Beruflich Dokumente
Kultur Dokumente
Doctoral Thesis
Author(s):
Kaufmann, Thomas
Publication Date:
2011
Permanent Link:
https://doi.org/10.3929/ethz-a-006532435
Rights / License:
In Copyright - Non-Commercial Use Permitted
This page was generated automatically upon download from the ETH Zurich Research Collection. For more
information please consult the Terms of use.
ETH Library
DISS. ETH No. 19622
ETH ZURICH
Doctor of Sciences
presented by
THOMAS KAUFMANN
2011
I may not have gone where I intended to go,
but I think I have ended up where I needed to be.
Douglas Adams
Abstract
Meshless methods are a class of numerical methods with unique advantages over classical
mesh-based methods. Instead of calculating the solution of physical equations on a mesh
topology, a solution is sought on a set of collocation nodes. A high flexibility in the
placement of the nodes allows to accurately model complex geometries.
Radial basis functions (RBFs) show superb approximation properties and are applied
in many fields of research where accurate interpolation of scattered data point is required.
Numerical solvers for physical equations, i.e. partial differential equations, with radial
basis functions gained in interest recently due to the ability of obtaining highly accurate
results for a relatively low computational effort. The high accuracy comes at the cost
of high matrix condition numbers especially for flat basis functions. Due to the radial
dependence of the basis functions, implementations in two or three dimensional settings
are formulated straightforwardly.
A special variant of RBF methods is the radial point interpolation method (RPIM). In
a preprocessing step, a new set of interpolating shape functions is calculated that shows
extraordinary properties, namely the ability to use explicit time-stepping and a beneficial
structure for matrix solvers. In a localized scheme on one hand, each collocation node is
only assumed to influence the surrounding neighbors in a so-called support domain. This
leads to sparse matrices, even for large domains. For global basis functions on other hand,
very high accuracy can be achieved, but the computational effort increases drastically
when large problems are solved. A possible remedy is to divide the domain into several
smaller support domains through a domain-decomposition approach.
In this thesis, the properties of the RPIM scheme are summarized and an algorithm
to optimize the flatness of the basis functions is introduced. A two-dimensional frame-
work in electromagnetics is developed for the first- and second-order form of Maxwells
equations. The first-order form yields solutions in Cartesian and cylindrical coordinates
in the time domain in a similar approach to the generalized finite-difference time-domain
method. This allows to simulate transient signals and to obtain results of a broad fre-
quency range in one simulation run. Important material boundary conditions and a
stability criterion for stable time iterations are presented. Investigations on the long-
time stability are performed by studying the corresponding eigenvalue problem. The
second-order implementation solves the wave equation in frequency domain using global
basis functions. First an RPIM eigenvalue solver for resonant cavities is developed and
compared to other preexisting RBF methods. Very similar numerical results are obtained
for all methods. Second a solver for source problems is implemented in the RPIM frame-
work and a solution is calculated through a matrix inversion. Domain-decomposition
strategies are proposed to handle material discontinuities.
Absorbing boundary conditions to simulate open problems are developed. For the first-
order framework, perfectly matched layers are introduced. In the second-order scheme,
v
Abstract
due to the high accuracy offered by global basis functions, high-order non-reflecting
boundary conditions can be directly implemented. An algorithm to optimize these bound-
ary conditions is presented. To improve the accuracy offered by the high-order boundary
conditions, a procedure to progressively increase the node density close to the boundary
is introduced. Numerical evaluations show a very good correspondence between theoret-
ically expected results and numerical solutions.
For the second-order RPIM eigenvalue solver, adaptive refinement strategies are intro-
duced. The collocation node distribution is refined iteratively based on two alternative
a posteriori error estimators on a set of test nodes. The first estimator calculates the
residual error on the test nodes and the second estimator indicates regions with large
jumps in the gradient. Nodes with large estimated errors are subsequently added to the
set of collocation nodes. This iteration is repeated until a stopping criterion is fulfilled.
Numerical experiments show the effectiveness of these two refinement strategies. Regard-
less of the chosen error estimation approach, both algorithms perform very similarly and
show much lower numerical errors than a naive uniform refinement.
The RPIM framework developed in this thesis is finally evaluated on a number of nu-
merical examples. For the time-domain solver, the input reflections of a waveguide bend
are calculated. Solutions with conformal node placement converge much faster than with
a rectangular setup with stair-casing effects. In cylindrical coordinates, the simulation
of a corrugated copper cable shows a very good correspondence with a commercially
available field solver. For the second-order eigenvalue solver, comparisons with differ-
ent finite-element codes reveal a much higher accuracy than lower-order finite-element
methods and even better performance than a high-order discontinuous Galerkin code
with curvilinear elements.
vi
Zusammenfassung
Gitterfreie Methoden gehoren zu einer neuen Art von numerischen Methoden, die einige
wichtige Vorteile gegenuber herkommlichen, gitterbasierten Methoden aufweisen. Git-
terfreie Methoden konnen die Losung eines physikalischen Gesetzes an frei wahlbaren
Kollokationsstellen berechnen und sind somit nicht an eine Gittertopologie gebunden.
Der grosse Vorteil einer flexiblen Platzierung dieser Punkte liegt in einem einfachen
Modellieren von komplexen Geometrien.
Radiale Basisfunktionen (RBF), die ausgezeichnete Eigenschaften in der Annaherung
von Funktionen mit unstrukturiert verteilten Stutzstellen aufweisen, sind fur gitterfreie
Methoden besonders geeignet. Numerische Verfahren zur Losung von Partiellen Differ-
entialgleichungen unter Anwendung von RBF haben in letzter Zeit wegen der sehr ho-
hen Rechengenauigkeit bei einem relativ kleinen Rechnenaufwand an Bedeutung gewon-
nen. Die Rechengenauigkeit wird dabei auf Kosten einer grossen Matrixkonditionszahl
erreicht, die durch flache Formen der Basisfunktionen bedingt ist. Dank der radialen
Abhangigkeit der Basisfunktionen ist eine Implementierung in zwei und drei Dimensio-
nen relativ unkompliziert.
Ein spezieller Typ von Methoden mit RBF ist die Radiale Punktinterpolations-
Methode (RPIM). In einem Vorverarbeitungsschritt werden vor dem Losen der Glei-
chungen neue Naherungsfunktionen berechnet, die nutzliche Eigenschaften aufweisen.
Die Vorteile liegen in der Moglichkeit eines expliziten Zeitschrittverfahrens und einer
gunstigen Struktur zur Matrizeninvertierung. In einer lokalen Formulierung wird davon
ausgegangen, dass jeder Punkt nur von seinen Nachbarn innerhalb eines kompakten
Raumgebiets beeinfluss wird. Dies fuhrt zu einer schwach besetzten Matrix. Globale
Formulierungen dagegen haben eine sehr hohe Rechengenauigkeit, jedoch nimmt der
Rechenaufwand drastisch zu, wenn grosse Probleme berechnet werden. Eine mogliche
Abhilfe bietet die Aufteilung des Rechengebiets in einzelne kleinere Teilbereiche mit
Hilfe einer Gebietszerlegungstechnik.
In dieser Dissertation wird eine Implementierung von RPIM fur elektromagnetische
Probleme vorgestellt. Zuerst werden die Eigenschaften von RPIM zusammengefasst und
ein Algorithmus eingefuhrt, welcher die Form der Basisfunktionen optimiert. Fur zwei-
dimensionale Probleme wird eine Grundstruktur zum Losen der Maxwell-Gleichungen in
erster und in zweiter Ordnung vorgeschlagen. Das Losungsverfahren fur Probleme erster
Ordnung wird im Zeitbereich fur kartesische und zylindrische Koordinatensysteme ahn-
lich der Finiten-Differenzen-Methode im Zeitbereich (FDTD) formuliert. Das ermoglicht
die Modellierung von zeitabhangigen Signalen und somit die Simulation eines breiten
Frequenzspektrums in einem einzigen Durchgang. Dazu werden wichtige Randbedingun-
gen und ein Stabilitatskriterium fur eine stabile Zeititeration erortert. Das Losungsver-
fahren fur Probleme zweiter Ordnung lost die Wellengleichung im Frequenzbereich mit
Hilfe von globalen Basisfunktionen. Zunachst wird ein Verfahren zur Berechnung der
vii
Zusammenfassung
Eigenwerte von resonanten Strukturen mittels RPIM eingefuhrt. Das Verfahren wird
mit anderen existierenden RBF Methoden verglichen, wobei sehr ahnliche Resultate er-
reicht werden. Danach wird ein Verfahren zum Losen der inhomogenen Wellengleichung
implementiert, das auf einer numerischen Matrixinvertierung basiert. Schliesslich werden
Gebietszerlegungstechniken diskutiert, die es ermoglichen Materialubergange korrekt zu
simulieren.
Sogenannte Absorbing Boundary Conditions (ABCs) dienen der Imitation von offenen
Rechengebieten. Fur das Verfahren erster Ordnung wird die Perfectly Matched Layers
(PML) Methode eingefuhrt. Im Verfahren zweiter Ordnung konnen ABCs hoher Ordnung
dank der hohen Rechengenauigkeit der globalen Basisfunktionen direkt eingebunden
werden. Der Reflexionkoeffizient wird dabei mit einer Optimierungsmethode reduziert.
Um die Interpolationsgenauigkeit der Ableitungen hoherer Ordnung zu verbessern, wird
ein Punkteplazierungsverfahren vorgeschlagen, welches die Kollokationsstellen nahe am
Rand des Rechengebietes anordnet. Bei beiden Implementationen wird eine sehr gute
Ubereinstimmung zwischen den numerischen Resultaten und den theoretischen Werten
erzielt.
Die Rechengenauigkeit kann durch zwei verschiedene adaptive Verfeinerungsstrategien
fur das Eigenwertverfahren zweiter Ordnung Schritt fur Schritt erhoht werden. Die Punk-
teverteilung wird iterativ verfeinert, indem eine von zwei a posteriori Schatzfunktionen
auf zusatzlichen Testpunkten ausgewertet wird. Die eine Schatzfunktion bestimmt das
Residuum auf den Testpunkten und die andere Schatzfunktion benutzt den Sprung im
Gradienten zwischen zwei Stutzstellen als Fehlerindikator. Testpunkte mit einem hohen
geschatzten Fehler werden zu den Kollokationsstellen hinzugefugt und die Iteration dann
so lange wiederholt bis ein Abbruchkriterium erreicht ist. In numerischen Experimenten
wird die Wirksamkeit dieser Strategien nachgewiesen. Beide Ansatze konvergieren deut-
lich schneller als eine einfache Strategie mit gleichformiger Punkteverfeinerung.
Schliesslich werden die vorgeschlagenen Verfahren an verschiedenen numerischen
Beispielen getestet. Fur das Zeitbereichsverfahren erster Ordnung werden die Eingangsre-
flexionen eines gekrummten Wellenleiters berechnet. Dank der konformen Annhaherung
der Kollokationsstellen an die Geometrie konvergiert die Losung schneller als auf einem
Rechteckgitter. Die Simulation eines Koaxialkabels mit gewellten Wanden zeigt eine
sehr gute Ubereinstimmung mit einem kommerziellen Programm, das auf der Methode
der Finite-Elemente-Methode (FEM) basiert. Vergleiche des Frequenzbereichsverfahrens
zweiter Ordnung mit verschiedenen FEM Codes zeigen, dass eine grossere Genauigkeit
erreicht werden kann als mit FEM niedriger Ordnung und sogar eine hohere Genauigkeit
als mit der diskontinuierlichen Galerkin Methode hoherer Ordnung und kurvenformigen
Elementen.
viii
Contents
Abstract v
Zusammenfassung vii
1. Introduction 1
1.1. Mesh-based Numerical Methods in Electromagnetics . . . . . . . . . . . . 1
1.2. Meshless Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3. Radial Basis Function Methods . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4. Overview over the Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
ix
CONTENTS
x
CONTENTS
7. Conclusion 143
7.1. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
7.2. Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
7.3. Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
Bibliography 159
Acknowledgement 171
xi
List of Figures
xiii
LIST OF FIGURES
xiv
LIST OF FIGURES
6.1. Physical geometry of the 90 H-plane waveguide bend. The model is fed
by a Gaussian broadband pulse and truncated by a perfectly matched layer.132
6.2. Comparison of conformal regular and node placement to model the waveg-
uide bend. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
6.3. Convergence for the VSWR in rectangular and conformal grid for increas-
ingly fine node distributions. . . . . . . . . . . . . . . . . . . . . . . . . . 134
6.4. Schematic of undisturbed and corrugated coaxial cable. . . . . . . . . . . 135
6.5. Comparison of the reflection coefficient E computed with RPIM and
COMSOL. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
6.6. Amplitude variation and phase difference of the corrugated test model for
the corrugations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
6.7. Convergence for increased node densities with uniform refinement. Com-
parison between RPIM, first-order h-FEM and high-order discontinuous
Galerkin finite element-method. . . . . . . . . . . . . . . . . . . . . . . . . 139
6.8. Computation times required to achieve a certain accuracy for the three
different methods. A comparison is shown between the RPIM scheme with
uniform refinement, a high-order discontinuous Galerkin finite-element
method with curvilinear elements and polynomial refinement and a first-
order continuous finite-element method with uniform mesh refinement.
The region with exponential convergence for the RPIM algorithm is
marked with a solid line. The computation time of the fifth eigenmode
for the discontinuous Galerkin method with a full matrix solver (marked
with a cross) is provided for a direct comparison. . . . . . . . . . . . . . . 140
xv
LIST OF FIGURES
6.9. Node distributions at different stages of the refinement iteration for the
square domain with T M11 mode and L-shaped domain for the 1st smooth
eigenmode. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
6.10. Convergence of several eigenvalues of a square and L-shaped domain. . . . 142
xvi
List of Tables
3.1. Boundary conditions for the electric and magnetic fields using perfect
electric and magnetic conductors. . . . . . . . . . . . . . . . . . . . . . . . 46
6.1. umber of degrees of freedom required to achieve an error below 105 for
the first and fifth eigenvalue. . . . . . . . . . . . . . . . . . . . . . . . . . 139
B.1. Benchmark results for numerical eigenvalues of the L-shaped domain. . . 153
xvii
List of Acronyms and Abbreviations
Numerical methods
IE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Integral Equation
BEM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Boundary Integration Method
MoM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Method of Moments
LOOCV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Leave-one-out-cross-validation
SPH . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Smooth Particle Hydrodynamics
FVTD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Finite-Volume Time-Domain
xix
LIST OF TABLES
Numerical terminology
CEM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . computational electromagnetics
BC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . boundary condition
Electromagnetics terminology
EM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . electromagnetic
TE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . transverse electric
TM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . transverse magnetic
xx
LIST OF TABLES
xxi
1. Introduction
Fast and accurate simulation tools are crucial for the progress in the research involv-
ing electromagnetic fields and in engineering electromagnetics. The design of antennas,
resonators, filters, waveguides and other microwave devices in a competitive industrial
environment is only possible when the time-to-market can be reduced. Before the avail-
ability of numerical tools, numerous prototype-measurement cycles were necessary to
obtain a design that fulfilled its specification. With the possibility to numerically predict
the behavior and optimize the design of such a structure, it was possible to embrace new
concepts and new technologies with drastically improved performance. Many modern
devices could not have been established without versatile simulation and optimization
techniques. One such example was development of the compact planar inverted F anten-
nas (PIFA) with multiband operation as commonly integrated in cellular phones today.
The extraction of parameters like the radar cross section (RCS) of an object is heavily
simplified with numerical tools in comparison to measurement setups. In the design of
an airplane for example the RCS is is an important far-field parameter that is sought
to be characterized for radar detection or minimized in military stealth applications.
Furthermore in the field of electromagnetic compatibility, the numerical prediction of
high surface currents caused by lightning strikes is a typical example of the success of
numerical methods.
Electromagnetics in nanoscale is one of the new frontiers in electromagnetic research.
In nano-physics, very small structures in the nanometer scale are investigated. The size
of these objects becomes comparable with or even smaller than the wavelength at opti-
cal frequencies, which make full-wave simulations a viable option for accurate modeling
of their interactions with electromagnetic fields. Novel process technologies enables to
realize structures at this scale for a wide range of applications. Sophisticated tools are
required to predict the material interaction with electromagnetic fields. The new chal-
lenges include the incorporation of advanced material models and quantum effects into
the electromagnetic modeling process.
Analytical approaches fail to give answers when non-canonical structures are investi-
gated. Therefore, with the ever increasing available computer power numerical tools are
gaining in importance.
1
1. Introduction
limited resources of a computer, the spatial domain of the structures has to be described
through a number of nodes or mesh cells. The time variable is either solved explicitly
in time domain, or it is assumed that the behavior is time-harmonic and the medium is
linear.
In the field of computational electromagnetics a distinction is made between two
classes of methods, boundary and domain discretization methods as illustrated in Fig. 1.1
and 1.2 respectively. Boundary discretization methods solve the wave equation with the
help of Greens functions, which usually describe the propagation of electromagnetic
waves in linear homogeneous materials. These types of methods belong to the class of
integral equation (IE) methods [1]. In a special variant, the boundary element method
(BEM), the boundary of an object is discretized through mesh elements on the object
surfaces. The most common implementation is called the method of moment (MoM) [2].
The memory requirement associated with the surface discretization is particularly
advantageous when the physical object is much smaller than the surrounding volume.
Electrically very large metallic structures in free space can be solved very efficiently since
only the structure surface is modeled. In general the method is well-suited to simulate
open problems, for example to predict the behavior of an antenna placed on a large
metallic object, or the scattering properties of highly conductive structures. The main
difficulty associated with boundary methods lies in the integration of possibly singular
functions and in the solution of equation systems with full matrices and sometimes high
condition numbers. Also for more complex structures with inhomogeneous lossy dielectric
materials many of the advantages of the method disappear and large numerical computa-
tions become necessary. The efficiency could be greatly improved with the development
of fast multipole methods (FMM) [3], where the computation of matrix-vector products
is accelerated through iterative integral methods.
In the class of domain discretization problems (Fig. 1.2), the complete domain is
discretized into elements or nodes. On all boundaries, suitable boundary conditions must
be applied. A solution can be either obtained in time or frequency domain. A time-domain
solver takes the original Maxwells equations with time derivatives into account, and
yields the solution in a time-marching iteration with given steps. The most prominent
form of these methods is the finite-difference time-domain (FDTD) method [4]. In FDTD,
the whole domain is split into a regular grid (Fig. 1.2b) and material parameters are
assigned to each grid cell. The method is based on the differential form of Maxwells
equations and uses linear basis functions. Staggering the locations of the magnetic and
electric field components in the Yee scheme [5] allows an elegant exploitation of the the
second-order accurate central difference scheme in time and space. Due to its structure
the method is fast and has been parallelized from early on for multi-core processors
and graphical processing units [6]. The accuracy on the other side is rather low and
difficulties arise when complex geometries have to be modeled, due to stair-casing effects
associated with the Cartesian discretization. Furthermore in the vicinity of corners, the
FDTD method requires very small grid sizes to accurately resolve rapid field variations of
singular solutions. Subgridding algorithms [7] partially solve this problem. A successful
extension of the FDTD method was the development of the finite integration technique
(FIT) [8]. Instead of the differential form of Maxwells equations, the integral form is
solved. On a rectangular grid both FDTD and FIT results in the same update equations.
2
1.1. Mesh-based Numerical Methods in Electromagnetics
Figure 1.1.: Illustration of the boundary discretization on a cylindric disc. The mesh
discretizes the surface of the physical object.
(a) (b)
Figure 1.2.: Example of two different types of domain discretizeation, through (a) a
rectangular grid and (b) an unstructured mesh.
However FIT shows a higher flexibility in the handling of boundaries and geometric
modeling when complex shaped objects are discretized.
Solvers in the time domain [9] have many beneficial properties, e.g. the ability to sim-
ulate broad frequency ranges in one simulation run or to model non-linear materials.
In non-linear structures, materials change their properties in dependence of the field
strength, which precludes the use of solvers that assume stationary fields, i.e. in the fre-
quency domain. The ability of time-domain methods to solve the behavior of transient
signals enables the analysis of new classes of problems. Moreover, for the design and anal-
ysis of complex structures, the possibility to observe the evolution of the electromagnetic
fields over time helps understanding and improving a setup.
3
1. Introduction
For both the FEM and FDTD method, numerous commercial packages exist that
conveniently guide users through the whole design process. These two methods have
become the de-facto standard for most engineering applications.
Among the other methods in this class of domain discretization methods, the finite-
volume time-domain (FVTD) method [9] can be mentioned. It updates the fields through
a time iteration in mesh cells by an approximation of electric and magnetic fluxes through
the mesh element surfaces. This combines the advantages of a time-domain approach with
the discretization in unstructured meshes. On the negative side, the accuracy is rather
low and the method suffers from dissipative effects.
The DG-FEM scheme in time domain has lately become very popular. In many ways,
this method can then be seen as a generalization of the FVTD method which allows
the use of high-order elements. It has been implemented and used by many research
groups around the world recently. Nevertheless, these methods rely on an explicit mesh
that has to be generated in the course of a preprocessing step. This process is often
cumbersome and the quality of a mesh influences the accuracy and convergence properties
of a numerical simulation.
The applications mentioned in the beginning of this chapter require fast and versatile
numerical solvers. All the previously mentioned methods show drawbacks, either in low
accuracy, in the complex data structures or in geometrical challenges involved in dealing
with a mesh topology. In this thesis, a node-based method is proposed that combines
the advantages of simplicity through node discretization with the ability of conformal
node placement around complex geometric object. The discretization can resolve fine
structures, and allows coarse node distributions in places where no fine discretization is
required. With the right choice of basis functions, very high accuracy can be achieved.
These so-called meshless methods are a promising class of methods in computational
electromagnetics.
4
1.2. Meshless Methods
5
1. Introduction
the interpolation accuracy can be improved by increasingly flat (i.e. extended) radial
basis functions. This theoretically leads to exponential convergence through adaptation
of the shape of the basis functions only, at no additional computational cost. There
is however a practical limitation to this convergence, reached at the point where the
problem becomes ill-conditioned and the scheme breaks down.
Apart from increased accuracy through a clever selection of the shape parameter,
increased spatial discretization is also expected to lead to a solution converging at a
much higher rate than low-order finite-element methods or the FDTD scheme [39]. An
estimated convergence rate [40] is in Fig. 1.4 compared to that of FDTD and low- and
high-order FEM with mesh refinement.
In a meshless node distribution unlike in a classical mesh structure, no geometrical
connectivity between collocation nodes exist. In meshless methods the basis functions
take on this role through a support domain of the functions that defines the area of
influence of each node. In a related approach, it has been shown that using global ba-
sis functions, i.e. where the field values at all node locations depend on each other,
renders superb interpolation quality. Unfortunately, this yields full matrices and the ef-
ficiency of the methods decreases dramatically with increased problem size. Two main
directions emerged to tackle this issue. First, the support domain of each node can be
restricted to a small number, usually in the range of 10 to 20 neighbors [41]. This yields
sparse matrices that can be solved very efficiently through iterative numerical solvers
or Marching-On-In-Time methods. To this end, new types of radial basis functions were
developed that intrinsically fulfill this property through a local compact support [28]. A
second approach is to retain the interpolation quality of the global basis functions, but to
divide a computational domain into several smaller subdomains, each one small enough
to yield small individual systems of linear equations that are solved quickly. A domain
decomposition method [42] is used at interfaces between sub-systems to iteratively find
an accurate global solution.
6
1.4. Overview over the Thesis
Figure 1.4.: Exemplified illustration of the highest theoretical convergence rates of rela-
tive errors when increasing the fineness of the spatial discretization. Shown
are low-order methods like FEM or FDTD, high-order FEM with mesh re-
finement and RPIM using global RBFs.
7
1. Introduction
spectrum and geometrical features. The treatment of several boundary condition types is
introduced. The method shows many similarities with the FDTD method and in fact for
specific parameters the two methods are identical. In order to investigate the stability of
the time-domain scheme, an eigenvalue solver is introduced and implemented. Numerical
experiments are performed to find a relationship between long-time stability, the choice of
shape parameter and support domain size. The second-order from of Maxwells equations
is implemented in frequency domain. Global basis functions have excellent interpolation
quality, even for very coarse node densities. Based on the Helmholtz equation, a scalar
formulation of an RPIM framework for eigenvalue solvers in cavities is developed in a
more theoretical context. The scheme is compared to the non-symmetric and symmetric
Kansa method in two numerical experiments. Next, a solver for classical source problems
for RPIM is implemented. Based on this framework, domain decomposition methods for
solving electrically large problems are discussed and a method is proposed to handle
material discontinuities.
Absorbing boundary conditions to simulate open problems are introduced in the fourth
chapter. For the localized first-order RPIM framework in time domain, a uniaxial PML
model is introduced and validated through numerical experiments. Low reflections from
the PML down to 80 dB are demonstrated. For the second-order RPIM framework using
global basis functions, high-order absorbing boundary conditions are implemented. Due
to the high accuracy of this scheme, highly efficient high-order non-reflecting boundaries
can be developed at very little additional cost. A global numerical optimization using
genetic algorithms is presented that finds an ideal set of coefficients for minimal reflec-
tions from the boundary. To further increase the accuracy of the boundary conditions,
8
1.4. Overview over the Thesis
9
2. Scattered Data Interpolation with
Radial Basis Functions
Abstract Interpolation of data point values with a suitable reproduction of the original function
is important in many fields of science. Radial basis functions are well-suited for highly accurate
interpolation of scattered data sets and therefore have been applied in many research areas. This
chapter considers a special form of the approximation denoted as radial point interpolation method
(RPIM). After a summary of the theory of globally supported functions, a localized approach
is introduced. It is found that the accuracy of the interpolation always comes at the cost of a
poorly-conditioned matrix. Numerical estimations on the accuracy of the first- and second-order
spatial derivatives of harmonic sinusoidal functions are given, and guidelines on choice of the
support domain size to achieve a predefined accuracy are developed. Since the method yields best
accuracy close to a numerical breakdown, an algorithm is presented that optimizes the interpolation
parameter for highest interpolation accuracy.
2.1. Introduction
Whenever intermediate data values have to be estimated between given data points a
suitable interpolation is required. In many practically relevant cases, the available data
points are not evenly distributed. Algorithms aiming at good reproduction of inter-
mediate data have been a field of research for a long time. Recently it emerged that
interpolation by radial basis functions (RBFs) has many advantages over existing meth-
ods [28, 27]. The ability to reproduce sufficiently smooth original functions with a very
high accuracy makes them a viable choice in many fields of science. Due to the radial
formulation, higher-dimensional formulations are straightforward and have been imple-
mented successfully. This is in contrast to many other methods, where formulations in
three dimensional space are cumbersome.
Common applications of RBFs can be found in the field computer vision, where a three
dimensional representation of a scanned object is often given by sample point clouds.
Here RBFs can produce a smooth approximation of the original surface [43]. In the field
of neural networks, high-dimensional data sets often need to be interpolated. There too
RBFs led to the highly successful implementation of radial basis function networks [44].
For the application of RBFs in numerical simulations of partial differential equations,
the often smooth basis functions struggle or are even unable to reproduce constant,
singular or non-smooth solutions. Recently combinations of RBFs with other types of
basis functions have been presented, which are summarized in the following. The most
direct approach to increase the accuracy for many applications is the extension with
11
2. Scattered Data Interpolation with Radial Basis Functions
monomial functions [36]. This method will be introduced in the subsequent section.
In computational mechanics a method based on smooth and singular basis functions
has been presented [45]. And in order to solve material discontinuities, an extension by
step functions has been successfully implemented for elasticity problems [46]. In another
numerical investigation, the behavior of the interpolation by RBFs near boundaries has
been studied in [47]. Several techniques to increase the accuracy of the interpolation near
boundaries have been suggested. Some of these techniques will be applied later in this
thesis.
The approach considered here is the radial point interpolation method (RPIM) [41].
Based on the RBFs, a point-matching procedure is applied that yields a new set of shape
functions. These shape functions are later used in an implementation to solve physical
problems, in contrast to other radial basis function methods where the RBFs are used
directly. The shape functions have properties that make them well-suited for an efficient
implementation in time and frequency domain. The implementations considered in this
thesis are all two-dimensional cases, thus the formulations will only take the Cartesian
x, y-components as spatial variables.
with the radial basis functions rn centered at the collocation nodes xn . Many types
of RBFs are known in literature (Tab. 2.1, from [36] and Fig. 2.1). In general, positive
definite basis functions are desired [30], the details will be discussed later in this chapter.
The last two basis functions listed in Tab. 2.1 are examples of Wendlands compactly
supported functions [28], i.e. the values take zero value outside of a domain of size d0 .
Many different forms of Wendlands basis functions exist. The functions are defined in
a recursive formulation to achieve a required smoothness. For all cases investigated in
this thesis, Gaussian basis functions proved to provide best accuracy while being last
prone to problems with singularities. Therefore, they used from now on.
The Gaussian basis functions
|xn x|2
rn (x) = exp c (2.2)
d2c
contain a shape parameter c that controls the flatness of the function. The normalization
factor dc can be defined using the physical size A of the computational domain as
A
dc = . (2.3)
N 1
12
2.2. Global Basis Functions
|xn x|
Basis Function Expression rn (x) with = dc Parameter
Gaussian exp c 2 c
q
Multiquadric 2 + C 2 C, q
Thin Plate Spline 2
5
Quintic d0 d0
( 2
6
(1 d0 ) (3 + 18 d0 + 35 d2 ), d0
Wendland C 4 0 d0
0, otherwise
( 2 3
8
6
(1 d0 ) (1 + 8 d0 + 25 d2 + 32 d3 ), d0
Wendland C 0 0 d0
0, otherwise
Table 2.1.: Alternative radial basis functions. The lower two Wendland functions have a
compact support and they are positive definite up to the 4th or 6th deriva-
tives.
It denotes the average distance of all nodes inside the computational domain.
The interpolation by radial basis functions (2.1) can be expanded in a so-called mono-
mial reproduction to provide accurate approximation of polynomial functions by using
monomial basis functions pm (x)
N
X M
X
hu(x)i = an rn (x) + bm pm (x) = r(x)T a + pT (x)b, (2.4)
n=1 m=1
with M denoting the number of terms in the function pT (x). This interpolation (2.4)
can also be seen as an extension of the point interpolation technique (PIM) [36], where
only the monomials pT (x) are applied as basis functions. This method often suffers from
very high condition numbers, especially for higher orders M . Also for low orders, singular
matrices can be encountered for some specific node positions. Sophisticated techniques
have to be applied to guarantee an invertible matrix. In this respect, the addition of
RBFs to PIM solves the often encountered problem of singular matrices arising from
certain sample node positions.
The interpolation parameters, a and b from (2.4), are calculated in a preprocessing
step as follows. A system is set up to interpolate the field values in the nodes Us =
13
2. Scattered Data Interpolation with Radial Basis Functions
5
Gaussian
4.5
Multiquadric Gaussian
4 Thin Plate Splines 1
Quintic
3.5
Wendland 0.8
3
0.6
2.5
0.4
2
0.2
1.5
1 0
0 1 2
0.5
0
0 0.5 1 1.5 2
Figure 2.1.: Illustration of the alternative radial basis functions in Tab. 2.1. For the
parameters typical values are chosen as follows: (Gaussian) c = 0.5, (Mul-
tiquadric) C = 1.05, q = 0.5, (TPS) = 2.5 and (Quintic, Wendland)
d0 = 1.5. The inset depicts the influence of the shape parameter c on the
flatness of the Gaussian basis function.
(u(x1 ), u(x2 ), . . . , u(xN ))T . This is done in a point-matching procedure, where all basis
functions (2.2) are evaluated at the node locations. For the radial basis functions the
following matrix is defined:
r1 (x1 ) r2 (x1 ) . . . rN (x1 )
r (x ) r (x ) . . . rN (x2 )
1 2 2 2
R0 =
.
. (2.6)
.. .. .. ..
. . .
r1 (xN ) r2 (xN ) . . . rN (xN )
N N
zT R0 z > 0, z RN (2.7)
and thus invertible when the basis functions rn are positive definite [30]. All radial basis
functions listed in Tab. 2.1 fulfill this constraint. For the monomial basis functions, an
14
2.2. Global Basis Functions
N M matrix results:
T
p (x1 ) 1 x1 y1 . . .
pT (x ) 1 x2 y2 . . .
2
Pm = = . (2.8)
.. ...
.. ..
. . . . . .
T
p (xN ) 1 xN yN ...
N M
Those two matrices (2.6) and (2.8) can be collected into a single system to interpolate
the field values Us
Us R0 Pm a a
= = G (2.9)
0 PTm 0 b b
which includes the constraint condition PTm a = 0. This condition leads to a square
N + M N + M matrix G, which is often called moment matrix [36]. The interpolation
parameters can be calculated as
a Us
= G1 (2.10)
b 0
a T Us
hu(x)i = [rT (x), pT (x)] = r (x), pT (x) G1 = (x)Us .
(2.11)
b 0
This procedure of obtaining the shape functions through point-matching is called the
radial point interpolation method (RPIM).
Based on this approximation of the field component functions, the approximation of
their spatial derivatives along the Cartesian coordinate axes = x, y can be expressed
15
2. Scattered Data Interpolation with Radial Basis Functions
as
Us
h u(x)i = [ rT (x), pT (x)]G1
= (x)Us
(2.12a)
0
Us
2 u(x) = [2 rT (x), 2 pT (x)]G1 2
= (x)Us (2.12b)
0
..
.
Us
hp u(x)i = [p rT (x), p pT (x)]G1 p
= (x)Us .
(2.12c)
0
max (G)
(G) = (2.13)
min (G)
the ratio between the maximum and minimum singular value of the moment matrix.
Common numerical matrix solvers using double precision usually suffer from round-off
errors when this condition number is larger than 1020 . Additional to the shape parameter,
practical factors can also lead to singular moment matrices:
Two or more nodes nearly coinciding, i.e. lying almost on the same location, lead
to nearly identical basis function values and thus to nearly identical rows (up to a
permutation of the entries) in the matrix.
These factors lead to a non-invertible moment matrix G regardless of the shape param-
eter c .
16
2.2. Global Basis Functions
As mentioned previously, it has been shown that low values for c , i.e. flat basis
functions, give better interpolation accuracy. Considering the form of the matrix G
in (2.9), it is understandable that the condition number of G becomes higher when the
basis functions become flatter, as all elements in R0 become close to unity. Fig. 2.2 shows
this behavior of the moment matrix structures for three different shape parameters.
For a small shape parameter (Fig. 2.2a) all entries in the matrix R0 in the moment
matrix are visually equal to one. For increasing values of the shape parameter, the basis
function becomes less flat and more and more entries have values less than one. For very
large values of the shape parameter in Fig. 2.2c, the matrix containing the radial basis
functions starts to resemble an identity matrix. More fundamental analysis showed that,
theoretically, when the shape parameter goes to zero c 0, the radial basis functions
span the same space as spherical harmonics. In an implementation for solving partial
differential equations, parallels could be drawn with the Fourier pseudo-spectral method
which has very good properties even when the solution is non-smooth [50, 51].
According to [36], the addition of monomial basis functions (M > 0) has several
positive effects. First, it ensures that the interpolation can pass the patch test, i.e.
the shape functions can exactly reconstruct a linear function. Secondly, it provides an
improved accuracy and makes the method less sensitive to the shape parameter c . The
following properties of the shape functions have been identified [36]:
1. Delta property:
(
1, i = j
i (xj ) = , (2.14)
0, i 6= j
Global basis functions take the whole domain into account. This leads to full matrices.
In principle, the computational effort can be reduced by the use of compactly supported
functions such as Wendlands functions [27], which have been applied in some RBF
implementations [52]. Here, due to the matrix inversion associated with (2.11), a full
moment matrix is the result nevertheless when applying the shape functions . This
detail will be further discussed in Chap. 3 where a solver for partial differential equations
based on RPIM is implemented.
A further decrease in computational cost is achieved in a localized implementation
with local RBFs where only field values in the vicinity of each node inside a local support
domain are considered, leading to an efficient algorithm using sparse matrices. This will
be introduced in the following section.
17
2. Scattered Data Interpolation with Radial Basis Functions
Figure 2.2.: Exemplary structure of moment matrix G in (2.9) for different shape pa-
rameters (a) c = 0.001, (b) c = 0.1 and (c) c = 10.
A nearest neighbor search can be efficiently implemented through a kd-tree [53]. This tree
organizes points in a k-dimensional space into a space-partitioning data structure and
allows a highly optimized nearest neighbor search. The support domain is of radius ds ,
and the local average node distance inside the support domain can now be approximated
with knowledge of the number of neighbor nodes nAs in the support domain as
As
dlc = . (2.16)
nAs 1
The term As = d2s is the physical area of the support domain with radius ds . In
the following the local support domain size ds = s dc is defined as a multiple s of the
average node distance inside the support domain. The expressions of R0 and Pm in (2.6)
and (2.8) are only evaluated on the nodes inside the local support domain instead of all
node locations. Fig. 2.3 shows this decomposition. The shapes of the truncated Gaussian
basis functions are depicted in Fig. 2.4.
The procedure can be summarized as follows:
18
2.3. Local Basis Functions
Figure 2.3.: Illustration of local support domain with radius ds . Average node distance
dc is used for normalization of radial basis functions.
Figure 2.4.: Influence of the shape parameter c and support domain size ds on Gaussian
radial basis functions. The solid lines shows a truncated Gaussian radial basis
function, defined locally in a support domain with radius ds .
Additional to the properties of the global basis functions, the following can be identified
for local basis functions:
19
2. Scattered Data Interpolation with Radial Basis Functions
1. Partition of unity:
N
X
n = 1 (2.17)
n=1
nAs 4 8 16 24 40
Table 2.2.: Relationship between number of neighbors and support domain size.
20
2.4. Numerical Evaluations
Figure 2.5.: Randomly disturbed node arrangements for different extents of the support
domain. Number of neighbors are (a) nAs = 4, (b) nAs = 8, (c) nAs = 16,
(d) nAs = 24 and (e) nAs = 40. Nodes are allowed to move up to 60% of
the average node distance based on an original regular distribution, allowing
overlapping of nodes.
For 30 different random node distributions, a plane wave is interpolated for various
angles of incidence and phases ranging within [0, 2]. An illustrative example is displayed
in Fig. 2.6. For all the variations, the maximum interpolation errors of the field value
and its first- and second order derivatives are recorded. The plane wave is a suitable
example to estimate the accuracy of the RBF interpolation for smooth and harmonic
solutions of an EM problem.
21
2. Scattered Data Interpolation with Radial Basis Functions
(a) (b)
Figure 2.6.: Illustration of harmonic plane wave test function. In (a) a two-dimensional
representation of a plane wave u(x, y) is shown with an incidence angle of
30 . In (b) also the first- and second-order derivatives along the x-direction
are illustrated for this example.
22
2.4. Numerical Evaluations
Figure 2.7.: Influence of the shape parameter c on the interpolation accuracy of a plane
wave. Average node distance is /10. Different monomial orders have been
considered with (a) M = 0, (b) M = 3 and (c) M = 6.
quadratic monomial basis functions (M = 6). Instead of a band as before, only the
maximum, worst cases are considered here. For each parameter set, the optimum shape
parameter is selected which gives the lowest maximum errors. For the first-order deriva-
tives at coarser discretizations, all methods behave almost identically. When the wave-
length increases, the harmonic test function becomes progressively flatter. The purely
radial basis functions fail to interpolate the almost constant function values. An error
plateau with a minimum of 108 is reached. The situation is significantly improved
when increasing the monomial order. But also the linear case M = 3 (dashed) shows
some degradation for larger wavelengths. This can be observed as the error curve gets
flatter. Best accuracy at large wavelengths is achieved for the second-order monomial
case.
The same arguments are valid for the second-order derivatives in Fig. 2.10. The error
plateau of the interpolation without monomial bases is at a higher level of 106 . An
interesting observation is the accuracy of nAs = 4 with M = 6. The curve remains flat
in the case of the second-order derivatives in Fig. 2.10 at a very large level with an error
of 60%. This is due to ill-condition of the moment matrix because of an insufficient
number of of data points (nAs < M ).
In many algorithms such as the finite-difference time-domain (FDTD), finite-volume
23
2. Scattered Data Interpolation with Radial Basis Functions
(a) (b)
Figure 2.8.: Results of spatial convergence study with linear monomial basis functions
(M = 3). Each shaded band represents the range of error for a number of
evaluated distorted node distributions. Shape parameters are (a) c = 104
and (b) c = 102 . The bands on the right overlap over the whole range of
average node distances.
24
2.5. Ill-Condition Due to Flat Basis Functions
Figure 2.9.: Interpolation accuracy of first-order derivative of plane wave for increasing
node densities, i.e. decreasing average node densities in terms of wavelengths.
Different monomial orders are compared with (a) M = 0, (b) M = 3 and
(c) M = 6.
25
2. Scattered Data Interpolation with Radial Basis Functions
Figure 2.10.: Interpolation accuracy of second-order derivative of plane wave for increas-
ing node densities, i.e. decreasing average node densities in terms of wave-
lengths. Different monomial orders are compared with (a) M = 0, (b)
M = 3 and (c) M = 6.
invert the moment matrix and thus again to very low interpolation accuracy. In the
finite-element method for Helmholtz problems, convergence for sufficiently fine meshes
is defined through Ceas Lemma [54]. In essence it can be expressed as
This means that the numerical error of the method converges only when a suitable
approximation algorithm exists and the scheme is stable at the same time. For the finite-
element method the first part is achieved by the right choice of the function space of the
polynomial basis functions. The second condition is roughly fulfilled when the mesh is
fine enough. This observation is valid for most numerical methods. For the application
of radial basis functions, the first criterion of approximability has been established in
the previous section, at least for smooth solutions. But regarding the second criterion
of stability it has been observed that numerical break-downs can endanger reliable or
stable reproduction of functions. In that sense for the radial point interpolation method
it is important to reliably deal with the ill-conditioned matrices, or to find a shape
parameter c that gives stable results. Another approach to defuse the problem is using
26
2.6. Application of the Leave-One-Out-Cross-Validation Algorithm
oversampling strategies [40]. In that case a least-squares problem would be solved instead
of (2.11). Nevertheless, numerical breakdowns are likely to occur for naive choices of the
shape parameter.
The question on how to deal with the issue of having a stable interpolation scheme is
intensively researched in mathematics with three main approaches.
First, the development of an efficient and stable preconditioning algorithm [38, 55]
is underway. It could facilitate the choice of the shape parameter. Only a suffi-
ciently small value has to be selected, and the preconditioner renders the problem
invertible. Often the methods, such as the Contour-Pade algorithm [56] fail when
the domain are larger than a thousand nodes.
Second, very recently new developments on adapting the basis functions arose.
In [57], the Gaussian basis function is expanded through a Taylor series. By fac-
toring out of an exponential term, the remaining Taylor terms are represented by
monomials (unrelated to the previous monomial basis functions). These monomi-
als were replaced with Chebyshev polynomials. The new basis functions lead to
stable results even for very flat basis functions. An RBF-QR algorithm was devel-
oped in [51] for nodes located on a sphere. Here, the basis functions that cause
the ill-conditioning are replaced with basis functions leading to lower condition
numbers. In [58], the efficiency of a fast Fourier transformation algorithm has been
exploited by applying a conformal mapping from an arbitrarily shaped domain
onto concentric circles. These subjects are heavily researched in the field of applied
mathematics and the recent results are very promising.
27
2. Scattered Data Interpolation with Radial Basis Functions
The cost function CF t of the optimization is based on the diagonals of the pseudo-
inverse of the moment matrix:
PN t
t j=1 pinv(R0 )k,j f (xj )
k = (2.18)
pinv(R0 )k,k
CF t = max |tk | (2.19)
k
with a test function f t (x). The inverse of the matrix R1 0 is approximated using the
Moore-Penrose pseudo-inverse algorithm pinv [63]. A bounded optimization algorithm
fminbnd, [64] is used to find the minimum of the cost function CF t . This optimizer sub-
divides a given interval with a golden section search and uses parabolic interpolation of
the cost function to find a local minimum. The LOOCV algorithm works very efficiently
for small systems. When the number of nodes increases, the test function becomes less
smooth and more local minima appear. Then the search interval must be preselected
small enough to ensure finding the global minimum.
The choice of a good test function is not trivial, since it should reflect the characteristic
of the solution. In [30], a sinc function (Fig. 2.11) has been used
t x y
f (x = (x, y)) = sinc sinc . (2.20)
d0 d0
Evaluations showed that choosing the scaling parameter d0 in the order of the problem
size, reliable results could be achieved. Nevertheless, it is not guaranteed that a shape
function found with this test function reliably works for higher oscillatory solutions.
An example of the relation between the condition number and the cost error function
is shown in Fig. 2.12a. The dashed line signifies the minimum found by the optimization
algorithm. When using the test function f t the behavior of the cost function can vary,
depending on the choice of the normalization term d0 . This sensitivity leads to different
results of the optimized shape parameter c , which is problematic when the solution is
unknown.
An adaptation of the algorithm has been summarized in [65]. Instead of a test function,
the residual error of a differential operator is calculated. In that case, the algorithm
becomes independent of a test function and only relies on the problem type that is to
be simulated. The new cost function thus becomes
PN L
L j=1 pinv(R0 )k,j (R0 )k,j
k = (2.21)
pinv(R0 )k,k
CF L = max |L
k |. (2.22)
k
Here the term (RL 0 )k,j stands for the entries k, j of the matrix resulting from the dif-
ferential operator applied on the basis functions and evaluated on all node positions. In
the present case the Laplace operator L = was used and thus
(RL
0 )k,j = rj (xk ). (2.23)
28
2.7. Summary
0.5
0
1
1
0 0
-1 -1
y x
The relationship between the cost function CF L and condition number is illustrated in
Fig. 2.12b. The condition number at the found optimized shape parameter is already at
1020 , but the interpolation with the selected shape parameter still delivered good results.
In the numerical experiments performed throughout this work the residual-based
LOOCV algorithm consistently showed better performance, but was less robust than
using sinc test functions. This means that highly accurate results can be achieved with
small adjustments to the search interval of the optimization for the residual-based al-
gorithm. When the method is applied in iterative schemes where no user interference is
possible, e.g. in a node adaptivity scheme, the first LOOCV based on the test function
showed better results.
The LOOCV algorithm has been developed to use purely radial basis functions (2.1)
with no monomial extension. Thus in the following chapters, the algorithm is only applied
for formulations of global RBFs with M = 0. Even though it has been shown that an
order of the monomial basis functions of M = 6 shows a better convergence behavior, the
application of the LOOCV algorithm in combination with purely radial basis functions
is a viable choice for stability of RBF approximation.
2.7. Summary
This chapter summarized the development of shape functions based on the radial point
interpolation method. Global radial basis functions were introduced and several strate-
gies to increase the computational efficiency have been discussed. In general, very high
interpolation accuracy can be achieved using very flat basis functions. These cause poor
condition numbers of the full moment matrix.
A localized approach of this interpolation has been further introduced. In this case,
only nodes within a predefined support domain are taken into account. Instead of using
a large poorly-conditioned system, a large number of small matrices is calculated. Using
numerical experiments, the performance of the local shape functions have been evalu-
ated in the interpolation quality of plane waves. Guidelines for the minimum number of
29
2. Scattered Data Interpolation with Radial Basis Functions
(a)
(b)
Figure 2.12.: Relation between the condition number of moment matrix and the cost
error function to be minimized in LOOCV algorithm for different shape
parameters c . The theoretical continuation of the condition number is
illustrated as a dashed line. Shown are (a) test function based on sinc
function (2.19) and (b) test function based on residual error (2.22).
30
3. Application of the Radial Point
Interpolation Method to Maxwells
Equations
Abstract The Maxwell equations are partial differential equations that describe the behavior of
electric and magnetic fields. By applying the point-wise sampling and interpolation technique intro-
duced in the previous chapter, the spatial derivatives of the field quantities can be approximated.
Two different approaches to solve the Maxwell equations are discussed in this chapter. In the first
part, the Maxwell system in first-order form is considered. Using the localized approach of RPIM,
an algorithm similar to the finite-difference time-domain (FDTD) method is developed. However,
in contrast to the regular grid of the standard FDTD implementation, the presented meshless
approach allows for discretization of complex geometries using arbitrary node distributions. After a
discussion of the time-domain formulation, further investigations of the scheme are performed on
the eigenvalue spectrum. The longtime stability properties are discussed in terms of the support
domain size and shape parameter choice based on the eigenvalue spectrum. In the second part,
implementations for the second-order system, i.e. the wave equation, are introduced in the frequency
domain. From a practical point of view, this formulation is comparable to the finite-element method
in its accuracy. A framework for an eigenvalue solver and for source problems are discussed. A
comparison between the RPIM implementation and two other RBF methods is performed in terms
of accuracy and efficiency. The treatment of material discontinuities with a domain decomposition
is discussed. This method can also be applied to greatly improve the efficiency of electrically large
systems. The frameworks introduced here will be used as a starting point for the next two chapters.
31
3. Application of the Radial Point Interpolation Method to Maxwells Equations
solved. This leads to a very similar formulation to FDTD in a regular grid, but allows the
modeling of arbitrary structures using geometrical information in the integration proce-
dure. Other algorithms are the generalized finite-difference method (GFDM) [66] where
a low-order Taylor expansion is used for approximating the finite difference operators of
arbitrary grid points. A similar ansatz was taken in the multi resolution time-domain
(MRTD) method [67], where wavelets are applied to arbitrary refined regular grids. In
this last method stability of the time iteration is an issue.
The scheme introduced here is based on an unstructured sample point distribution.
This type of method is referred to as meshless or meshfree method, because the data
structure does not rely on any connectivity between the grid- or node-locations. Using
more sophisticated basis functions, such as the previously introduced radial point inter-
polation shape functions, a much higher accuracy can be achieved than for the usually
low-order polynomial basis functions. The flexibility of unstructured node placement
allows for the modeling of geometries with small, complex details. In the first imple-
mentation, the algorithm solves the Maxwell system in time domain. This brings the
advantage of feeding a model with a wideband pulse, allowing to solve the problem for
many frequencies in one simulation run making use of Fourier analysis. In this formu-
lation, non-linear materials can be introduced where the model parameters can change
depending on the amplitude of the fields at a given time.
The RPIM scheme is applied with a geometrical pattern that staggers nodes where
magnetic and electric field components are stored in separate locations. Using local basis
functions, a formulation very similar to the FDTD scheme can be achieved, however with
more geometrical flexibility and a higher accuracy. The staggered node arrangement for
RPIM in the first-order system has the following advantages:
The interpolation of first-order spatial derivatives when using localized basis func-
tions is more accurate than that of second-order derivatives. This is demonstrated
in a numerical experiment in Sec. 2.4.
Staggering of magnetic and electric field nodes brings the dual neighbors closer to
a given node, i.e. a higher local node density results. This leads to a higher local
interpolation accuracy compared to a co-located grid structure where both field
types are stored in the same location.
In the next parts of this section, different aspects of the introduced meshless time-domain
implementation of Maxwells equations are described. As starting point, the relevant par-
tial differential equations are recalled, and the general solution through the method of
lines [68] is introduced. Later, the discretization of the computational domain is summa-
rized, including the procedure for generating a dual node distribution and the application
of the radial basis functions. Then, the discretization in time is discussed. The update
equations for the RPIM time-domain method for first-order systems is developed in local
and in matrix form. A condition to find stable time steps is introduced. The treatment
32
3.1. First-Order Systems
The two Gauss laws (3.1c) and (3.1d) explain the behavior of the electric flux density in
the presence of charges , and the non-existence of magnetic monopoles. Amperes (3.1a)
and the Maxwell-Faraday equations (3.1b) explain the connection of a change in time and
a rotation in space between electric and magnetic fields through the curl operator. The
curl operator describes the infinitesimal rotation of a vector field with the nabla operator
. The electric current density J and the charge density satisfy the conservation of
charges
t + J = 0. (3.2)
The relationship between magnetic field and flux density as well as between the electric
field and flux density is traditionally expressed on macroscopic level where the following
linear material models are considered:
Z t
D(x, t) = E = (x, t s)E(x, t)ds (3.3a)
Zt
B(x, t) = H = (x, t s)H(x, t)ds (3.3b)
Z t
J(x, t) = E + Je = (x, t s)E(x, t)ds + Je (x, t). (3.3c)
33
3. Application of the Radial Point Interpolation Method to Maxwells Equations
= 0 r , = 0 r . (3.4)
D = 0 r E (3.5a)
B = 0 r H (3.5b)
J = E + Je . (3.5c)
The system (3.1) contains only spatial and temporal derivatives of at most order one
and therefore is called a first-order system.
The approach taken here is using the shape function (2.12) from Chap. 2 to approx-
imate the spatial derivatives of arbitrary node locations. The shape function is gained
from scattered interpolation of the fields based on radial basis functions. A similar ap-
proach for a meshless time-domain method for electromagnetics was developed in [70]
as the smoothed particle hydrodynamics method. The two-dimensional approach using
RPIM has been introduced by the author of this thesis in [71, 72] for a Cartesian transver-
sal electric (TE) mode and in [73] for a rotationally symmetric transverse magnetic (TM)
mode.
The first-order system (3.1) is implemented in a two-dimensional setup for two coordi-
nate systems. Firstly, Cartesian coordinates are used to implement problems with the
domain infinitely extended in z-direction (Fig. 3.1a). In a TE formulation the magnetic
field vector points in the x, y-plane, and the electric field vector points in z-direction.
Propagation of waves eventually occurs in the x, y-plane with the electric field perpen-
dicular to the propagation plane. Second, a formulation in cylindrical coordinates with
a symmetry in -direction is introduced. A TM-mode is implemented where the elec-
tric field components lie in the r, z-plane and the magnetic field component H points
perpendicularly to the computational domain. This last formulation is well-suited to
simulate objects that are rotationally symmetric. It is assumed in this formulation that
the magnetic field is constant along the -direction. This arrangement is illustrated in
Fig. 3.1b.
In the Cartesian coordinate systems x = (x, y, z)T and cylindrical coordinate systems
34
3.1. First-Order Systems
(a) (b)
Figure 3.1.: Implementation in (a) Cartesian and (b) cylindrical coordinates. The gray
plane signifies the computational domain. The field components are (a) Hx ,
Hy and Ez and (b) Er , H and Ez .
35
3. Application of the Radial Point Interpolation Method to Maxwells Equations
The source term Jz (x, t) can be imposed to model infinitely extended current line sources.
In the case of the cylindrical TM formulation, the computational domain lies on the rz-
pane and the field components are at positions x = (r, z). The Maxwells equations then
take the form
1
t H (x, t) = (r Ez (x, t) z Er (x, t)) (3.9a)
1
t Er (x, t) = z H (x, t) (3.9b)
11 1 1
t Ez (x, t) = r (rH (x, t)) = H (x, t) + r H (x, t) . (3.9c)
r r
These formulations are continuous in time and space. In order to solve the equations
with a computer, both domains have to be discretized. The subsequent derivation of the
first-order algorithm follows the general method of lines [68]. The idea is to reduce the
partial differential equation with spatial and temporal derivatives to a system of ordinary
differential equations (ODEs). This is achieved by keeping the time variable continuous
while the spatial derivatives are discretized analogously to time-independent problems.
The time domain is discretized using an appropriate time integration scheme. Commonly
used methods are the Euler, the Crank-Nicolson or the Runge-Kutta scheme. This brings
the formulation into explicit form. A solution is obtained through an iteration in time
which only requires linear operations at each time step with values from previous time
steps.
Here the space is discretized by approximating a physical geometry by node locations.
Analogously to the FDTD method, a staggered arrangement of nodes for the magnetic
field components (H-nodes) and the electric field components (E-nodes) is sought. The
creation of this staggered node arrangement is discussed in the following section.
36
3.1. First-Order Systems
Figure 3.3.: Different methods to calculate first-order derivatives. Shown are (a) forward
derivatives, (b) backward derivatives and (c) central differences. The dashed
line shows the analytical solution as a tangent and the solid line is the
approximated value.
the staggering of the nodes lifts the order of the method by simply shifting the grid
locations using (3.10c).
An adaptation of the Yee scheme to arbitrary node distributions can be achieved
through a Voronoi tessellation [74]. This type of tessellation builds a cell around each
node. All points within a cell are closer to this node than any other node. The dual
graph of a Voronoi tessellation is the Delaunay triangulation. Connecting the centers
of the Delaunay triangles generates the Voronoi cells. A Delaunay triangulation will be
introduced and applied later in this thesis.
The two sets of staggered nodes are generated as follows: first an arbitrarily distributed
set of nodes is generated to discretize the whole volume (or area in the present 2D
case) of a given geometry. Then the Voronoi tessellation is applied to this first set and
creates a cell around each node. In a two-dimensional setting these cells are described
through polygons. On the edge center of each polygon, a new node is placed. All these
edge-centered nodes form a new dual set of nodes. The implementation of the TE-
mode suggests that the first set of nodes contains the locations of the electric field
components. In this way, the E-nodes which contain only one field component can be set
on the boundaries. This allows for a simple implementation of boundary conditions. The
37
3. Application of the Radial Point Interpolation Method to Maxwells Equations
nodes are placed in a circular way to precisely reconstruct the geometrical details of
the bend. The resulting ratio is N NE = 1.9. It can clearly be seen that for the straight
H
section with regular distribution before and after the bend, the dual H-node distribution
exactly corresponds to a Yee scheme. In a last example (Fig. 3.6c), a cylindrical cavity
is discretized with concentrically placed E-nodes. In that case, fewer dual H-nodes are
the result with NNE = 1.7.
H
Using the localized RPIM scheme, only the nAs H-neighbors of each node i inside a
support domain of size s dc are taken into account, where dc represents the average
node distance. The elements of the vector kHi correspond to the k = [1, . . . , nAs ]
neighboring H-nodes within the support domain of E-node i. Vice versa, the spatial
derivatives of the electric fields are centered at H-node location j.
n
Ej = [ 1Ej , . . . , EAj s ]. (3.12)
Here, the vector elements correspond to the neighboring E-nodes. The calculation of the
shape functions in the preprocessing step is very efficient, since only small matrices of
the size nAs nAs are inverted. Due to the normalization term, the shape parameter c
in the basis functions can be set as a global parameter, whose influence will be discussed
later.
38
3.1. First-Order Systems
Figure 3.6.: Examples of dual node distributions for (a) a waveguide with local increase
of the node density, (b) a waveguide bend and (c) a cylindrical cavity. E-
nodes are denoted with a dot and H-nodes with a cross .
The system (3.8) can now be written as a set of ordinary differential equations in a
39
3. Application of the Radial Point Interpolation Method to Maxwells Equations
The indices j sum over all the neighbors within the support domain of node i. An
overview of the nomenclature of the indices can be found in Appendix A. In the case
of the rotationally symmetric cylindrical T M formulation (3.9), the ordinary differential
equations are
nA nA s
1 Xs X
t H,i (t) = r jEi Ez,j (t) z jEi E,j (t) (3.14a)
j=1 j=1
nA
1 Xs
t Er,i (t) = z jHi H,j (t) (3.14b)
j=1
nA nAs
1 1 Xs j X
t Ez,i (t) = H,j (t) + r jHi H,j (t) . (3.14c)
i j=1 Hi j=1
All shape functions are applied on the field components separately, thus acting as scalar
basis functions. These sets of equations only depend on a continuous time variable. The
discretization in time will be discussed in the next section.
where n = [0, 1, 2, . . . ]. For time integration several schemes are known. In the present
implementation the temporal derivatives in (3.8) and (3.9) are discretized by central
40
3.1. First-Order Systems
differences in time:
H (n+1/2) H (n1/2)
t H (n) = (3.16a)
t
E (n+1) E (n)
t E (n+1/2) = . (3.16b)
t
Due to the staggering in time, accuracy of second order can be obtained without ad-
ditional operations or additional cost in memory. This method is called a second-order
leapfrog scheme as shown in Fig. 3.7. Using the relationship between the magnetic and
electric fields in (3.8) and (3.9), explicit update equations can be derived. For the Carte-
sian setup, these are
nA
(n+1/2) (n1/2) t Xs (n)
Hx,i = Hx,i y jEi Ez,j (3.17a)
j=1
nA
(n+1/2) (n1/2) t Xs (n)
Hy,i = Hy,i + x jEi Ez,j (3.17b)
j=1
nAs nAs
(n+1) (n) t X (n+1/2)
X (n+1/2) (n+1/2)
Ez,i = Ez,i + x jHi Hy,j y jHi Hx,j + Jz,i .
j=1 j=1
(3.17c)
The indices j sum over all the neighbors [1, . . . , nAs ] within the support domain of node
i. In the case of the rotationally symmetric cylindrical formulation, the update equations
become
nAs nAs
(n+1/2) (n1/2) t X (n)
X (n)
H,i = H,i + r jEi Ez,j z jEi E,j (3.18a)
j=1 j=1
nAs
(n+1) (n) t X (n+1/2)
Er,i = Er,i z jHi H,j (3.18b)
j=1
nAs nAs
(n+1) (n) t 1 X (n+1/2)
X (n+1/2)
Ez,i = Ez,i + j H,j + r jHi H,j . (3.18c)
i j=1 Hi j=1
Here, the shape functions are calculated in a cylindrical coordinate system. The spatial
derivatives r and z of the shape functions are then applied in a straightforward manner.
The variable i corresponds to the radial location of E-node i.
The application flow is illustrated in Fig. 3.8. The two staggered node distributions
are generated and the shape functions are calculated in a preprocessing step. The electric
and magnetic fields are updated iteratively. The update of the three field components
is implemented as a sum over all nodes inside each support domain. The end of the
simulation is reached when the energy has passed through the system.
41
3. Application of the Radial Point Interpolation Method to Maxwells Equations
Figure 3.7.: Leapfrog time discretization for the staggered first-order system.
end no
reached?
yes
post-processing
Figure 3.8.: Flowchart of the procedure to solve the first-order RPIM Maxwells equa-
tions in time-domain.
The implementation of the summation over the neighboring nodes can be achieved
through an appropriate data structure which contains pointers to neighboring field com-
ponent nodes. Nevertheless, it requires a large number of memory lookups in different
regions of the computer RAM due to the unstructured nature of the node distribu-
tion. In FDTD this is done very efficiently owing to the regular structure of the grid.
The data is processed sequentially and the caching algorithm in the computer hard-
ware heavily reduces the number of page lookups. In order to increase the speed of
the unstructured node distributions for RPIM, another approach is the use of sparse
42
3.1. First-Order Systems
matrices. The libraries of mathematical tools such as MATLAB are highly optimized
for such cases. Thus, a formulation for the implementation in Cartesian coordinates is
presented in matrix form for sparse matrices. The electric field is stored in a vector
Ez = [Ez,1 , . . . , Ez,NE ]T and the magnetic fields are stored in Hx = [Hx,1 , . . . , Hx,NH ]T
and Hy = [Hy,1 , . . . , Hy,NH ]T . The discretized differential operators correspond to en-
tries in real valued matrices x Le , y Le , x Lh , y Lh . For the derivatives in = x, y
direction, they take the following form
with the locations of the j th E-node xE (j) and H-node xH (j). The update equations in
matrix form are expressed as
t
H(n+1/2)
x = H(n1/2)
x y Lh E(n)
z (3.20a)
t
Hy(n+1/2) = H(n1/2)
y + x Lh E(n)
z (3.20b)
t h i
E(n+1)
z = E(n)
z + x Le H(n+1/2)
y y Le Hx(n+1/2) J(n+1/2)
z . (3.20c)
In this matrix form, the update of the fields is computed for ever time step by one sparse
matrix multiplication and one vector addition per field component.
This estimate is based on the distance dmini to the closest neighbor of node i and on
the material properties in the domain of i. The limit proves to be a good, although
43
3. Application of the Radial Point Interpolation Method to Maxwells Equations
conservative estimate for stable simulations. In [79], the authors numerically calculate
the time-step from an eigenvalue problem. The system (3.20) is brought into second-order
form and the maximum eigenvalue so max is calculated. The time step then becomes
2
t p (3.22)
so
max
This approach requires the computation of eigenvalues from sparse matrices. Numerical
evaluations showed that in general, the time step (3.22) is slightly larger than (3.21) [79].
Therefore, it is advantageous for larger problems to use the second approach involving
an eigenvalue computation. Then, the general computation time can be decreased.
As a stopping criterion three alternative methods are viable. When the time required
for a pulse to completely pass through a physical model can be approximated beforehand,
the number of time steps is calculated through
tend
Nsteps = . (3.23)
t
Alternatively, the energy in the computational domain is observed, and the simulation
ends when the total electromagnetic energy has decreased below a given limit. For simula-
tions in resonant structures, the Fourier spectrum is observed. The simulation is stopped
when the spectrum has stabilized.
Sources
The injection of energy into the computational domain is crucial for time-domain sim-
ulations. In this work, two concepts are applied as illustrated in Fig. 3.9. First, using
hard sources, field values are forced to specific values by a source function fs (x, t). Sec-
ond, currents are injected as soft sources, i.e. energy is accumulated to the computed
field components. Hard sources are used to inject energy from a boundary, for example
44
3.1. First-Order Systems
(a) (b)
Figure 3.9.: Modeling of sources to inject energy into time-domain simulation. On the
left is a hard source for electric fields placed on the boundary. On the right
is an infinitely extended line current as a soft source.
for waveguide problems. The source function for a mth order waveguide mode in the
geometry Fig. 3.9a takes the following form
m(y y0 )
fs (x = (x, y)) = sin
dW G
2 ! (3.24)
nt 4source
exp sin (2f0 (nt 4source )) .
2source
This function has a temporal and spatial dependence. The first sinus function imposes
the waveguide mode and spans along the y-axis. The waveguide starts at y = y0 and has a
width dW G . The temporal dependence generates a broadband pulse with the width factor
source . It is centered at a frequency f0 and the half-power bandwidth is restricted by the
minimum and maximum frequency fmin and fmax . The width factor can be calculated
as
1
source = . (3.25)
|fmax fmin |
The source function can be easily modified to adapt to a change of the source boundary
orientation. By using hard sources in these applications, no absorbing boundaries in the
source region are required. The source function depicted in Fig. 3.9a corresponds to a
second-order T E20 mode.
The soft sources are suitable to inject energy inside the computational domain. These
sources are placed inside the computational domain, at a distance to the boundaries. For
the 2D-TE case, these sources are modeled as infinitely extended current line sources
in z-direction. A broadband soft source then takes the following form with the same
parameters as above
2 !
nt 4source
Jz = exp sin (2f0 (nt 4source )) . (3.26)
2source
For the investigations conducted in this thesis in the cylindrical TM framework, no soft
sources were required and thus they are omitted here.
45
3. Application of the Radial Point Interpolation Method to Maxwells Equations
Perfect Conductors
Other types of boundary conditions that can enclose a domain or represent a material
boundary are perfect electric and magnetic conductors (PECs and PMCs). For PECs, it
is assumed that an unlimited number of free charges exist inside this artificial material.
Hence no electric field can exist, since it is automatically compensated by the free moving
electrons. No such material exists in real life, but the model is often sufficiently accurate
for good conductors up to microwave frequencies. The perfect magnetic conductor is the
analogon to the PEC, but here the magnetic field is zero. The conditions for the electric
and magnetic fields at the interface to these two perfect conductors are summarized in
Tab. 3.1. The subscripts n and t stand for the normal and tangential component of the
fields, respectively. For PECs, the condition for the electric field follows directly since
no electric field can exist. The condition for the normal magnetic field is a consequence
of the conservation of charges, and assuming that no free currents exist on the surface.
For PMCs the normal electric field En and transverse magnetic field Ht are zero. An
implementation of these boundary conditions depends on the placement of the nodes.
In the TE case, the electric field Ez points perpendicularly to the domain (x, y). The
boundary condition can then be naturally implemented by forcing the Ez components
to zero at the PEC boundary. Fig. 3.10a illustrates this concept. The same applies for
the TM case, where the perpendicular magnetic fields can simply be set to zero to fulfill
the PMC boundary conditions. For these cases, the update equations (3.17) and (3.18)
are extended by the following equations
It is sufficient to enforce the boundary conditions just by setting the transverse fields to
zero on the electric boundary nodes in the TE case and the magnetic boundary nodes for
the TM case. The conditions for the normal field components are an intrinsic property
of the Maxwells equations and are fulfilled implicitly.
In the opposite TE-PMC and TM-PEC cases, a concept using mirror nodes has been
implemented. The idea to increase the accuracy near boundaries by adding a node on
the outside of the computational domain has been introduced in [47] as Not-a-Knot. A
node is placed on the outside of the domain boundary. The necessary condition for PEC
and PMC can subsequently be enforced. A setup using these mirror nodes for a PEC in
PEC Et = 0 Hn = 0
PMC En = 0 Ht = 0
Table 3.1.: Boundary conditions for the electric and magnetic fields using perfect electric
and magnetic conductors.
46
3.1. First-Order Systems
(a) (b)
Figure 3.10.: Treatment of (a) PEC boundaries for the TE case by forcing the trans-
verse electric field to zero and (b) concept of mirror nodes to implement
conditions on normal field components for cylindrical TM-case.
the TM-case is depicted in Fig. 3.10b. The E-node i with the electric field Ei is mirrored
onto the outside of the computational domain at a distance d of the boundary. The
boundary can be arbitrarily oriented and its direction is defined by the normal vector n.
The condition Et = 0 can be implemented by adding
to the set of update equations [73]. This leads the average field in the middle between the
two nodes to have a zero tangential component [80]. For the TE case, an implementation
of PMC boundary conditions is done analogously.
Material Discontinuities
Inside a piece-wise homogeneous material, the material parameters and in (3.17)
and (3.18) are simply set to the appropriate value. This change affects the stability
criterion (3.21) and can lead to smaller time steps. Inhomogeneous materials can be
modeled with continuous variations of the material properties by the appropriate choice
of the material parameters at the collocation nodes.
On the interface between two materials, the fields must fulfill the transmission con-
ditions. They are formulated for the transversal and normal components of the electric
and magnetic fields as follows [81]:
for the magnetic fields with the free surface current t = t f . In the following both free
surface charges and currents are assumed to be zero.
47
3. Application of the Radial Point Interpolation Method to Maxwells Equations
At the boundary interface the first spatial derivative of the electric field is discon-
tinuous. This affects the approximation of the spatial derivative. As continuous basis
functions only accurately interpolate continuous fields, oscillations of the approximated
field can be observed in the vicinity of the discontinuity. This effect is called the Gibbs
phenomenon [47]. Literature suggests local adaptations of the basis functions, e.g. by
introducing jump functions [82] or by stretching of the radial basis functions [83]. In
general, for very low shape parameters c 0 the non-smooth fields are expected to be
interpolated accurately [51]. Unfortunately, numerical instabilities in the standard imple-
mentation render this approach unpractical. Instead the use of extended basis functions
from Sec. 2.5 might be advantageous.
Some preliminary results in the RPIM framework were presented in [84]. An adap-
tation of the basis functions led to an increased accuracy with a material-dependent
stretching parameter. The author of this thesis published another approach in [85]. A
numerical optimization was used to find a correction function applied to the update
equations (3.17).
For non-linear materials, complex higher-order material models exist. In the time do-
main, models such as Debye or Lorentz media can be integrated through the formulation
of convolution terms in the update equations [4].
48
3.1. First-Order Systems
1.5
0.75
-0.75
-1.5
Figure 3.11.: Stability regions from [78] for the (a) leapfrog scheme, the (b) staggered
third-order Adams-Bashforth algorithm and (c) the staggered fourth-order
Runge-Kutta scheme.
is brought into frequency domain by replacing the temporal derivatives t by their sta-
tionary form j. The system (3.8) can then be written in matrix form as
Ez Ez
jM Hx = L
Hx
(3.32)
Hy Hy
with the curl operators in the differential operator matrix L and the material parameters
in the diagonal matrix M.
For an eigensolver, the eigenvalues correspond to the resonance frequencies = j
and the eigenvectors represent the field distribution of each mode. The system (3.32)
can be discretized and written as
Ez Ez
h
M
Hx
= L Hx .
(3.33)
Hy Hy
The stiffness matrix Lh contains the spatial derivative matrices in x- and y- direction
from (3.19). Since the shape functions fulfill the delta property, the mass matrix M is a
49
3. Application of the Radial Point Interpolation Method to Maxwells Equations
For local RBFs these matrices are sparse with the number of entries per row equal to
the number of neighbors considered in the support domain. In experiments later on, a
comparison will be conducted with global RBFs. There the stiffness matrix is block-wise
full. A visual comparison on the impact of the two basis function types on the stiffness
matrix structure is shown in Fig. 3.12. The eigenvalues and eigenfunctions are calculated
using an eigenvalue solver for general eigenvalue problems. The ARPACK library [87]
is an efficient implementation of such a numerical solver for sparse matrices, and the
LAPACK solver [88] is suitable for full matrices.
In the following, the eigenvalue distribution is investigated in a cylindrical domain. A
study on the influence of the shape parameter is conducted for local and global basis func-
tions. First, the existence of spurious modes at low and high frequencies is investigated.
This quantitative analysis gives implications on the time-domain stability depending on
the interpolation parameters. The results are confirmed with time-domain simulations
in longtime simulations.
The eigenvalue distribution is computed using an eigenvalue solver for full matrices.
This solver computes a number of eigenvalues and eigenfunctions equal to the matrix
size, in contrast to sparse eigenvalue solver which generally finds eigenvalues in a given
interval through an iterative method. This second type can be more problematic in
terms of convergence, but is generally much more efficient when dealing with large sparse
matrices with few entries.
The eigenvalues are plotted in Fig. 3.14 and are symmetric with respect to the imag-
inary and real axis. It can be observed first that DC-modes, i.e. at = 0, exist. The
number of these zero-valued eigenvalues corresponds to the number of magnetic field
component nodes, i.e. 2NH . This null space is not surprising as the divergence condition
of Maxwells equations is not explicitly enforced. A modification of the basis functions
50
3.1. First-Order Systems
(a) (b)
Figure 3.12.: Structure of the stiffness matrix Lh using (a) local and (b) global basis
functions. Non-zero elements are indicated by dots.
Figure 3.13.: Model of the cylindrical cavity with staggered E- and H-node distributions.
In time-domain simulations, energy is injected into the system through a
soft source at an off-center location denoted in the figure.
can solve this problem. Additionally, the matrix structure shows a rank deficiency due
to the unequal number of E- and H-nodes, leading to a non-symmetric block matrix [89].
This is expected to disappear in a 3D-implementation. Furthermore, at low frequencies
no unphysical modes are observed, i.e. the eigenvalues are purely imaginary and each
corresponds to a physical mode. This is a very important observation. In other methods
such as the discontinuous Galerkin method, significant effort must be taken to suppress
non-physical modes at low frequencies [90].
At higher frequencies non-zero real parts begin to occur. Even if these unstable eigen-
modes are at very high frequencies, beyond the highest simulated frequency, eigenvalues
with non-zero real parts are known to cause instabilities in time-domain simulations.
51
3. Application of the Radial Point Interpolation Method to Maxwells Equations
(a) (b)
Figure 3.14.: Eigenvalue distribution for the parameter set c = 0.4, M = 3. On the left
are local RBFs with a support domain size of ds = 1.8dc , on the right are
global RBFs. The dark bar represents the bandwidth of the time-domain
simulation.
Due to numerical noise the modes get excited and eventually cause instabilities because
of their non-physical amplification.
Next, the magnitudes of these non-zero real parts are analyzed for multiple sets of
parameters. It is expected that the larger the magnitudes, the more amplification of the
non-physical modes is experienced. Here, the relationship between the parameters and
the magnitude of the real parts, i.e. how quick a time-simulation is expected to become
unstable, is established. The relative magnitude is calculated according to
Re(i )
R = max . (3.35)
i Im(i )
Fig. 3.15 shows these magnitudes for local and global RBFs and the matrix condition
numbers of the corresponding moment matrices G. With local basis functions and no
monomial basis functions (M = 0), the magnitude of the non-zero real parts decreases
for increasing shape parameter c , which indicates longer time stability. For M > 0
though, the magnitudes R remain quite high, with a minimum at approximately c =
1. For global RBFs, the magnitude of the non-zero real part goes to zero, when the
matrix condition number is significantly below the numerical limit of 1020 at c 0.1.
This means that time-domain simulations will always be longtime stable, provided the
condition number is sufficiently low. This also applies to all cases with a monomial basis
function. When the size of the support domain is increased for local RBFs, the condition
number of the moment matrix is driven up, but the non-zero real parts decrease in the
region where cond(G) < 1020 .
52
3.1. First-Order Systems
Figure 3.15.: Distribution of non-zero real parts and matrix condition number of the
respective interpolation moment matrix. Comparison between local RBF
with ds = 1.5dc (a,d), local RBF with ds = 1.8dc (b,e) and global RBFs
(c,f). The non-zero real part is shown in the top row and the condition
number on the bottom row.
53
3. Application of the Radial Point Interpolation Method to Maxwells Equations
Time-Domain Stability
In the following, time-domain simulations are performed for very long times to relate
the previous findings on spurious modes to long-term stability. These results have been
presented by the author of this thesis in [91]. The total energy in the system is observed
for time-domain simulations in the lossless resonator in Fig. 3.13. The energy in the
system has to remain constant after having been introduced into the system. When
non-physical modes with eigenvalues off the imaginary axis are excited, the field values
explode and the simulation becomes unstable. Observing the total energy allows longtime
stability and possible dissipation effects to be investigated. Only a rough estimation of
the energy is required, since the spurious modes increase exponentially in magnitude
over time.
Assuming a homogeneous node distribution with approximately equal areas A0 sur-
rounding each node, the normalized energy of the discretized problem is approximated
as
Z X NH X
E = 0 |E|2 + 0 |H|2 dV 0 A0 |Ez,i |2 + 0 A0 (|Hx,i |2 + |Hy,i |2 ). (3.36)
i
N E i
The ratio N NE of number of H- and E-Nodes estimates the size difference between E
H
and H cell areas. The approximation is justified for the present problem (Fig. 3.13),
as the Ez -field is zero at the boundary, the only place where the node distribution is
inhomogeneous.
The same node distribution as for the eigenvalue investigations is chosen. Energy is
injected into the system over one soft source node in the bandwidth f = [1 GHz, 2 GHz].
This frequency band is designated as a black bar in Fig. 3.14. The simulation has been
run for 100 000 periods at the center frequency. The number of time steps required was
Nsteps = 5350 490. In Fig. 3.16 the energy over time is shown for different parameters.
Local RBFs with a support domain size of ds = 1.8dc are compared against global RBFs.
Again, dc stands for the average node distance inside the support domain. It is clearly
observed that the higher the non-zero real part of the eigenvalue, the earlier instabilities
occur. In the case of local RBFs, for higher values of the shape parameter c 5, this
magnitude becomes very small and longtime stability is ensured. As expected, global
RBFs perform very well for c 0.1. For smaller values of c , non-physical modes
dominate the time-domain simulations from early on. It is also noted that no dissipative
effects have been observed for any set of parameters.
54
3.1. First-Order Systems
(a) (b)
The spectral properties of the scheme were further investigated in an eigenvalue anal-
ysis. No spurious modes were experienced at low frequencies, therefore physical solu-
tions are expected in time-domain simulations. A large null-space was observed. This
is explained partly by a rank-deficiency of the stiffness matrix. Another reason is that
the scheme does not explicitly enforce the divergence condition of Maxwells equations.
Therefore static solutions are allowed. This can be circumvented by using divergence-free
basis functions. Instead of calculating scalar shape functions for each field component,
matrix-valued basis functions are applied to interpolate the vector field directly [92]:
A condition is enforced that the divergence of the interpolated vector field has to be
zero. This was demonstrated for a Navier-Stokes problem in fluid dynamics in [93]. New
matrix-valued shape functions then implicitly enforce the divergence condition.
It was found that for some sets of parameters, the high frequency eigenvalues have non-
zero real parts. These eigenvalues correspond to non-physical instabilities that become
apparent after long simulation time (hundreds of periods). It was found that a shape
parameter with a sufficiently large value leads to stable simulations over a very long time.
This comes at the cost of accuracy. In the case of local shape functions, large support
domains increase the longtime stability of the scheme. Especially when using global
basis functions, the longtime stability issue completely disappears for sufficiently large
shape parameters. In this framework, the choice of shape parameters and order of the
monomials is a trade-off between accuracy and longtime stability. The shape parameters
55
3. Application of the Radial Point Interpolation Method to Maxwells Equations
are chosen beforehand as a global parameter. Due to the precarious dependence of the
stability on this parameter, it cannot be optimized through the LOOCV algorithm. An
optimization through the eigenvalue distribution nevertheless remains an option.
In the future, several approaches can be taken to improve longtime stability in com-
bination with high accuracy:
1. Higher-order schemes that include non-zero real parts of the eigenvalue distribution
in the stability region such as the Adams-Bashforth algorithm (Fig. 3.11b) can be
applied. The high-order schemes result, however, in higher computational costs in
time and memory. They require scaling of the time-step to fit all non-zero valued
eigenvalues into the stability region.
2. In a hybrid formulation between time-domain finite elements and the FDTD
method [94], the implementation of a digital lowpass filter into the time itera-
tion is proposed to suppress unstable modes. This method could be applied here
as well. Unfortunately this adds dispersion to the numerical solution.
3. Very recently, a method for RPIM has been proposed where the time domain is
discretized using Laguerre polynomials [95]. Due to the decreased weighting of the
functions at later time the scheme is unconditionally stable.
The combination of local basis functions with a second-order time integration scheme
is an efficient and versatile framework with the possibility of simple hybrid formulations
with the FDTD or other node-based method. In Chap. 6, several numerical examples
are presented for time-domain problems.
56
3.2. Second-Order Systems
jumps in between the elements, but these discontinuities are minimized using penalty
functions [12]. Both the standard finite-element method and the discontinuous Galerkin
method have sparse matrices, i.e. most of the entries in the system matrix are zero. This
is very advantageous for numerical solvers, since many highly efficient algorithms exist
to compute solutions iteratively.
The RPIM scheme presented here for frequency-domain applications is based on col-
location. Instead of computing the problem on elements, the solutions are sought on
collocation nodes only. In this approach, no mesh is required, which greatly decreases
the computational overhead of creating a mesh and handling the associated data struc-
ture. Using radial basis functions, it has been shown that surprisingly accurate results
can be found [31]. The RPIM method is a special case of radial basis function methods,
since shape functions are computed in a preprocessing step, leading to particular advan-
tages. In contrast to the localized approach, global basis functions are used here. This
yields full matrices that are much less efficient to solve than the sparse structures of local
basis functions, and the matrices in the FEM schemes. The higher computational costs
however are outranked by the superior accuracy for frequency-domain problems [39].
Additionally global basis functions allow the application of the LOOCV algorithm to
find a global optimized shape parameter.
In the following text, RPIM algorithms for two different problem types are introduced.
First, an eigenvalue solver yields all resonance frequencies as eigenvalues and the field
patterns associated with each frequency. Solvers of this type are important to analyze
structures such as waveguides cross sections, and their solutions are commonly used
for port definition in a three-dimensional setup. Also to analyze resonances in cavities,
eigenvalue solvers are employed in a simplified fashion by only solving one field com-
ponent. This second type is introduced for RPIM and compared to two other variants
of radial basis function methods. Second, an RPIM algorithm is introduced for solving
arbitrary problems with a known power source in frequency domain. This type is called
source problem. The setup here is currently limited to structures with symmetries that
allow the simplification to two dimensions. A next step would be the extension to three
dimensions. In order to increase the efficiency of the schemes, a domain decomposition
algorithm is laid out. Splitting the domain into smaller parts with homogeneous material
properties leads to small matrices and also allows for the accurate treatment of material
discontinuities.
57
3. Application of the Radial Point Interpolation Method to Maxwells Equations
Practically this is achieved through a multiplication of the phasor expression with ejt
and taking the real part. The electric field in time domain for example would be formed
as a function of the complex valued phasor E(x, ) as
Re(Ex (x, ) ejt )
jt
E(x, t) = Re(E(x, ) e ) = Re(Ey (x, ) ejt )
. (3.40)
Re(Ez (x, ) ejt )
The third and fourth equations in the Maxwells equations in time-domain (3.1) are
implied by the stationary form and can be recovered by taking the divergence of the
expressions (3.41). Applying the material models (3.5), the following frequency-domain
equations result:
with the optionally imposed current source Je . The two terms for the electric field can
be combined by using a complex permittivity
c
=+ (3.43)
j
and the stationary form of Maxwells equations eventually is
In the following, the stationary wave equations for the magnetic and electric fields are
derived [97]. Taking the curl of (3.44b) yields
( E) = j (H) (3.45)
| {z }
E+(E)
58
3.2. Second-Order Systems
Here, only homogeneous materials are considered. Thus, the permeability is constant
and thus (H) = H. It can be shown that inside such a material in stationary
form, no free charges can exist by taking the divergence of (3.44a). But for external
current sources Je , the conservation of charges dictates a source charge e such that
c
Je = j e . (3.48)
Therefore the divergence of the electric field contains this additional source term
e
E= . (3.49)
The expression (3.45) simplifies to
1
E ( Je ) = j H (3.50)
jc
The following expression is obtained by multiplying (3.44a) by j:
The equations (3.50) and (3.51) are brought together in the next step. To simplify the
notation, a wave number is defined through k 2 = 2 c = 2 j. Considering
forward traveling waves, i.e. waves moving away from its source, the wave number has a
positive real part and is expressed as
p
k = + 2 j. (3.52)
The two equations (3.50) and (3.51) combined yield the inhomogeneous Helmholtz wave
equation for the electric field
1
E k 2 E = j(Je + ( Je )). (3.53)
k2
Alternatively, calculating the curl of (3.44a) and inserting into (3.44b) results in the
wave equation for the magnetic field:
H k 2 H = Je . (3.54)
59
3. Application of the Radial Point Interpolation Method to Maxwells Equations
In the following the wave equation is implemented for a scalar two-dimensional prob-
lem in Cartesian coordinates. A computational domain on the x, y-plane corresponds
to the cross section of a structure that is infinitely extended in z-direction. This corre-
sponds to a TM-mode, since only electric fields exist in normal direction to the cross
section. The scalar expression can be formulated with a general source function f inside
a homogeneous linear material as
Ez k 2 Ez = f. (3.55)
When seeking solutions to for a TE problem, the wave equation for the scalar magnetic
field Hz is solved.
Boundary Conditions
The PEC and PMC boundary conditions from Tab. 3.1 are implemented here as well.
Since only one type of field is explicitly calculated in this formulation, both PEC and
PMC have to be enforced on the considered field type. The boundary of the computa-
tional domain is partitioned into the mixed boundary conditions
Ez = 0 on P EC . (3.57)
For the case of a PMC, the condition for the magnetic field Ht = 0 is considered. Using
the normal vector, the condition in z-direction can be expressed as
n H = nx Hy ny Hx = 0. (3.58)
z
1
Hx = y Ez , (3.59a)
j
1
Hy = x Ez . (3.59b)
j
60
3.2. Second-Order Systems
Figure 3.17.: Computational domain in the x, y-plane with unit transverse vector t
and unit normal vector n pointing outwards to the boundary .
Sources
Sources can be modeled similarly as in the first-order cases presented in Sec. 3.1.7. Cur-
rent sources Je = (Jx , Jy , Jz )T can be imposed inside the domain through the right-hand
side expression of the inhomogeneous wave equations (3.53) and (3.54). The simplified
source function f in (3.55) represents the effect of the imposed current Je on the z-
component of the electric field and can be calculated as
1
f = j(Jz + (x z Jx + y z Jy + z2 Jz )). (3.64)
k2
In usual applications the line current source Je has a constant amplitude along the
infinitely extended z-axis. Then the expression further reduces to
f = jJz . (3.65)
The source function f has to be sufficiently smooth and must be defined inside the
computational domain . This source type corresponds to soft sources.
For sources imposed on the boundary Source , e.g. a waveguide source in a port, addi-
tional boundary conditions apply. A source function g describes the stationary amplitude
of the electric or magnetic field component in an inhomogeneous Dirichlet condition
Ez = g(x) on Source . (3.66)
61
3. Application of the Radial Point Interpolation Method to Maxwells Equations
For a waveguide source, this source function takes a similar form to the source function
for the first-order system (3.24). A T Em mode in a waveguide with a width of dW G and
the boundary along the y-axis starting at y0 can be excited through
m(y y0 )
g(x = (x, y)) = sin . (3.67)
dW G
For the excitation of a plane wave, a constant function value along a straight boundary
Source is chosen
g(x) = 1. (3.68)
In the analogy to the first-order RPIM time-domain implementation, this boundary type
characterizes hard sources.
The term Bi (x) is a basis function with its ith term multiplied with a coefficient i . This
general expression is not limited to radial basis function, and in the present derivation,
it can take other forms. The expressions for the discretized Ezh become
( k 2 )Ezh = f in (3.70a)
Ezh =0 on . (3.70b)
In this discretized form, the variable Ezh can generally not solve the equations exactly.
A residual error is the result, and it can be formed as
R = ( k 2 )Ezh f in (3.71a)
R = Ezh on (3.71b)
If the approximation is exact, the residuals R and R are zero for all x . However,
a perfect approximation is generally not available and the residuals depend on the basis
62
3.2. Second-Order Systems
functions. The goal of the computation is then to minimize the residuals for a given
approximation. This is done in an average sense by setting the weighted integrals to zero
Z Z
Wi R dV + Vi R dA = 0. (3.72)
The first part of the expression is a volume integral over the whole domain with the
volumetric variable of integration dV . The second part represents the surface integral
over the boundary and the integration variable is dA. The functions Wi and Vi are a
set of of weight functions with i = 1, . . . , N . By including the differential equations, the
following formulation is obtained
Z Z
2 h
Vi Ezh dA = 0.
Wi ( k )Ez f dV + (3.73)
63
3. Application of the Radial Point Interpolation Method to Maxwells Equations
The delta function (xi x) is zero everywhere except at the collocation node xi and
the integral over the function is equal to one. When inserting the new weight functions
into (3.74), the function becomes
Z Z
i (xi x)R dV + i (xi x)R dA =
Z Z (3.77)
2
i (xi x) ( k )B f dV + i (xi x)BdA = 0.
Due to the sampling property of these delta functions, the integrals in the weighted
residual formulation evaluate the basis functions at the collocation node locations.
In functional analysis, this form is often referred to as the strong form, since derivatives
of the basis functions need to exist up to the maximum order in the differential equations
(second-order here). The collocation method minimizes the residual error of the PDEs
exactly on the collocation nodes xi . In between these nodes, the magnitude of the error
depends on the choice of a suitable basis function. The RPIM approach is expected to
yield small errors in the intermittent locations.
In collocation methods no connectivity between the collocation node locations has to
be known beforehand. No elements or cells are associated with the nodes. This property
is the core of the method, simplifies node adaptation, and greatly reduces the compu-
tational overhead usually associated with mesh-based methods. It is therefore a truly
meshfree method.
In the previous derivation only the Dirichlet condition was discussed. The extension
to other types of boundary conditions can be implemented in the same way and will
be discussed later. In the following first step, an eigenvalue solver for the TM-mode is
implemented to compute the resonant frequencies and the corresponding field patterns.
Later in a second step, a source problem solver will be implemented. Again the treatment
will consider only solving for the Ez component of the electric field, but the terminology
is different. For the source problem, this component corresponds to a TE-mode, since
the propagation direction lies on the x, y-plane with the perpendicular electric fields.
64
3.2. Second-Order Systems
Ez k 2 Ez = f in (3.79a)
Ez = 0 on . (3.79b)
To formulate the eigenvalue problem, eigenvalues = k 2 are defined inside the compu-
tational domain. The domain is filled with a homogeneous material and no sources are
present, i.e. f = 0. The differential equations (3.79) are rewritten as
Ez = Ez in (3.80a)
Ez = 0 on . (3.80b)
A = B (3.81)
with a vector of length equal to the number of nodes NI +NB . A is the stiffness matrix
for the Laplacian and B the mass matrix. Due to the boundary condition, the mass
matrix contains NB rows with zero entries. This renders it singular, which is problematic
for eigenvalue solvers. By rearranging these two matrices the method can be further
modified to only solve for the interior nodes using modified A0 , B0
A0 0 = B0 0 , (3.82)
where the numerical solver computes the reduced vector 0 . The length NI corresponds
to the number of degrees of freedoms ndof . This step leads to a non-singular matrix B0 .
As the derivation takes different forms in different frameworks, it needs to be performed
separately for each method. A generalized eigenvalue solver computes the eigenvalues
and eigenvectors 0 . Since the matrices are full, the package LAPACK [88] for full
matrices yields fast results. It is a direct solver which is very efficient for small ma-
trices. For larger problems nevertheless, the computational effort increases with a rate
between O(N 2 ) and O(N 3 ) [88]. In contrast to sparse eigenvalue solvers where only a
small number of eigenvalues are computed, the direct solver yields ndof eigenvalues and
corresponding eigenvectors.
In the following two variations of RBF methods, the Kansa method [31, 32] and
its symmetrical form [34], are summarized. Subsequently, an algorithm for eigenvalue
problems with RPIM is presented. A comparison between these methods is performed
in terms of possible applications and numerical accuracy and efficiency.
65
3. Application of the Radial Point Interpolation Method to Maxwells Equations
gradients, which gives better accuracy, especially for Neumann boundary conditions. The
governing equation (2.1) subsequently becomes
NI
X NB
X
hu(x)i = an rn (x) + an (x xn )rn (x) (3.83)
n=1 n=NI +1
with NI interior and NB boundary nodes (N = NI +NB ). The eigenvalue problem (3.80)
is discretized and written in the matrix form (3.81). The following notation is used:
i = [1, NI ]
L
[A1 ]i,j = rj (xi ), (3.84a)
j = [1, NI ]
i = [1, NI ]
L
[A1 ]i,j = (xi xj )rj (xi ), (3.84b)
j = [NI + 1, NB ]
i = [NI + 1, NB ]
[A1 ]i,j = rj (xi ), (3.84c)
j = [1, NI ]
i = [NI + 1, NB ]
[A1 ]i,j = (xi xj )rj (xi ), (3.84d)
j = [NI + 1, NB ]
i = [1, NI ]
[B1 ]i,j = rj (xi ), (3.84e)
j = [1, NI ]
i = [NI + 1, NB ]
[B1 ]i,j = (xi xj )rj (xi ), . (3.84f)
j = [1, NB ]
66
3.2. Second-Order Systems
A1 (1,NI ) + A
1 (NI +1,NI +NB ) = 0. (3.86)
This relationship can be used to express the eigenvector at the boundary nodes
(NI +1,NI +NB ) in terms of the interior nodes (1,NI ) as
1
(NI +1,NI +NB ) = A
1 A1 (1,NI ) (3.87)
The system (3.81) solves for the interpolation coefficients = [a1 , . . . , aNI ]T . The ex-
traction of the eigenmodes of the electric field is done via (3.83) and involves a matrix
inversion. The NI NI matrices A0 , B0 are not symmetric and counter examples show
that A0 may not be invertible [99]. However, it should be noted though that for a large
number of investigated numerical examples, this did not cause any adverse effect on the
results [100]. A thorough theoretical investigation on this issue was performed in [40]
and is was shown that it can generally be solved through oversampling.
The formulation of the NS-Kansa method can be brought to a symmetric form by mod-
ification of the basis functions [30, 34]. The Gaussian interior RBFs rn are replaced
by their Laplacians xn rn where the derivatives act on the collocation nodes xn . The
modified governing equation takes the following form
NI
X NB
X
hu(x)i = an (xn rn )(x) + an rn (x). (3.89)
n=1 n=NI +1
The stiffness and mass matrices can be constructed in the same way as for the NS-Kansa
Method. Here, the Laplace operator is applied analogously as in (3.84a) onto the interior
xn
nodes (ALL
2 ) and leads to fourth-order derivatives of the original RBFs (xn rn ).
The operator applied to the boundary nodes yields AL 2 with rn similar to (3.84b). The
xn xn
RBFs Lxn rn applied to (3.84c) and (3.84e) lead to AL 2 and BL
2 respectively. The
basis functions in (3.84d) and (3.84f) are replaced with rn and yield the matrices A2
and B2 . The listing of all matrix entries can be found in Appendix C.2.
67
3. Application of the Radial Point Interpolation Method to Maxwells Equations
Due to the expansion (3.89), the stiffness matrix A2 is now of symmetric form with
xn
Lxn
fourth-order derivatives in ALL
2 and second-order derivatives in AL
2 and A2 . In the
lower right in A2 , the symmetric matrix A2 contains the RBFs in basic form rj (xi ). It is
important to note that this modification of the basis functions only leads to a symmetric
form for the Dirichlet boundary conditions. In case of Neumann or mixed conditions,
the basis functions have to be modified accordingly.
The same transformation as in NS-Kansa brings both matrices into non-singular form:
h i
xn
Lxn 1
A02 = ALL
2 AL 2 A2 A2 , (3.91a)
h i
Lxn 1
B02 = BL
2 B2 A 2 A 2 (3.91b)
As a result both the mass matrix A02 and the stiffness matrix B02 have become symmetric
and non-singular, thus making the formulation well-posed. Additionally this symmetric
form is beneficial for numerical solvers.
RPIM
In both Kansa methods, the field quantities are extracted in a post-processing step by a
matrix inversion. For RPIM, this step is moved to a preprocessing step. In contrast to
the direct application of radial basis functions in the previous two methods, the RPIM
shape functions from (2.12) are used. n (x) fulfills the delta property and therefore
has several numerical advantages. The discretized form of (3.81) is obtained analogously
to the previous two approaches, but this time the Laplace operator is applied to the
shape function n (x) from (2.12). In order to solve the problem on the interior nodes,
and to bring the mass matrix into non-singular form shape functions are calculated on
all NI + NB node locations, but only evaluated at the interior points NI . The Laplace
operator applied on the RPIM shape functions is
68
3.2. Second-Order Systems
The eigenfunctions = [u1 , . . . , uNI ]T in (3.81) now correspond directly to the field
values. It is interesting to note that B03 is the identity matrix. This is advantageous for
the numerical eigenvalue solver or when the method is applied to time-domain problems
where an explicit time stepping, e.g. in the localized approach [72], leads to an efficient
scheme.
69
3. Application of the Radial Point Interpolation Method to Maxwells Equations
(a) (b)
Figure 3.18.: Models considered in numerical experiments for the second-order eigen-
value solver. Node distributions are shown for (a) a rectangular geometry
with randomly disturbed node locations and (b) a circular geometry with
concentric node locations.
Fig. 3.19 shows the structures of the stiffness and mass matrix for the three methods.
It can be seen that both Kansa mass matrices in Fig. 3.19a and 3.19b are full, whereas
the mass matrix for RPIM is the identity matrix (Fig. 3.19c). Additionally Fig. 3.19b
illustrates the symmetry of the structure in the S-Kansa method.
The analytical eigenvalues for both domains are based on the resonance frequency
m,n
fm,n = 2 of order m, n summarized in Appendix B.1. The relative eigenvalue error
is then calculated as
| (mn )2 |
rel = . (3.94)
(mn )2
This investigation focuses on the first and fifth eigenmodes (in the numerical order of
the eigenvalues) are further looked upon. For both domains, these are the T M11 and the
T M13 modes as illustrated in Fig. 3.20.
Parameter Convergence
Two fixed node distributions are generated, with NI = 49 for the rectangular domain
and NI = 155 for the circular domain. Full matrices of sizes 4949 and 155155 are the
result. That size is equal to the number of degrees of freedom ndof . According to [39],
the accuracy of the methods should increase for lower values of the shape parameter
c . This is clearly visible in Fig. 3.21 where the relative error for the first and fifth
eigenvalues is displayed. In Fig. 3.21, the relative error for the first and fifth eigenvalues
is displayed. The results clearly show the expected decrease of the error for the shape
70
3.2. Second-Order Systems
Figure 3.19.: Matrix structures for (a) the non-symmetric Kansa Method (3.88), (b) the
symmetric Kansa Method (3.91) and (c) RPIM (3.93). The stiffness matrix
A0 is shown on top and the mass matrix B0 on the bottom.
parameter decreasing down to c 102 . The relative errors decrease down to very low
minimum of 107 and 104 , respectively for both domains. However when the value of
the shape parameter becomes too small, the accuracy of the calculated eigenvalue de-
generates quickly. This numerical breakdown occurs at c 0.05 due to ill-conditioning
of the matrix. The results of all three RBF methods are almost identical despite of the
differences in the formulation. Only in the case of the rectangular domain, the NS-Kansa
method performs slightly worse than the other two methods.
Spatial Convergence
To evaluate the convergence rate and achievable accuracy of the method, the error for the
selected two eigenvalues has been calculated for increasing node densities. For each node
distribution, the LOOCV algorithm is applied to find an optimized shape parameter
c . Fig. 3.22 shows the result in both domains. For both cases, a uniform refinement
has been conducted. In the case of the rectangular domain, the regular distribution has
been again disturbed with an amplitude equal to the average node distance. For the
circular domain, the rather neat uniform concentric distribution is chosen. The meshless
approaches show the expected fast convergence up to a point where the rate slows down.
This is explained by the inability of the current LOOCV implementation to find the
best global shape parameter c for these larger node distributions. Again, all three
RBF methods are almost identical. Only in the case of the NS-Kansa method for the
unstructured rectangular domain, the results fail to converge. Numerical instabilities in
that case could be solved though by the aforementioned concept of oversampling. In
general it can be observed that even for very coarse discretizations with ndof < 50, very
71
3. Application of the Radial Point Interpolation Method to Maxwells Equations
(a) (b)
(c) (d)
Figure 3.20.: Investigated modes for the comparison between the three meshless eigen-
value solvers. The top graphs show the first eigenvalue (a) and the fifth
eigenvalue (b) for the rectangular domains and the bottom graphs show
the first eigenvalue (c) and the fifth eigenvalue (d) for the circular domain.
Off-Axis Eigenvalues
In order to quantify the eigenvalue distribution, an experiment is conducted similarly as
for first-order problems in Sec. 3.1.8. For the lossless resonator, the eigenvalues = k 2 are
72
3.2. Second-Order Systems
(a) (b)
Figure 3.21.: Convergence for decreasing values of the shape parameter c . The number
of degrees of freedom is fixed to ndof = 49 for the rectangular domain
and to ndof = 155 for the circular domain. The results are for (a) the
rectangular and (b) circular domain.
expected to be real-valued. Spurious modes for low values of the shape parameters can
be experienced analogously as in Fig. 3.14. Here, a positive imaginary value corresponds
to an eigenmode with an unphysical negative conductivity.
Fig. 3.23 shows the maximum relative imaginary part for the eigenvalues for varying
shape parameters using the previously fixed node distribution. A picture similar as the
one for first-order problems can be observed. For low values of c , very large imaginary
parts indicate the existence of spurious modes polluting the solution. For larger values,
this imaginary part decreases. In the highly unstructured rectangular domain, these
spurious modes exist up to a shape parameter of c = 1. As in the previous experiments,
numerical instabilities of the NS-Kansa method degrade the solution here as well. For
the circular domain, due to a more homogeneous node distribution the imaginary parts
of the eigenvalues vanish to a numerical zero for c > 0.1.
If the shape parameter c is too small, ill-conditioned matrices will produce spurious
modes. This observation is not only valid for the second-order RPIM eigenvalue frame-
work, but also for the non-symmetric and symmetric Kansa method. Again, conclusions
on the stability can be drawn for a possible time-domain implementation using a finite
second-order derivative scheme for the time variable. Also here, the shape parameter
73
3. Application of the Radial Point Interpolation Method to Maxwells Equations
(a) (b)
Figure 3.22.: Convergence for increased node densities with uniform refinement. The
methods are NS-Kansa (solid ), S-Kansa (dashed ) and RPIM (dash-
dot ). The graph shows the results for (a) the rectangular domain and
(b) the circular domain.
should be large enough to ensure that spurious modes would not cause instabilities.
Higher Eigenvalues
For the same fixed node distribution the first eight eigenvalues are observed. In this case,
the shape parameter is fixed to c = 0.05. In both the rectangular and circular examples,
higher-order eigenfunctions have faster oscillating spatial variations of the fields. As a
consequence, the node density per oscillation is reduced. This results in lower accuracy
of the eigenvalues. The relative eigenvalue errors for the first eight eigenvalues ordered
according to their resonance frequency are displayed in Fig. 3.24. On the left the errors
are shown for the rectangular problem. The trend of increasing errors for higher-order
eigenvalues is visible. The lowest eigenvalue error is at rel = 106 , increasing up to
rel = 3 104 . As in the previous numerical experiments, the NS-Kansa method error
shows a higher volatility. Due to the symmetry in the rectangular domain, the eigenvalues
2, 3, as well as 5, 6 and 7, 8 come in pairs. All the errors are nevertheless quite low, given
the very low number of degrees of freedom of ndof = 49.
The right-hand side graph of Fig. 3.24 shows the eigenvalue errors for the circular
74
3.2. Second-Order Systems
(a) (b)
Figure 3.23.: Relative magnitude of off-axis imaginary parts of the eigenvalues for differ-
ent shape parameters c . The rectangular domain (a) is shown on the left
and the circular domain (b) is on the right.
domain. The errors start at rel = 106 again for the first eigenvalue and increase to
rel = 102 for the highest considered mode. In this case the results of the three methods
are much more similar than for the rectangular domain. This is expected due to the more
homogeneous node distribution of this discretization.
Efficiency
For the node distributions investigated in the spatial convergence studies, the computa-
tion time of the full eigenvalue solver has been recorded . The code is run on a multi-core
64-bit Linux machine in a MATLAB implementation. For very short computation times,
the uncertainty in the time measurement is very large and fluctuations must be ex-
pected, since it cannot be guaranteed that the operating system scheduler will direct
the full processing power to the computation. On the left-hand side graph in Fig. 3.25,
starting at a very low ndof = 9, very short the computations time of order a fraction
of a millisecond are recorded for the rectangular domain. For increasingly larger prob-
lems, the computation time increases monotonously. The rate is between O(n2dof ) and
O(n2dof ). At a matrix size of 289 289 the computation of the eigenvalues takes 0.34 s,
which is still very fast. The larger circular domain shows the same behavior. Here, the
75
3. Application of the Radial Point Interpolation Method to Maxwells Equations
(a) (b)
Figure 3.24.: The first eight eigenvalues for a fixed node distribution and constant shape
parameter. On the left are the eigenvalues for (a) the rectangular domain,
on the right for (b) the circular domain.
computation time is increased to one second for a size of about ndof 400. For both
examples, to achieve an accuracy below 106 for the first eigenvalue, a computation of
less than 102 s is required.
All results for the three methods lie very close together. A comparison with other
methods will be presented for these numerical examples in Chap. 6.
Three meshless methods based on radial basis functions have been presented. The NS-
Kansa method leads to a non-symmetric form and does not necessarily guarantee unique
solutions. Brought into symmetric form, the method becomes well-posed (S-Kansa). Both
algorithms solve for the interpolation coefficient, and to obtain the field values, a matrix
inversion must be performed in a post-processing step. The radial point interpolation
(RPIM) method performs this step in preprocessing.
Despite the different formulations, all algorithms appear to yield numerically almost
identical results. In time-domain schemes where an explicit mass matrix is required,
RPIM appears to be best suited. The S-Kansa method is beneficial in its ability to
guarantee unique solutions. Nevertheless, for the NS-Kansa method the approach of
76
3.2. Second-Order Systems
(a) (b)
Figure 3.25.: Computational cost for different number of degrees of freedom ndof for (a)
the rectangular domain and (b) the circular domain.
Ez k 2 Ez = f (x) in (3.95a)
Ez = 0 on P EC (3.95b)
n Ez = 0 on P M C (3.95c)
Ez = g(x) on Source . (3.95d)
77
3. Application of the Radial Point Interpolation Method to Maxwells Equations
p
The wave number k = 2 j depends on the material properties and the chosen
frequency of the stationary signal.
The physical domain is described through a set of collocation nodes xi . Sufficiently
many nodes are placed inside the domain and on the boundaries P EC , P M C and
Source . In order to solve the physical problem (3.95) with the RPIM collocation ap-
proach, the shape functions (2.12) are calculated in the preprocessing step. The shape
functions are (xi ) = [1 (xi ), . . . , N (xi )] with N = NI + NB . The shape functions
and their spatial derivatives are chosen as the basis functions B(xi ) in (3.69). The new
set of equations becomes
The differential operators are discretized through (xi ) = x2 (xi ) + y2 (xi ) and
(xi ) = [x (xi ), y (xi )]T . The vector ni = (nx,i , ny,i )T is the normal unit vector
at the boundary node i. The vector Ez is of length N = NI + NB . The equations (3.96)
can be assembled in a system of the form
BEz = M. (3.97)
The terms in (3.96) containing (xi ) and its spatial derivatives correspond to the ith
row in the stiffness matrix B as
2
j (xi ) k j (xi ),
i , j = [1, . . . , N ]
(x ),
j i i P EC , j = [1, . . . , N ]
[B]i,j = (3.98a)
x j (xi )nx,i + y j (xi )ny,i , i P M C , j = [1, . . . , N ]
j (xi ), i Source , j = [1, . . . , N ]
f (xi ), i , j = [1, . . . , N ]
[M]i = g(xi ), i Source , j = [1, . . . , N ] (3.98b)
0, otherwise
Ez = B 1 M. (3.99)
The stiffness matrix B is full and not symmetric, although individual blocks inside the
matrix using a Dirichlet condition remain symmetric. The reason is that these entries
require no geometrical information, thus are insensitive to node locations. An example
of the stiffness matrix structure is shown in Fig. 3.26. Since the shape functions of
the RPIM setting fulfill the delta property, most entries are zero at the PEC and source
boundaries in B. This structure is advantageous for full matrix solvers. The transforma-
tion (3.88) is not necessary here, since no invertible mass matrix is required.
78
3.2. Second-Order Systems
Figure 3.26.: Sample structure of stiffness matrix B. The domain is enclosed by source,
PMC and PEC boundaries.
For the second-order RPIM framework, normalized units are used, i.e. 0 = 1, 0 = 1,
c0 = 1. Even though the scaled units are applied here, scaling back to physical units is
straightforward.
Taking a closer look at (3.96a), it can be observed that it contains the wave number
k. Large values of k dominate in the stiffness matrix B, thus leading to higher condi-
tion numbers. For very high wave numbers, a numerical limit is reached and the prob-
lem (3.97) becomes non-invertible. Practically this limit is not reached, because in these
cases the discretization is too coarse to accurately reproduce solutions with such high
spatial oscillations. Choosing a finer discretization leads to larger values of the Laplacian
of the shape functions, which balances the magnitudes again.
In contrast to other RBF schemes such as the Kansa method, the RPIM approach
solves directly for the field values. The matrix inversion required to compute the solutions
for the Kansa methods is moved to the preprocessing. This facilitates the application
of the LOOCV algorithm, an important step to directly use the method without tuning
parameters for each setup. In the next chapter, uniaxial absorbing boundary conditions
will be introduced to model open waveguide problems and other structures enclosed by
open rectangular boundaries.
79
3. Application of the Radial Point Interpolation Method to Maxwells Equations
the following.
The problem can be solved by decomposing the domain into piecewise homogeneous
domains. Fig. 3.27 illustrates such a situation. In this case the domain contains two
materials. The first one with the properties 1 , 1 and 1 and the second one with 2 , 2
and 2 . The domain is subsequently split in to two subdomains 1 and 2 such that
= 1 2 . The interface between the materials is the intersection of the boundaries
of both subdomains
DD = 1 2 . (3.100)
In this setup a domain decomposition method can be applied. In literature, two con-
cepts were introduced for general RBF methods in arbitrary PDEs [42]. Here, they are
applied to the second-order RPIM implementation for source problems. Both concepts
are illustrated in Fig. 3.28.
First, the Schwarz method uses overlapping domain regions to solve two problems
in an additive or multiplicative fashion [102]. The overlapping region is part of both
subdomains. On both subdomains the problem is solved separately. On the interface in
between, the discontinuity conditions are enforced and iteratively, a solution is found.
The balancing between the two subdomains is achieved through Dirichlet or Neumann
conditions on the interface. The formulation applied here for the overlapping scheme uses
Dirichlet conditions. In the additive scheme, the interface values are updated based on
both field values of the previous iteration. In the multiplicative algorithm on the other
hand, the field values at the boundary of the second domain 2 are set to the already
updated values at the first domain boundary 1 . The multiplicative scheme has been
found to converge faster than the additive method. For the interior domains and the
boundary DD , the equations to be solved are
k 2 Ez1
n
=f in 1 12 (3.101a)
n n1
Ez1 = Ez2 on DD (3.101b)
k 2 Ez2
n
=f in 2 12 (3.102a)
n n
Ez2 = Ez1 on DD . (3.102b)
for the second subdomain. The continuity of the transverse electric field is guaranteed
in this expression. The wave number is chosen with the appropriate material parameters
inside the domains as
p
k1 = 2 1 1 j1 1 (3.103)
p
2
k2 = 2 2 j2 2 . (3.104)
80
3.2. Second-Order Systems
(a) (b)
Figure 3.28.: Overlapping (a) and non-overlapping (b) decomposition schemes. The over-
lapping region 12 is part of both 1 and 2 .
and setting the boundary values according to (3.101b) and (3.102b). It has been reported
that the solution converged after less than 10 iterations.
Second, the non-overlapping so-called iterative substructuring method uses disjoint
regions 1 2 = 0 (Fig. 3.28b). For RBFs, a method using Dirichlet conditions on
one side and Neumann conditions on the other side of the interface have been presented
in [42] such that in the first subdomain
k12 Ez1
n
=f in 1 (3.105a)
n
Ez1 = n1 on DD (3.105b)
81
3. Application of the Radial Point Interpolation Method to Maxwells Equations
k22 Ez2
n
=f in 2 (3.106a)
n 2 n1
n Ez2 = n Ez1 on DD . (3.106b)
1
The factor 21 is derived from (3.31) and (3.59) and ensures that the transverse magnetic
field is continuous. The boundary term n1 is updated in each iteration. It is the
weighted average between the previous values at the boundary in the first subdomain
and the updated terms from the second subdomain
n = n1 + (1 )Ez2
n
on DD . (3.107)
The conformal derivative n is taken along the normal direction to the interface DD .
The iteration coefficient is (0, 1). A value of = 0.5 as a starting point is a natural
choice. In this setup, the Gibbs error can be avoided since the wave vectors k1 and k2
are constant for each subdomain. Experiments have been performed in [42] and it was
found that the accuracy did not significantly deteriorate when the points at the interface
do not match. Hence it was suggested to generate non-matching node distributions due
to the simplified node generation procedure. The value at the interface of the opposing
subdomain can be interpolated using the RPIM shape functions. Again, the two stiffness
matrices B1 and B2 are set up and the solution can be achieved iteratively.
Using a domain decomposition scheme is not only expected to be helpful in solving
domains containing multiple materials. In general, the computational effort for larger
matrices increases with O(N 2 ). It is therefore desired to keep the matrix size to a min-
imum. Domain decomposition is a good tool to reduce the overall computational effort
by splitting up the problem into several smaller subproblems that are solved much more
efficiently than the global one. It has been shown in [42] that the error of a bench-
mark example remains in the same range for up to 60 subdomains while reducing the
computation time by a factor of 100.
Attempts to implement the second domain decomposition scheme for matching node
points for the second-order RPIM scheme lead to an unstable iteration. The problem
lie in a high sensitivity on the shape parameter. Further techniques have been proposed
in [103], based on Neumann-Neumann transmisstion conditions. It is assumed that this
condition with a combination of non-matching grids and a more dense node distribution
close to the interface will lead to more successful results.
82
3.3. Summary
for numerical solvers, whereas a symmetric formulation of the Kansa method guarantees
unique solutions.
An implementation of RPIM for source problems has been presented. The subdivi-
sion of the computational domain into several subdomains that can be solved through
a domain decomposition scheme have been discussed. It has been noted that this would
greatly improve the efficiency and allow for the exact treatment of material discon-
tinuities. This implementation is the basis for the application of absorbing boundary
conditions in the next chapter.
3.3. Summary
The first-order Maxwell time-domain radial point interpolation method has been intro-
duced. It can be seen as a generalized form of the FDTD method, giving more freedom in
modeling complex structures and higher accuracy due to the freedom of node placement
and choice of the size of the local support domain. It retains the desired properties of the
time-domain method for simulating large frequency bandwidths in a single computation
and the facilitated implementations of non-linear material structures, even though the
latter point has not been performed here. The important aspects of the method, starting
from the staggered node arrangement with the creation of the dual node distribution,
to the treatment of boundaries, sources and the choice of a stable time step have been
discussed. It has been shown through an eigenvalue spectrum analysis that the method
requires a sensible choice of the shape parameter to reach an optimal trade-off between
stability and accuracy.
In a second part, the second-order wave equation was implemented using RPIM. Firstly
for eigenvalue problems the method was compared to other RBF methods. The ad-
vantageous properties of the RPIM shape functions lead to an identity mass matrix.
Experiments showed that the numerical performance is almost identical to other well-
established RBF methods. Finally an RPIM implementation for source problems was
presented and important boundary conditions were discussed. The importance of us-
ing domain decomposition methods to increase the efficiency and to accurately enforce
boundary conditions on material boundaries has been discussed. This method might
become a competitor to the finite element method for many applications. Performance
comparisons will be presented in numerical examples in Chap. 6.
In the next chapter, an implementation of the uniaxial perfectly matched layer ab-
sorbing boundary condition will be presented based on the RPIM time-domain imple-
mentation. After which high-order absorbing boundary conditions will be introduced in
the second-order RPIM framework.
83
4. Absorbing Boundary Conditions
Abstract Simulations of open problems are an important task in computational electromagnetics.
Scattering or radiating problems are typical types of open problems that demand highly efficient
absorbing boundary conditions (ABCs) to allow truncation of the finite computational domain.
The requirements for ABCs are low extra computational cost and low reflections. These low
reflections from the absorbing boundary allow for reducing the size of the computational domain
and increasing the accuracy of the solutions. In this chapter two types of ABCs for the first- and
second-order RPIM are introduced. First, a review of existing concepts in time- and frequency
domain is given. Following this, perfectly matched layers (PMLs) in uniaxial formulation are
applied to the time-domain first-order RPIM scheme. The profile of the conductivity term inside
the layer is an important simulation parameter. A study is performed for best choices of this
coefficient and the PML is validated for waves impinging at normal and off-normal incidence. In a
second implementation high-order non-reflecting boundary conditions are derived for second-order
frequency-domain problems. This formulation relies on high-order derivative approximations, which
are an intrinsic property of the RPIM shape functions. In order to increase the accuracy, a special
node arrangement close to the boundary is developed. The implementations are validated with
numerical experiments in rectangular waveguides.
4.1. Introduction
Open problems are important challenges in computational electromagnetics. The deter-
mination of scattering from objects, the radiation property of antennas or coupling of
several antennas are typical tasks in this category. In open problems the electromagnetic
properties of a given physical geometry is investigated, this requires the domain to be
extended to infinity. In boundary discretization methods such as the method of moments
or boundary element method, this is a feature of the algorithms. Unfortunately, these
algorithms are often not suitable when complicated dielectric, non-linear or anisotropic
materials have to be simulated. For domain discretization methods, open space is ex-
plicitly discretized and the domain has to be enclosed by a boundary that acts as an
absorber. This concept is analogous to an anechoic measurement chamber absorbing wall
shown in Fig. 4.1. A larger chamber yields better measurement results, but often space
and money restrictions limit the size. Better absorbers allow for smaller chambers or
more accurate results. Similarly, the quality of absorbing boundary conditions (ABCs)
in numerical simulations determines the accuracy of the results. The lower reflections
from the boundary become, the closer the ABCs can be placed to a scatterer and the
smaller the overall computational cost. A typical setup is shown in Fig. 4.2 where a
85
4. Absorbing Boundary Conditions
Figure 4.1.: Absorbing walls in anechoic measurement chamber for open problems.
86
4.1. Introduction
Figure 4.2.: Simulation of an open problem with a scatterer placed within a waveguide
and absorbing boundary.
(a) (b)
Figure 4.3.: Concept of uniaxial absorbing boundary conditions (a) and perfectly
matched layers (b). The boundary is placed along the x-direction and the
formulation is derived for TE-modes.
This so-called global Dirichlet-to-Neumann map replaces the Sommerfeld condition (4.1)
with an infinite series of Hankel functions. In its original formulation this operator is non-
local, leading to full blocks in the stiffness matrix and is inefficient for large problems.
Further adaptations lead to a more efficient local formulation that fits very well into
the framework of the finite element method [110], but truncation errors can degenerate
solutions.
Later, the development of physically motivated perfectly matched layers (PMLs) took a
different approach. They were first introduced in the time domain for the FDTD method
and will be summarized in the following section. Recently PMLs gained importance in
frequency-domain solvers as well.
87
4. Absorbing Boundary Conditions
88
4.2. PMLs for First-Order Time-Domain Problems
4.2.1. Formulation
In the original framework introduced in the previous chapter, the material models are
restricted to isotropic linear homogeneous media inside the computational domain (3.4).
This is extended in this section to anisotropic material properties. In contrast to the
SPML, where a non-physical magnetic conductivity is added, matching and absorption
is achieved through the inclusion of a frequency-dependent complex valued term. The
permittivity and permeability tensors in frequency domain are written as follows
= [] , = [] (4.4)
with a diagonal matrix . For achieving uniaxial absorption in x-direction the matrix
[] takes the form
1/a 0 0
[] =
0 a .
0 (4.5)
0 0 a
In order for the material to be matched to the linear isotropic materials inside the domain
and to act as an absorber, the terms in the diagonal material tensors are chosen as [115]
x
a=1+ (4.6)
j
where x represents the loss term or conductivity for absorption along x and denotes the
permittivity. In the UPML formulation, this loss inducing frequency-domain parameter
is included into the system of partial differential equations. In three dimensions it yields
additional equations for the physical magnetic flux density B and the electric flux density
D. Absorption along other Cartesian axes can be achieved similarly by rearranging the
diagonal elements in [].
89
4. Absorbing Boundary Conditions
In the following the PML is developed for the two-dimensional TE-mode with absorp-
tion in x-direction. Then the component of B in the anisotropy direction based on the
material properties (4.4) and the frequency-domain term a (4.6) takes the form
Bx = x Hx . (4.7)
1 + j
Incorporated into the set partial differential equations (3.8), the following frequency-
domain equations result [115]:
jBx = y Ez (4.8a)
1 x
jHx = y Ez + Bx (4.8b)
1 x
jHy = x Ez Hy (4.8c)
1 x
jEz = (x Hy y Hx ) Ez . (4.8d)
Note that the x-component of the magnetic field Hx contains the expression of the
magnetic flux density Bx (4.7). This vector field component is added as an auxiliary
partial differential equation.
When transforming (4.8) to time domain, the following equations are obtained
t Bx = y Ez (4.9a)
1 x
t Hx = y Ez + Bx (4.9b)
1 x
t Hy = x Ez Hy (4.9c)
1 x
t Ez = (x Hy y Hx ) Ez . (4.9d)
Analogously to the previous chapter, the method of lines is used to derive the anisotropic
PML update equations for the first-order time-domain RPIM scheme.
4.2.2. Implementation
The implementation of the set of partial differential equations (4.9) into the RPIM
scheme follows the same steps as in Sec. 3.1. First space is discretized through a staggered
node arrangement scheme and Bx is stored at the same node locations as the magnetic
field. The shape functions x,y Hi and x,y Ej are calculated in the preprocessing step
according to (3.11) and (3.12). As before time is discretized with constant time steps
and the magnetic and electric fields are staggered in time. Again, a leapfrog scheme is
applied to derive the electric and magnetic field update equations from (4.9).
The magnetic flux density Bx is updated via central differences in time. Due to its
dependence on the spatial derivatives of the electric fields (4.9a), this field component
is stored at the same temporal positions as the magnetic fields. The partial differential
90
4.2. PMLs for First-Order Time-Domain Problems
equation for the magnetic field Hx (4.9b) is expressed by the magnetic flux Bx . In
the discretized scheme, the leapfrog scheme makes use of the magnetic field component
stored at the time steps n 1/2 and n + 1/2. The central difference method requires the
right-hand side to be evaluated at the intermediate point in time n. Therefore a time
average of the fields Bx n+1/2 and Bx n1/2 is computed [112]. After some simplifications,
the local update equations can be calculated as:
nAs
(n+1/2) (n1/2) (n)
X
Bx,i = Bx,i t y jEi Ez,j (4.10a)
j=1
The summations for each node are over all nAs neighbors inside the support domain. The
time step is chosen according to (3.21) as introduced in the previous chapter. The loss
term x in (4.10b-4.10d) varies over the space. Outside the PML, in the bulk of the com-
putational domain, the conductivity x is set to zero. In this region, the update equations
are identical on a discrete level to the original form for isotropic materials (3.17).
In the present formulation the PML is absorbing in x-direction and a layer of thickness
0 is placed at the end of the computational domain as shown in Fig. 4.3b. Absorption
is achieved by choosing a positive conductivity x as a continuous function in space.
The shape of the loss term is crucial to ensure a well performing PML. It has been
shown [112] that the step size, i.e. change of x in the progression of the PML, must be
small where most of the energy is absorbed to achieve low numerical reflections. For a
PML placed at the right end of the computational domain, starting at xP M L , the loss
term here calculated on discrete collocation nodes xi = (xi , yi ) takes the polynomial
form ( p
0,p xi x0P M L , xi xP M L
x,i = . (4.11)
0, xi < xP M L
The profile function is described by the order of the polynomial p = 1, 2, . . . and maxi-
mum loss coefficient 0,p . Fig. 4.4 illustrates the different profiles of the conductivity x .
91
4. Absorbing Boundary Conditions
Figure 4.4.: Illustration of the absorption coefficient x with profiles (4.11) with p =
1, . . . , 5 and the maximum absorption 0,p . All curves share the same theo-
retical reflection coefficient th (0).
becomes [111]
R 0
th () = e2(cos /c) 0 x (x)dx
. (4.12)
This means that the physical thickness 0 , the magnitude of the loss term x and the
angle of incidence all influence the absorption of waves in the PML region. In practical
applications, a theoretically expected reflection coefficient at normal incidence th (0) is
preset and the conductivity is chosen accordingly. For polynomial profiles, the loss term
is
p+1
c0 ln th (0)1
0,p = (4.13)
20
based on the theoretical reflection coefficient and on the polynomial order p. The theo-
retical performance of the PML for waves impinging at angle can be derived as
The performance degrades for higher angles and results in total reflection at oblique
incidence. Practically this means that waves can propagate unaffectedly along the PML.
In [4], a review of a vast number of numerical experiments was presented to give a
rule of thumb for optimal values of 0,p for different PML thicknesses 0 for the FDTD
method. Since these rules do not necessarily apply to RPIM, a comparison between the
theoretical and numerical reflection is performed to find practical values for the profile
and theoretical reflection coefficient th . In general, reflections in the order of = 80 dB
are a good result for practical simulations.
4.2.3. Validation
The PML model is validated through numerical experiments. The performance at nor-
mal and off-normal incidence is evaluated in two separate simulations. In both cases a
rectangular waveguide structure is used. To investigate normal incidence, a parallel plate
waveguide with PMCs on the boundaries is simulated. In such a structure, plane waves
with the propagation direction along the PMC boundary can be excited (Fig. 4.5a).
To investigate the case of off-normal incidence, a rectangular waveguide, enclosed by
92
4.2. PMLs for First-Order Time-Domain Problems
(a) (b)
Figure 4.5.: Models to simulate (a) normal and (b) off-normal incident waves. In the
first case a plane wave is excited and in the second case the fundamental
T E1 -mode corresponds to the superposition of two plane waves traveling
with wave vectors at angle .
PECs is used. According to [97], the fundamental T E1 mode in such a waveguide can
be represented by two symmetrical plane waves, both traveling at an angle of
s
2
fc
(f ) = cos1 1 . (4.15)
f
93
4. Absorbing Boundary Conditions
Figure 4.6.: Nodal distribution for the PML time-domain verification. A discretization
of x = min /8 is used for the plane wave approach and x = min /16 for
the simulation at off-normal incidence.
the discretization, this thickness leads to different number of nodes along the x-direction
inside the PML and thus influences the size of the steps of the conductivity x .
The electric field at a number of representative sensor nodes in Fig. 4.6 is recorded.
These sensor nodes are placed at a distance ds from the domain boundary in two ad-
joining rows. The distance ds = 2min is chosen as two times the minimum wavelength
of the largest simulated frequency. For all experiments, a reference solution has been
obtained using a domain extended in x-direction by a factor of five in an otherwise un-
altered node distribution. The sensor nodes are at exactly the same locations as before.
The simulation is stopped before any backward traveling wave is reflected at the source
and pollutes the recorded data at the sensor node locations. The reflected field from
the PML can then be obtained by subtracting the solution of the PML model from the
reference solution. The sensor data is transformed to frequency domain via a discrete
Fourier transform. The maximum numerical reflection coefficient is then calculated as
|F(EP M L,i )(f ) F(Eref,i )(f )|
num (f ) = max . (4.17)
i sensors |F(Eref,i )(f )|
A comparison between the theoretically expected reflection coefficient and the nu-
merically obtained reflection at normal incidence is performed for different orders p in
the profile (4.11). For each set of parameters the loss term x is computed according
to (4.11) and (4.13). The model is discretized with x = min /8 of the smallest sim-
ulated free-space wavelength. At this discretization, the thickness 0 corresponds to 10
layers of PML nodes. The maximum reflection coefficient within the given bandwidth is
plotted against its theoretical values ranging from th (0) = 140 dB to th (0) = 20 dB
in Fig. 4.7.
For all profiles the numerical reflection coefficient num follows the theoretically ex-
pected value th to approximately th = 60 dB. For smaller values of th (0), the
results start to deviate. Large steps in the loss term x occur for small orders of the
profile p in the beginning of the PML, and for large orders p in the latter region of
the PML. As expected, these steps are detrimental to the performance. A cubic profile
p = 3 shows lowest reflections. The numerical reflection reaches a plateau slightly below
num = 80 dB, even for much lower values of th . For a theoretical reflection coeffi-
94
4.2. PMLs for First-Order Time-Domain Problems
Figure 4.7.: Comparison between theoretically expected reflection th and the numeri-
cally measured reflection coefficient num for different profiles p of the ab-
sorption coefficient x . The theoretical reflection coefficient is shown as a
dashed line.
cient of th (0) = 80 dB, the lowest reflections achieved are num = 72 dB. For all
other orders p of the polynomial profiles, the breakdown occurs earlier and ultimately
leads to a deterioration of the PML performance for lower th . Generally, the numerical
reflections for the PML can be minimized by either refining the node density or choosing
a thicker PML model in terms of nodes.
Simulations for off-normal incident plane waves are performed in a model with a dis-
cretization of x = min /16. This finer node density is chosen to assure that numerical
discretization errors are minimized. The model described in Fig. 4.6 has the same size,
simulation bandwidth and physical thickness as before. For the finer discretization the
PML thickness corresponds to 20 nodes, leading to smaller step sizes of the conductiv-
ity. Lower numerical reflections and therefore a better correspondence to the theoretical
reflection coefficient th is expected to validate the PML implementation.
The cutoff-frequency at the fundamental mode for this structure is fc = 0.67 GHz. The
angles in this setup range from 15.5 to 90 . The results with a fixed theoretical reflection
coefficient of th (0) = 80 dB for the different profiles p is shown in Fig. 4.8. The group
velocity of the T E1 mode close to the cutoff-frequency fc is very close to zero. Therefore
not enough energy is injected into the waveguide in this frequency range to achieve
reliable results for the numerical reflection coefficient. This region in Fig. 4.8 is marked
with a gray box. Above this critical frequency the numerical reflections show the expected
95
4. Absorbing Boundary Conditions
-10 p=1
p=2
-20 p=3
p=4
-30
p=5
-40
-50
-60
-70
-80
-90
0.5 1 1.5 2 2.5
Frequency [GHz]
Figure 4.8.: Performance of the PML model at various angles of incidence for a given
theoretical reflection coefficient th (0) = 80 dB for different profiles p of
the absorption coefficient x (x). The dashed line represents the theoretically
expected value.
behavior. The dashed line shows the theoretical reflection coefficient according to (4.14).
The performance of the PML is about 10 dB higher than the theoretical values. Best
results are achieved for quadratic and cubic profiles, which again confirms the previous
results.
In the following, the PML thickness and the node density are varied. The polynomial
order of the conductivity profile is fixed to p = 3. Two different node distributions for the
rectangular waveguide model are considered. One coarse discretization uses an average
node density of x = min /8. A finer discretization is generated with a node density of
x = min /16. The thicknesses of the PML models are chosen with an identical physical
size for both models. The used values 0 = [0.5, 0.75, 1, 1.25, 1.5]2 GHz are based on the
free-space wavelength at 2 GHz. This leads to different thicknesses in terms of nodes for
the two different densities. For the coarse discretization min /8, the thickness ranges
from 5 to 15 nodes. In the case of the finer discretized model with min /16, the PML is
between 10 and 30 nodes thick. Fig. 4.9 shows the simulation results.
It can be clearly seen in both graphs that increasing the thickness drastically improve
the PML performance. For the coarse discretized model in Fig. 4.9a the 50% thicker
model with 0 = 15x yields 10 dB lower reflections than the original model using a
10-node-thick layer. Models thinner than the original model yield a degenerated perfor-
mance. This can again be explained by the size of the step in the conductivity over the
96
4.2. PMLs for First-Order Time-Domain Problems
(a) (b)
Figure 4.9.: Comparison of different PML thicknesses for a coarse discretized ((a), x =
min /8) and fine discretized ((b), x = min /16) waveguides at different
angles of incidence. The dashed line is the theoretically expected reflection
coefficient and the gray area signifies the region where the guided wave
velocity is too slow for meaningful results.
progression of the PML model. In conclusion, for rather low discretizations the perfectly
matched layer should be generated with a thickness of at least 10 nodes.
In the case of a finer discretized model (Fig. 4.9b), the performance of the 10 node
thick model is at the same level as for the lower discretization, around 50 dB. This is a
rather small layer in terms of wavelengths and the dominant step sizes of x are compa-
rable to the conductivity profile of the previous experiment for 10 nodes. Increasing the
thickness again improves the performance. At a thickness of 30 nodes or 0 = 1.52 GHz ,
the reflections are in the range of the theoretically expected values. Fluctuations never-
theless occur in the simulation results. This can be explained by the measurement setup.
The frequency range is larger than the single-mode regime and higher-order modes are
likely to be excited due to numerical noise and pollute the solution. Additionally standing
wave patterns are recorded. More elaborate recording methods, such as an S-parameter
extraction to determine the energy flow in forward- and backward-direction for the fun-
damental mode [124] is expected to give smoother results.
97
4. Absorbing Boundary Conditions
Thickness 0.52 GHz 0.752 GHz 2 GHz 1.252 GHz 1.52 GHz
Table 4.1.: Number of additionally introduced field components by the PML formulation
for the investigated thicknesses and node densities.
98
4.3. High-Order ABCs for Second-Order Problems
4.3.1. Theory
High-order non-reflecting boundary conditions in the time domain were derived previ-
ously in [106, 108] for uniaxial boundaries
J
Y
(cos p t c0 x ) Ez = 0 on ABC . (4.18)
p=1
The boundary ABC is placed along x = x0 and absorbs plane waves with an angle of inci-
dentce p as depicted in Fig. 4.10. Each factor in the product (4.18) causes the absorption
of waves impinging at angles p = 1 , . . . , J . The cos p term represents shrinking of
the x-component of the wave vector k for off-normal incidence. A non-reflecting bound-
ary condition (4.18) of order J absorbs plane waves from J incident angles. Therefore,
the boundary condition contains temporal derivatives and spatial derivatives of at most
order J. In time domain, the existence of high-order time derivatives can be tackled by
99
4. Absorbing Boundary Conditions
Figure 4.10.: Absorbing boundary placed along ABC at x = x0 . The plane wave prop-
agation is described through the wave vector k and the angle of incidence
at the boundary is p .
the introduction of auxiliary variables in a recursive scheme [126], however at the cost
of additional computational effort.
When considering frequency-domain solvers, the high-order temporal derivatives trans-
form to high-order j terms. For such stationary problems the boundary condition can
be expressed as
J
Y
(cos p j c0 x ) Ez = 0 on ABC . (4.19)
p=1
In this formulation, only derivatives of the spatial variable of order up to J are required.
For simulations at high frequencies, the term J or matrix entries for high-order deriva-
tives of the shape functions can nevertheless become large which could cause numerical
problems due to high condition numbers.
For the finite element method, non-reflecting boundary conditions became feasible
by the use of auxiliary variables to reduce the order of the spatial-derivatives to low
orders [127]. Alternatively, special finite elements with high regularity can in principle
be used [128]. However, the complexity of these elements makes this approach unpractical
for high-order ABCs.
Even though the presented implementation is limited to absorption in x-direction,
a generalization to other directions of the boundary is straightforward using confor-
mal derivatives. In addition, boundaries placed along a circle (or spheres in a three-
dimensional setting) can be achieved through a formulation in cylindrical coordi-
nates [129]. The type of boundary condition introduced here only absorbs plane waves,
comparably to the uniaxial PML types. For PMLs, extensions to absorb evanescent
waves exist using complex frequency shifted PML formulations [119]. Correspondingly,
high-order ABCs that take evanescent waves into account were recently introduced [130].
Instead of the real valued cos p terms, a number of complex valued coefficients are then
chosen.
In the following, the absorbing boundary conditions from order one to four for the
100
4.3. High-Order ABCs for Second-Order Problems
continuous problem is calculated through the direct evaluation of the product (4.19) as
J =1: ((j) cos 1 c0 x )Ez = 0 (4.20a)
2
J = 2 : ((j) cos 1 cos 2 (j)(cos 1 + cos 2 )c0 x + c2 x2 )Ez =0 (4.20b)
3
J =3: ((j) cos 1 cos 2 cos 3
(j)2 (cos 1 cos 2 + cos 1 cos 3 + cos 2 cos 3 )c0 x (4.20c)
+(j)(cos 1 + cos 2 + cos 3 )c20 x2 c30 x3 )Ez =0
4
J =4: ((j) cos 1 cos 2 cos 3 cos 4
(j)3 (cos 1 cos 2 cos 3 + cos 1 cos 2 cos 4
+ cos 1 cos 3 cos 4 + cos 2 cos 3 cos 4 )c0 x
(4.20d)
+(j)2 (cos 1 cos 2 + cos 1 cos 3 + cos 1 cos 4
+ cos 2 cos 3 + cos 2 cos 4 + cos 3 cos 4 )c20 x2
(j)(cos 1 + cos 2 + cos 3 + cos 4 )c30 x3 + c40 x4 )Ez = 0.
The angles p each represent perfect absorption for plane waves impinging from this
direction. These four expressions are now discretized and brought into the RPIM frame-
work. This is achieved by applying the collocation method introduced in the previ-
ous chapter. Collocation nodes are placed along the absorbing boundary. The equa-
tions (4.20) are discretized according to (2.12) through the spatial derivative shape func-
tions x1,...,4 (xi ). The discretized form of (4.20) is
J =1: (j) cos 1 (xi ) c0 x (xi ) (4.21a)
2
J =2: (j) cos 1 cos 2 (xi )
(j)(cos 1 + cos 2 )c0 x (xi ) (4.21b)
+c2 x2 (xi )
3
J =3: (j) cos 1 cos 2 cos 3 (xi )
(j)2 (cos 1 cos 2 + cos 1 cos 3 + cos 2 cos 3 )c0 x (xi )
(4.21c)
+(j)(cos 1 + cos 2 + cos 3 )c20 x2 (xi )
c30 x3 (xi )
J =4: (j)4 cos 1 cos 2 cos 3 cos 4 (xi )
(j)3 (cos 1 cos 2 cos 3 + cos 1 cos 2 cos 4
+ cos 1 cos 3 cos 4 + cos 2 cos 3 cos 4 )c0 x (xi )
+(j)2 (cos 1 cos 2 + cos 1 cos 3 + cos 1 cos 4 (4.21d)
+ cos 2 cos 3 + cos 2 cos 4 + cos 3 cos 4 )c20 x2 (xi )
(j)(cos 1 + cos 2 + cos 3 + cos 4 )c30 x3 (xi )
+c40 x4 (xi ).
The vectors represent the ith row in the stiffness matrix B for ABC node i. An extension
to higher orders J > 4 is straightforward, but for the sake of simplicity has not been
101
4. Absorbing Boundary Conditions
considered here. For each boundary node a new row is added to the stiffness matrix
B (3.97) and a zero entry is added to the vector M at the right-hand side of the discretized
system (3.97).
In [108], a theoretical reflection coefficient for the boundary conditions, which depends
on a preset selection of J absorbing angles p was derived as
J
Y cos p cos
ABC
th () = . (4.22)
p=1
cos p + cos
According to (4.22), the reflection coefficient is zero when the incoming angle matches
one of the absorbing angles p . Additionally ABC th goes exponentially to zero when
increasing the order J regardless of the choice of p . However, the reflection coefficient
can further be improved for a fixed J by a sensible choice of the absorbing angles.
The condition number of the stiffness matrix B is dependent on the angular frequency
. Fig. 4.11 shows the condition number for a numerical example in comparison to the
angular frequency using normalized units. In a usable frequency range limited by an
acceptable approximation accuracy however, the J terms do not dominate. Rather the
high-order spatial derivatives of the shape functions contribute to large condition num-
bers. For the fourth-order absorbing boundary condition, the condition number exceeds
1014 . For higher orders J, this is expected to rise further. In all investigated examples
in this theses nevertheless, no numerical problems were encountered.
with a fixed discrete set of equally spaced incoming angles i = 0 , . . . , 90 . The op-
timized absorbing angles range within 0 p 90 . The first absorbing angle 1 is
always set to zero, since in a practical environment boundaries will usually be mod-
eled perpendicularly to a major incidence. In order to avoid singularities of the cost
function (4.23), the output of the theoretical reflection coefficient has been limited to a
practically feasible minimum of 104 .
The optimized absorbing angles are shown in Tab. 4.2 and the resulting reflection
coefficient as a function of the incident angle is displayed in Fig. 4.12. In this graph the
results are compared to the theoretical reflection coefficient of a PML based on (4.14)
with th (0) = 80 dB. As introduced in the previous section, setting a practically rel-
evant theoretical reflection coefficient is sufficient for a comparison, since this term is a
102
4.3. High-Order ABCs for Second-Order Problems
Figure 4.11.: Condition number of the stiffness matrix B as function of the angular fre-
quency and the order of the absorbing boundary condition J.
J 1 2 3 4
1 0
2 0 22.3
3 0 30.2 43.2
Table 4.2.: Absorption angles resulting from a global optimization of the average theo-
retical reflection coefficient for ABCs of first to fourth order.
function of the thickness and conductivity of the PML model. When using this set of
absorbing angles the non-reflecting boundaries of orders J = 3 are comparable to the
theoretical PML. For higher orders, these absorbing boundary conditions outperform the
theoretical PML.
103
4. Absorbing Boundary Conditions
Figure 4.12.: Theoretical reflection coefficient based on the optimized absorbing angles
for first to fourth-order absorbing boundary conditions.
In order to achieve a smooth transition from the node distribution with average node
distance x to the Chebyshev zone the thickness L0 is chosen based on the previous
parameters as
x
L0 = . (4.25)
sin 2N C
With this node distribution, the node density increases near the boundary. Fig. 4.14
shows the node distribution for increasing numbers of Chebyshev nodes. Each line of
104
4.3. High-Order ABCs for Second-Order Problems
Figure 4.13.: Increased node density close to the boundary with a Chebyshev-like node
spacing.
dots illustrates a section of the resulting node distribution for NC = 1 to 10. Larger
values of NC bring nodes closer to the boundary. The interpolation accuracy of the
higher-order spatial derivatives is therefore expected to increase.
105
4. Absorbing Boundary Conditions
(a) (b)
Figure 4.15.: Parallel plate waveguide (a) to estimate numerical reflections at normal
incidence and rectangular waveguide to estimate the numerical reflection
at off-normal angles (b).
106
4.3. High-Order ABCs for Second-Order Problems
(a) (b)
Figure 4.16.: Node distribution and field plots of plane wave model (a) and off-normal
waveguide model with T E1 mode (b) at a frequency of = 2. The box
highlights the selected sensor nodes.
Figure 4.17.: Increased node density with a Chebyshev-like node spacing near the bound-
ary to increase the accuracy of the high-order spatial derivatives. Distri-
butions for NC = 1 (a), NC = 4 (b), NC = 7 (c) and NC = 10 (d) are
investigated.
the theoretically expected values. Here, only the numerical results for the high-order ab-
sorbing boundaries are presented. Instead the reflections computed with the second-order
RPIM scheme are compared to the theoretical PML reflection coefficient (4.12) in the
following numerical experiments.
Normal Incidence
In the first experiment, the numerical reflections at normal incidence are evaluated. A
parallel plate waveguide shown in Fig. 4.15a is used to excite a plane wave. The source
107
4. Absorbing Boundary Conditions
NC 1 4 7 10
Table 4.3.: Number of degrees of freedom for different numbers of Chebyshev-like node
spacings.
Grazing Angles
Off-normal incidence is simulated in the rectangular waveguide shown in Fig. 4.15b. This
model is enclosed by perfect electric conductors (P EC ) to guide a waveguide mode. The
source is modeled as g(x = (x, y)) = sin ya as in (3.67), thus exciting a T E1 mode. The
dimensions and node distributions are identical to the previous model, leading to a cutoff
frequency of c = .
Numerical results are shown in Fig. 4.19 for orders of the ABC from one to four.
The reflection coefficients are compared to the theoretically expected values and to the
theoretical PML reflection coefficient. Similarly as for the case of normal incidence, the
first and second-order ABCs (J 2) in an unaltered grid perform well and match the
theoretically expected values as shown in Fig. 4.19a. The performance of the third-
order ABC does not match ABC th , but nevertheless is comparable to that of a PML.
In the case of J = 4 the accuracy of the fourth-order spatial derivative breaks down,
leading to higher reflections. This is again improved by increasing the node density near
the boundary, i.e. NC > 1. In the case of NC = 7 in Fig. 4.19c, the reflections of all
investigated ABCs get very close to the theoretical values. This is even further improved
in the case of NC = 10. The numerical reflections of the fourth-order ABCs in Fig. 4.19d
are superior to the idealized PML model over the whole simulated frequency bandwidth.
Especially at frequencies close to the cutoff frequency numerical reflections are at least
108
4.3. High-Order ABCs for Second-Order Problems
(a) (b)
(c) (d)
J=1 J=3
J=2 J=4
Figure 4.18.: Numerical experiments measuring the reflection at normal incidence. The
Chebyshev-like node spacings are NC = 1 (a), NC = 4 (b), NC = 7 (c) and
NC = 10 (d). The discretization in terms of wavelengths is on the top of
each figure.
20 dB smaller.
In Fig. 4.20, these numerical results using a Chebyshev-like node distribution of
NC = 10 are plotted against the angle of incidence. At frequencies close to the cut-
off frequency, the length of the waveguide is not sufficient to resolve a whole period of
the guided wavelength. Therefore the SWR at the corresponding angles very close to
109
4. Absorbing Boundary Conditions
Inf 40.0 20.0 13.3 10.0 Inf 40.0 20.0 13.3 10.0
0 0
-20 -20
-40 -40
-60 -60
-80 -80
-100 -100
-120 -120
-140 -140
0 0
(a) (b)
Inf 40.0 20.0 13.3 10.0 Inf 40.0 20.0 13.3 10.0
0 0
-20 -20
-40 -40
-60 -60
-80 -80
-100 -100
-120 -120
-140 -140
0 0
(c) (d)
J=1 J=3
J=2 J=4
PML
Figure 4.19.: Numerical experiments measuring the reflection for off-normal incidence.
The numbers of Chebyshev nodes are NC = 1 (a), NC = 4 (b), NC = 7
(c) and NC = 10 (d). The solid lines represent the theoretical reflection
coefficients, and the theoretical PML model is shown as a dashed line.
oblique incidence cannot be computed reliably. But for lower angles up to 75 , a very
good agreement between the theoretical, optimized absorbing angles and the numerical
reflections is achieved.
110
4.3. High-Order ABCs for Second-Order Problems
-20
-40
-60
-80
-100 J=1
J=2
-120 J=3
J=4
-140
0 15 30 45 60 75 90
Figure 4.20.: Comparison between theoretical and numerical reflection coefficient for dif-
ferent angles of incidence. NC = 10 Chebyshev nodes are used.
Convergence Analysis
To analyze the behavior of the errors caused by the absorbing boundaries in numerical
simulations, a convergence study is performed. The numerical errors consists of two dif-
ferent types of errors. Approximation errors in the whole domain dominate for coarse
discretizations. They are reduced in the progressively finer node distribution. The second
type of errors is caused by the reflections at the absorbing boundary. It has been estab-
lished in the previous section that these non-zero reflections naturally appear when the
absorbing angles p are not matched exactly. For the finer discretizations, these errors
dominate the approximation errors and the overall errors remain on a constant level in
this convergence study.
Errors are computed based on the theoretical field values inside the waveguides and
take the values at all collocation node positions into account. The exact fields at posi-
tion x = (x, y) for the parallel plate waveguide for plane wave excitation, and for the
rectangular waveguide with the T E1 mode excitation are [97]
111
4. Absorbing Boundary Conditions
TE-mode is s
2
c
kg = k 1 . (4.29)
112
4.3. High-Order ABCs for Second-Order Problems
Figure 4.21.: Convergence study the average error for a plane wave impinging at normal
incidence. The node density is increased at the absorbing boundary. The
curves denote absorbing boundaries from order one to four. The average
node density in terms of free-space wavelength is shown on top.
On the right in Fig. 4.22b, the higher frequency corresponds to a shorter guided wave-
length and therefore denser node distributions are required for the solution to converge.
The first-order formulation leads to errors of L2 = 3.8%. For second and third orders,
the solutions converge quickly to a constant level after ndof = 300 of 5104 and 3104 ,
respectively. For more than ndof = 400 nodes, the fourth-order boundary condition is
satisfactorily converged below 2 104 . The average errors inside the computational do-
main correspond very well to the theoretical reflection coefficient, since the dominant
error arises from the numerically reflected plane waves at the absorbing boundary.
113
4. Absorbing Boundary Conditions
(a) (b)
Figure 4.22.: Convergence of the average error for increased node densities for a T E1
mode. The orders of the absorbing boundary conditions range from one to
four. On the left are the results for an angular frequency of 1 = 1.1, and
on the right for 2 = 3.2. On the top of the figures are the corresponding
average node densities for the guided wavelengths.
can be significantly higher [4, 132]. Thus, the computational overhead of the presented
method is typically much smaller than for PMLs with similar performance while the
performance can be increased.
114
4.4. Summary
that for only a slight increase of the node distribution in the boundary region, when
increasing the overall node density the numerical errors of the computational domain
quickly decrease to the level expected by the theoretical absorption coefficient. In prac-
tical simulations for e.g. a scattering problem, less optimal results than the presented
test cases are expected, since the angle of incidence is hard to predict. Nevertheless, the
proposed absorbing boundary conditions combined with the absorbing incidence angles
are expected to perform much better than a PML model.
4.4. Summary
Two types of absorbing boundary conditions were implemented for the first-order RPIM
framework in time-domain and the second-order frequency-domain formulation of the
RPIM scheme for source problems. In the localized formulation of the time-domain radial
point interpolation method, a perfectly matched layer based on an anisotropic material
model has been implemented. A layer of a highly lossy material that is matched to the
surrounding material is placed in the vicinity of the computational domain boundary.
As illustrated for this case, uniaxial absorbers in x-direction for the two-dimensional
TE-mode have been implemented. Numerical validation showed the effectiveness and
accuracy of the PML model. Minimal reflections could be achieved by a selection of a
cubic profile for the conductivity term. Thicker layers with larger number of layers
lead to lower reflections due to smaller steps in the tapered conductivity profile.
For the second-order frequency-domain RPIM solver, high-order absorbing boundary
conditions were implemented. In contrast to the PML model, the absorber acts on the
boundary directly. Due to the high accuracy of spatial derivative interpolation of smooth
functions using global basis functions, the RPIM framework is very well-suited for the
direct application of this type of non-reflecting boundary condition. The implementation
has been introduced for the scalar wave equation with the uniaxial absorbing boundaries
placed along the y-axis. Boundary conditions up to fourth-order have been introduced.
Optimized absorbing angles have been numerically calculated to minimize the average
reflections for unknown incident angles. Numerical evaluations at normal and off-normal
incidence showed a very good agreement with the theoretically expected reflection coef-
ficient. A Chebyshev-like node distribution close to the boundary was demonstrated to
increase the approximation accuracy for the higher-order derivatives.
115
5. Adaptive Node Refinement
Abstract The possibility to change a given node arrangement without computationally expensive
adaptation of a mesh topology is one of the main advantages of the meshless RPIM method. This
chapter discusses an adaptive node refinement algorithm for the second-order eigenvalue solver.
First, an initial solution on a coarse node distribution is computed. Then the algorithm is applied
iteratively. A set of new potential collocation nodes is found in between existing node locations.
These nodes are called test nodes. Two different error estimators are introduced, where the first
estimator is based on the residual error of the physical problem to solve. The second estimator
purely relies in field variations in the solution and selects regions where the difference in gradients
between collocation nodes is large. The error functions are evaluated on the test nodes. A criterion
is defined to determine which nodes will be added to the set of collocation nodes. The iteration is
repeated until a required accuracy level is reached. Numerical experiments are conducted to give
guidelines for the simulation parameters and the two adaptive schemes are compared to a uniform
refinement.
5.1. Introduction
In numerical methods, the behavior of the numerical error for increasing node densities
is an important indicator of whether the solution is sufficiently accurate. The simplest
approach is to do a uniform refinement, where the mesh or node density is uniformly
decreased step by step. When the difference between the solutions is small enough in
two successive steps, a solution is assumed to be converged. The uniform refinement is in
most cases not an efficient strategy, because the discrete model is refined at all positions,
also at locations where the error is already small enough. More elaborate approaches
use an adaptive refinement to minimize the number of degrees of freedom added for
convergence of the solution.
Two main strategies are known. Firstly in an a priori geometry dependent refinement,
the discretization is refined in certain areas based on information about geometrical
features. For example close to a tip or corner, singular solutions can be expected and thus
the density is increased in the vicinity of this region. Secondly in a geometry independent
approach, a posteriori refinement strategies increase a mesh or node density based on the
last available solution. Starting from a coarse discretization with large local numerical
errors, the mesh or node density is increased at locations where large errors are present.
In areas of small errors, the discretization can be assumed to be sufficiently fine. Both
these approaches save memory and computation time because the number of degrees of
freedom is minimized.
117
5. Adaptive Node Refinement
The meshless collocation approach has the advantage of free node placement. The
nodes do not require an explicit connectivity, in contrast to mesh-based algorithms. The
procedure introduced here calculates a solution on an initial set of collocation nodes.
Since the algorithm is based on the collocation method, the residual error of the solution
is always zero on the collocation nodes. The residual error is therefore calculated on
a separate set of test nodes. When the error on some of the test nodes is above a
certain threshold, these nodes are selected and added to the set of collocation nodes.
Iteratively, a more and more accurate solution is obtained. When the overall error on
the test nodes is below a certain limit, the numerical solution is assumed to be sufficiently
accurate. Theoretically the test nodes can be placed randomly inside the domain. For a
more efficient implementation however, the algorithm proposed in this chapter generates
test nodes at mid-distance between the existing collocation nodes based on a Delaunay
tessellation.
In the following, the adaptivity algorithm is demonstrated for the second-order RPIM
eigenvalue solver due to the simplicity of the formulation. For validation, the comparison
of the eigenvalues to their theoretically expected values provides a good method to test
the reliability of a numerical method. The difficulty with this test lies in consistently
selecting the same eigenmode during the iteration to achieve a continuous adaptive
refinement of one selected solution.
118
5.2. Error Estimation
jump in the gradient between two mesh cells or nodes is an estimator of field varia-
tions that are not resolved well in the current discretization. This approach proved to
be very successful for the finite-element method [138]. For a general meshless node gen-
eration scheme in one dimension a gradient-based algorithm has been introduced very
recently [139]. For the element-free Galerkin method in computational mechanics, a re-
finement strategy based on gradients has been presented in [140]. In the field of computer
vision, a gradient-based algorithm was developed to add points for better visual repre-
sentation of node-based geometrical objects [141]. For problems with time-dependent
geometries, an adaptive meshless algorithm that allows for the movement of nodes was
introduced in [142]. An error estimator for radial basis functions based on the residual
error on test nodes has been introduced for general RBF methods [136] and has been
specialized for a localized mechanical RPIM scheme [143]. This estimator function relies
only on the location of the nodes and will be adapted to the RPIM collocation framework
for electromagnetic problems in the following.
In the following sections the refinement algorithm is developed for the two-dimensional
second-order RPIM eigenvalue solver introduced in Sec. 3.2. In a novel approach, the
LOOCV algorithm is applied at every iteration to find an optimized shape parameter and
to ensure an ideal solution for each node distribution. The residual-based estimator can
theoretically be applied on an arbitrary node distribution, also on randomly placed node
locations. To avoid calculating the errors on a too large set of nodes though, a procedure
to place nodes in between the collocation nodes is introduced. Afterwards, two error
estimators are presented. An error based on the residual is introduced. Furthermore, in
a new approach an error estimator is presented that takes the gradient of a solution into
account. This strategy is based on the jump in the gradient of the solution between two
existing node locations. Afterwards, the two error estimators are evaluated in numerical
experiments.
In order to increase the convergence behavior of the method, the set of test nodes is
generated at intermediate locations between collocation nodes Nc = [x1 , . . . , xN ]. Based
on the initial set Nc , a Delauney tessellation [144] is generated. The optimal set of
triangles tiling the existing node distribution is obtained by maximizing the minimum
angle for all triangles. As a result, triangle edges connect all collocation nodes. The new
set of test nodes Nt = [y1 . . . yT ] is then generated by placement of nodes on the edge
centers. Fig. 5.1 illustrates the procedure. For each test node, the distance between the
collocation nodes dyi is recorded. The directions of the left and right neighbors are stored
as unit vectors nlj and nrj .
Additionally on the boundary, nodes are placed in the center between preexisting
boundary nodes. To better conform to the original geometry, the curvature has to be
taken into account in the test node placement. For a circular boundary for example the
test nodes are placed at the half-angle in the arc segment between two collocation nodes.
119
5. Adaptive Node Refinement
(a) (b)
Figure 5.1.: Generation of test nodes based Delauney triangulation of a set of collocation
nodes. The principle is shown in (a) and an example on a circular domain
is illustrated in (b).
R = Ezl l Ez in (5.1a)
R = Ezl on (5.1b)
are evaluated on the set of test nodes Nt . To discretize (5.1) based on (2.12), the shape
functions (yj ) are calculated on the test nodes using no monomial basis functions
The moment matrix R0 has been calculated on the collocation nodes to obtain the
solution Ezl and is reused here.
The error estimator is then computed as
(
1 d2yj |(yi )(yj )Elz l (yi )Elz |2 (interior residual)
(yj ) = . (5.3)
+2 d2yj |(yi )Elz |2 (boundary residual)
The first term on the right-hand side of this equation is based on the residual error R .
It is applied to interior test nodes generated through the Delaunay triangulation. The
second term on the right-hand side of (5.3) is weighted by a factor 2 and represents
the residual error on the Dirichlet boundary nodes R . Large residuals on the inter-
nal or boundary nodes indicate unreliable solutions. The ratio of the weighting factors
120
5.2. Error Estimation
1 , 2 in (5.3) is used to adjust the different scales of the internal and boundary test
node residuals. For best results in the test cases 2 is usually chosen several orders of
magnitude larger than 1 . A discussion about these parameters follows in the numerical
investigations. The term d2yj in (5.3) approximates the area that is covered by the test
node yj .
In an iterative process, nodes with a residual error larger than a predefined threshold
(yj ) max (yl ), (5.4)
l
where (0, 1) is a constant, are subsequently added to the set of collocation nodes.
The choice of the threshold parameter determines how many nodes are added at
each iteration. For every new node distribution, a new optimized shape parameter c is
obtained through the LOOCV algorithm, and the eigenpair Elz , l is computed again.
A global residual sX
g = (yj ) (5.5)
j
can be used as a variable to formulate a stopping criterion. After each iteration, the
current global residual g is compared against the previous, and when the difference is
below a certain limit, it can be assumed that the solution has converged. Alternatively,
a stopping criterion based on the maximum global residual can be formulated [136]. The
maximum global error mg is recorded throughout the iteration. When the criterion
g(i) < mg . (5.6)
is fulfilled at the iteration step i, the iteration is stopped. The stopping parameter 0 <
< 1 is preset as a global parameter in the simulation. Fig. 5.2 illustrates the process.
The whole algorithm can be summarized in the following steps:
The discretization for the initial solution must be fine enough to ensure a sufficient
approximation quality at the intermediate test node locations.
121
5. Adaptive Node Refinement
Figure 5.2.: Illustration of the stopping criterion for the refinement algorithms.
i.e. the normal unit vectors along a line connecting two collocation nodes, is used to
compute the jump of the gradient of a previous solution in between two nodes. On a
discrete level in a homogeneous domain, large jumps are an indicator for rapid field
variations that are not resolved well with the given discretization. A section of the node
distributions is shown in Fig. 5.3. The jump in the gradient a numerical solution is
illustrated.
For the finite-element method, the estimator is applied on the interface between two
mesh cells [134]. In the meshless approach, the connectivity between the left and right
collocation nodes (xlj and xrj ) of a test node yj is exploited. The gradients are estimated
along the unit vectors nlj and nrj connecting these nodes. When the jump of the gradients
between the left and right neighbor is large, this test node j is subsequently added to
the set of collocation nodes. The estimated error is small either when only slow field
variations occur, or when the discretization is fine enough in a region with rapid field
variations.
The error estimator calculated for the RPIM scheme is
(yj ) = d2yj |(xlj )Elz nlj (xrj )Ez nrj |2 (5.7)
Again, the distance dyj between the nodes is used for estimating the local area surround-
ing the test node. In the finite-element estimator, this would be the size of the interface.
The gradients of the shape functions (xlj ) and (xrj ) are calculated on the left
and right neighbors of the test node location yj . Using conformal derivatives along the
connecting line between xlj and xrj , the jump is estimated on the test node.
The error estimator (5.7) can be applied successfully on the internal and boundary
nodes. This approach does not require weighting terms 1 , 2 as in (5.3). The iteration
is performed in the same way as in the previous residual-based method. A threshold
determines the ratio of test nodes added in each iteration and a global estimated error
g can be used in a stopping criterion.
122
5.3. Numerical Experiments
Figure 5.3.: Illustration of the jump in the gradient based on a previous numerical solu-
tion.
123
5. Adaptive Node Refinement
in (5.4). Second the number of test nodes added to the set of collocation nodes is
evaluated in dependence of the threshold . Afterwards the influence of this threshold
on the convergence behavior of the eigenvalue error is illustrated. The ability of the global
error estimator to predict whether the solution has converged is investigated in a next
step. Finally, a comparison between both adaptivity algorithms to a uniform refinement
is performed.
The evolution of the number of nodes over many iterations and for a large range of
the threshold parameter is shown in Fig. 5.5. The number of degrees of freedom ndof
is compared over 20 iteration steps for both the residual- and the gradient-based error
estimator. For both error estimators (5.3) and (5.7), the parameter determines how
124
5.3. Numerical Experiments
(a) (b)
Figure 5.4.: Histogram of the distribution of the error indicators (yj ) for the 1st eigen-
mode of the circular domain. The frequency of the error estimators is shown
for (a) the residual-based and (b) the gradient-based scheme.
many nodes are added at each iteration. Different values of lead to different rates of
added nodes. This can be determined by the steepness of the curves in Fig. 5.5.
A large value of the threshold parameter, i.e. 0.6 signifies that only a few nodes
are added per iteration. This means that many iterations have to be performed to have
a significant gain in accuracy. On the other hand, if the threshold parameter is small,
e.g. = 103 , the majority of the test nodes are added at each iteration. Choosing the
threshold parameter = 0 corresponds to a uniform refinement when a homogeneous
initial node distribution was chosen.
Confirming the observations from Fig. 5.4, different rates of node additions are seen
in Fig. 5.5a for the residual-based estimator and for the gradient-based estimator in
Fig. 5.5b. For various values of in a logarithmic scale, the even distribution of the errors
in Fig. 5.4a leads to curves in Fig. 5.5a with evenly distributed slopes. In Fig. 5.5b, the
rates of added nodes for = 103 to = 0.1 lie very close together, i.e. the curves almost
overlap. Higher values of the threshold parameter must be chosen for the gradient-based
estimator to slow down the addition of nodes. This is due to the clustering of the nodes
towards high values of the error estimator in Fig. 5.4b.
125
5. Adaptive Node Refinement
(a) (b)
Figure 5.5.: Evolution of the number of nodes over the iterations for different values of
for the refinement of a T M01 mode in a circular domain. Each marker
represents an iteration step. The residual-based refinement algorithm (a) is
shown for lower values of than the gradient-based algorithm (b).
126
5.3. Numerical Experiments
(a) (b)
Figure 5.6.: Influence of the threshold parameter on the convergence rate. The relative
eigenvalue error for the T M02 mode in the cylindrical unit disc is shown for
(a) the residual-based and (b) the gradient-based scheme.
127
5. Adaptive Node Refinement
(a) (b)
Figure 5.7.: The estimation of the global error g for (a) the residual-based and (b) the
gradient-based method.
estimator g is nevertheless able to show the evolution of the eigenvalue error, even
though the estimations are on a different scale of magnitude. An application of this
error estimator as a stopping criterion is acceptable, since the characteristics of the error
can be resolved. Alternatively, the residual-based error estimator could be applied to
calculate the global estimated error.
128
5.4. Conclusion
(a) (b)
Figure 5.8.: Convergence rate of the relative eigenvalue errors on the cylindrical unit disc
for (a) the T M01 mode and (b) the T M02 mode. The dashed lines represent
the two node refinement algorithms and the solid line stands for the uniform
refinement as a comparison.
5.4. Conclusion
In this chapter, iterative adaptive node refinement strategies for the second-order RPIM
eigenvalue solver were presented. The general procedure is as follows: An initial solution
is calculated on a coarse set of collocation nodes. On a second set of test nodes, the local
errors are estimated. The test nodes that yield errors larger than a preset threshold are
subsequently added to the set of collocation nodes. A global error threshold can be used
as stopping criterion of the iteration.
The first error estimation algorithm is based on the residual of the Laplace problem
on the internal test nodes and the residual of the Dirichlet condition on the boundary
nodes. A weighting factor is used to bring the two residuals to the same scale. In a second
129
5. Adaptive Node Refinement
algorithm, a gradient-based scheme was introduced. There, the jump of the gradient
between two collocation nodes is used as an indicator of the speed of the field variations.
A larger jump indicates regions in nodes densities that do not sufficiently resolve the
field variations. In this algorithm, no weighting factors are required.
In various numerical examples, the distribution of the error functions were evaluated
and the influence of the threshold parameter on the iterative refinement was investigated.
Moreover, guidelines on the choice of the threshold parameter were developed. The two
refinement strategies were compared to a uniform refinement, where the node density was
increased step by step. Both refinement algorithms showed a clearly improved conver-
gence behaviour compared with a uniform refinement strategy. For practical simulations
this means that a problem can be solved more accurately with given limited computa-
tional resources and the global estimated errors can be used as a reliable indicator of the
quality of a solution.
130
6. Numerical Examples
Abstract In this chapter the previously introduced frameworks are illustrated with numerical
examples in a two-dimensional setting. For the first-order time-domain RPIM scheme, two practical
examples are simulated. For the implementation in Cartesian coordinates an H-plane waveguide
bend is simulated and the standing wave ratio (SWR) at the input is computed. Perfectly
matched layers are applied to truncate the computational domain. The advantage of conformal
node placement is illustrated in a comparison with a Cartesian rectangular grid. For cylindrical
coordinates, a corrugated copper cable is simulated and the results are compared to a commercial
available numerical solver. For the second-order RPIM scheme, a comparison of the presented
eigenvalue solver with a high-order discontinuous Galerkin code and two finite-element solvers is
performed. Lastly, the adaptive node refinement algorithm is applied to a rectangular and a L-
shaped domain. The evolution of the node distribution through the adaptivity iteration is illustrated.
6.1.1. Waveguide-Bend
Setup
The physical properties of the chosen waveguide are depicted in Fig. 6.1. The waveguide
WR-229 operates in a frequency band from 3.22 GHz to 4.90 GHz with a thickness of
a = 5.82 cm. The bend radius is chosen as r = 7.62 cm.
In order to evaluate the performance of the first-order RPIM time-domain scheme, a
conformal node distribution is applied using a constant node spacing in the r- and the
131
6. Numerical Examples
Figure 6.1.: Physical geometry of the 90 H-plane waveguide bend. The model is fed by
a Gaussian broadband pulse and truncated by a perfectly matched layer.
-direction of the bend (Fig. 6.2b). Before and after the bend, a regular rectangular
node distribution is applied. Using such a node placement allows correct modeling of the
curved perfect electric conductor (PEC) at the waveguide boundaries. This conformal
nodal distribution is compared to a fully regular rectangular node distribution (Fig. 6.2a).
The local support domain of the basis functions has been set to include at least 4
nodes. Linear monomial basis functions (M = 3) are used and the shape parameter
is fixed to c = 0.6. The computational domain is truncated by a perfectly matched
layer as described in Sec. 4.2. The source is modeled through a sine-modulated Gaussian
pulse (3.24) that excites a T E1 mode in the desired frequency range from 3.22 GHz to
4.9 GHz.
Both the solutions in the Cartesian and conformal node distributions are expected to
converge with increased node density. For the conformal model reliable results are an-
ticipated for coarser discretizations already. For regular node distributions stair-casing
effects will pose problems for coarse discretizations since the geometry is not resolved
accurately. Nevertheless comparable results are expected for very fine discretization
where the physical geometry is properly approximated. The voltage standing wave ratio
(VSWR) is computed by recording the electric field values at a sensor location. The
sensor node is placed at a distance of h = 7.14 cm from the bend as shown in Fig. 6.1.
The reflected fields are obtained by subtracting the numerical solutions from a reference
solution provided by a very long straight waveguide in positive y-direction. The VSWR is
computed through the relationship (4.27).Simulations were performed for discretizations
[/8, /17, /26, /34, /43, /51] where corresponds to the free-space wavelength
of the highest operation frequency.
Results
Fig. 6.3a shows the convergence for the regular node distribution. Due to stair-casing
effects, a very fine discretization of at least /43 is required to achieve reliable simulation
results. The results in Fig. 6.3b on the other hand converge much faster. For discretiza-
132
6.1. First-Order Time-Domain Simulations
(a) (b)
Figure 6.2.: Comparison of (a) regular and (b) conformal node placement to model the
waveguide bend.
tions finer than /17 the solution has already reached a reasonably stable state.
These results confirm that the ability for arbitrary node placement significantly in-
creases the simulation accuracy over the usage of conventional rectangular grids. The
VSWR converges much faster in the case of conformal node placement compared to the
results in the rectangular grid where stair-casing effects impair the simulation accuracy.
Setup
Fig. 6.4 shows the physical geometries of the plain coaxial cable acting as a reference
solution (Fig. 6.4 a)), and of a cable with a corrugated part (b). The diameter of the
outer conductor is chosen as Douter = 14.28 cm and of the inner conductor as dinner =
7.14 cm, respectively. These dimensions relate to commercially available products. The
corrugations are modeled as sinusoidal functions with the amplitude varied between
0.0002% and 4% of the thickness of the dielectric (Douter dinner )/2. In the model of
the corrugated copper cable Fig. 6.4b), the shape variations are applied over a length of
lC = 14.28 cm and include 4.5 sinusoidal periods.
A reference solution is computed on the plain model Fig. 6.4a. For the internal media
free-space material properties are chosen. Using a staggered node distribution, the models
are discretized with an average node density of /28. For the reference model, a regular
node arrangement is optimal. In the vicinity of the sinusoidal corrugations of model
Fig. 6.4b, unstructured distributions are generated to resolve the geometry. The inner
and outer conductors are modeled as perfect electric conductors through mirror nodes
as presented in (3.29). The Gaussians basis functions (2.2) are fixed with the global
133
6. Numerical Examples
(a)
(b)
Figure 6.3.: Convergence for the VSWR in rectangular (a) and conformal grid (b) for
increasingly fine node distributions.
shape parameter c = 0.6. Linear monomial basis functions (M = 3) are selected and
the support domain size s has been chosen to have at least nAs = 6 nodes inside each
support domain.
The coaxial cables are operating in the single-mode regime below the cutoff frequency
of the first T M1 mode at fc = 4.2 GHz. Hence, a simulation frequency of f = 3 GHz
is chosen to ensure that only the transverse electromagnetic mode can propagate. The
134
6.1. First-Order Time-Domain Simulations
a) coaxial line b)
(PEC) dA lC
Douter
dinner
A
z
Figure 6.4.: Schematic of (a) undisturbed and (b) corrugated coaxial cable
Douter
H () = sin(2f t) (6.1)
2
Results
Numerical reflections from the corrugations are computed through
Er
E = (6.2)
Ei
where Er and Ei are the reflected and incident energy flowing through a cable cross-
section A in Fig. 6.4b) over one period. The plane A is placed at dA = 10 cm before the
start of the corrugations. This is sufficiently far to ensure that evanescent higher-order
modes have decayed enough and do not affect the computations.
The energy flow integrated over one period is defined through the Poynting vector Sav
as
Z 2 Z Douter
1 T
2
Z
E= E (, t) H (, t) dt d d (6.3)
0
dinner T
2 | 0 {z }
Sav
E (, t) and H (, t) are the field values at time t and position in the plane A. The field
distribution for the reflected wave is obtained by subtracting the numerical reference so-
lution from the field pattern in the corrugated cable. The incident energy Ei corresponds
to the reference solution.
Fig. 6.5 shows the numerical results of the reflections for various sizes of the sinu-
soidal corrugations. A comparison with the finite element method package COMSOL
in frequency-domain is provided. In COMSOL, the model is terminated with a port
that directly extracts the scattering parameters. The relationship between the reflected
energy (6.2) and the scattering parameter S11 is
E = |S11 |2 . (6.4)
135
6. Numerical Examples
Figure 6.5.: Comparison of the reflection coefficient E computed with RPIM and COM-
SOL.
A very good agreement between the two simulation results can be observed. The numer-
ical reflections show a linear dependence on the amplitude of the corrugation in a log-log
scale. For large corrugations, these reflections amount to a considerable magnitude. In
commercially available cables, the corrugations have amplitudes in the range of 4% of the
thickness of the dielectric, leading to realistic reflections of up to 26 dB. As expected,
the reflection coefficient converges to zero as the size of the corrugations vanishes.
To study the behavior of the electromagnetic fields in the presence of the corrugations,
the amplitude and phase of a harmonic wave at 3 GHz is analyzed. The amplitude of the
corrugations for this investigation represents 2% of the thickness of the dielectric. Fig. 6.6
shows the amplitude and the phase of the standing wave pattern at the corrugation for
the total field including incident and reflected wave. The standing wave ratio (SWR)
in Fig. 6.6a) before the corrugations (section A1 ) characterizes the reflection coefficient
E (4.26). The maximal amplitude in section A3 after the corrugations is smaller than
the reference amplitude, which is consistent with the fact a part of the energy is reflected.
Fig. 6.6b) shows the phase difference between the plain cable reference model and
the corrugated model. The phase variations in section P1 are again consistent with the
standing wave pattern. The phase difference in section P3 is negative. This shows a
retardation caused by a slightly increased characteristic impedance in the corrugated
segment P2 .
136
6.2. Second-Order Eigenvalue Simulations
(a)
(b)
(c)
Figure 6.6.: Amplitude variation (a) and phase difference (b) of the corrugated test model
for the corrugations shown in (c).
age.
137
6. Numerical Examples
138
6.2. Second-Order Eigenvalue Simulations
(a) (b)
Figure 6.7.: Convergence for increased node densities with uniform refinement (a). Com-
parison between RPIM, first-order h-FEM and high-order discontinuous
Galerkin finite-element method using curvilinear boundaries. For the first
eigenvalue, the convergence behavior for an adaptively refined solution (rect-
angular box) is shown as a comparison to the uniform refinements (cross).
The two eigenfunctions are shown in (b) for reference.
Table 6.1.: Number of degrees of freedom required to achieve an error below 105 for
the first and fifth eigenvalue (EV).
method is provided where both solutions are obtained with a full matrix solver. There
the computation times are comparable and the RPIM algorithm is generally faster.
139
6. Numerical Examples
0
exponential reduced rate
10
RPIM
DG-FEM
h-FEM
-2
10
Rel. Eigenvalue Error
-4
10
-6
10
-8
10 -3 -2 -1 0 1
10 10 10 10 10
Figure 6.8.: Computation times required to achieve a certain accuracy for the three dif-
ferent methods. A comparison is shown between the RPIM scheme with uni-
form refinement, a high-order discontinuous Galerkin finite-element method
with curvilinear elements and polynomial refinement and a first-order contin-
uous finite-element method with uniform mesh refinement. The region with
exponential convergence for the RPIM algorithm is marked with a solid line.
The computation time of the fifth eigenmode for the discontinuous Galerkin
method with a full matrix solver (marked with a cross) is provided for a
direct comparison.
eigenmodes of an L-shaped domain, both enclosed with perfect electric conductors, are
computed. Both problems have smooth non-singular sinusoidal solutions. Starting with
a coarse node distribution, the presented adaptive algorithm is applied to increase the
accuracy. The chosen parameters for the weighting of the internal and boundary test
nodes (5.3) are 1 = 1, 2 = 102 . However, the algorithm is robust for a large range of
the weighting factor 2 .
The two computational domains with node distributions at different times of the
adaptive iteration (5.4) with = 0.1 are displayed in Fig. 6.9. In the case of the square
cavity, the adapted node distributions are illustrated for the T M11 eigenmode. In the
case of the L-shaped domain, the adapted mode shown is the first smooth eigenmode.
In both cases, the final node distributions show that the added nodes tend to cluster
close to the boundary. This correlates with the findings in [47] where it was observed
that computations using RBFs become more accurate when increasing the node density
close to to boundaries.
140
6.2. Second-Order Eigenvalue Simulations
(a)
(b)
(c)
Figure 6.9.: Node distributions at different stages of the refinement iteration for the
square domain with T M11 mode (left) and L-shaped domain for the 1st
smooth eigenmode (right). The solutions are shown for (a) the initial dis-
tribution, (b) after one iteration and (c) final node distribution after six
iterations.
The eigenvalues of the first six distinct modes for the square domain are compared to
the analytical values from (B.3) and the first two smooth eigenmodes for the L-shaped
domain are compared to benchmark solutions [148] summarized in Tab. B.1. In Fig. 6.10
the error is plotted against the number of degrees of freedom (ndof ). The relative error
of the eigenvalues decreases rapidly by a factor of at least 103 for doubling the number
of nodes. For all computed modes, the eigenvalue error decreased below at least 106
after less than six iterations.
141
6. Numerical Examples
(a) (b)
Figure 6.10.: Convergence of several eigenvalues of a square (a) and L-shaped domain
(b).
142
7. Conclusion
7.1. Summary
In this thesis the meshless radial point interpolation method (RPIM) was implemented
for the numerical solution of electromagnetic problems. This collocation method dis-
cretizes a physical geometry by an arbitrary node distribution. In contrast to classical
mesh-based methods, this has the advantage of simple modeling of complex structures by
conformal node allocation along the geometrical features of a physical model. Also, the
method might be combined with other physical differential equations without extensive
effort. The radial basis functions have very good approximation properties compared
to traditionally utilized polynomial basis functions. In the RPIM scheme, a new set
of shape functions is computed based on the radial basis functions. These shape func-
tions approximate field components and spatial derivatives based on the values on the
surrounding nodes. This formulation allows explicit time-stepping schemes. Due to the
simplicity of a node-based collocation method with radially dependent basis functions,
the RPIM algorithm can be implemented for different forms of Maxwells equations in
a very versatile framework. A highly accurate scheme involving global basis functions
in the frequency domain and a more efficient extension using local support domains for
the time-domain implementation was derived. The advantages of both schemes were se-
lectively exploited forthe integration of absorbing boundary conditions to terminate the
computational domain in open problems. For the localized scheme perfectly matched lay-
ers have been implemented, and for the global scheme high-order non-reflecting boundary
conditions have been developed. The simplicity of free node placement was exploited in
a node adaptation scheme. This clever approach showed a much faster decrease of the
numerical errors in comparison with a naive uniform refinement. Comparisons to other
state-of-the-art methods showed much faster convergence rates and higher accuracy even
at very low node densities. The meshless radial point interpolation method for electro-
magnetics showed bright prospects as a versatile and highly accurate numerical scheme
in computational electromagnetics.
In Chap. 2, the theoretical background of interpolation by the RPIM scheme was
reviewed. The derivation of the shape functions has been first presented for global ba-
sis functions. In a point-matching procedure, interpolation coefficients were computed
through a matrix inversion. The high accuracy of this formulation comes at the cost of
full matrices which are inefficient for numerical solvers in large problems. A more effi-
cient formulation has been achieved by assuming a local support domain for each node.
Only nodes inside this compact domain are incorporated into the equations. Instead of
one large system, many small matrices are inverted very efficiently. This efficiency comes
as a trade-off with numerical accuracy. Numerical experiments were performed to give
143
7. Conclusion
guidelines on the size of the support domain to nevertheless achieve acceptable accuracy.
It is found that the approximation error is decreased for increasingly flat basis functions,
however at the cost of very high matrix condition numbers. Limitations of numerical
solvers constrain the achievable interpolation quality. A good choice of the flatness of
the basis functions is crucial for reliable results. Thus, an optimizing algorithm was pre-
sented that minimizes the approximation error based either on a test function or on a
residual error.
In Chap. 3 the application of these shape functions was presented to first- and second-
order formulations of the Maxwells equations in two dimensions. In a first-order scheme,
a time-domain solver with localized RPIM shape functions was introduced that, when ap-
plied in a Cartesian node distribution, shares striking similarities to the FDTD method.
The electric and magnetic field components are stored in a staggered node arrangement.
In contrast to the FDTD method based on rectangular grids, this scheme brings the
additional flexibility of free node placement for modeling complex geometries. A numer-
ical solution is found through a time-marching iteration in the time domain. Important
material boundary conditions, source terms and a criterion for stable iteration have been
derived. Based on the time-domain formulation, an eigenvalue solver was developed to
investigate the spectral properties of the method. Investigations on the eigenvalue dis-
tribution have shown that long-time stability can be an issue depending on the flatness
of the basis functions. The choice of the parameter controlling the flatness is a trade-
off between stability over very long time iterations and accuracy. The RPIM scheme
was furthermore developed for the second-order wave equation for stationary problems.
This scheme shows practical similarities to advanced high-order finite-element methods
in terms of accuracy and applicability. In a first step, an eigenvalue solver with global
radial basis functions for Laplace problems has been introduced. The numerical scheme
was compared to preexisting radial basis function methods, and striking similarities in
the numerical results were observed despite significantly different formulations. This
second-order RPIM eigenvalue scheme has an identity mass matrix, which is advanta-
geous for numerical solvers and implementations in time-domain. Again, the spectral
properties were investigated and it was observed that the RPIM scheme always gave
reliable results in these numerical examples in contrast to other standard formulations.
In a second step, a second-order frequency-domain solver for source problems was intro-
duced. The Helmholtz equations were discretized using global basis functions. Different
boundary conditions and a strategy to efficiently solve large problems through a domain-
decomposition method were discussed.
Based on the previously developed frameworks, absorbing boundary conditions for out-
going waves were developed in Chap. 4. For the first-order system, a perfectly-matched
layer based on an artificial anisotropic material model has been implemented. This has
the advantage of efficient absorption of waves from nearly all angles of incidence (ex-
cept grazing incidence). An investigation on different tapering functions of the loss term
showed that large step sizes can lead to numerical reflections. Guidelines for a reliable
choices of parameters were developed in numerical experiments. That also confirmed the
effectiveness of these absorbing boundary conditions. For the second-order frequency-
domain framework, high-order non-reflecting boundary conditions were developed. In
this type of boundary conditions, for every order, plane waves incident at one given
144
7.1. Summary
angle are perfectly absorbed. Higher orders of the boundary condition yield in theory
lower reflections but require high-order spatial derivatives. The accuracy of the deriva-
tive can become degraded with increasing order. The RPIM scheme using global basis
functions exhibits high accuracy for high-order spatial derivatives. Those derivatives can
be directly applied to the absorbing boundary conditions. A global optimization scheme
was introduced to find optimal sets of absorbing incidence angles for different orders of
the absorbing boundary condition. It was found that the numerical errors of the absorb-
ing boundary condition could be reduced by an increasingly denser Chebyshev-like node
distribution close the absorbing boundary. In numerical experiments the high-order ab-
sorbing boundary conditions showed a performance very close the theoretically expected
behavior and proved to yield more accurate results than a theoretical PML model.
The advantages of the RPIM scheme include simplified node placement, which suggests
the application of adaptive refinement strategies. This was introduced for the second-
order eigenvalue solver in Chap. 5. In an iterative process, an a posteriori error is esti-
mated on a set of test nodes, based on a previous solution. In this thesis, a residual-based
estimator that computes the error of the physical equations, and a gradient-based esti-
mator that identifies areas exhibiting large jumps in the gradient have been introduced.
Test nodes with estimated errors above a selected threshold are added to the set of col-
location nodes. Then, the flatness of the basis functions is optimized and a new solution
is calculated. This process is repeated until a sufficiently accurate numerical solution
is calculated. Guidelines for the simulation parameters have been developed and the
two error estimators have been compared in numerical experiments. Even though both
estimators operate on very different principles, similar results were achieved.
In Chap. 6, several numerical examples were presented for the first-and second-order
RPIM frameworks. In time-domain simulations, two practical illustrations have shown
the advantages of the RPIM formulation. For a 90 waveguide bend, the benefits of
conformal node placement along the circular bend were illustrated in a comparison to a
regular node distribution. Much faster convergence has been observed. In another numer-
ical example a corrugated copper cable was simulated. The physical reflections caused
by sinusoidal corrugations of the inner and outer conductor were studied. A comparison
with a commercial finite-element solver showed very good agreement. For the second-
order eigenvalue solver, the scheme was compared to different finite-element packages. It
has been established that the RPIM scheme using global basis computes the eigenpairs
of the Laplacian for a test case much more accurately and at a lower computational
cost than a high-order discontinuous Galerkin method for a given discretization. This
statement is even more valid for comparisons to standard low-order finite-element meth-
ods in commercial solvers. The adaptive refinement scheme has been further illustrated
in two last numerical examples. The iterative process has been highlighted through an
investigation of the refined node distribution with the corresponding field plots. The
accuracy in the calculation of several eigenmodes of a square and L-shaped domain was
significantly improved through a residual-based refinement.
145
7. Conclusion
7.2. Discussion
This has thesis built the foundation for meshless RPIM schemes in electromagnetics. It
has been shown in different frameworks, that the flexibility of free node placement can
be highly advantageous for accurate modeling of complex structures. Adaptive node re-
finement algorithms can exploit the simplicity of unconnected node distributions. With-
out large efforts, the accuracy can be increased by making the discretization finer at
places where necessary. For the time-domain implementation, the introduction of per-
fectly matched layers has allowed the simulation of open problems. Another kind of ab-
sorbing boundary conditions has been implemented in the frequency-domain formulation
using global basis functions. Due to the high accuracy of the interpolation of higher-order
spatial derivatives, high-order absorbing boundary conditions have be implemented at
theoretically additional computational cost. Only a slightly increased node density close
to the absorbing boundary was necessary to achieve results superior to the classical
perfectly-matched layer model. This approach not been previously implemented in other
methods and is particularly well suited for the simulation of open problems because no
additional variables are introduced.
This work is meant to be a starting point for continued research on the RPIM scheme in
electromagnetics. Two main directions seem feasible. First, with to the local formulation
of the time-domain method a very efficient method is available that is able to outperform
classical time-domain solvers. The accurate implementation of material discontinuities
and nonlinear media is a natural next step. Second, the frequency-domain solver has
the advantage of highly accurate modeling. The advantages combined with a reliable
domain-decomposition algorithm will allow for the simulation of large problems at small
computational cost.
Mesh-based methods are usually developed in complex frameworks that deal with
the mesh topologies and data structures. The programming is often cumbersome and
complicated. In contrast, new concepts can be implemented in this meshless formulation
with a very low effort in programming, which allows to evaluate new principles at much
higher speed. This was partly the reason why several formulations of the RPIM scheme,
adaptivity strategies and absorbing boundary conditions could be initiated and evaluated
in this thesis. Also, recent publications by several research groups showed a very fast
progress on many aspects of RPIM in electromagnetics.
In todays rich environment of methods in computational electromagnetics, commer-
cial solvers provide a streamlined simulation process, which guides a user through the
technical drawing of a model, setting up of the simulation parameters and postpro-
cessing of the data. Often parallel solvers are included. For classical applications, this
convenience is a reason for a decision for many engineers. Having said that, for structures
in nano-physics and very complex geometries, these programs are not sufficient anymore.
This is the place where novel methods such as the meshless RPIM scheme fill a valuable
niche. More effort is required to challenge the established methods, but the algorithms
introduced in this thesis - in a proof of concept sense - can be the origin of a successful
class of new methods for broad applications.
146
7.3. Future Work
147
7. Conclusion
consist of the same node positions. Further away the RPIM node distribution grad-
ually moving towards unstructured node locations. Furthermore, a hybridization
with the discontinuous Galerkin method in time-domain would be highly advanta-
geous - it would allow to incorporate the flexibility of the RPIM scheme into this
well-established method.
In conclusion, due to all the advantages associated with the meshless formulation the
radial point interpolation method has great potential as a highly accurate and flexible
domain discretization method in computational electromagnetics.
148
A. Mathematical Notation
In this section, the mathematical notations are summarized. For partial derivatives, the
following simplified terminology was applied
= (A.1)
for spatial coordinates = x, y and the time variable = t.
In the update equations (3.17) and (3.18), the summation runs over all nAs nodes in
the computational domain. The field components at node position xi and at time step
n are designated as
(n)
Ez,i = Ez (xi , n). (A.2)
For the approximation of the spatial derivatives of the fields in = x, y or = r, z
direction, the shape functions centered at H-node location xi were applied in (3.17a)
and (3.17b) as
n
Ei = [ 1Ei , . . . , EAi s ]. (A.3)
The spatial derivative is calculated through the linear combination with the surrounding
E-neighbors at positions xj .
For the E-nodes, the shape function is centered at the E-node position xi is denomi-
nated as
n
Ei = [ 1Ei , . . . , EAi s ].jH,i . (A.4)
The elements of the vector are multiplied in (3.17c) with the surrounding H-nodes at
positions xi .
149
B. Solutions to the Numerical
Benchmark Tests
B.1. Circular Disc
The analytical eigenvalues for the circular domain with radius R (Fig. B.1) are based on
m,n
the resonance frequency fm,n = 2 of order m, n [97] when enclosed by a PEC:
1 pnm
fm,n = (B.1)
2 R
where pmn denotes the nth zero of the mth order Bessel function Jm (pmn ) = 0.
The TM-eigenmodes of the scalar electric field component Ez are expressed as
pmn
Ez (, ) = E0 Jm ( ) cos(m). (B.2)
R
151
B. Solutions to the Numerical Benchmark Tests
152
B.3. L-shaped Domain
Eigenvalue # Eigenvalue #
1 9.6397238445 7 44.948488
2 15.19725192 8 49.34802201
3 19.73920880 9 49.34802201
4 29.5214811 10 56.709610
5 31.9126360 11 65.376535
6 41.4745099 12 71.057755
Table B.1.: Benchmark results for numerical eigenvalues of the L-shaped domain.
153
C. Matrix Entries for Second-Order
Eigenvalue Problems
Tn the following, the matrix entries are summarized for the Non-Symmetric and Sym-
metric Kansa second-order eigenvalue solver. The generalized eigenvalue problem is
A0 0 = B0 0 . (C.1)
155
C. Matrix Entries for Second-Order Eigenvalue Problems
i = [NI + 1, NB ]
[A
1 ]i,j = (xi xj )rj (xi ), (C.5d)
j = [NI + 1, NB ]
i = [1, NI ]
[B1 ]i,j = rj (xi ), (C.5e)
j = [1, NI ]
i = [NI + 1, NB ]
[B
1 ]i,j = (xi xj )rj (xi ), (C.5f)
j = [1, NB ]
The Laplacian is applied on the interior collocation nodes for a symmetric scheme. The
generalized eigenvalue problem is stated as
h i
1
A01 = AL L
1 A1 A1 A1 , (C.7a)
h i
1
B01 = B1 B1 A1 A1 (C.7b)
156
C.2. Symmetric Kansa Method
xn
i = [1, NI ]
[BL
2 ]i,j = xn rj (xi ), (C.8e)
j = [1, NI ]
i = [NI + 1, NB ]
[B2 ]i,j = rj (xi ), . (C.8f)
j = [1, NB ]
157
Bibliography
[1] R. Kress, Linear Integral Equations, ser. Applied Mathematical Siences. New
York, NY, USA: Springer Verlag, 1989.
[2] R. Harrington, Field computation by moment methods. Wiley-IEEE Press, 1993.
[3] L. L. Thompson, A review of finite-element methods for time-harmonic acoustics,
The Journal of the Acoustical Society of America, vol. 119, no. 3, pp. 13151330,
2006.
[4] A. Taflove and S. C. Hagness, Computational Electrodynamics: The Finite-
Difference Time-Domain Method, 3rd ed. Norwood, MA, USA: Artech House,
Inc., June 2005.
[5] K. Yee, Numerical solution of initial boundary value problems involving maxwells
equations in isotropic media, IEEE Trans. Antennas and Propagation, vol. 14,
no. 3, pp. 302307, May 1966.
[6] W. Yu, R. Mittra, T. Su, Y. Liu, and X. Yang, Parallel Finite-Difference Time-
Domain Method. Norwood, MA, USA: Artech House, Inc., 2006.
[7] M. Okoniewski, E. Okoniewska, and M. Stuchly, Three-dimensional subgridding
algorithm for FDTD, IEEE Trans. Antennas and Propagation, vol. 45, no. 3, pp.
422429, Mar. 1997.
[8] T. Weiland, Time domain electromagnetic field computation with finite-difference
methods, International Journal of Numerical Modelling: Electronic Networks, De-
vices and Fields, vol. 9, no. 4, pp. 295319, 1996.
[9] S. Rao, Time domain electromagnetics. Academic Press, 1999.
[10] J. Jin, The finite element method in electromagnetics. New York, USA: John
Wiley & Sons, 2002.
[11] I. Babuska and M. Suri, The p and h-p Versions of the Finite Element Method,
Basic Principles and Properties, SIAM Review, vol. 36, no. 4, pp. 578632, 1994.
[12] J. Hesthaven and T. Warburton, Nodal Discontinuous Galerkin Methods: Algo-
rithms, Analysis, and Applications, Springer Texts in Applied Mathematics ed.
New York, NY, USA: Springer Verlag, 2008, vol. 54.
[13] S. Beissel and T. Belytschko, Nodal integration of the element-free Galerkin
method, Computer Methods in Applied Mechanics and Engineering, vol. 139, no.
1-4, pp. 4974, 1996.
159
Bibliography
[15] S. N. Atluri and T. Zhu, A new Meshless Local Petrov-Galerkin (MLPG) ap-
proach in computational mechanics, Computational Mechanics, vol. 22, pp. 117
127, 1998.
[16] S. Atluri and S. Shen, The meshless local Petrov-Galerkin (MLPG) method, ser.
Contemporary research on emerging sciences and technology. Tech Science Press,
2002.
[20] X. Zhang, X.-H. Liu, K.-Z. Song, and M.-W. Lu, Least-squares collocation mesh-
less method, International Journal for Numerical Methods in Engineering, vol. 51,
no. 9, pp. 10891100, 2001.
[23] T. Fries and H. Matthies, Classification and overview of meshfree methods, Tech-
nische Universitat Braunschweig, Braunschweig, Tech. Rep. 2003-3, 2003.
160
Bibliography
[33] E. Larsson and B. Fornberg, A numerical study of some radial basis function based
solution methods for elliptic PDEs, Computers & Mathematics with Applications,
vol. 46, no. 5-6, pp. 891902, 2003.
[35] G. R. Liu and Y. T. Gu, An Introduction to Meshfree Methods and their Program-
ming. Dortrecht, The Netherlands: Springer, 2005.
[36] G. Liu, Mesh Free Methods: Moving Beyond the Finite Element Method. Boca
Raton, FL, USA: CRC Press, 2003.
[37] W. Madych, Miscellaneous error bounds for multiquadric and related interpola-
tors, Computers & Mathematics with Applications, vol. 24, no. 12, pp. 121138,
1992.
[38] G. E. Fasshauer and J. Zhang, On choosing optimal shape parameters for RBF
approximation, Numerical Algorithms, vol. 45, no. 1, pp. 345368, August 2007.
161
Bibliography
[41] G. R. Liu and Y. T. Gu, A Local Radial Point Interpolation Method (LRPIM)
for Free Vibration Analyses of 2-D Solids, Journal of Sound and Vibration, vol.
246, no. 1, pp. 2946, Sept 2001.
[42] J. Li and Y. C. Hon, Domain decomposition for radial basis meshless methods,
Numerical Methods for Partial Differential Equations, vol. 20, no. 3, pp. 450462,
2004.
[44] P. Yee and S. Haykin, Regularized radial basis function networks: Theory and ap-
plications. New York, NY, USA: John Wiley & Sons, Inc., 2001.
[45] M. Kindelan and F. Bernal, Radial Basis Function (RBF) Solution of the Motz
Problem, Progress in Industrial Mathematics at ECMI 2008, pp. 907912, 2010.
[51] B. Fornberg and C. Piret, A stable algorithm for flat radial basis functions on a
sphere, SIAM Journal on Scientific Computing, vol. 30, pp. 6080, 2007.
[52] P.-L. Jiang, S.-Q. Li, and C. H. Chan, Analysis of elliptical waveguides by a
meshless collocation method with the Wendland radial basis functions, Microwave
and Optical Technology Letters, vol. 32, no. 2, pp. 162165, 2002.
162
Bibliography
[55] R. Beatson, J. Cherrie, and C. Mouat, Fast fitting of radial basis functions: Meth-
ods based on preconditioned GMRES iteration, Advances in Computational Math-
ematics, vol. 11, no. 2, pp. 253270, 1999.
[57] B. Fornberg, E. Larsson, and N. Flyer, Stable Computations with Gaussian Radial
Basis Functions in 2-D, Dept. of Information Technology, Uppsala University,
Uppsala, Sweden, Tech. Rep. 2009-020, 2009.
[59] Y. C. Hon, R. Schaback, and X. Zhou, Adaptive Greedy Algorithm for Solving
Large RBF Collocation Problems, Numerical Algorithms, vol. 32, pp. 1325, 2003.
[60] R. Schaback and H. Wendland, Adaptive greedy techniques for approximate so-
lution of large RBF systems, Numerical Algorithms, vol. 24, no. 3, pp. 239254,
2000.
[61] S. Rippa, An algorithm for selecting a good value for the parameter c in radial
basis function interpolation, Advances in Computational Mathematics, vol. 11,
pp. 193210, 1999.
[64] R. Brent, Algorithms for minimization without derivatives. Mineola, NY: Dover
Publications, 2002.
[66] P. S. Jensen, Finite difference techniques for variable grids, Computers & Struc-
tures, vol. 2, no. 1-2, pp. 1729, 1972.
163
Bibliography
[78] M. Ghrist, B. Fornberg, and T. A. Driscoll, Staggered Time Integrators for Wave
Equations, SIAM Journal on Numerical Analysis, vol. 38, no. 3, pp. 718741,
2000.
[79] Y. Yu and Z. D. Chen, An Unconditionally Stable Radial Point Interpolation
Method for Efficient Meshless Modeling in Time Domain, in 25th Annual Review
of Progress in ACES, Monterey, USA, March 2009.
[80] K. Krohne, P. Gi-Ho, and E. Li, Waveguide Mode Computation through
Smoothed Particle Electromagnetics, in 24th Annual Review of Progress in ACES,
Niagara Falls, Canada, 2008, pp. 882887.
164
Bibliography
[82] C. Herault and Y. Marechal, Boundary and interface conditions meshless methods
[for EM field analysis], IEEE Trans. Magnetics, vol. 35, no. 3, pp. 14501453, May
1999.
[84] Y. Yu and Z. D. Chen, Dielectric boundary conditions with the meshless radial
point interpolation method, in 14th International Symposium on Antenna Tech-
nology and Applied Electromagnetics the American Electromagnetics Conference,
jul. 2010, pp. 14.
[87] R. Lehoucq, D. Sorensen, and C. Yang, ARPACK Users Guide: Solution of Large-
Scale Eigenvalue Problems With Implicityly Restarted Arnoldi Methods. Philadel-
phia, PA, USA: Society for Industrial and Applied Mathematics, 1998.
[89] C. D. J. Meyer, Generalized Inverses and Ranks of Block Matrices, SIAM Journal
on Applied Mathematics, vol. 25, no. 4, pp. pp. 597602, 1973.
[90] E. Gjonaj and T. Weiland, A divergence cleaning method for the high order DG-
FEM in the time domain, in International Conference in Electromagnetics on
Advanced Applications, 2010, pp. 581584.
165
Bibliography
166
Bibliography
[107] F. Kang, Finite Element Method and Natural Boundary Reduction, in Proc. of
the International Congress of Mathematicians, Warsaw, Poland, 1983, pp. 1439
1453.
[109] G. J. Fix and S. P. Marin, Variational methods for underwater acoustic problems,
Journal of Computational Physics, vol. 28, no. 2, pp. 253270, 1978.
[111] J.-P. Berenger, A perfectly matched layer for the absorption of electromagnetic
waves, Journal of Computational Physics, vol. 114, no. 2, pp. 185200, 1994.
[115] S. Gedney, An anisotropic perfectly matched layer-absorbing medium for the trun-
cation of FDTD lattices, IEEE Trans. Microwave Theory and Techn., vol. 44,
no. 12, pp. 16301639, December 1996.
[119] J. Roden and S. Gedney, Convolution PML (CPML): An efficient FDTD imple-
mentation of the CFS-PML for arbitrary media, Microwave and Optical Technol-
ogy Letters, vol. 27, no. 5, pp. 334339, December 2000.
167
Bibliography
[126] D. Givoli and B. Neta, High-order non-reflecting boundary scheme for time-
dependent waves, Journal of Computational Physics, vol. 186, no. 1, pp. 2446,
2003.
[128] D. Givoli and J. B. Keller, Special finite elements for use with high-order boundary
conditions, Computer Methods in Applied Mechanics and Engineering, vol. 119,
no. 3-4, pp. 199213, 1994.
168
Bibliography
[133] R. Verfuhrt, A review of a posteriori error estimation and adaptive mesh refinement
techniques, ser. Advances in Numerical Mathematics. Wiley-Teubner, 1996.
[136] B. B. T. Kee, G. R. Liu, G. Y. Zhang, and C. Lu, A residual based error estimator
using radial basis functions, Finite Elements in Analysis and Design, vol. 44, no.
9-10, pp. 631645, 2008.
[137] T. Rabczuk and T. Belytschko, Adaptivity for structured meshfree particle meth-
ods in 2D and 3D, International Journal for Numerical Methods in Engineering,
vol. 63, no. 11, pp. 15591582, 2005.
[138] O. Zienkiewicz and J. Zhu, The superconvergent patch recovery (SPR) and adap-
tive finite element refinement, Computer Methods in Applied Mechanics and En-
gineering, vol. 101, no. 1-3, pp. 207224, 1992.
[139] P. Brito and A. Portugal, Adaptive Collocation Methods for the Solution of Par-
tial Differential Equations, in Innovations in Computing Sciences and Software
Engineering, T. Sobh and K. Elleithy, Eds. Springer Netherlands, 2010, pp. 499
504.
[140] Y. Luo and U. Combe, An adaptivity procedure based on the gradient of strain
energy density and its application in meshless methods, Meshfree methods for
partial differential equations, pp. 267279, 2003.
[142] S. A. Sarra, Adaptive radial basis function methods for time dependent partial
differential equations, Applied Numerical Mathematics, vol. 54, pp. 7994, June
2005.
169
Bibliography
[144] A. Okabe, B. Boots, K. Sugihara, and S. Chiu, Spatial tessellations: Concepts and
applications of Voronoi diagrams. New York, USA: John Wiley & Sons, 2000.
[145] T. Kaufmann, , C. Engstrom, and C. Fumeaux, A Comparison of Three Mesh-
less Algorithms: Radial Point Interpolation, Non-Symmetric and Symmetric Kansa
MethodThe Meshless Radial Point Interpolation Method for Time-Domain Elec-
tromagnetics, in IEEE MTT-S International Microwave Symposium Digest. Bal-
timore, MD, USA: IEEE, June 2011, pp. 11, accepted for publication.
[146] T. Kaufmann, C. Engstrom, and C. Fumeaux, Residual-Based Adaptive Refine-
ment for Meshless Eigenvalue Solvers, in International Conference in Electro-
magnetics on Advanced Applications. Sydney, Australia: IEEE, Sept. 2010, pp.
244247.
[147] C. Engstrom and M. Wang, Complex dispersion relation calculations with the
symmetric interior penalty method, International Journal for Numerical Methods
in Engineering, vol. 84, no. 7, pp. 849863, 2010.
[148] M. Dauge, Benchmark computations for maxwell equations for the approxima-
tion of highly singular solutions, http://perso.univ-rennes1.fr/monique.dauge/
benchmax.html.
170
Acknowledgement
In the completion of my PhD at the Laboratory of Electromagnetic Fields and Microwave
Electronics (IFH), I would like to express my deepest gratitude to the many persons
involved in the success of this endeavor. I had the opportunity to work on a unique
project that taught me many interesting aspects of computational electromagnetics.
The first thank you goes to Prof. Rudiger Vahldieck who due to his tragic illness is
not among us anymore. He took me on the boat after finishing my Master thesis at IFH.
I had all the freedom and means to pursue my own ideas and to present my progress in
many conferences. Unfortunately, he cannot be here today to see the fruit of this work.
I will always keep his openness and his pragmatism in my memory.
As his successor, Prof. Christian Hafner did a great job to ensure the continuation
of my work. If it aint broke, dont fix it. In that sense he allowed me to continue
seamlessly in my project, but was always available for guidance and assistance in times
of trouble. Usually when I passed by his office, it was to ask for money for yet another
trip, institute event or barbecue. No was a reply I never heard from him. I am especially
grateful for his willingness to wrap up this thesis rather quickly and allow me to pursue
new paths.
Prof. Christophe Fumeaux was my initial direct supervisor at IFH. Initially sharing
the same office, he was available at all times for my sometimes rather benign questions
and offered advice in uncountable situations. After his departure for Adelaide to start
his position as Associate Professor we kept in contact and it was him eventually that
allowed my first visit to the University of Adelaide. As it turned out it was only the
first trip of many. He recently received a fellowship grant that lead to a position for me
as lecturer and post-doctoral researcher under his guidance in Adelaide. Also he took
tremendous efforts to allow this document to be finished in time. Additionally, he could
squeeze in attending my defense as co-examiner despite his full schedule. Christophe, I
cannot be grateful enough for all your efforts, ideas and initiatives.
My direct supervisor for the second half of my PhD was Dr. Christian Engstrom. He
brought his own vision and helped me change focus of the project. He persuaded me to
follow his course at the Seminar of Applied Mathematics, which helped me to understand
more theoretical work. This opened new insights and initiated new approaches (which
turned out to be very successful). Also his relentless effort in teaching me the funda-
mentals of the finite-element methods helped me gaining a deeper understanding of the
whole field. During the writing of this document, he allowed me to give him preliminary
versions and he would rapidly give meaningful comments. At one point shortly before the
deadline, he jokingly asked me whether I was planning to kill him - he but gave valuable
feedback nevertheless. Christian, in your career and personal life everything seems to be
on track; I wish the best of luck for the future.
During my PhD I had the opportunity to work with Christoph Bocklin, Thomas Merz
171
Acknowledgement
and Jan Hafner who did their semester theses under my guidance. They all brought their
own ideas into the project and each made valuable contributions. They taught me a lot
about my work and how to guide a project.
Also people not directly involved contributed to my work: all the other inhabitants of
IFH. Too many to name them all, I want to express my extended thanks to a few selected
persons. The technical support I received from Aldo Rossi and Pascal Leuchtmann made
this thesis possible. Pascal also helped in many ways in his position as a lecturer in
electromagnetic field theory. Especially grateful I am to Barbara Schubeck-Wagner, and
later Eva Knobel who are both solid as a rock in stormy weather. With my former
colleague Hannes Grubinger I had uncountable discussions and shared gossip over many
beers. Dirk Baumann was always available for fruitful discussions and provided a lot of
assistance and moral support in the setup of this thesis and the defense. My late office
mates Mengyu Wang and Dimitra Psychogiou were both very special characters that
made my time here very enjoyable. Your helpfulness has always been appreciated and I
wish you both all the best in your own projects.
Also in my personal live many people (too many to name them all) did their share
to sufficiently distract me from my thesis, e.g. during many Furobebiere on Friday
evenings. Also my old friends from Kanti Zofingen gave a different perspective to life
when I needed it. A special thank goes to Lukas Bolliger who has been my closest friend
since we were teenagers - we shared apartments, dancing classes and boozy jollity - and
yet he is losing another friend to a foreign country. I hope we manage to stay in contact
regularly even though the distance is getting larger in the future.
A special thanks goes to my girlfriend Claudia List who I met during my first stay in
Adelaide. After realizing that the distance from Switzerland to Australia might be a little
large, we finally manage to move together onto the same continent. Again, Christophe
was not totally innocent that this could happen. Claudia, I cannot thank you enough
for everything you have done for me and being there when I needed it. Additionally,
I could persuade you to proof-read all my work, which improved the quality of this
thesis significantly.
Last but not least, my family was the pillar I can always rely on. Anneke and Markus
were the perfect hosts in Stockholm and Brussels. I enjoy spending time with you, albeit
the distance will be greater in the future. My parents knew from the start that I would
end up where I am now (although they only told me so afterwards). They made it possible
for me to study at such a good institution and supported me during all my trips to very
diverse places around the world. I am very grateful for that. Even though I did not visit
or call as often as I should have, you stood behind me for every important decision in
my life.
172
List of Publications
Journal papers
Conference papers
173
List of Publications
174
Curriculum Vitae
Personal data
Name: Thomas Kaufmann
Nationality: Swiss
Date of birth: February 8, 1982
E-mail: thomas.kaufmann@ifh.ee.ethz.ch
Professional experience
03/07 present: ETH Zurich, Zurich, Switzerland
Laboratory for Electromagnetic Fields and Microwave Electronics
Research Assistant
06/04 12/04: Invodane Engineering, Toronto, Canada
Internship
08/99 05/06: TK Computing, Zofingen, Switzerland
Build-up of a company in part-time to provide IT solutions and services
to individuals and small businesses.
Education
03/07 present: ETH Zurich, Zurich, Switzerland
Laboratory for Electromagnetic Fields and Microwave Electronics
The Meshless Radial Point Interpolation Method in Electromagnetics.
(PhD thesis)
06/06 12/06: ETH Zurich, Zurich, Switzerland
A Comparison of Two Perfectly Matched Layer Models for Finite-
Volume Time-Domain Method.
(Master thesis)
10/01 05/06: ETH Zurich, Zurich, Switzerland
Master of Science ETH in Electrical Engineering
08/05 12/05: Norges teknisk-naturvitenskapelige universitet (NTNU),
Trondheim, Norway
Design of an Omnidirectional Antenna for Sounding Rockets
(Semester thesis)
08/97 05/01: Kantonsschule Zofingen, Zofingen, Switzerland, High School
Development of a CNC Drilling Machine for Printed Circuit Boards
(Final year project)
175