., vol. 11, no. 2, February 1999,
pp. 239-241.
[70] Kuznetsov. M., "Cascaded coupler Mach-Zehnder channel
dropping filters for wavelength-division-multiplexed opticalsystems," J. Lightwave Technol., vol. 12, no. 2, February1994, pp. 226-230.
[71] Dragone, C, "Efficient N x N star couplers using fourieroptics" J. Lightwave Technol., vol. 7, no. 3, March 1989,
pp. 479-489.
[72] Fischer, U., T. Zmke, J.-R. Kropp, F. Arndt, and K.
Petermann, "0.1 dB/cm waveguide losses in single-modeSOI rib Waveguides," IEEE Photon. Technol. Lett., vol. 8,no. 5, May 1996, pp. 647-648.
[73] Soref, R.A., J. Schmidtchen, and K. Petermann, "Largesingle-mode rib waveguides in GeSi-Si and Si-on-Si02,"IEEE J. Quantum Electron., vol. 27. no. 8, August 1991,pp.1971-1974.
[74] Kohl, A.. R. Zengerle, H.J. Bruckner, H.J. Olzhausen, R.
Müller, and K. Heime, "Fabrication of buried InP/InGaAsP
waveguide-wires for use in optical beamwidth
transformers," Microelectronic Engineering, vol. 21, April1993, pp. 483-486.
[75] Kawano, K., M. Kohtoku, N. Yoshimoto, S. Sekine, and Y.
Noguchi, "2 x 2 InGaAlAs/InAlAs multiquantum well
(MQW) directional coupler waveguide switch modules
integrated with spotsize convertor^s," IEE Electron. Lett.,
vol. 30, no. 4, February 1994. pp. 353-355.
References 101
[76] Wenger, G., L. Stoll, B. Weiss, M. Schienle, R. Müller-
Nawrath, S. Eichinger, J. Müller, B. Acklin, and G. Müller,
"Design and fabrication of monolithic optical spot size
transformers (MOST's) for highly efficient fiber-chip
coupling," J. Lightwave Technol., vol. 12, no. 10, October
1994, pp. 1782-1790.
[77] Mitomi, 0., K. Kasaya, and H. Miyazawa, "Design of a
single-mode tapered waveguide for low-loss chip-to-fiber
coupling," [EEE J. Quantum Electron., vol. 30, no. 8,
AugusU994, pp. 1787-1793.
[78] de Ridder, R.M., R.A. Wijbrans, H. Albers, J.S. Aukema,P.V. Lambeck, H.J.W.M. Hoekstra, and A. Driessen, "A
spot-size transformer for fiber-chip coupling in sensor
applications at 633 nm in silicon oxynitride," Proc.
LEOS'95, vol. 2, October 1995, pp. 86-87.
[79] Zengerle, R., O. Leminger, W. Weiershausen, K. Faltin, and
B. Hubner, "Laterally tapered InP-InGaAsP waveguides forlow-loss chip-to-fiber butt coupling: a comparison ofdifferent configurations" IEEE Photon. Technol. Lett.,vol. 7, no. 5, May 1995, pp. 532-534.
[80] Zengerle, R., B. Jacobs, WT. Weiershausen, K. Faltin, and A.
Kunz, "Efficient spot-size transformation using spatiallyseparated tapered InP/InGaAsP twin waveguides," J.
Lightwave Technol., vol. 14, no. 3, March 1996, pp. 448-
453.
[81] Mitomi, 0.. K. Kasaya, Y. Tohmori, Y. Suzaki, H. Fukano,Y. Sakai, M. Okamoto, and S.-I. Matsumoto. "Optical spot-size converters for low-loss coupling between fibers and
optoelectronic semiconductor devices," J. LightwaveTechnol., vol. 14, no. 7, July 1996, pp. 1714-1720.
[82] Paatzsch, T., I. Smaglinski, M. Abraham, H.-D. Bauer, U.
Hempelmann, G. Neumann, G. Mrozynski, and W.
Kerndlmaier. "Very low-loss passive fiber-to-chip couplingwith tapered fibers," Appl. Opt., vol. 36, no. 21, July 1997,
pp.5129-5133.
[83] Kawano, K., M. Kohtoku, M. Wada. H. Okamoto, Y. Itaya,and M. Naganuma, "3-D semivectorial beam propagationanalysis of a spotsize-converter-integrated laser diode in the
1.3-pm-wavelength region," IEEE Photon. Technol. Lett.,vol. 9, no. 1, January 1997, pp. 19-21.
102 References
[84] Weissmann, Z., and I Hendel. "Analysis of periodically
segmented waveguide mode expanders" J. LightwaveTechnol., vol. 13, no. 10, October 1995, pp. 2053-2058.
[85] Chou, M.H., M.A. Arbore, and M.M. Fejer, "Adiabatically
tapered periodic segmentation of channel waveguides formode-size transformation and fundamental mode
excitation," Opt. Lett., vol. 21, no. 11, June 1996, pp. 794-
796.
[86] Weissmann, Z., and A. Hardy, "Modes of periodicallysegmented waveguides," J. Lightwave Technol., vol. 11,
no. 11, November 1993, pp. 1831—1838.
[87] Weissman, Z., and A. Hardy, "2-D mode tapering via
tapered channel waveguide segmentation," ÏEE Electron.
Lett., vol. 28, no. 16, July 1992, pp. 1514-1516.
[88] Liang, T., and R.W7. Ziolkowski, "Mode conversion of
ultrafast pulses by grating structures in layered dielectric
waveguides," J. Lightwave Technol., vol. 15, no. 10, October
1997^pp. 1966-1973.
[89] Offrem, B.J., G.L. Bona, R. Germann, I. Massarek, and
H.W.M. Salemmk, "High contrast and low loss SiON
optical waveguides by PECVD," Proc. IEEE/LEOS Symp.,Benelux Chapter, November 1996, pp. 290-293.
[90] Hoffmann, M., P. Kopka, and E. Voges, "Low-loss fiber-matched low-temperature PECVD waveguides with small-
core dimensions for optical communication systems," IEEE
Photon. Technol. Lett., vol. 9, no. 9, September 1997,
pp.1238-1240.
[91] Elyutin, P.V.. and F. Ladouceur, "Adiabaticity and random
wave propagation," Opt. Lett., vol. 22, no. 16, August 1997,
pp. 1241-1243.
[92] Ladouceur, F., "Roughness, inhomogeneity, and integratedoptics," J. Lightwave Technol., vol. 15, no. 6, June 1997,
pp.1020-1025.
[93] Goh, T., S. Suzuki, and A. Sugita, "Estimation of waveguidephase error in silica-based waveguides," J. LightwaveTechnol, vol. 15, no. 11, November 1997, pp. 2107-2113.
[94] Holland, J.H, "Adaptation m natural and artificialsystems" The University of Michigan Press, 1975, 183 p.
References 103
[95] Grefenstette, J.J., and Baker, J.E.. "How genetic algorithmswork: a critical look at implicit parallelism," Proc. Third
Intern. Conf. on Genetic Algorithms, June 1989, pp. 20-27.
[96] Winter, G., J. Périaux. M. Galân, and P. Cuesta, eds.,"Genetic algorithms in engineering and computer science,"
Wiley, New York, Chichester, 1995, 464 p.
[97] Fröhlich, J., "Evolutionary optimization for computational
electromagnetics" Diss. ETH No. 12232, 1997.
[98] Moosburger. R., C. Kostrzewa, G. Fischbeck, and K.
Petermann, "Shaping the digital optical switch usingevolution strategies and BPM," Proc. ECIO'97, April 1997,
pp. 22-25.
[99] Johnson, J.M., and Y. Rahmat-Samii, "Genetic algorithmsin engineering electromagnetics" IEEE Ant. Prop. Mag.vol. 39, no. 4, August 1997, pp. 7-25.
[100] Goldberg, D.E.. "Genetic algorithms in search, optimization,and machine learning," Addison-Wesley, Reading, MA,
1989, 412 p.
[101] Syswerda G., "A study of reproduction in generational and
steady-state genetic algorithms" m Foundations of Genetic
Algorithms, edited by G.J.E. Rawlins, Morgan Kaufmann
Publishers, San Mateo, CA, 1991, pp. 94—101.
[102] Orta, R., P. Savi, R. Tascone, and D. Trinchero, "Synthesis
of multiple-ring-resonator filters for optical systems," IEEE
Photon. Technol. Lett., vol. 7, no. 12, December 1995,
pp.1447-1449.
[103] Jinguji, K,, "Synthesis of coherent two-port optical delay-line circuit witJi ring waveguides" J. Lightwave Technol,vol. 14, no. 8, August 1996. pp. 1882-1898.
[104] Oppenheim. A.V.. and R.W. Schäfer. "Discrete-time signalprocessing," International Edition, Prentice-Hall,
Englewood Cliffs. NJ, 1989, 879 p.
[105] Dragone, C, C.H. Henry, LP. Kaminow, and R.C. Kistler,
"Efficient multichannel integrated optics star coupler on
silicon"
IEEE Photon. Technol Lett., vol. 1, no. 8, August1989, pp. 241-243.
[106] Dragone, C, "An N x N optical multiplexer using a planararrangement of two star couplers," IEEE Photon. Technol.
Lett., vol 3, no. 9, September 1991, pp. 812-815.
104 References
[107] Jinguji, K., and M. Kawachi, "Synthesis of coherent two-port
lattice-form optical delay-line circuit," J. Lightwave
Technol, vol. 13, no. 1, January 1995, pp. 73-82.
[108] Barbarossa, G., A.M. Matteo, and M.N. Armenise,
"Theoretical analysis of triple-coupler ring-based optical
guided-wave resonator," J. Lightwave Technol, vol. 13,
no. 2, February 1995, pp. 148-157.
[109] Barbarossa, G., A.M. Matteo, and M.N. Armenise,
"Compound triple-coupler ring resonators for selective
filtering applications in optical FDM transmission systems,"Proc. SPIE, vol. 2401, February 1995, pp. 86-94.
[110] Monaco, V.A., and P. Tiberio, "Automatic scattering matrix
computation of microwave circuits," Alta Frequenza,
vol. 39, no. 2, February 1970, pp. 165-170.
[Ill] Epprecht, G., "Die S-Matrix zusammengesetzter Systeme",
Skript, Kurs 35-136, Mikrowellen-Laboratorium ETHZ,
1974, pp. 1-9.
[112] Marcuse, D., "Tlieory of dielectric optical waveguides,"Academic Press, Inc., San Diego, CA, 1991, 380 p.
[113] Working Group I, COST 216, "Comparison of different
modelling techniques for longitudinally invariant integrated
optical waveguides," IEE Proc, Part J, Optoelectronics,vol. 136, no. 5, October 1989, pp. 237-280.
[114]Yariv, A., "Coupled-mode theory for guided-wave optics,"IEEE J. Quantum Electron., vol. 9, no. 9, September 1973,
pp. 919-933.
[115] Taylor, H.F., and A. Yariv, "Guided wave optics," Proc.
IEEE, vol. 62, no. 8, August 1974, pp. 1044-1060.
[116] Hardy, A., and W. Streifer. "Coupled mode theory of
parallel waveguides," J. Lightwave Technol, vol. 3, no. 5,
October 1985, pp. 1135—1146.
[117] März, R., "Integrated optics, design and modeling," Artech
House, Inc., Norwood, MA, 1995, 336 p.
[118] Little, B.E., "Filter syntliesis for coupled ivaveguides," J.
Lightwave Technol, vol. 15. no. 7, July 1997, pp. 1149-
1155.
References 105
[119]Madsen, CK., and G. Lenz, "Optical all-pass filters forphase response design with applications for dispersioncompensation," IEEE Photon. Technol. Lett., vol. 10, no. 7,
July 1998, pp. 994-996.
[120] Soref, R.A., and B.E. Little, "Proposed N-Wavelength M-
Fiber WDM CrossConnect switch using active microringresonators," IEEE Photon. Technol. Lett., vol. 10, no. 8,
August 1998, pp. 1121-1123.
[121] Oda, K., N. Takato, and H. Toba, "A wide-FSR double-ringresonator for optical FDM transmission systems," J.
Lightwave Technol, vol. 9, no. 6, June 1991, pp. 728-736.
[122] Chu, S.T., W. Pan, S. Sato, T. Kaneko. B.E. Little, and Y.
Kokubun, "Wavelength trimming of a, microring resonator
filter by means of a UV sensitive polymer overlay" IEEE
Photon. Technol. Lett., vol. 11, no. 6, June 1999, pp. 688-
690.
Acknowledgments
Many people were involved in this project. It is nearly
impossible to name all of them. Nevertheless I want to give some
special thanks.
First of all I want to thank my professor. Werner Bächtold for
his support and for giving me the possibility to make this work
at the IFH. I like also to thank René Dändliker and Gian-Luca
Bona for their interest in this work and for acting as co-
examiners. Special thanks go to Daniel Erni, who always had the
right answer and kept away from me a lot of administrative
stuff. He gave me a lot of support and 1 am very grateful to him.
I want to mention my colleagues and friends Nicolas Piller and
Eric Nicolet. Nicolas gave me the motivation to pursue the
studies at the ETH, without him I wouldn't stand here. I am also
very grateful to Eric. Not only was he a good friend during the
studies, our semester project was the basis for this Ph.D. thesis.
Many thanks go to my colleagues at the laboratory. EspeciallyI want to mention Martin Schmatz, with whom I had many
fruitful discussions, Hansruedi Benedickter. who was the movingspirit of the institute. Ray Ballisti was always present to helpwith the computer problems. Many thanks go to Jürg Fröhlich,
whose genetic algorithm was the bootstrap of my research
activities. I also thank the colleagues of the optics group,
especially Dorothea WTesmann, for the good ambiance and many
discussions. Finally I want to thank Charlotte Biber for the final
reading of this text.
From the IBM Research Laboratory I want to give specialthanks to Gian-Luca Bona. Bert Offrein, Roland Germann and
Folkert Horst. Their support and the many fruitful discussions
had a big influence for the successful outcome of this work.
I also want to thank my mother and my family for their
continuous support and love during my whole studies.
Finally, I want to thank my wife Eva for the infinite patience,help and love she gave me during this work.
107
Curriculum Vitae
Name: Michael—Martin SPÜHLER
Date of birth: April 30, 1969 in Fribourg, Switzerland
Nationality: Swiss, citizen of Zürich and Wasterkingen ZH
Married to Eva, one son Simon, born on
December 22nd 1999.
Education:
8.1976-7.1984: Primary and secondary school
8.1984-6.1988: Electromechanical apprenticeship10.1988-11.1991: Studies in electrical engineering at the Ecole
d'Ingénieurs de Fribourg.11.1991: Diploma (HTL) in electrical engineering;
diploma thesis: "Visualisation et commande
d'un processus chimique".11.1992—4.1996: Studies in electrical engineering at the Swiss
Federal Institute of Technology (ETH), Zurich.
4.1996: Diploma in electrical engineering; diplomathesis: "Charakterisierung von unscharfen
blobförmigen 3-D Strukturen.
10.1994-10.1997: Studies in teaching (Teaching Certificate in
Higher Education) at the ETH Zurich.
5.1996-1.2000: Research assistant and doctoral student at the
Swiss Federal Institute of Technology (ETH),
Laboratory for Electromagnetic Fields and
Microwave Electronics.
Professional Experience:
10.1989-12.1991: Electronic developments at Contrinex SA,
Fribourg, Switzerland.
1.1992—10.1992: Full-time software engineer at Sintro
Electronics AG, Interlaken, Switzerland.
1.1993-8.1995: Software developments at Sintro Electronics
AG, Interlaken. Switzerland.
since 10.1996: Partial-time lecturer at the University of
Applied Sciences (former HTL), Zurich,Switzerland.
109
Publications
Journal Papers and Chapters in Books:
[1] Spühler, M.M., D. Erni, and J. Fröhlich, "An evolutionaryoptimization procedure applied to the svnthesis of integrated
spot-size converters," Opt. and Quantum Electron., vol. 30,
no. 5/6, May 1998. pp. 305-321.
[2] Erni, D., M.M. Spuhler, and J. Fröhlich. "Evolutionaryoptimization of non-periodic coupled-cavity semiconductor
laser diodes," Opt. and Quantum Electron., vol. 30, no. 5/6,
May 1998, pp. 287-303.
[3] Spuhler, M.M., B.J. Öftrem, G.L. Bona, R. Germann, I.
Massarek, and D. Erni, "A very short planar silica spot-sizeconverter using a non-periodic segmented waveguide," J.
Lightwave Technol, vol. 16, no. 9, September 1998,
pp.1680-1685.
[4] Spuhler, M M., D. Wiesmann, P Freuler, and M. Diergardt,"Direct computation of higher-order propagation modes using
the imaginary-distance beam propagation method," Opt. and
Quantum Electron., vol. 31, no. 9/10, October 1999, pp. 751-
761.
[5] Erni, D., D. Wiesmann, M.M. Spuhler, S. Hunziker, B.
Oswald, J. Fröhlich, and C Hafner, "Evolutionaryoptimization algorithms in computational optics," Chapter mRecent Research Developments in Optical Engineering,Research Signpost. Trivandrum. India. 1999, pp. 19-36.
[6] Spuhler, MM., and D. Erni. "Towards structural
optimization of planar integrated lightwave circuits," Opt.and Quantum Electron
.in press
[7] Erni, Ü., D. Wiesmann, M.M. Spuhler, S. Hunziker, E.
Moreno, B. Oswald. J Fröhlich, and C. Hafner, "Applicationof evolutionary optimization algorithms in computationaloptics," ACES Journal submitted
111
112 Publications
Conferences and Workshops:
[8] Erni, D., M.M. Spühler, and J. Fröhlich, "A generalized
evolutionary optimization procedure applied to waveguidemode treatment in non-periodic optical structures," Proc.
Eurpean Conf. on Integrated Optics (ECIO), Stockholm,
Sweden, April 1997, pp. 218-221.
[9] Spühler, M.M., D. Erni, and J. Fröhlich, "Topological
investigations on evolutionary optimized non-periodic optical
structures," Int. WTorkshop on Optical Waveguide Theory and
Numerical Modeling, Twente, the Netherlands, September1997.
[10] Spühler, M.M., B.J. Offrein, G.L. Bona, D. Erni, and I.
Massarek, "Design and implementation of short optical spot-
size converters," SPG Jahrestagung, Bern, Bulletin
SPG/SSP, vol. 15, no. 1, February 1998, p. 94.
[11] Spuhler, M.M., B.J. Offrein, G.L. Bona, R. Germann, and D.
Erni, "Compact spot-size converters using non-periodic
segments for high refractive index contrast planar
ivaveguides," Conf. on Lasers and Electro-Optics (CLEO),
Glasgow, UK, Technical Digest, September 1998, p. 234.
[12] Spuhler, M.M., Ü. Wiesmann, P. Freuler, M. Diergardt, and
D. Erni, "Accelerated computation of higher-order
propagation modes using the imaginary-distance BPM" Int.
Workshop on Optical Waveguide Theory and Numerical
Modeling, Hagen, Germany, September 1998.
[13] Spuhler, M.M.. and D. Erni, "A design and optimization
platform for integrated optica devices" SPG Jahrestagung,Bern, Bulletin SPG/SSP, vol. 16, February 1999. p. 43.
[14] Spuhler, M.M., D. Wiesmann, and D. Erni, "Evolutionaryoptimization in computational optics," invited, Proc.
Progress in Electromagn. Research Symp. (PTERS), Taipei,Taiwan, March 1999, p. 765.
Publications 113
[15] Neuhold, S.M., M.L. Schmatz. M. Hässig, M.M. Spühler, and
G. Storf, "Combined broad and narrow band multichannel
PD measurement system with high, sensitivity for GIS" Proc.
Eleventh Int. Symp. on High-Voltage Engineering (ISH),
London, UK, vol. 5, August 1999, pp. 152-155.
[16] Spühler, M.M., and D. Erni, "Structural optimization in
planar integrated optics," Int. Workshop on Waveguide
Theory and Numerical Modeling. Saint-Etienne, France,
September 1999.