Calculation of frequency dependent power losses in inductive

systems with litz wire conductors by a coupled numeric

approach
Berechnung von frequenzabhängigen Leistungsverlusten in induktiven
Systemen mit Litzenkabeln mittels eines gekoppelten numerischen
Ansatzes

Der Technischen Fakultät

der Friedrich-Alexander-Universität Erlangen-Nürnberg
zur
Erlangung des Doktorgrades Dr.-Ing.

vorgelegt von
Andreas Helmut Rosskopf
aus Nürnberg
 Als Dissertation genehmigt von der Technischen Fakultät der
Friedrich-Alexander-Universität Erlangen-Nürnberg

Tag der mündlichen Prüfung: 12. März 2018

Vorsitzender des Promotionsorgans: Prof. Dr.-Ing. Reinhard Lerch
Gutachter: Prof. Dr. rer. nat. Lothar Frey
Prof. Dr. rer. nat. Eberhard Bänsch

ii
 Abstract
In power electronic systems, inductive components with litz wire conductors are widely used to
reduce power losses at higher frequencies. Due to hundreds of strands, isolated against each other,
and an appropriate twisting strategy, litz wires enable a significantly higher system performance
in comparison to solid conductors. However, common simulation approaches fail to calculate the
complex superposition of the electromagnetic effects on different geometry levels. Therefore, a
new simulation approach for the calculation of the frequency dependent losses of inductive com-
ponents with litz wires has been developed within this thesis.
This approach bases on the division of the simulation process: The simulation of the system
behavior with regard to the magnetic field distribution is separated from the resistance and loss
calculation on the litz wire level. First, the equivalent external field distribution of stranded and
solid conductors in case of a homogenous current density distribution is simulated by a finite el-
ement simulation of the full system. The magnetic field distribution is calculated based on the
exact position of the conductors and by considering the electric and magnetic material properties
of surrounding parts. Based on these results, the external magnetic field is evaluated and extracted
on a high number of 2D cuts along the length of the conductor. In a second step, these data are
used to calculate the frequency dependent resistance on litz wire level by analytic, numeric, and
measurement-based methods. The most sophisticated approach corresponds to the numeric partial
element equivalent circuit method, which analyzes litz wire conductors with regard to the electro-
magnetic behavior based on their twisting structure. The resulting data are applied to all 2D cuts
and enable the prediction of the frequency dependent winding losses of the complete inductive
component.
The verification of the simulation approach is based on measurements of different experimental
setups corresponding to common problems in the design process of inductive components. The
results demonstrate the accuracy of the new coupled simulation method and the simulations well
agree with the measured data up to 1 MHz for complex air coils. However, ferrite material within
the setups significantly decreases the quality of the simulation results due to additional core losses.
The treatment of the corresponding effects and also the improvement of the accuracy of the numer-
ical models with regard to tightly packed litz wire structures are topics for research in subsequent
work.

iii
 Zusammenfassung
In leistungselektronischen Systemen ist die Nutzung von Litzenkabeln weit verbreitet und er-
möglicht eine Verringerung der Verluste bei höheren Frequenzen. Im Vergleich zu Massivleitern
können derartige Kabel, bestehend aus hunderten gegeneinander isolierter Einzeladern und einer
geeigneten Verdrillungsstrategie, die Leistungsfähigkeit des Gesamtsystems deutlich steigern. Die
gängigen Simulationsansätze scheitern jedoch aktuell daran, die komplexe Überlagerung der elek-
tromagnetischen Effekte in den unterschiedlichen Geometrieebenen aufzulösen. Deswegen wurde
im Rahmen dieser Doktorarbeit ein neuer Simulationsansatz entwickelt, der die Berechnung der
frequenzabhängigen Verluste von induktiven Komponenten mit Litzenkabeln ermöglicht.
Der Ansatz basiert auf der Aufteilung des Simulationsprozesses: Die Simulation der magnetis-
chen Feldverteilung auf Systemebene wird separat von der Widerstands- und Verlustberechnung
auf Litzenebene behandelt. Im ersten Schritt wird mit der Finite-Elemente-Methode das Gesamt-
system simuliert und ausgenutzt, dass die externe Feldverteilung im Litzenkabel identisch mit der
Verteilung bei einem gleichmäßig bestromten Massivleiter ist. Die magnetische Feldverteilung
wird dabei unter Berücksichtigung der genauen Leiterposition und den elektrischen und magnetis-
chen Materialeigenschaften der umgebenden Bauteile berechnet. Basierend auf diesen Ergebnis-
sen werden die externen Felder ausgewertet und auf einer großen Anzahl an 2D-Schnitten entlang
des Leiters extrahiert. Im zweiten Schritt werden diese Daten benutzt, um mit Hilfe von analytis-
chen, numerischen und messungsbasierten Methoden die frequenzabhängigen Widerstände auf
Litzenebene zu berechnen. Die ausgefeilteste Methode ist dabei der numerische Partial-Element-
Equivalent-Circuit-Ansatz, bei dem das elektromagnetische Verhalten von Litzenkabeln rein über
deren Verdrillungsstruktur bestimmt wird. Die daraus gewonnen Ergebnisse werden auf alle 2D-
Schnitte übertragen und ermöglichen die Vorhersage der frequenzabhängigen Windungsverluste
in der kompletten induktiven Komponente.
Die Verifikation des Simulationsansatzes erfolgt mit Hilfe von Messungen an unterschiedlichen
experimentellen Aufbauten, die gängige Fragestellungen aus dem Entwicklungsprozess von in-
duktiven Komponenten widerspiegeln. Die Ergebnisse bestätigen die hohe Genauigkeit des neuen
gekoppelten Simulationsansatzes und zeigen gute Übereinstimmung von Simulation und Messung
für Frequenzen bis 1 MHz bei Luftspulen. Bei Aufbauten mit Ferriten zeigt sich jedoch ein deut-
licher Abfall bei der Qualität der Simulationsergebnisse auf Grund der zusätzlichen Kernverluste.
Die Auflösung der damit verbundenen Effekte in der Simulation und ebenso die Genauigkeit der
numerischen Modelle bei eng gepackten Litzenstrukturen sind Themen weiterer Forschungen.

iv
 Danksagung
Mit der Frage - ’Kannst du das mal simulieren?’ - begann für mich Ende 2012 eine Reise in die
Tiefen der Ingenieurswissenschaften, die in diesem Werk ihren Abschluss fand.

Der besondere Dank gebührt dabei Herrn Professor Frey und Herrn Professor Bänsch für die
Betreuung, die hilfreichen Diskussionen und Ratschläge, sowie die schlussendliche Beurteilung
dieser interdisziplinären Arbeit.

Die Bewältigung der mannigfaltigen Herausforderungen, die diese Dreiheit aus Elektrotechnik, In-
formatik und Mathematik mit sich brachte, wäre jedoch ohne die Hilfe und Unterstützung meiner
beiden Kollegen Eberhard Bär und Christopher Joffe nicht möglich gewesen: Sie begleiteten mich
durch viele Iterationen der Planung, Durchführung und Auswertung von Simulationen und Mes-
sungen und halfen mir mit konstruktiven Ideen, Rückmeldungen und Korrekturen auf dem steini-
gen Weg der wissenschaftlichen Erkenntnis.

Mein herzlicher Dank geht darüber hinaus insbesondere an Clemens Bonse für die Implemen-
tierungen in Python, an Anne Heubner für die Telefonseelsorge bei allen (un)mathematischen
Fragestellungen, an Jürgen Lorenz für die wissenschaftliche Freiheit und das Vertrauen über die
ganzen Jahre und an Christoph Friedrich Bayer, der mich am Institut und darüber hinaus stets un-
terstützt und motiviert hat.

Ich bedanke mich außerdem bei all den Kollegen aus der Simulation, Leistungselektronik und
IT, die mir bei meinen (oft sehr exotischen) Fragen und Problemen mit Rat und Tat zur Seite
gestanden haben.

Zu guter Letzt danke ich meinen Eltern, meinem Bruder und den Helden, sowie meiner Freundin
Andrea für den Rückhalt, die Unterstützung und Motivation.

...und all den anderen auf ihrem schweren Weg: NO SURRENDER!

v
vi
 Contents

1 Introduction 1

2 Physical background and solving approaches 7

2.1 Maxwell’s equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2 Solving approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3 Analytical methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3.1 Solid conductors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.3.2 Stranded conductors . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.3.3 External effect of conductors . . . . . . . . . . . . . . . . . . . . . . . . 19
2.4 Finite Element Method (FEM) . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.4.1 Algorithmic and mathematical basis . . . . . . . . . . . . . . . . . . . . 23
2.4.2 Comparison of FEM and analytical results . . . . . . . . . . . . . . . . . 26
2.5 Partial Element Equivalent Circuit (PEEC) . . . . . . . . . . . . . . . . . . . . . 31
2.5.1 Algorithmic and mathematical basis . . . . . . . . . . . . . . . . . . . . 34
2.5.2 Comparison of PEEC with FEM . . . . . . . . . . . . . . . . . . . . . . 39
2.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

3 Multi-scale algorithm for litz wire systems 43

3.1 Multi-scale approaches in the physical domain . . . . . . . . . . . . . . . . . . . 44
3.2 SlicerPro approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.3 Preparations on CAD level . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.4 Field simulation with FEM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.5 Extraction of magnetic field data . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.6 Calculation of external magnetic field Hext . . . . . . . . . . . . . . . . . . . . . 53
3.7 Litz wire simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.7.1 Analytic approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.7.2 PEEC approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.7.3 Measured factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.8 Realization and coding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
3.9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

4 Experimental verification 75
4.1 Requirements on the experiments . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.2 Definition of experimental setups . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.3 Measurement settings and error estimation . . . . . . . . . . . . . . . . . . . . . 77
4.3.1 Simulation vs. experiment - Choke . . . . . . . . . . . . . . . . . . . . 82

vii
Contents

4.3.2 Simulation vs. experiment - Bundle structure . . . . . . . . . . . . . . . 86

4.3.3 Simulation vs. experiment - Ferrite variation . . . . . . . . . . . . . . . 90
4.3.4 Simulation vs. experiment - Coil system . . . . . . . . . . . . . . . . . . 93
4.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96

5 Summary and outlook 99

A Appendix 103
A.1 Definition of a homogenous current density distribution in a conductor . . . . . . 103
A.2 Export of the spatial magnetic field distribution (Ansys EMag) . . . . . . . . . . 104
A.3 Standard output file of the PEEC method for litz wires . . . . . . . . . . . . . . 105

B List of publications 107

viii
 1
Introduction

Electrical components and systems spread in more and more fields of daily life. Due to a steady
progress in the performance, functionality and robustness, electrical devices became an indis-
pensable helper in all situations. Moreover, additional requirements considering flexibility and
connectivity are driven by the increasing importance of mobility in our society. Along with this
development, the storage, transfer and transformation of electric energy is becoming increasingly
crucial.
Nowadays, for many electronic systems there is a clear trend to inductive instead of conductive
connections within the application and to the grid. One of the pioneering devices in this con-
text was the electric toothbrush several decades ago using inductive power transfer. However, the
scope of applications for such inductive systems is much larger and contains mobile and wearable
devices as well as inductive charging for electric vehicles.
In the last decades the performance of basic components in the domain of power electronics
P o w e r d e n s ity [k W /d m ³]

1 0 0

1 0

1
2 0 0 0 2 0 0 5 2 0 1 0 2 0 1 5
Y e a r

Figure 1.1: Increase of the power density of (DC-DC) power converters at the Fraunhofer IISB
[1].

and accordingly the entire system performance (figure 1.1) showed significant progress. In this
process, the most remarkable development is taking place in the field of semiconductor devices.
Common insulated-gate bipolar transistors (IGBT) enable switching frequencies in the range of

1
CHAPTER 1. Introduction

Figure 1.2: Dominating parameters in an inductive component influencing the power loss
distribution of the three major parts: circuit, core, and winding losses (visualisation not
quantitative).

hundreds of kilohertz or even gigahertz for applications of several kilowatts. In combination with
new materials and improvements in the manufacturing process of capacitors, ferrites, and conduc-
tors, the performance of power electronic devices reaches new limits.
Apart from a sole power increase, the reduction of the installation space and higher efficiency of
the devices became one of the driving forces for research and development. In this regard, the
total power losses belong to the most crucial factors in the design of inductive systems. In gen-
eral, these losses are commonly separated in circuit, core, and winding losses (figure 1.2). The
first component corresponds to all losses generated on the circuit board, such as switching losses
of the transistor or losses in the diodes. Both the core and winding losses occur in the inductive
component - the former in the ferritic core caused by the temporal change of the magnetization,
the latter resulting from the ohmic resistance of the conductor.
In reality, the quantitative determination of the different loss components is very complex, be-
cause most of the losses are transformed into heat. A measured distinction of which part of an
inductive system is heated up by which parasitic effect, is very time consuming and costly, and
nearly impossible in terms of realistic working conditions. Consequently, a targeted optimization
of inductive systems is very challenging, especially in an early development status or in case of
new design concepts.
However, apart from measurements and basic analytical formulas, engineers are increasingly
supported by a variety of simulations covering the electrical, mechanical, and thermal domain.
Considering the separation of the total power losses mentioned above, the different loss compo-
nents require different solving approaches.
In the following, the current status of the simulation methods is summarized for all three cases (cir-
cuit, core, and winding losses) and analyzed with a focus on the basic strengths and weaknesses
with regard to the resolution of the losses for inductive systems.

• The circuit losses are mostly resolved in electronic circuit simulators such as SPICE, PLECS,
or Simplorer. They all have in common, that the circuit topology is transferred to a set of
differential equations. The most challenging task in this context is not the mathematical
evaluation of the equations, but the accurate description of the electrical components. Com-
monly the information provided in the data sheets of the manufacturers does not cover the
entire complexity of the real devices. Due to the strong non-linear behavior of diodes or tran-
sistors, marginal deviations in only one characteristic value yield totally different results for

2
 1.0.

Figure 1.3: Various shapes of litz wires for different electrical, thermal, and mechanical loads [7].

the entire circuit, especially with regard to the time dependent quantities. Consequently, in
terms of circuit losses, substantial research efforts are invested [2], [3] to improve the char-
acterization of the electrical components, while the enhancement of the solving approach
and the basic algorithms has only minor relevance.
• The basic effects of core losses have already been studied one century ago by K. Steinmetz
[4]. The magnetization and demagnetization of ferromagnetic material due to an exposure
to a time dependent magnetic field causes so called hysteresis losses. With the help of the
Steinmetz formula these losses can be approximated for a defined material. In common
commercial programs using the finite element methods (FEM) the Steinmetz approach is
implemented and applied on each element taking into account the local magnetic field [5].
In general, the required Steinmetz coefficients (or relevant curves for calculating them) base
on measurements of a certain geometric shape, commonly a toroid. However, verification
measurements have proven a significant deviation of these coefficients depending on the fre-
quency and the geometry of the parts [6]. Therefore, due to the lack of transfer approaches
from toroidal shape to arbitrary geometries, these measurements are required for all geo-
metric modifications of the ferrites.
Moreover, at increasing frequencies the electrical conductivity of the ferrite material rises.
Even though the absolute values are quite small (0.2 - 2.0 mS ), common FEM simulations
enable the resolution of the resulting eddy currents which lead to additional losses in the
ferrite cores. Similar to the hysteresis losses, an accurate prediction of these effects in the
simulation requires a good characterization of the material properties.
From a numerical point of view, an accurate calculation of the core losses is feasible with
common methods, but it requires detailed information about the material which is very
costly and time intensive in case of non-standard geometries.
• The winding losses (also known as copper losses) include all ohmic losses within the con-
ductor. Based on the geometry and the material of the conductor, the direct current (DC)
losses can be calculated for a given excitation. In case of alternating currents (AC) the cur-
rent density distribution inside the conductor is significantly changed due to induced eddy
currents. First, the alternating current itself yields an increasing current density at the outer
layers of the conductor (skin effect). Second, time varying external magnetic fields induce
eddy currents within the conductor, resulting in an inhomogeneous concentration of the cur-
rent density within the cross section (proximity effect). In fact, these two effects reduce the
effective area of the conductive cross-section and therefore cause the dominating share of
the winding losses in common applications.

3
CHAPTER 1. Introduction

Figure 1.4: Inner structure and associated terms of litz wire conductors.

The effects of skin and proximity losses are well-known for more than a century and an-
alytical formula for simplified solid conductors in 2D belong to the standard repertoire of
electrical engineers. However, the shape of the conductors has changed within the last few
decades. Instead of solid wires, common high frequency applications use conductors with
several hundreds of strands isolated against each other and twisted in a sophisticated way
on varying geometry levels (figure 1.3).
Unfortunately, the terminology of the so called “litz wires” is not yet the same in literature
and industry. Figure 1.4 provides all relevant technical terms used in this thesis based on a
sketch of a litz wire conductor, starting with the entire conductor, the bundle structure, and
the strands as well as the pitch lengths on varying geometry levels.
These special conductors significantly reduce the winding losses, but also complicate their
simulation due to the high geometric complexity and the superposition of varying elec-
tromagnetic effects on different geometry levels (strands, bundle, or conductor). Common
methods use measured factors of the real conductor or try to approximate the stranded struc-
ture by homogenization. For the last few years, first numerical approaches have demon-
strated their potential even for complex litz wire structures based on the partial element
equivalent circuit (PEEC) method [8].
All these approaches have in common that they are limited to the litz wire itself. However,
apart from the requirement for further improvements of these individual methods, the main
challenge results from the usage of these complex litz wire conductors in power electronic
systems: The micron-sized inner structure of the litz wire influences the power losses on the
macroscopic system level.
Up to now, there is no algorithmic approach available, which enables the calculation of such
a multi-scale problem for an entire inductive system such as simple air coils or chokes with
ferrite cores.
It should be noted that this listing does not provide a conclusion of all activities with regard to
research and potential for improvement, but rather demonstrates which bottlenecks prevent simu-
lations in the domain of power electronics from being more precise [9].
At present, the main limitation considering the determination of the circuit and core losses is given
by the insufficient characterization of the electrical components and of the magnetic materials. If
these input data are well known the simulation algorithms would achieve a sufficiently high accu-
racy in both cases. In contrast to that, the calculation of the winding losses primarily fails due to
the lack of a suitable simulation methodology. Even if there is an accurate characterization of a
litz wire conductor provided, no strategy or simulation workflow would be available to use these

4
 1.0.

data for the determination of winding losses in a power electronic device or system.
Consequently, the scope of this thesis covers the design and implementation of a simulation work-
flow, enabling the calculation of winding losses in inductive systems with litz wire conductors.
The primary objective thereby aims to resolve the complexity of real applications: On one hand,
the detailed geometry of the inductive components and on the other hand the inner structure of litz
wire conductors.
Apart from that, further requirements have to be taken into account:

• For verification reasons, the fundamental physical effects need to be identified and solved
separately. Results of different solving approaches have to be rated with regard to their
accuracy and effort, and compared to analytical methods.

• Verifications with measurements of the entire systems have to prove the accuracy of the
simulation workflow or to indicate which adaptions and improvements are additionally re-
quired.

• The simulation workflow has to fit into common development processes of inductive com-
ponents and provide interfaces to standard CAD file formats.

• A graphical user interface (GUI) providing the handling of files and the visualization of
all relevant results, allows a larger user community to use the simulation workflow. The
resulting feedback enables a user-driven development process accompanied by a customized
adaption and improvement of the workflow.

Even if all listed items are fulfilled, the establishment of a new simulation approach or even its
acceptance in the engineering community is not guaranteed at all. However, it would help to over-
come or even reduce the gap between simulation results and measurements.
In the following, chapter 2 starts with a presentation of the physical background and an analysis
of the relevant solving approaches for the calculation of litz wire losses. The limits of different
analytical and numeric approaches are demonstrated and rated, from a single solid conductor up to
complex litz wires. Moreover, the commonalities and differences of solid and stranded conductors
are carved out with regard to their behavior on system level. Based on these findings, chapter 3
presents a coupled simulation approach for the multi-scale problem of litz wire systems. First, a
general overview of all different modules is given. Subsequently, details about the settings, param-
eters, interfaces and exchange files are described. Special attention is paid to the requirements of
the different solving approaches on litz wire level and their realization in the code. At the begin-
ning of the chapter 4 the experimental verification is considered by discussing the general demands
on the experimental setups. Based on that, definitions and requirements are worked out to guar-
antee a high quality standard of the measurement results. The verification of the new simulation
workflow with the measurements is realized by four test setups. They all have in common, that
they are closely related to common central issues in the design of inductive components. Apart
from varying ferrite geometries and bundle structures of the litz wire, entire chokes or coil systems
are studied. Finally, chapter 5 provides conclusions on the results and findings of this thesis and
presents an outlook on further research activities.

5
CHAPTER 1. Introduction

6
 Physical background and solving
2
approaches

In this chapter the theoretical background of electromagnetic field problems is presented with
a focus on power losses in litz wire systems. Analytical and numerical solving approaches are
discussed and benchmarked with regard to their practical applicability.

7
CHAPTER 2. Physical background and solving approaches

2.1. Maxwell’s equations

In the domain of power electronics the main focus is on component and on system level. The
electrical behavior of a device is mostly very complex - however, it results out of the superposition
of different physical effects. In this chapter, basic mathematical equations describing the physical
system in the domain of electromagnetics are presented.
In the first half of the 19th century, basic scientific investigations of electric resistivity and electrical
losses were carried out by Georg Simon Ohm [10] and James Prescott Joule [11]. The resulting
Ohm’s and Joule’s law enable the calculation of electrical losses for direct currents (DC):
The power losses in a line conductor of length l with a cross section A and an electrical resistivity
ρ are given by
l
PDC = I ·U = I 2 · RDC = I 2 ρ (2.1.1)
A
for a direct current I.
In common applications of power electronic systems, these calculations need to be extended to
the domain of alternating currents (AC). Due to the superposition of time-varying electric and
magnetic fields additional effects appear.
The general description of the physics behind electromagnetic fields is first given by the Maxwell’s
equations [12]. These equations are standard for more than a century, consequently the derivation
is well documented and can be studied in various publications [13],[14].
From the mathematical point of view, Maxwell’s equations are a coupled system of linear partial
differential equations (PDE) for the electromagnetic field variables ~E, H, ~ ~B, and ~D, and the current
~
and charge densities J and ρ, respectively. Moreover, constant material properties of ε, µ, and σ
are taken into account.
Two common formulations of Maxwell’s equations are presented below describing how currents
and charges generate and interact with electromagnetic fields:
Differential formulation Integral formulation
!
~ ~D
~ = J~ + ∂D
I Z
~ · d~l = ∂
∇×H H J~ + · d~A (2.1.2)
∂t L A ∂t
!
∂~B ~B
I Z
~E · d~l = − ∂
∇ × ~E = − · d~A (2.1.3)
∂t L A ∂t
I Z
∇ · ~D = ρv ~D · d~A = ρv dv (2.1.4)
A v
I
∇ · ~B = 0 ~B · d~A = 0 (2.1.5)
A

and additionally
~D = ε~E ~B = µH
~ J~ = σ~E (2.1.6)
with

~E V - Electric field intensity, ~ A - Magnetic field intensity,

   
H
m  m 
~D C2 - Electric flux density, ~B W2 - Magnetic flux density,
 mC   mA 
ρv m3 - Volume charge density, ~
J m2 - Electric current density,
F 
µ H
 
ε m - Capacitivity of the medium, m - Inductance of the medium,
 1 
σ Ωm - Electrical resistivity.

8
 2.2. Solving approaches

In principle, both formulations describe the same physical system and are suitable for solving all
kinds of electromagnetic problems. However, the mathematical and computational effort varies
strongly from case to case, depending on the chosen formulation. Details of the mathematical
theory of the system of equations are provided in [15], especially with a focus on uniqueness and
existence of the solution.
Simplifications of Maxwell’s equations are common in the application on real systems and often
used in case of dominating effects:

• Magnetostatic case
~ ~
The time-varying electric ( ∂∂tD ) or magnetic ( ∂∂tB ) field terms are neglected, resulting in a
separation of the equations for the electric and the magnetic field.

• Electrostatic case
~
Only fixed electrical charges are considered and consequently time-varying field terms ( ∂∂tD
∂~B
and ∂t ) vanish on equation level.

• Quasi-stationary case
Full Maxwell’s equations in which displacement currents (due to time-varying electric fields
∂~D
∂t ) are neglected.

Based on this theoretical background, solving approaches are discussed in the following chapters
with a focus on power electronic systems using litz wire conductors.

2.2. Solving approaches

The most common way to “solve” complex electromagnetic problems and therefore Maxwell’s
equations, is an experimental setup and measurements. However, in the scope of experiments, the
separation of effects is very difficult or even impossible. Moreover, especially in industries, the
amount of experiments and prototypes during the development process needs to be reduced due to
cost and time reasons.
Over a long period, the only alternative to experiments were analytic approaches solving Maxwell’s
equations for simplified geometries and dominating effects. The most common methods use sep-
aration of variables or series expansions for solving the corresponding PDE [15]. Some familiar
formulas for skin and proximity losses in 2D basing on separation of variables are presented in
chapter 2.3.1. Although these results are an exact solution of Maxwell’s equations, the usage on
system level is limited due to the restriction on very basic geometries.
In the 2nd half of the 20th century, new methods based on numerical approaches became popular
[16], [17]. Details about first historical methods up to state of the art approaches are chrono-
logically provided in [18]. In contrast to analytical methods, numerical approaches calculate an
approximation of the exact solution. In the domain of electromagnetics, two main approaches are
established differing considerably in their numerical techniques and the underlying mathematical
formulation:

• In differential based techniques, the entire structure - including the surrounding medium
(i.e. air) - has to be discretized. In addition to the definition of suitable boundary conditions
at the border of the simulation domain, the numerical system has to be gauged to yield a
unique solution for realistic setups. The solution consists of field values (~E, H,
~ and J)
~ at the
nodes of the elements.

• Integral approaches in electromagnetics only take conductive materials and their excitation
into account. Elements of the resulting mesh are treated like elements of an electrical circuit.

9
CHAPTER 2. Physical background and solving approaches

Table 2.1: Overview of some important properties of FEM and PEEC method.

FEM PEEC
Formulation Differential Integral
Solution
Field Circuit
variable
Solution
Time domain, frequency domain Time domain, frequency domain
domain
Complex materials, reduced model Reduced complexity for solving fre-
possible by using symmetric bound- quency sweeps, reduction of the num-
Advantages
ary conditions, standardized tools and ber of elements by restriction to con-
CAD import available ductive materials
Evaluation of Green’s function, “non-
Large sparse matrices, resolution of
Disadvantages sparse” matrices, problems with mag-
air
netic materials

After combining all elements to a huge electrical circuit, the system can be simplified and
solved on circuit level (i.e. by electronic circuit simulation tools). The resulting solution is
expressed by integral quantities like L, R, and C.

In the domain of electromagnetic simulations, the differential formulation is used in the finite dif-
ference method (FDM) and the finite element method (FEM).
The integral formulation enables the usage of boundary elements (BE) methods, such as in the
method of moments (MoM) and in the partial element equivalent circuit (PEEC) method.
Details about all four methods are provided in [19] including a benchmark concerning the accu-
racy and the resolution of complex geometries. In the following, the description is limited to the
FEM and PEEC method, because these methods are the most advanced ones in both formulations
described above. Based on [19], some relevant properties of both methods are summarized in table
2.1.
The listed pros and cons of the FEM and PEEC method are very general and differ strongly on the
implementation and the regarded geometry setup. In the following subchapters, the elaborations
focus on solid and stranded conductors in power electronic systems. The relevant state of the art
approaches are presented considering the analytical methods and the numerical techniques FEM
and PEEC in chapter 2.3, 2.4 and 2.5, respectively. In all cases, the approaches are first introduced
in general, followed by a description of their adaption on litz wire systems. Exact analytical results
for a single solid conductor are used to determine the requirements for the FEM method to reach
a sufficiently high accuracy of the solution (chapter 2.4.2). These requirements for the FEM are
imposed on geometries of much higher complexity and the FEM results are compared with the
results of the PEEC method in chapter 2.5.2.
The specific benefits of the PEEC method in comparison to the FEM approach are worked out
with a focus on litz wire systems and all findings are summed up in a separate conclusion section
at the end of this chapter.

2.3. Analytical methods

In all physical domains analytic approaches are common, but their complexity differs strongly
between different physical systems. While the equations of motion are solvable on high school

10
 2.3. Analytical methods

level, even the existence and smoothness of 3D solutions of the Navier-Stokes equations has not
yet been proven.
In the domain of electromagnetics, the formulation of Maxwell’s equations (2.1.2)-(2.1.5) de-
scribes all interactions between currents, charges, and (electric and magnetic) fields. However,
analytical solution approaches for full 3D systems are very complex and only available for some
theoretical examples [15]. Consequently, in general, there are no exact analytic approaches for
solving real 3D electromagnetic systems, but rather sectoral specific rules of thumb or approxima-
tion formulas [20], [21].
However, without loss of accuracy the complexity of some real systems can be reduced and there-
fore simplifies the system of equations:
• Neglecting minor effects like displacement currents results in vanishing terms on equation
level (see simplifications described in the last chapter).

• Symmetric systems can be reduced to problems with lower complexity or even lower di-
mension which enable less complex mathematical operations and a greater variety of math-
ematical approaches.
A very common simplification in the domain of transmission lines [22] is the usage of quasi static
models on a 2D geometry. The physical system is assumed to be equal along the length of the
conductor with two dominating effects:
The time-varying excitation causes eddy currents in the conductor itself and also in adjacent con-
ductors. At higher frequencies, the current density distribution tends to be higher in regions close
to the surface of the conductor. Due to this increasing current density at the surface of conductors,
this effect is called skin effect. Moreover, alternating currents effect alternating magnetic fields
outside the conductors. These fields induce eddy currents in adjacent conductors and cause ad-
ditional modifications of the current density distribution by the so called proximity effect. Both
effects on the current density distribution can be calculated analytically, starting with the curl
operation on both sides of equation (2.1.3):

d (∇ × ~B)
∇ × ∇ × ~E = −
dt
d (∇ × µH)~
=−
dt
~
eq.2.1.2 d (J~ + ddtD )
= −µ . (2.3.1)
dt
d ~D
In the next step, equation (2.3.1) is modified by neglecting the displacement current dt (quasi
~
static solution) and assuming a homogenous medium ~E = J : κ

d J~
∇ × ∇ × J~ = −µκ . (2.3.2)
dt
In case of a circular line conductor in z-direction, the coordinate system can be switched to cylin-
drical coordinates, resulting in a current density

J~ = J(r,
~ ϕ, z) = Jr~er + Jϕ~eϕ + Jz~ez .

By focusing on the skin effect, only a radial dependency of the current density (which has only a
component in z-direction) has to be taken into account:

J~ = Jz (r)~ez . (2.3.3)

11
CHAPTER 2. Physical background and solving approaches

The curl operation in the corresponding cylindrical coordinates [23] yields

dJz
∇ × J~ = −~eϕ , (2.3.4)
dr
and therefore
2J
 
1 dJz d z
∇ × ∇ × J~ = − + 2 ·~ez (2.3.5)
r dr d r
for the left side of equation (2.3.2). The time variation of Jz depends on the signal of the excitation.
Assuming a sine waveform, the current density is given by

Jz (r,t) = Jˆz (r) e jωt , (2.3.6)

which can be applied to the combination of equations (2.3.2) and (2.3.5):

d 2 Jˆz (r) e jωt 1 d Jˆz (r) e jωt d(Jˆz (r) e jωt
+ = µκ
d2r r dr dt
= µκ jωJˆz (r) e jωt
d 2 Jˆz (r) 1 d Jˆz (r)
⇒ + − µκ jωJˆz (r) = 0. (2.3.7)
d2r r dr
This partial differential equation can be transformed to Bessel’s differential equation [24] by sub-
stitution (by z). The resulting simplified system
d 2 Jˆz d Jˆz
z2 + z + (z2 − 0)Jˆz = 0 (2.3.8)
d2z dz
can be solved by two main approaches using Kelvin or Bessel functions. The implementation is
very technical and can be studied in various publications [14], [25], [24].
In this thesis, approaches based on modified Bessel functions of zero (I0 ) and first (I1 ) kind are
used to solve (2.3.8). Details about the mathematical derivation and the adaption of the method to
take into account also the proximity effect caused by the external magnetic field Ĥ are provided in
[26].
Assuming a solid conductor with a diameter of 2a, the resulting current density distribution at a
given radius r is given by
ˆ
ˆ = I αaI0 (αr)
J(r) (2.3.9)
2πa2 I1 (αa)
for the skin effect, and
ˆ ϕ) = 2Ĥ αaI1 (αr) sin ϕ
J(r, (2.3.10)
a I0 (αa)
for the proximity effect. Both equations enable the calculation of the (z-directed) current density
in 2D. However, in case of the proximity effect, the distribution is not rotational symmetric any
more and depends on the direction of the amplitude of the sinusoidal external magnetic field Ĥ:
More precise, ϕ (in equation 2.3.10) corresponds to the angle between the coordinate of a point
inside the conductor (in cylindrical coordinates) to the direction of the external magnetic field the
conductor is exposed to. The frequency dependency is included in both cases in
1+ j
α= (2.3.11)
δ
by the material dependent skin depth r
ρ
δ= , (2.3.12)
πfµ

12
 2.3. Analytical methods

(a) (b)

Figure 2.1: Current density distribution due to the skin effect at 100 kHz (a) and proximity effect
caused by an external magnetic field Ĥ = 500 mA (in the direction of the arrows) at the same
frequency (b). The distributions have been calculated analytically based on equations (2.3.9) and
(2.3.10) for a copper conductor of 0.25 mm diameter and an excitation current of I =1 A.

using the electrical resistivity ρ and the magnetic permeability µ.

In figure 2.1 the current density distribution of skin (100 kHz) and proximity effect (100 kHz with
A
an external magnetic field amplitude of 500 m ) is visualized for a conductor with 0.25 mm diame-
ter.
To sum up, the analytical method described above provides an exact solution of the spatial current
density for the skin and proximity effect, considering the following limitations:

• The physical model is simplified to the quasi-static case (equation 2.3.2).

• The electrical properties of the medium are assumed to be homogenous (equation 2.3.2).

• The geometry of the conductor is assumed to be an infinitely long line with only a radial
dependency of the current density distribution (equation 2.3.3).

• The excitation is sinus-shaped (equation 2.3.6) and the input quantity corresponds to the
current I and the amplitude of the (external) magnetic field Ĥ, respectively.

In the following chapters, this approach is used for analytical loss calculations and benchmarks of
the numerical methods.

2.3.1. Solid conductors

The simplest case of a wire system is a single solid line conductor. Assuming an infinitely long
conductor, the exact solution of the current density distribution in this 3D system can be calculated
by using a quasi static solving approach applied on the 2D cross section (see last chapter).
In real systems, especially in the domain of power electronic systems, the conductor structure is
much more complex and in particular not straight-lined. However, all physical effects which occur
in conductors - even in HF litz wires with hundreds of strands - can be locally approximated by
this simplified model.
The great benefit of analytic approaches is the possibility to separate physical effects and to deter-
mine their relevance compared to the entire amount of losses.
In case of circular conductors and homogenous external magnetic fields, skin- and proximity losses

13
CHAPTER 2. Physical background and solving approaches

can be calculated separately due to their orthogonality [27], [28], [29]. Based on the direct current
losses PDC (equation 2.1.1), the entire frequency-dependent power losses are given by

Ploss = PDC + Pskin + Pprox . (2.3.13)

A very convenient method to calculate frequency-dependent losses bases on the integration of

equation (2.3.9) and (2.3.10) over the cross section of the conductor [30]. The resulting equations
provide a closed formulation for skin- and proximity losses in line conductors.
The skin losses Pskin in a conductor with radius a are given by
2
Pskin = IRMS RDC · (Fs -1), (2.3.14)

with the skin factor  

1 I0 (αa)
Fs = Re αa . (2.3.15)
2 I1 (αa)
The highlighted factor -1 in equation (2.3.14) is often omitted - in such cases the formula represents
both, the frequency-dependent skin losses including the DC-losses.
Assuming a line conductor of length l which is exposed to a homogenous magnetic field Ĥ
(orthogonal to the direction of l), the proximity losses are given by

l
Pprox = Ĥ 2 · Ds (2.3.16)
κ
with the proximity factor  
I1 (αa)
Ds = 2πRe αa (2.3.17)
I0 (αa)

0.26
skin losses
plus additional proximity
0.24 losses by dotted curves

1000A/m
Ploss [mW]

500A/m
0.22 200A/m
100A/m
increasing H ext
0.20

0.18
Pdc

0.16
1k 10k 100k 1M
f [Hz]

Figure 2.2: Frequency-dependent losses in a solid conductor with 0.25 mm diameter based on
equations (2.3.14) and (2.3.16) for an excitation current of 1 A. Skin losses are visualized by a
black solid line, additional proximity losses for varying external magnetic fields are plotted by
dotted curves.

14
 2.3. Analytical methods

and the specific conductance κ.

In figure 2.2, power losses in the frequency range up to 1 MHz are visualized for a conductor of
0.25 mm diameter using the analytic formulas based on an excitation current of 1 A. In addition to
the basic loss share of RDC , frequency-dependent losses of Pskin are plotted as a black solid line.
Moreover, Pprox for varying external magnetic fields in the regarded frequency range is visualized
by dashed colored lines. With increasing external magnetic fields the proximity losses get the
dominating factor of the frequency-dependent power losses, because of the quadratic influence of
Ĥ (in equation 2.3.16).
Considering both formulas for the skin and proximity losses, the losses are driven by the excitation
current IRMS and the (external) magnetic field Ĥ, respectively. However, in real systems increasing
power densities are often accompanied by using conductors closely located to each other, such as
tightly packed windings in a coil. In such a setup, the external magnetic fields and consequently
Pprox directly depend on the excitation current.
Up to now, all physical effects in this simple line conductor are resolved by analytical and closed
formulas. Consequently, in the following this method is used as a benchmark for the numerical
approaches with regard to their accuracy. The computational effort is not considered, because the
time for evaluating the analytical formulas is negligible, as even thousands of analytical calcula-
tions require only a fraction of a second.

2.3.2. Stranded conductors

In many applications in the domain of power electronics, solid conductors are substituted by litz
wires. Apart from higher flexibility (smaller curvature radii), the electrical behavior at higher fre-
quencies is improved by using isolated strands with a smaller diameter (figure 2.3 a). Commonly,
the dimension of the strands is less than the skin depth and therefore the skin losses are signifi-
cantly reduced.
However, power losses vary strongly even for litz wires with the same amount of strands depend-
ing on the inner twisting structure. To get an impression of the variation of resulting power losses,
this section focuses on both limiting cases:
The “worst” litz wire conductor consists of parallel strands, while in a perfectly twisted litz wire
each strand is located on each possible position of the cross section along one pitch length. Both
cases are later named parallel (par) and ideal (id) litz wire.
In general, the same basic electromagnetic effects occur in solid as well as in litz wire con-
ductors. However, the geometric structure is much more complex, due to the twisting, varying
strand and bundle levels, and soldering. Consequently, resulting losses can be split according to
the different levels of accruement, dependencies, and dimensions. In litz wires, standard direct

Table 2.2: Overview of all relevant effects in a litz wire conductor separating between excitation
current dependent losses and power losses caused by an external magnetic field Hext . Moreover,
technical terms, major dependencies and geometric dimensions are listed.

 Technical term Dependencies Dimension


 RDC I 1D 
Pskin_strand I, f 2D


Pskin

I dep. Pskin_bundle I, f, solder 3D 
Pint_prox_strand I, f 2D


Pint_prox


I, f, solder 3D

 Pint_prox_bundle 
Pext_prox_strand Hext , f 2D
Hext dep. Pext_prox
Pext_prox_bundle Hext , f , solder 3D

15
CHAPTER 2. Physical background and solving approaches

Hint
Hext

(a) (b)

Figure 2.3: Full 3D model (a) of one pitch length of a litz wire conductor with radius rc
consisting of 19 strands with a diameter of 0.25 mm. The 3D model is reduced to 2D (b) and the
internal (black arrows) and external (marked in red) magnetic fields are visualized.

current and strand losses (calculable with 1D and 2D approaches just like for solid conductors)
are increased by induced loop currents (visualization in figure 2.4 by red arrows) excitated by the
external magnetic field in adjacent strands conductively connected via the solder joint (requiring a
3D resolution).
Unfortunately, the distinction of the technical terms in the domain of litz wires varies in publica-
tions of the leading groups, such as [31], [32]. In table 2.2 all different shares of power losses in
litz wire conductors are provided based on [31], [8]. In contrast to other definitions, these terms
distinguish between losses depending on the excitation (skin losses and internal proximity losses)
and losses caused by external magnetic fields (proximity losses). Because the first share is caused
by (and consequently dependent on) the excitation current in the conductor, the resistance term R
is often used to quantify the power loss at a given current. For benchmarking different litz wire
types with regard to their power losses both shares are required. In this thesis, normalized power
losses are calculated by summing up skin losses with an excitation current of 1 A (RMS) at a fix
frequency and proximity losses at a given external magnetic field.
In the following course of this section, a detailed description of the varying shares of power losses
is presented based on [31]. The elaboration uses and explains the condensed information of table
2.2 and provides analytic approaches for both extreme cases of parallel (par) and ideal (id) litz
wire conductors.

Excitation dependent losses (skin effect and internal proximity effect)

Commonly, the direct current resistance of litz wires is provided by the data sheets of the manu-
facturer. Based on these data, measurements of real applications can be used to verify the quality
of the solder joint.
Moreover, the direct current resistance of a litz wire conductor can be approximated analytically.
Assuming N copper strands with a radius of rs and length l, the DC resistance is given by

l
RDC = . (2.3.18)
Nrs2 πσcu

An additional factor often mentioned in the data sheets of litz wires is the so called copper fill
factor δcu . Regarding the structure of a litz wire conductor (figure 2.3 a) with a complete radius

16
 2.3. Analytical methods

Figure 2.4: Sketch of both shares of the external proximity losses caused by the external
magnetic field Hext within a litz wire. Apart from opposed directed loop currents on strand level
(highlighted in green) also one directed loop currents (red color) are induced. The latter are
caused by conduction loops between adjacent strands and mainly depend on the size and
orientation of the area spanned between adjacent strands. In case of perfectly twisted litz wires,
strands are twisted in a special manner that these one directed loop currents vanish.

rc , the factor δcu is defined by the ratio of the copper area to the conductor area:
 2
Nrs2 π rs
δcu = =N . (2.3.19)
rc2 π rc

Assuming an ideal litz wire, strands are twisted in a special manner, that all induced eddy currents
in loops between strands (red curves in figure 2.4) are canceled out along the entire length of
the conductor (Rskin_bundle and Rprox_int_bundle vanish). Consequently, the excitation current IRMS is
equally divided among all N strands.
To simplify the distinguish of varying radii, the normalized skin depth parameter for copper strands
is defined as
eq.2.3.11 p
xS = α · rs = (1 + j) π f µ0 σcu · rs . (2.3.20)

Assuming all strands having the same length, the resistance of each strand is given by RDC_strand =
RDC N and the direct current loss of the stranded conductor
 2
IRMS 2
PDC_strand = N (RDC N) = IRMS R0 (2.3.21)
N

is equal to the one of a solid conductor with resistance R0 .

Based on equation (2.3.14), the power losses due to the skin effect in an idealized stranded
conductor can be calculated by
2
Pskin_strand = IRMS RDC · (Fskin_strand − 1) (2.3.22)

with the skin factor

1 n I0 (xS ) o
Fskin_strand = Re xS . (2.3.23)
2 I1 (xS )
on strand level.
With regard to the proximity loss, each strand in the litz wire can be considered as a solid conductor
with N − 1 adjacent conductors. Depending on the position of a strand within the conductor, the
superposition of the magnetic field caused by neighboring strands differs. However, in case of

17
CHAPTER 2. Physical background and solving approaches

litz wires with a high amount of strands, the magnitude of the magnetic field at a radius ρ can be
approximated by
Iρ
H(ρ) = (2.3.24)
2πrc2
for any excitation current I. This spatial distribution of the internal magnetic field can be applied
to equation (2.3.16) to evaluate internal proximity losses [33]. Resulting equation

1 n r2 I1 (xS ) o
2
Pint_prox_strand = IRMS RDC · Re xS N 2 s2 (2.3.25)
2 rc I0 (xS )

can be simplified to

2 1 n I1 (xS ) o
Pint_prox_strand = IRMS RDC · Re xS Nδcu , (2.3.26)
2 I0 (xS )

and yields
2 1 n I0 (xS ) I1 (xS ) o
Pid,skin = IRMS RDC · Re xS + Nδcu xS (2.3.27)
2 I1 (xS ) I (x )
| {z } | {z 0 S }
Pskin_strand Pint_prox_strand

for the entire skin losses of an ideal litz wire conductor.

In case of litz wires without twisting, the conductor corresponds to a system of parallel strands.
By soldering both ends of such a litz wire, the behavior of the system can be approximated by
a homogenized solid conductor [25] of the same shape, but with an adapted conductivity with
regard to the power losses. Assuming the same cross section, litz wires have a reduced electrical
conductivity compared to solid conductors. Therefore, the specific averaged electrical conductivity
σlitz of a litz wire can be calculated by

rc2 πσlitz = Nrs2 πσcu ⇔ σlitz = δcu σcu . (2.3.28)

Based on this homogenized material, the normalized skin parameter on conductor level is given
by p
xC = (1 + j) π f µ0 σlitz · rc . (2.3.29)
Using such an artificial solid conductor, no substructure has to be taken into account. Conse-
quently, neither further losses due to the bundle structure nor internal proximity losses occur.
Combining the standard equation for skin losses (equation 2.3.14) with definition (2.3.29), yields

2 1 n I0 (xC ) o
Ppar,skin = IRMS RDC · Re xC (2.3.30)
2 I1 (xC )

for the entire skin resistance in a parallel stranded conductor.

External magnetic field dependent losses (external proximity effect)

Up to now, effects in a single litz wire line conductor have been studied. In real applications, con-
ductors are used in windings closely packed in inductive components such as coils and chokes. In
these systems, frequency-dependent magnetic fields cause additional proximity losses in adjacent
conductors.
For calculation reasons, the entire litz wire conductor is assumed to be exposed to a homogenous
external magnetic field Ĥext (which is orthogonal to the direction of the conductor). In contrast
to Pint_prox which is caused by adjacent strands inside the conductor, external proximity losses are

18
 2.3. Analytical methods

caused by eddy currents induced by the magnetic field depending on neighboring conductors, air
gaps, or ferrite material. In figure 2.3 b) the eddy currents causing Pext_prox_strand and Pext_prox_bundle
are visualized for a simplified setup by black and red arrows, respectively.
In case of circular conductors, losses caused by internal and external magnetic fields can be calcu-
lated separately due to the symmetric structure of the internal magnetic fields [8]. This assumption
matches with the geometric structure of common HF litz wires consisting of a high number of
strands. Similar to the calculation of skin losses, both normalized factors xS and xC are used for
the calculation of the proximity losses.
In case of an ideal twisted litz wire, all induced eddy currents in the loops between strands
(Pext_prox_bundle ) are canceled out (highlighted in red in figure 2.4). Consequently, the external
proximity losses on strand level are equal for each strand and yield
 
l 2 I1 (xS )
Pid,prox = Ĥ · NDid with Did = 2πRe xS (2.3.31)
σcu ext I0 (xS )
for the entire ideal litz wire conductor.
The parallel stranded conductor is considered by an equivalent solid wire with adapted conductiv-
ity σlitz just as in the previous section. Losses can be approximated based on [31] by
 
l 2 2π I1 (xC )
Ppar,prox = Ĥ · NDpar with Dpar = Re xC . (2.3.32)
σlitz ext N I0 (xC )
The total losses of a litz wire conductor are split in excitation dependent losses (skin) and external
magnetic field dependent losses (prox). Both shares are depicted for the limiting cases of an
ideal twisted litz wire (id) and a conductor of parallel strands (par). However, in contrast to
analytical results for solid conductors, the resulting formulas (2.3.27) and (2.3.30) - (2.3.31) only
approximate the behavior for both extreme litz wire types. Moreover, in case of real litz wires a
reliable predictions of the power losses is impossible based on these formulas without additional
information (chapter 3.7.1) or measurements (chapter 3.7.3). Consequently, these formulas cannot
be used for verifying the accuracy of numerical methods, but allow us to demonstrate the potential
of improvement by using perfectly twisted litz wires.

2.3.3. External effect of conductors

The both effects investigated in the last chapter only have influence inside the conductor. How-
ever, due to the time varying current in the litz wire, alternating external fields are caused in the
surroundings. In the following, the electromagnetic behavior of solid and stranded conductors is
investigated.
From an analytical point of view, the magnetic field strength H ~ can be determined in the en-
tire domain by using Ampere’s circuital law of Maxwell’s equations (2.1.2). Assuming a current
density J~ passing through a connected surface A with a boundary c, the magnetic field H ~ satisfies
I ZZ
~ =
Hdc ~
JdA. (2.3.33)
c A

In the domain of litz wires, the total current I is commonly assumed to be distributed equally
among all strands. Consequently, equation (2.3.33) can be evaluated at the radius rc of the con-
ductor, yielding
~ c ) =~eϕ I
H(r (2.3.34)
2πrc
with a tangential direction ~eϕ . This is even true in the case of non-perfect joints [34] and corre-
sponding rotationally symmetric current density distribution within the litz wire conductor.

19
CHAPTER 2. Physical background and solving approaches

(a) (b)

Figure 2.5: The left figure shows a simplified solid conductor with the same radius as the
420 x 0.1 mm litz wire on the right side. As long as the current density distribution in a wire is
homogenous (or at least rotationally symmetric), the magnetic field outside of the wire depends
only on the total current I in the wire.

Consequently, in case of a rotationally symmetric or homogenous current density distribution

along the cross section of the conductor, the magnetic field surrounding a conductor (figure 2.5) is
independent of its inner structure and only depends on the total current I.
In figure 2.6 the magnetic field along the contour of the circular surface of a solid and a stranded
conductor is visualized. In both cases, a round conductor loop with an excitation current of 1 A is
simulated in a magnetostatic FEM simulation (details about the numerical method are provided in
the next chapter). In case of the litz wire, the conductor model consists of 420 strands à 0.1 mm

250
Sketch:
Simulation of a loop with loop diameter d
225 diameter d, w=1.5p 1mm
conductor 1.5mm
w=p w=0
3mm
radius 1.5mm
200 6mm
and 1A. w=0.5p
solid wire = solid line
litz wire = dashed line
(420x0.1mm)
175
d
H[A/m]

150

125

100

75

50

0.0p 0.5p 1.0p 1.5p 2.0p

w on the surface of the conductor

Figure 2.6: Comparison of the magnetic field strength for a solid (figure 2.5 a) and a stranded
(figure 2.5 b) conductor loop. On the x-axis, the magnetic field along the boundary of the
conductor is plotted, with its maximum at π corresponding to the inner side of the loop with
varying diameters d (visualized by different colors).

20
 2.4. Finite Element Method (FEM)

diameter. The cross section of this conductor is equal to the one chosen for the solid conductor
model. The magnetic field distribution on the circular boundary (ω = 0 − 2 π) of both conductor
surfaces is almost equal for all diameters d of the loop. Some minor differences occur due to small
differences in the shape of the conductor surfaces: The arrangement of 420 strands in the cross
section of the litz wire does not yield an exact circular surface.
Based on these elaborations, the resolution of the complex inner structure of litz wires can be ne-
glected, if only the magnetic field distribution on the surface (and therefore in the surroundings) is
required.

2.4. Finite Element Method (FEM)

A very common simulation approach for all kinds of physical problems is the finite element
method (FEM). This numerical method enables an approximation of the solution of a large variety
of partial differential equations (PDE), like Laplace equations for diffusion problems, Navier-
Stokes equations for fluid dynamics, or Maxwell’s equations for the electromagnetics.
From a mathematical point of view, the FEM approach is limited to differential equations D(u) = f
(with additional boundary conditions) in Ω, with a certain sense of equivalent weak formulation,
given by
a(u, ϕ) = F(ϕ) ∀ϕ ∈ X(Ω). (2.4.1)
The bilinear form a(u, ϕ) is the weak formulation of the differential operator D(u) on the domain
Ω, and usually corresponds to an energy minimization problem. Due to the validity of equation
2.4.1 for any test function ϕ out of the function space X(Ω), these kind of formulations are also
known as variational formulation.
In the following, the Ritz-Galerkin approach [35] is applied on equation (2.4.1) to transform the
continuous operator problem in a discrete problem, which can be solved by a linear system of
equations.
First, the function space X(Ω) is approximated by the finite dimensional subspace Xh (Ω) ⊂ X(Ω)
with h > 0, resulting in

a(uh , ϕh ) = F(ϕh ) ∀ϕh ∈ Xh (Ω). (2.4.2)

Assuming a basis φ1 , ..., φN of the finite dimensional Xh (Ω), this problem is equivalent to

a(uh , φi ) = F(φi ) with i = 1, ..., N. (2.4.3)

Defining
N
uh = ∑ v jφ j (2.4.4)
j=1

yields a linear system of equations

N
∑ a(φ j , φi )v j = F(φi ) with i = 1, ..., N (2.4.5)
j=1

for the unknown coefficients v j .

For solving this system efficiently, the computation of matrix A = (Ai j ) = a(φ j , φi ) has to be sim-
ple and its structure has to be sparse. Both requirements are met by the finite element approach
by using local basis functions. First, the entire domain Ω̄ is separated in a finite number of “cells”
S
τ by Ω̄ = τ∈Xh (Ω) . These “cells” are commonly known as elements in the FEM and consist of
triangles or rectangles in 2D, or tetrahedrons or polyhedrons in 3D. All basis functions φi are

21
CHAPTER 2. Physical background and solving approaches

piecewise polynomials with a support consisting of some few of τ. Due to the local support of the
basis functions, a(φ j , φi ) is zero for most combinations i and j, and the associated linear system of
equations (2.4.5) can be solved much more easily since A is sparse.
A detailed description of the mathematical background and the adaption of the FEM approach on

P
X W

M 6 4 O
B
U V
N
Y 5 3 A
L
T S
Z
I 2 1 K

Q R
J

(a) (b)

Figure 2.7: In the left picture several types of elements (basic polyhedron and corresponding
tetrahedral, pyramid, and prism option [36]) for electromagnetic simulations are depicted. Due to
additional nodes (numerated by letters) in the middle of the edges a higher accuracy of the
solution is reached without an increasing amount of elements. Different types of elements can be
used in the same mesh to improve the resolution of complex geometric objects or physical effects
(i.e. a thin layer of prism elements to resolve the skin losses).

varying differential equations are provided in [35], [37], [38], [39], [40].
Since the early 1960s, when Turner [16] and Clough [41] first used approaches which were later
named finite element method, an improvement process started and has not yet finished. All basic
developments in the FEM, caused by significant progress of the mathematical basis and of the
numerical algorithms as well as the influence of the steady increase of the computing performance
are well documented in [42]. While the application of the method was first limited to structure
simulations for avionic and automotive, the FEM became common for all kinds of engineering do-
mains and even further. Especially the development from individual programs and university codes
to commercial software packages like ANSYS, COMSOL, or ABAQUS lead to a wide spreading
of simulations in industry. Due to increasing accuracy of the solution and implementation of more
and more physical effects, FEM based simulation tools got established and accepted. Moreover,
easy handling of the simulation programs and good post processing tools lead to an expansion of
the user group. In addition to simulation and algorithmic experts, more and more application and
design engineers use FEM software.
Another important part of simulation tools is the meshing algorithm subdividing the simulation
domain. The quality of the elements has a big influence on the convergence behavior of FEM
algorithms, thus for the accuracy of the results. By now, different complex approaches like auto-
matic local refinement (and even coarsening) and high order element types (like in figure 2.7 a)
are state of the art. Consequently, modern simulation tools can handle even complex geometries

22
 2.4. Finite Element Method (FEM)

like in figure 2.7b very efficiently with a low number of elements.

In the next subchapter the mathematical background of a common FEM algorithm for electromag-
netic simulations is presented. Because of the complexity of such solvers, the elaboration in this
thesis is limited to the essential parts of the numerical workflow to understand influencing factors
and sensitivities with regard to the FEM methodology. In chapter 2.4.2, the numeric results of the
FEM are verified by exact analytic results of chapter 2.3.1 and later in chapter 2.5.2 the FEM is
used as a benchmark for the litz wire approach using the PEEC method.

2.4.1. Algorithmic and mathematical basis

Nowadays, there is a large number of different FEM based software products for solving elec-
tromagnetic problems. In most cases the dominating catchwords are HPC (high performance
computing), GPU (graphics processing unit), adaptive meshing, or multiphysics, while the core
of the software - the solving algorithm - is never mentioned or even explained. In most cases, the
documentation only shows little details on the implementation, although the mathematical back-
ground has been known for nearly half a century and is similar in most products.
In most physical problems, dominating effects permit one to reduce the electromagnetic sys-

Electromagnetic fields
(Maxwell’s equations)
. &
Stationary Transient
fields
 fields

. y & y &
Electrostatic Stationary Magnetostatic Eddy Full wave
fields eddy fields Currents
 currents
  
y y y y
Φ Φ or ~T Φ or ~A ~A+~T , ~T +Ω
or ~A+Φ

Table 2.3: Different potential approaches for partial formulations of Maxwell’s equations with
electric / magnetic scalar potential (Φ, Ω) and electric / magnetic vector potential (~T , ~A).

tem described by the Maxwell’s equations to partial problems. Due to these simplifications the
calculation time and hardware resources are reduced significantly. Consequently, modern FEM
tools offer different specialized algorithms, such as electro-static, magneto-static, or magneto-
quasistatic solvers (also knowns as eddy current solver) apart from the full wave solver for the full
time-discretized Maxwell’s equations. On implementation level (see figure 2.3 for Ansys) differ-
ent combinations of the scalar potential (Φ, Ω) and the vector potential (~T , ~A) are used for the
electric and magnetic quantities, respectively [43], [44].
In the domain of physics, a potential is a function whose derivative yields fields. While fields are
related to forces, potentials are associated with energy. Details about various approaches to define
the different potentials in the fields of electromagnetics are provided in [45].
In the following the A-V (also known as A-Φ) potential approach for the magneto-quasistatic
Maxwell’s equations is sketched based on [46], [47]. The name of this implementation refers to
the degrees of freedom used in the solving approach: the magnetic vector potential ~A and the elec-
tric scalar potential Φ. This method is implemented in the algorithms of Ansys Emag and Ansys
Maxwell2D (the 3D solver bases on a ~T − Ω formulation). Further information on this finite ele-
ment scheme is provided in [48], while details on extending this method to a full wave solver are

23
CHAPTER 2. Physical background and solving approaches

described in [49].
Regarding the magneto-quasistatic electromagnetic problem, the Maxwell’s equations of chap-
ter 2.1 are simplified by neglecting the time depending influence of the electric fields. This is a
quite common case in simulations of systems having a small size compared to the electromagnetic
wavelength associated with a dominant effect of frequency driven eddy currents [50]. Conse-
quently displacement currents and charges are neglected and the system of equations is reduced
to
~ = J,
∇×H ~ (2.4.6)
d~B
∇ × ~E = − , (2.4.7)
dt
∇ · ~B = 0. (2.4.8)

The constitutive relations of the material properties (equation 2.1.6) retain their validity.
Equation (2.4.8) implies that the magnetic flux ~B can be written as the curl of a vector potential ~A
with
~B = ∇ × ~A. (2.4.9)
Substituting equation (2.4.9) into (2.4.7) results in

d(∇ × ~A) d~A

∇ × ~E = − =⇒ ∇ × (~E + ) = 0. (2.4.10)
dt dt
~
Since the field ~E + ddtA is irrotational and exists (thus ensures a non trivial solution), it is given by
a pure source field. Consequently, a scalar potential V exists with

d~A
− ∇V = ~E + , (2.4.11)
dt
solving equation (2.4.10), because curl of any gradient field is zero.
Because both potentials ~A and V are defined implicitly by vector operators (curl and gradient), the
system has to be gauged to reach uniqueness. Assuming a unique solution of ~A, the case for V
is simple. Since the electric scalar potential is defined by its gradient, a unique solution can be
reached by defining V fix at one point in the domain (i.e. by a boundary condition).
Regarding the magnetic vector potential ~A and equations (2.4.6 - 2.4.8), all physical dependencies
(for corresponding ~B) are already defined. Consequently, there is no need for a physical gauging, it
would even destroy a physical solution. Mathematical approaches like the Coulomb (∇ · ~A = 0) or
Lorenz (∇ · ~A + c12 dV
dt = 0) gauging result in a unique solution (see realization in [46]), nevertheless
they are difficult to realize in 3D finite elements.
In 2D simulations no additional gauging process is required. The magnetic vector potential con-
tains only a z-component and due to equation (2.4.9), the magnetic flux density is given by
 
~e ~e ~e 
 x y z ~ ~
~B = ∇ × A~ z =  ∂ ∂ 0  =~ex ∂Az −~ey ∂Az . (2.4.12)
∂x ∂y
∂y ∂x
 0 0 ~Az 
 

Consequently, this one component vector potential ~A satisfies the divergence free constraint and
therefore automatically the Coulomb gauging.
In general, the magnetic flux Φ through a surface element f is defined by
ZZ ZZ I
Φ= ~B · d~f eq.2.4.9
= ∇ × ~A · d~f = ~A · d~s (2.4.13)
f f ∂f

24
 2.4. Finite Element Method (FEM)

Φ
A6 A5
tree
A4 co-tree
A3
1 3

A1 A2

Figure 2.8: Sketch of the gauging used in Ansys Emag for a tetrahedron with edges concerning
the tree (Ai 6= 0) and co-tree (Ai = 0). The magnetic flux Φ is calculated by equation (2.4.15) and
the integrated values of the magnetic vector potential along the corresponding edge (A2 , A4 , and
A5 ).

with ∂ f being a positively oriented, piecewise smooth, simple closed curve of f (sketched in figure
2.8).
In the FE space, the magnetic vector potential is stored as a scalar value
Z
Ai = ~A · d~ei (2.4.14)
ei

in the i-th node in the middle of its corresponding edge ei . Consequently, equation (2.4.13) can be
written as
Φ = ∑ Ai (2.4.15)
i
for all nodes i on the edges of the surface f . In nodes with magnetic vector potential in the direc-
tion of integration, the sign of AI is positive, otherwise negative.
In case of tetrahedrons, the physical system is over-determined: the magnetic vector potential on
six edges is used to calculate three magnetic fluxes. Due to equation (2.4.8), the fourth flux is
defined by the fluxes through the first three faces. To reach a unique solution in the entire domain,
the number of fluxes has to agree with the number of edges with non-zero potential values. Conse-
quently, three out of six values of Ai have to be defined to be zero. However, the associated fluxes
shall not vanish.
By transferring this task into the field of mathematics, the problem corresponds to the search of
the maximal sub-graph without closed paths. In the graph theory such a sub-graph is named tree,
while all other paths (corresponding to edges which are defined to zero) are named co-trees (figure
2.8). Algorithms for finding such structures are commonly summarized under the term “tree gaug-
ing” and differ strongly with regard to their complexity [51], [52]. The results are not unique and
the quality of the gauging process significantly influences the non-zero structure of the resulting
sparse matrix which is used in the solving process of the FEM. In Ansys Emag, there is no further
documentation of the implementation, but it is similar to [53], [52].
In a 3D domain with element types like the ones presented in figure 2.7a, the gauging process is
even more complex. With up to six fluxes and 20 nodes per element, the gauging process in stan-
dard simulations requires approximately 20 - 30 % of the entire calculation time. Due to the special
structure of the gauging problem, common algorithms are serial and high performance computers
are not able to take advantage of their multicore structure. Moreover, the special geometric shape

25
CHAPTER 2. Physical background and solving approaches

of litz wire models even worsens the convergence of these tree gauging algorithms - starting with
a very minor amount of boundary elements (at the contact), the tree has to grow far deep inside
the conductor model.
In the next step of the simulation process, a system of equations based on substituting the ini-
tial equations (2.4.6)-(2.4.8) into equation (2.4.9) and (2.4.10) is built up in consideration of the
boundary conditions (full documentation in [46]). The solving process for this linear system of
equations requires most of the computation time in standard simulations and therefore provides a
high potential for speed up. A lot of commercial software packages offer special high performance
licenses for solving this step fast in parallel: for instance parallel iterative solvers like multi-level
algebraic multi-grid solver (AMG) [54] reach a speedup of factor four with eight cores compared
to single core (in Ansys Emag).
The quantities calculated in FEM simulations are the (electric and magnetic) field and the current
density. Integral values like the resistance of a conductor or the self inductance of a coil need
additional post processing. In case of huge simulations, this calculation requires several hours,
because relevant data are stored in several millions of elements.
To sum up, the FEM enables a good integration in the design workflow of power electronic sys-
tems and commonly provides reliable interfaces to import CAD data. Moreover, materials with an
inhomogeneous electric or magnetic permeability can be used in the electric simulation, including
the resolution of complex physical effects, such as hysteresis. The computational efforts can be
reduced due to artificial boundary conditions and specialized solver types. However, in terms of
litz wires, the meshing (including also the free-space) and the gauging process is very costly and
results in very large linear systems of equations requiring a huge amount of memory. Finally, the
calculation of integral values such as resistance and inductance needs additional efforts.

2.4.2. Comparison of FEM and analytical results

The accuracy of numerical results compared to the exact physical solution described by Maxwell’s
equations depends on the solving algorithm (sketched in last chapter) and on the meshing. Assum-
ing a very fine meshing, the accuracy of the results is very high and even reaches dimensions of the
machine epsilon [34], [55]. These high quality results are purchased by a high amount of elements
and require huge computational resources. For practical cases, such as for litz wire simulations,
the accuracy needs to be discussed in relation to the required effort, and so do the number of
elements. Inside a litz wire conductor, the skin and the proximity effect are the relevant effects.
Without loss of generality the regarded structure can be limited to a single solid line conductor

(a) (b) (c)

Figure 2.9: Sketch of a solid conductor (a) with radius a and the frequency-dependent skin layer
δ. Corresponding discretization in Ansys Maxwell2D (b) with n = 10 edges and one additional
skin layer, and the resulting current distribution (c).

26
 2.4. Finite Element Method (FEM)

1E+00

conductor of 0.25 mm
diameter at 1 MHz
1E-01

relative error to analytic solution

1E-02

1E-03

Meshing
1E-04 without skinlayer
skinlayer

n=10
n=15
1E-05 n=20
with one skinlayer
n=10
1E-06 n=15
n=20

1E-07
0 1 2 3 4 5
a/δ

Figure 2.10: The relative error (k P Psim − 1k) of skin losses is plotted versus the dimensionless
analytic
factor aδ for different geometry and mesh setups. Different discretisations of the circular
conductor are highlighted by different colors and meshes with and without skin layer are
distinguished by solid and scattered lines, respectively. The green vertical line corresponds to the
limiting case of realistic applications of a copper conductor with 0.25 mm diameter at 1 MHz. In
systems with diameters and frequencies up to this limit, meshes without additional skin layers
and only 15 elements (N = 15) reach a sufficient accuracy with a deviation from analytical
values of less than 0.5 %

which is the smallest substructure in litz wires. This simple system includes all relevant physical
effects and can even be reduced to a 2D model in which analytical approaches (such as those in
section 2.3.1) are used to benchmark FEM results.
In this case, Ansys Maxwell 2D is used with quadratic shape functions on element level. The
regarded frequency and magnetic field range should take the realistic conditions inside of power
electronic systems into account: common frequencies in litz wire applications are up to 1 MHz
and magnetic field strengths can be up to 1 kA m for strands with a maximum diameter of 0.25 mm.
Based on these boundary conditions, variations on the meshing are needed to evaluate the accu-
racy of FEM results depending on the amount of elements.
First, the discretization of the surface has to be defined. In finite element approaches, the circular
copper conductor of radius a is approximated by a polygonal curve. The number of edges n and
the number of subdivisions along the radius determines the number of elements (see also [34]).
For simplification, only basic triangular elements are considered in the following. In figure 2.9 b)
the mesh for n = 10 with one subdivision at a radius of a − δ is visualized. This configuration
results in 30 elements for the discretization of the conductor.
For obtaining normalized quantities, we proceed as follows: A dimensionless quantity aδ is used
instead of a fixed radius or frequency. Moreover, the accuracy of FEM results is analyzed using
Psim
the norm Panalytic − 1, characterizing the relative error of the simulated power losses to the analytic
results based on equation (2.3.14) and equation (2.3.16) respectively. To get comparable values
even for different area sizes in the FE discretization of the conductor cross section, power losses

27
CHAPTER 2. Physical background and solving approaches

1.3

1.2

1.1

1.0
J /J 0

0.9

0.8
analytic res ult
without s kinlayer
0.7 with s kinlayer

0.6
0.5 1 0.5 1

0.5
0.0 0.2 0.4 0.6 0.8 1.0
r/a

Figure 2.11: Normalized analytic (green) and simulated (red) results of the current density (J/J0 )
versus normalized radius ar for the limiting case ( aδ = 1.9). The dashed red curve shows the
simulation result based on a meshing without a skin layer using only one polygon (of second
order), while an additional skin layer (red symbols) uses two of them. They are joined at the skin
depth (approximately at 0.55) and show a significantly better agreement with the exact solution.

Psim are normalized with respect to the conduction area A = a2 π. In the following sections, both
relevant physical effects (skin and proximity) inside a solid conductor are investigated separately
according to the accuracy with respect to the amount of elements.

Skin effect
In the 2D simulation model for the skin effect, an excitation current (1 A) is applied to the con-
ductor surrounded by a plenum limited by symmetric boundary conditions.
Power losses inside the conductor are calculated for a wide range of frequencies and are deter-
mined in dependence of the dimensionless factor aδ . In view of the extreme configuration (fre-
quencies up to 1 MHz for a conductor of 0.25 mm diameter) the range of interest is limited by
a
δ = 1.9.
In figure 2.10 the relative error for three different discretizations (n = 10, 15, and 20) of the circu-
lar conductor is visualized. Furthermore, the results for different meshing setups with and without
skin layer are shown. The skin layer is targeted to a radius of a − δ, but is enforced to be at least
a a
4 to avoid degenerated elements. Consequently, up to δ = 1 there is nearly no improvement by the
additional skin layer.
All curves of figure 2.10 show significant outliers (i.e. for n = 10 without skin layer at aδ = 2.5).
To get a better understanding of their origin, the basic FEM results need to be investigated. In fig-
ure 2.11 the dependence of the simulated normalized current density JJ0 on the normalized radius
is shown. The green curve shows the analytic results while the red curves show the FEM results
with and without a skin layer for n = 20.
In both latter cases, the shape of the curves shows a good illustration of the quadratic basis func-
tion used for the numerical approach on the edges of the triangles. The mesh without skin layers

28
 2.4. Finite Element Method (FEM)

(a) (b)

Figure 2.12: (a) shows the geometric setup of a solid conductor in 2D with boundary conditions:
At the upper and bottom side, the magnetic vector potential ~A is set opposed directed (with ~A in
the direction of the drawing plane). On the left and right side, the magnetic flux density is defined
orthogonal to the boundary. The resulting vectorized magnetic fields and the current density
distribution inside the conductor are visualized in (b).

uses only one quadratic interpolation (dashed red line), while the additional skin layers results in
two quadratic curves (lines with red symbols), linked at δ.
With increasing frequency the current density distribution gets more and more inhomogeneous
in the conductor and shows an exponential decrease at the surface. The approximation by one
quadratic basis function reaches its limit and requires more elements such as additional skin lay-
ers. In outliers (i.e. aδ = 1.75 for n = 20 with skin layer in figure 2.10) a minor over- or underesti-
mation of the spatial current density distribution is canceled out which yields significantly higher
accuracy in the calculation of power losses.
Regarding the relevant range up to aδ = 1.9 in figure 2.10 an additional skin layer does not lead
to a consistently better accuracy (apart from n = 20, aδ > 1.5). Only at higher ratios ( aδ > 3), the
increasing number of elements due to an additional skin layer provides benefits for all discretiza-
tions and reaches a significantly better accuracy of more than a factor of 10 in average.
Moreover, a higher amount of edges for the approximation of the circular conductor results in
more accurate results. However, a limitation is given due to the quality of the elements which gets
worse with increasing aspect ratio for increasing refinement.
To sum up, power losses due to the skin effect can be resolved very accurately without skin layers
for relevant conductor diameters (up to 0.25 mm) in the regarded frequency range up to 1 MHz. A
discretization of the circular structure of the conductor with 15 elements yields sufficient accuracy
with a numerical error less than 0.5 %.

Proximity effect
Based on the orthogonality of skin- and proximity effect [28], the resulting power losses can be
investigated separately. Within the analytical approach, equation (2.3.14) and (2.3.16) enable a
separate calculation of both shares. In the numeric simulation, the model of the previous section
can be reused with regard to geometry and mesh, but needs to be adapted by new boundary con-
ditions. A direct implementation of external magnetic fields is neither derivable from the basic

29
CHAPTER 2. Physical background and solving approaches

1E+00
conductor of 0.25 mm
diameter at 1 MHz
1E-01

relative error to analytic solution

1E-02

1E-03
Meshing
without skinlayer
1E-04 n=10
n=15
n=20
1E-05
with one skinlayer
n=10
1E-06
n=15
n=20

1E-07
0 1 2 3 4 5
a/δ

Figure 2.13: Relative error of normalized proximity losses versus aδ based on the same variations
as in figure 2.10 with regard to geometry and meshing. The differences between the relative
errors for different discretisations (N = 10, 15, 20) of the circular conductor is not as large as for
the skin losses investigated previously. Even with N = 20 and additional skin layers the accuracy
is limited to only 3 %.

formulation (A-Φ in figure 2.3) nor realized in Maxwell2D. However, by defining an opposed
directed magnetic vector potential on both opposite boundaries (up and bottom in figure 2.12 a) a
homogenous magnetic field is caused in the entire simulation domain. To avoid numerical outliers
at the corners of the plenum, both remaining boundaries (on the left and right side) need to be de-
fined as symmetric (flux normal). Applying the frequency sweep which is used in the last section,
the alternating magnetic field induces eddy currents inside the conductor and therefor causes an
inhomogeneous current density distribution (visualized in figure 2.12 b).
Power losses caused by the proximity effect are directly proportional to the square of the magnetic
A
field. In the following investigations, normalized power losses for H = 1 m are considered. The
relative error of these normalized losses is compared to analytic results of equation (2.3.16) and
visualized in figure 2.13 up to aδ = 5.
All curves show a continuous increase over the entire aδ range with a slightly increasing improve-
ment by additional skin layers. Moreover, the relative error increases much steeper than for the
skin losses with increasing aδ and the absolute values are significantly higher. Even setups with
the highest amount of elements (60 triangles at N = 20 and additional skin layer) already reach
an relative error of 1 % for aδ = 1.2.
In summary, the resolution of the proximity losses requires much more elements in FEM simula-
tions compared to the resolution of the skin losses to achieve a similar high accuracy. Especially
in case of high frequencies the proximity effect yields eddy currents resulting in a very inhomoge-
neous current density distribution within radial and tangential direction. Consequently, either the
frequency range (and therefore aδ ) is limited, or the amount of circular discretisations and therefore
the number of elements has to be increased. By neglecting the computational efforts, an adaptive
refinement would be suitable to investigate the varying current density distributions within a fre-

30
 2.5. Partial Element Equivalent Circuit (PEEC)

quency sweep.
These increasing computational resources should be taken into account, especially with regard to
the dominance of the proximity losses at higher frequencies (figure 2.2).

2.5. Partial Element Equivalent Circuit (PEEC)

Apart from the FEM, there are further approaches to solve partial differential equations (PDE).
A common method bases on the boundary value problem, which is associated with a PDE and
which can be evaluated by the so called boundary integral equations. Approaches calculating a
numerical approximation of the solution for these boundary integral equations are summed up as
boundary element methods (BEM). One sophisticated BEM in the domain of electromagnetics is
named partial element equivalent circuit (PEEC) method.
In contrast to the FEM (section 2.4), there is no straight solution strategy or just a closed for-
mulation for the BEM covering all kinds of PDE [56]. In general, for a given boundary value
problem (corresponding to one PDE and a defined dimension of the solution space) there exist
different boundary integral equations and for each of them several numerical approximation meth-
ods. Many of these solving approaches have still a lack of mathematical accuracy (considering
convergence, uniqueness, or error estimation) and are subject of current studies.
However, a typical solving approach of the BEM can be sketched as follows:
• Mathematical model
In general, the usability of the BEM is limited to linear partial differential equations with
constant coefficients, because the fundamental solution of the PDE has to be known explic-
itly (for determining the boundary integral equations in subsequent steps). The mathematical
model solved in the BEM is described by the boundary value problem of the PDE.
In case of a Laplace equation with Dirichlet boundary conditions, the PDE is given by
∆u = 0 in Ω ⊂ Rn , (2.5.1)
u=g on Γ := ∂Ω, (2.5.2)
with the fundamental solution
1
G(x, y) = with x, y ∈ R3 in n = 3, (2.5.3)
4πkx − yk
commonly known as Green’s function.
• Representation formula
In the next step, the solution of the PDE within the calculation domain has to be transferred
to a boundary potential formulation. Regarding classical boundary value problems of math-
ematical physics (such as Green’s third identity for potential theory or Betti’s formula for
elasticity theory) these so called representation formulas are already known [56] and enable
the calculation of u within the domain Ω in the aforementioned example.
• Boundary integral equation and elements
In addition, the function values at points on the boundary have to be evaluated based on the
representation formulas. Commonly, these formulas are inserted into the boundary condi-
tions and yield the boundary integral equations. However, the formulation of these equations
is not unique and an integrable solution on the entire boundary of Ω requires sophisticated
mathematical approaches, especially in case of non Dirichlet problems [57]. The evaluation
of the boundary integral equations is often realized by finite element functions. The bound-
ary Γ is divided in boundary elements and the numerical solution is calculated for each of
them.

31
CHAPTER 2. Physical background and solving approaches

• Equations and linear system

Based on the results which approximate the solution along the discretized boundary, a sys-
tem of (discrete) equations can be generated by collocation, the Galerkin’s method, or the
least squares method [56]. The system of equations can then be transferred to a linear sys-
tem of equations. In contrast to the FEM, the resulting system of equations is not sparse and
therefore limits the amount of suitable solvers, especially in case of large systems.

• Postprocessing
The results of the BEM provide the solution of the boundary integral equation and thus only
on Γ. Solutions in the interior of the domain Ω additionally require the evaluation of the
representation formula.

The points presented above only sketch the strategy of the BEM. Details and examples of this
method are provided in [56], while the mathematical aspects are discussed in detail in [57].
In the domain of electrical engineering, a specialized BEM has been developed half a century ago
to solve the Maxwell’s equations. Initially, this so called partial element equivalent circuit (PEEC)
method was invented by A. E. Ruehli in the 1970s at the IBM T.J. Watson Research Center for
interconnect problems [17], [58], [59]. The first approaches were limited to the calculation of
partial inductances on circuit boards. Nowadays the scope of the method includes the calculation
of all electromagnetic parameters in the time and frequency domain.
From the mathematical point of view, the method is based on the integral formulation of the
Maxwell’s equations (right hand side of equations (2.1.2) - (2.1.5)). This formulation is applied
on discretised volumetric elements of the electrically conductive parts of the regarded system.
Each of these elements is electromagnetically coupled with all other elements by self- and mutual-
impedance terms [60]. This coupling structure is used for the design of an equivalent electrical
circuit consisting of inductance, resistance, and capacitance elements.
Assuming a simple electric structure consisting of two connected and conductive bars, the ge-
ometric setup can be transferred to an equivalent electrical circuit (figure 2.14). Connecting this
structure to a voltage source at point 1 and 3, both endings are set to two different electric poten-
tials. Based on the definition of this voltage drop, the electrical system is uniquely defined and all
circuit elements can be evaluated by the PEEC. First, the resistance values R1 and R2 are given by
the material specific conductivity and the geometrical shape of the solid cells. The determination
of the inductances (L11 and L22 ), as well as the capacitances (C11 , C22 , and C33 ) bases on the total
electric field in the regarded (geometric) element calculated by the magnetic and electric potentials
in relation to the surrounding (details in subsequent chapter).
In general, depending on the specific application, the electrical system can be simplified on circuit
level and solved by common circuit analysis tools like SPICE [61].
The most well-known implementations of the PEEC method were realized by White and cowork-
ers with Fastcap [62] and Fasthenry [63], [64] in the early 1990s. Based on these tools, the
extraction of capacities and inductances for very-large-scale integration (VLSI) chips became fea-
sible in a wide frequency range.
The strict requirements on the geometric shape of the elements are very prejudicial for the import
and the coupling of CAD geometries compared to other numeric approaches (like FEM or FD). Up
to the first decade of this century, the subdivided elements had to have a rectangular structure. In
2003 the PEEC approach was improved due to the non-orthogonal formulation [65] allowing the
usage of hexahedral and triangular [66] elements. Nevertheless, there is still no meshing strategy
for complex geometries available [67], which is in contrast to FEM tools with established meshing
engines for almost arbitrary geometries.
Another problem over a long time period was the integration of permeable materials in the simu-
lation model. While dielectric materials were already integrated in the 1990s [68], approaches for

32
 2.5. Partial Element Equivalent Circuit (PEEC)

1 2 3

R L R L
1 1 1 1 2 2 2 2 3

C 1 1
C 2 2
C 3 3

Figure 2.14: A simplified case of a conductor (i.e orthogonal metal stripe) is discretized by two
solid cells with three nodes (top). The nodes are referenced by their numbers in the
corresponding PEEC circuit (bottom).

materials with magnetic permeability (µr  1) were first presented in the last decade [69], [8].
The great potential of the PEEC approach is the flexibility in the formulation of the method. For
the rigorous full-wave version of the PEEC method, the (Lp , P, R, τ) notation is used with partial
inductance Lp , potential coefficient P (inverse of capacitance), resistance R, and delay τ. The latter
is necessary, when the rise times of the signal waveform decrease and the corresponding frequency
increases, or the geometric dimensions become larger (compared to the vacuum wavelength) [70].
In most electromagnetic systems, the complexity of the PEEC model can be reduced without loss
of physical accuracy [71]:
• In cases of system structures much smaller than the wavelength of the highest frequency,
the delay term τ can be omitted resulting in a reduced (Lp , P, R) - PEEC model.

• In applications with a dominating influence of the inductances (i.e. calculations on a circuit

board), the complexity of the simulation model can be additionally reduced: In the simu-
lation tool Fasthenry a pure magneto-quasistatic approach (corresponding to a Lp - PEEC
model) is used, neglecting all capacitances between the elements.

• In 2014 a specialized approach for the frequency-dependent loss calculation of litz wires
was presented by Richard Y. Zhang with a tool called Fastlitz [8]. Similar to Fasthenry,
the Maxwell’s equations are simplified to the magneto-quasistatic case, but additionally the
frequency dependent resistance of the conductor is evaluated. The corresponding (Lp , R) -
PEEC model is implemented in Matlab and discussed according to its relevance for the
design of coils and inductive systems in [72], [73].
However, experts in the domain of the PEEC method considered the status of this numeric ap-
proach as being in an early development phase [67]. Currently, the most pressing challenges
contain increasing speed-up, accuracy, and stability, as well as interfaces to different methods and
programs. Even though the evolution of the PEEC methodology has shown significant progress

33
CHAPTER 2. Physical background and solving approaches

in the last decades, an extensive usage in the simulation domain is not likely within the next few
years. The FEM is well established in all parts of industry and research for all kinds of physical
problems. Moreover, the increasing number of application engineers requires an interface to CAD
tools and visualization, which is not yet foreseeable in the PEEC implementations.
To sum up, the PEEC method provides an alternative approach for solving the Maxwell’s equa-
tions [74], but only in some specialized domains with a focus on the reduction of the complexity
of the physical model, such as calculations of antennas, circuit boards, or litz wires. In the fol-
lowing chapter, the algorithm and the mathematical basis of the PEEC implementation Fastlitz
is sketched. In 2.5.2 this numeric approach is compared with FEM for a simulation model of a
simple litz wire conductor.

2.5.1. Algorithmic and mathematical basis

In 2014 a new approach for the calculation of litz wire resistances depending on the frequency and
external magnetic fields was presented by Zhang [8]. The associated program Fastlitz is the latest
version of several programs using the PEEC method all developed by White and co-workers at the
Massachusetts Institute of Technology (MIT).
The mathematical background of the PEEC method [17], [75] bases on the Maxwell’s equations
(described in chapter 2.1) and the definition of the total electric field ~E. At any observation point
~r, ~E is given by the magnetic vector potential ~A and the electric scalar potential Φ [71] by

~
~E(~r,t) = − ∂A (~r,t) − ∇Φ(~r,t). (2.5.4)
∂t
Assuming a harmonic excitation

∂~
E(~r,t) = ~E0 e− jωt ~E(~r), (2.5.5)
∂t
equation (2.5.4) can be written as

~E(~r) = − jω~A(~r) − ∇Φ(~r). (2.5.6)

In litz wires, the dominating share of losses is caused by the current driven skin and proximity
effects. Consequently, displacement currents (caused by time varying electric fields) and charge
~
accumulations are considered second order effects ( ∂∂tD = 0 and ∇ · ε~E = 0, respectively), and
therefore equation (2.5.6) can be simplified to

~E(~r) = − jω~A(~r). (2.5.7)

Combining Ampere’s circuital law (equation 2.1.2) with the definition of the magnetic vector
potential (equation 2.4.9), resulting equation can be simplified to a Laplace problem by using the
Coulomb gauging:
∇ × ~B = µJ~

∇ × (∇ × ~A) = µJ~ Coulomb ∆~A = −µJ. ~
~B = ∇ × ~A −−−−−→
Based on that, equation 2.5.7 can be transformed to
Z
~E(~r) = jωµ G(r, r0 ) J~exc (r0 ) dv0 , (2.5.8)
ΩV 0

by taking into account the explicit solution of the Laplace equation in R3 using the Green’s func-
tion (chapter 2.5). The total electric field in r is expressed by the excitation current density J~exc of

34
 2.5. Partial Element Equivalent Circuit (PEEC)

y
2 0 0
1 .0 0 k = 1 0

0 .7 5 1 5 0

n u m b e r o f f ila m e n t s
y /r

0 .5 0
0 .2 5 1 0 0
0 .0 0 x
- 1 .0 - 0 .5 0 .0 0 .5 1 .0 5 0
x /r
k = 3
0
1 2 3 4 5 6 7 8 9 1 0
d is c r e tis a tio n le v e l k

(a) (b)

Figure 2.15: Sketch of the discretization algorithm in Fastlitz using a square root based
distribution along the radius r of the conductor for the discretization level k = 2 (blue) and k = 3
(green) (a). Corresponding number of filaments versus k are provided in (b) for example
geometries k = 3 and k = 10 with 13 and 181 filaments respectively.

the source point r0 along the corresponding integration domain ΩV 0 . Assuming a metal stripe de-
scretized by quadratic elements (such as in figure 2.14), the integration domain corresponds to the
cross section of the conductor. In such a case, the integration term in equation 2.5.8 is equivalent
to the current value at the corresponding position of the electric circuit. Based on that, all mutual
inductances between conductive parts can be evaluated based on the current values (provided by
the electric circuit) and the distances (corresponding to r and r0 in equation 2.5.8).
Due to the usages of the integrale values provided by the electric circuit, the BEM approach used
in the PEEC implementation is limited to so called brick shape functions. In case of a spacial
resolution of the current density, the integration domain has to be discretized in multiple ele-
ments (forming the entire conductor), which are electrically connected in parallel on circuit level.
Based on that, even a strong inhomogeneous current density distribution within the cross section
of a conductor (caused by the skin- and proximity effect) can be resolved. In the Fastlitz al-
gorithm, conductors are discretized in filaments (rectangular parallelepiped solids) generated by
nested rectangles approximating the circular cross section (figure 2.15 a). On geometry level, the
discretization of the simulation domain in the PEEC method is similar to the FEM. However, the
filaments used in Fastlitz only provide constant current values, whereas the elements in the FEM
enable a spatial resolution of the current density by shape functions of first and second order (fig-
ure 2.11).
From a numerical point of view, a comparable high accuracy of the PEEC solution (compared to
the FEM) can only be achieved by using a higher amount of filaments, especially in the relevant
section. Based on the discretization parameter k the refinement algorithm generates k rectangu-
lar solids (using a square root based distribution along the radius of the conductor), which are
intersected with each other (figure 2.15 a). The resulting amount of filaments has a quadratic de-
pendency on the discretization parameter (appr. 2k2 ) and provides an improved resolution of the
outer area of the cross section and therefore of the skin- and proximity losses (figure 2.15 b).
Within the PEEC method the electromagnetic model is transformed to a circuit network model.
Regarding the loss calculation in a litz wire conductor from a theoretical point of view, it results
in an impedance extraction for a defined voltage difference at the terminals.
In general, the voltage difference between~r1 and~r2 is given by

Z ~r2
V= ~E(~s) · d~s, (2.5.9)
~r1

35
CHAPTER 2. Physical background and solving approaches

and can be applied to each pair of the filaments considering equation (2.5.8).
Assuming an electric circuit with n branches, the branch voltage vector Vb = [V1 ,V2 , V3 , ...,Vn ]T
and branch current vector Ib = [I1 , I2 , I3 , ..., In ]T are related by

Vb = ZIb with Z = R + jωL [76]. (2.5.10)

The complex impedance matrix Z ∈ Cn×n consists of a matrix R with resistance values of each
filament on the diagonal, and a dense matrix L including all partial inductances [63]
µ0
Z Z
Li j = [G(~ri ,~r j )]~li (~ri ) ·~l j (~r j ) d 3 xi d 3 x j (2.5.11)
ai a j f ili f il j

with ~li , ~l j the unit vectors in the direction of current flows through the cross sections ai and a j of
the corresponding filaments f ili and f il j .
Moreover, the connections and terminals in the entire circuit problem have to be taken into account.
Kirchhoff’s voltage and current laws are represented by

Vs = MVb (2.5.12)

and
M T Im = Ib . (2.5.13)
with the mesh matrix M ∈ Rn×m , the vector of the source branch voltages Vs and the vector of the
mesh currents Im (both Rm ). The calculation of the mesh matrix is well documented in [77], [78],
and enables the mapping of n branches to m closed current loops (known as mesh). Combining
equation (2.5.10) with (2.5.12) and (2.5.13) yields

MZM T Im = Vs . (2.5.14)

The system of equations consists of a full matrix (due to the mutual inductances in L) and the
right-hand side vector Vs with most entries being zero except for those rows corresponding to
terminal voltage excitations. In case of small systems, equation (2.5.14) can be solved directly
(with a complexity of O(n3 )) and resulting Im can be applied to equation (2.5.10) and (2.5.13) to
evaluate the impedance. A significant progress with regard to the calculation time was achieved in
the 1990s by the implementation of Fasthenry [64], [79] using a GMRES algorithm. This Krylov
subspace method was combined with an additional preconditioner and enables the evaluation of
equation 2.5.14 with a complete effort of O(n2 ). However, in the domain of high frequency litz
wires with hundreds of strands and a fine discretization of the cross section of the strands, several
tens of millions of filaments are reached. Consequently, in the Fastlitz implementation additional
sophisticated mathematical methods are used to reduce the computational efforts:

• Precorrected fast Fourier transformation (pFFT): In general the pFFT approximates a dense
matrix vector operation with a complexity of O(n2 ) by a summation of multiple smaller
matrix vector operations O(n log (n)) designed in a special manner [80]. In the case being
considered, the mutual inductance matrix Z is separated in two sparse matrices consisting
of the nearby and distance interactions [81]. Due to the neglecting effect of distance inter-
actions in the regarded electric system, the sparse matrix consisting the nearby interactions
is used for solving equation 2.5.14.

• Fast frequency sweep approach: The evaluation of equation (2.5.14) at varying frequencies
is enhanced by using a frequency-independent and a frequency dependent matrix [8]. Based
on the frequency independent term, adaptations due to different frequencies are calculated
iteratively.

36
 2.5. Partial Element Equivalent Circuit (PEEC)

First, we take a look at the pFFT method in general for a given dense matrix [A] (such as the
matrix Z consisting of the interactions between all pairs of filaments). Any n × n matrix [A] can
be approximated by

[A] = [S] + [C][H][P], (2.5.15)

with a sparse matrix [S] representing all nearby interactions and a pre-correction term [C][H][P]
for the distance interactions.
In the regarded electromagnetic system, the nearby interactions [S] correspond to filaments with a
strong mutual inductive coupling. Commonly, the entries of this sparse matrix cannot be estimated
in advance, because distance, current, and orientation of the filaments have a significant influence.
However, the distance interactions [C][H][P] can be calculated with less effort by using a coarse
grid (sketched in figure 2.16), approximating distance effects within the complex filament system
(compexity of O(n log (n))):
Initially, the matrix [P]m×n performs a projection of all basis functions (corresponding to the indi-
vidual current values of all n filaments) to a grid consisting of m points. From a global point of
view, the grid is comparable to a rough discretisation of the simulation domain with each node rep-
resenting dozens of filaments in its surrounding and therefore neglects local (dominating) effects.
The actual convolution (with the Green’s function) is computed by [H]m×m with much smaller
efforts (due to m  n). Following, the interpolation and pre-correction matrix [C]n×m transfers the
grid based results back to the filaments again (by a so called pre-correction term).

Projection on a uniform grid Initial guess of the current density

(a)

Calculation of the electric field New current density distribution

(b) (c)

Figure 2.16: The initial current density distribution of all filaments is projected by the matrix [P]
(a) on a uniform grid. Locally, the current density values of the filaments are summed up in the
grid points with regard to their distances. Following the convolution is calculated by the matrix
[H], yielding the electric field in all nodes of the grid (b). These results are interpolated (on all
filaments) and pre-corrected by the matrix [C] to gain the new current density distribution (c).

37
CHAPTER 2. Physical background and solving approaches

Based on equation (2.5.15), the sparse matrix [S]n×n can be calculated for each index i and j
of the given matrix [A] by this projection-convolution-interpolation workflow by

Si, j = Ai, j −Ci, : [H]P:, j . (2.5.16)

Even though the computation of [S] is very costly, the procedure is only required once and provides
a sparse matrix consisting of all dominating interactions of the impedance matrix Z. Based on that,
the complexity of (sparse) matrix-vector operations in the subsequent solving process of the linear
system is reduced to O(n).
Additionally, the pFFT method is combined with the fast frequency approach. In [8] one has
demonstrated, that a simulation domain containing different materials (such as air, copper, and
ferrite) can be calculated by the PEEC method with the help of multilayer Green’s functions.
These functions enable a split of the calculation in a free-space term and additional material terms.
However, in case of quasistatic simulations, the material term can be interpretated as a frequency
term. Thus, the effect of an increasing frequency on the current density distribution in a solid
conductor can be caused just the same by changing the material property in a special manner.
Consequently, equation (2.5.15) can be enhanced to

[A( f )] = [S( f )] + [C][H( f )][P], (2.5.17)

with a decomposition in frequency independent (free ) and dependent (add ) terms

[A( f )] = [Afree ] + [Aadd ( f )], (2.5.18)

free add
[S( f )] = [S ] + [S ( f )], (2.5.19)
[H( f )] = [H free ] + [H add ( f )]. (2.5.20)

In case of a sufficiently small distance of the nodes of the mesh mentioned above, frequency-
dependent terms are dominated by distance interaction and therefore [Sadd ( f )] (corresponding to
nearby interactions) vanishes [8]. Consequently, for a given matrix [Sfree ] only the projection-
convolution-interpolation workflow [C][H( f )][P] (figure 2.16) has to be performed for all varying
frequencies f to approximate the dense matrix [A( f )].
In Fastlitz the constant (pre-calculated) sparse matrix [Sfree ] is used in an iterative solving process
combing the frequency dependent correction term with the GMRES algorithm for solving the
linear system of equations (2.5.14). The corresponding pseudo code is provided in [8] and enables
the calculation of multiple frequencies in a significantly reduced time.
Resulting frequency-dependent impedance Z includes losses due to the skin effect and additionally
the internal proximity effect. In this case, “internal proximity” describes the influence on the
current density distribution in one strand caused by currents in adjacent strands. The influence
of external magnetic fields - resulting in external proximity losses - is calculated in the “power”
function as follows.
As a physical basis, Faraday’s law of induction for a wire loop is used to calculate the power losses
due to external magnetic field. Based on equation (2.1.3) a time varying external magnetic field
~Bext through a surface A causes a voltage drop of U, with
Z
∂Φ ∂ ~Bext d~A.
U =− =− (2.5.21)
∂t ∂t A

Regarding the harmonic case with a constant external magnetic field along the length of the con-
ductor, the induced voltage can be calculated by the dot product of ~Bext and the contour of the area
~A.

38
 2.5. Partial Element Equivalent Circuit (PEEC)

This approach is applied on all enclosed areas between strands [82], assuming a conductive con-
nection at both ends (such as a soldered joint). The resulting voltage drop for all strands [U] is
used for the calculation of the induced eddy currents [I] by

[I] = [Y ][U] (2.5.22)

with the admittance matrix [Y ] (equal to Z −1 out of equation (2.5.14)).

Based on that, the power dissipated in each strand is calculated for different frequencies and vary-
ing external magnetic fields.
In Fastlitz this PEEC approach is applied to a section of the litz wire, for example covering one
pitch length. Based on a handy front end, the litz wire structure and its discretization k is visu-
alized. The frequency-dependent power losses of the conductor are calculated, and additionally
the influence of external magnetic field is taken into account. Costly subroutines like the pFFT
method are implemented in C++ and integrated in the Matlab code.
To reduce license fees, the Fastlitz approach was implemented using the free software package
Octave within this PhD work. Although the calculation time of the new PEEC implementation
increases because of slower algorithms in the open-source program, parallel computation strate-
gies significantly reduce the amount of time. Moreover, the interfaces were adapted and enhanced
(details in chapter 3.7.2) to fit better to the entire workflow for calculating the power losses of litz
wire systems.

2.5.2. Comparison of PEEC with FEM

In this section, the PEEC method is compared to the FEM with the focus on the simulation of
litz wire conductors. The comparison of both methods is very challenging due to fundamentally
different approaches behind both numerical methods. In case of the FEM, simulations have been
performed by Ansys Maxwell3D, while the open source implementation FastLitz has been used
for the PEEC method.
In the following, the main challenges are sketched and the consequences for the comparison setup
are presented subsequently:

1. Geometry and meshing

PEEC: The geometry definition is specialized on litz wires. The position of each strand
is defined by analytic formulas, and all strands are piecewise discretized by cuboid-shaped
filaments. The circular structure of the cross section is resolved coarsely and approximated
by smaller elements at the boundary (figure 2.15).
FEM: CAD geometries of litz wires can be imported or analytically defined by functions,
but the meshing is very problematic. The cross section at the beginning and the end has to
be plain, resulting in slightly oval cross sections of the strands (because of the curvature due
to the twisting) which causes a lot of irritations in the meshing process. Moreover, adjacent
strands require a bigger distance compared to real systems to achieve the resolution of the
free-space.

2. Physics and boundary conditions

PEEC: The filaments and thus the corresponding litz wire geometry are transferred to the
circuit level. All adjacent filaments in a strand are conductively connected just as all strands
at their ends. The latter corresponds to a parallel connection of the strands within the con-
ductor and enables the resolution of bundle level effects due to loop currents between neigh-
boring strands. Proximity losses are calculated for a constant external magnetic field and
therefore depend only on the geometry and corresponding spanned area between strands.

39
CHAPTER 2. Physical background and solving approaches

FEM: All physical quantities are solved in the simulation domain depending on the bound-
ary conditions. The parallel connection of the strands can be approximated by setting all
surfaces at both ends on the same electrical potential. However, this definition does not pro-
vide a correct spatial resolution of the current flow between adjacent strands, which would
be needed to realize loop currents.
The definition of a constant opposed directed magnetic flux (similar to the 2D case in figure
2.12) on both opposite surfaces of the surrounding plenum results in a very homogenous
magnetic field in the entire simulation domain. However, this definition yields an overde-
termination of the physical system at the contact faces with the boundary condition for the
electrical potential. Consequently, a homogenous magnetic field distribution in the simula-
tion domain is not feasible with additional potential constrains at the boundary.
The issues described above demonstrate that a comparison of litz wire geometries does not
proceed very easily for both methods with regard to skin and proximity losses. The implemen-
tation of external magnetic fields in addition to the potential definition fails in Ansys Emag due
to the conflict between the boundary conditions. However, the centerpiece of the PEEC method -
the circuit approach - can be verified against the FEM based on the AC resistance (including RDC ,
skin and inner proximity effect) without an external magnetic field .
To depict a realistic system, one pitch length (30 mm) of a litz wire consisting of 19 copper strands
with 0.25 mm diameter is used as comparison setup. The exact position of each strand is defined
by a function and the distance of adjacent strands is set to 10 % of the strand diameter (small
picture in figure 2.17). The latter reduces the influence of varying discretizations of the circular
structure and enables a sufficient number of elements between the strands in the FEM.
In both numerical methods the excitation current is set to 1 A and the resulting frequency-

PEEC Model - 119.000 elements - 1.5 GB RAM

FEM Model - 860.000 elements - 42.3 GB RAM

(a) (b)

Figure 2.17: Absolute values of the frequency-dependent resistance (a) and corresponding
resistance factor FR = RRDC
AC
(b) for a litz wire geometry (upper picture) calculated by the PEEC
method (with 119 000 elements and 1.5 GB RAM) and the FEM (with 860 000 elements
requiring 42.3 GB RAM).

40
 2.6. Conclusion

dependent losses are calculated for frequencies up to 1 MHz. The absolute values of the resistance
for the FEM and the PEEC method show a good agreement over the entire frequency range (figure
2.17 a). The deviation of both curves even at low frequencies indicates that the geometric shape of
the litz wire (which is also influenced by the discretization and meshing) is not equal. Considering
a gauging of both curves by its RDC , the resistance factor FR yields a very good agreement up to
1 MHz for the FEM and PEEC method (figure 2.17 b).
Due to the FFT approach in Fastlitz, the calculation of the resistance requires only minor com-
putational resources for varying frequencies. 30 evaluation points in the range up to 1 MHz are
calculated in less than 10 minutes on a standard computer with negligible amount of memory (ap-
proximately 1.5 GB). In the FEM, the resolution of the twisted strands and the surrounding air
requires more than 40 GB of RAM and 5 minutes calculation time for a single frequency value.
This large difference between FEM and the PEEC method is caused by a much higher amount
of elements in the FEM discretization (factor 7 - 8), and the calculation and storage of vectorized
quantities (such as ~A and J)
~ instead of scalar integrated quantities (such as I, U, and R) used in the
circuit method.
This comparison of the FEM and the PEEC method does not represent a general benchmark or
rating of both numerical methods. However, in terms of litz wire simulations the PEEC method
reaches a good agreement with the FEM results, but requires only a small fraction of the com-
putational resources and time. Since common conductors in power electronic systems consist of
several hundreds of strands, the FEM is inappropriate to calculate the frequency-dependent power
losses for such systems.

2.6. Conclusion
In this chapter, the physical background of power electronic systems and corresponding solving
approaches have been presented with focus on the winding losses. Analytical methods as well as
the numerical approaches FEM and PEEC have been derived from the Maxwell’s equations and
applied to varying conductors.
Based on the exact analytical results of a single solid line conductor, the settings of the FEM have
been adjusted to reach a sufficiently high accuracy of the numerical results. With the help of these
settings, the power losses of conductors with a much higher complexity have been calculated and
used for the verification of the PEEC method.
Considering the accuracy and limitations of the solving approaches, the following findings are
essential for the resolution of real litz wire losses in power electronic systems:

• Analytic formulas provide an exact solution of the skin and proximity losses in 2D, but only
for single solid line conductors.

• For litz wire conductors the analytical approaches only provide an accurate approximation
of the power losses for both limiting theoretical cases: parallel stranded conductors and
ideal twisted litz wires.

• The FEM method is well established for arbitrary geometries and materials (including hys-
teresis effects). The discretization of the entire simulation domain, including the free space,
is required due to the basic mathematical formulation of the method.

• The resolution of stranded conductors in the FEM is possible, but requires a high amount
of elements and thus a huge amount of computational resources (due to the storage of all
field quantities in each element). Consequently, the meshing and especially the calculation
of tightly packed litz wires is not feasible.

41
CHAPTER 2. Physical background and solving approaches

• The integral-based PEEC method is limited to conductive materials (no ferrites) and is re-
stricted to basic geometrical shapes.

• The PEEC implementation Fastlitz enables the design of complex litz wire structures and
calculates the power losses based on corresponding circuit problems. The results agree with
reference values of the FEM, but require less than 5 % of the computational efforts of FEM.

• Even though the PEEC method is much more efficient than the FEM, a single litz wire sam-
ple of several centimeters and some hundred of strands requires several gigabytes of mem-
ory during the calculation. Consequently, the resolution of power electronic components
consisting of several meters of litz wire is not feasible or foreseeable in the near future.

The investigations show that there is a need for new numerical approaches in case of power elec-
tronic system with a complex litz wire structure. The FEM has clear advantages in terms of the
functionality in the engineering domain, the geometry import, and the consideration of electric
and magnetic permeability, but fails to resolve tightly packed litz wires. In contrast, the PEEC
approach enables the calculation of litz wire samples, while magnetic materials or arbitrary geo-
metric shapes are an insurmountable obstacle for this method.
Based on the presented solving approaches, one single numerical method cannot achieve the ob-
jectives with regard to the calculation of winding losses in litz wire systems.

42
 Multi-scale algorithm for litz wire
3
systems

Based on a combination of FEM simulations on system level and analytical, numerical, or mea-
sured based approaches on litz wire level, a new multi-scale algorithm for power electronic sys-
tems has been developed.

43
CHAPTER 3. Multi-scale algorithm for litz wire systems

3.1. Multi-scale approaches in the physical domain

In many physical domains, simulations have to face the fact that effects in regions several orders
of magnitude smaller than the whole domain have major influence on the entire system. This kind
of problem is neither new nor rare and often referenced as a multi-scale problem.
Some well-known problems are in the domain of antenna, nuclear, or fluid dynamics [83]. In the
latter case, the boundary layer between fluid and solid has a large influence on the velocity and
energy distribution of the entire simulation domain. A high resolution of this region is necessary,
especially to resolve turbulences or vortexes. Many multi-scale methods are available such as an-
alytic or heuristic approaches on boundary elements to calculate an approximation of the physical
solution based on a much coarser mesh in the rest of the domain.
In the electrotechnology, first approaches combining different numerical and analytical methods
have been investigated with regard to multi-scale problems at the end of the last century [18]. In
most cases, these approaches were limited to 2D geometries and only have minor significance for
real applications. A very convenient approach in the domain of system simulation was presented
for a shielded induction heater [84]. Due to the usage of a mixed approach of the finite elemente
method (FEM) and the boundary element (BE) method, the amount of elements and therefore the
calculation time were significantly reduced. It has been demonstrated that in particular the speed
up factor (coupled method vs. standard FEM method) increases by using an increasing amount of
elements. In other words, especially in the case of models with a high complexity the usage of
such coupled approaches gets more advisable.
In the domain of power electronics, a lot of inductive components like coils, chokes, and trans-
formers use conductors with a highly complex substructure. Measurements have proven a signif-
icant influence of this inner structure on the amount and distribution of power losses on system
level. Common simulation programs fail to resolve all relevant physical effects with regard to the
complete 3D system.
In the following a mixed solving approach is presented for calculating power losses in 3D by tak-
ing the complex litz wire structure into account. In chapter 3.2 a general overview of the workflow
and its implementation in the so called SlicerPro approach is provided. In the following chapters,
details of all parts of the realization are presented, such as the preparation of the CAD data, the
configuration of the FEM and the extraction of the relevant field data. In chapter 3.7.1 - 3.7.3
strengths and weaknesses of all three different approaches on litz wire level (analytic, PEEC, and
measured factors) are discussed with regard to their results for the power losses on system level.

3.2. SlicerPro approach

Simulations of litz wire systems have been reported to significantly deviate from measured data
[85], [31]. In the frequency range of power electronic systems (50 - 500 kHz) the simulation error
reaches up to 50 % and thus common simulation approaches provide no serious support during the
design period of such systems.
In chapter 2.2 and the corresponding subchapters different approaches for solving Maxwell’s equa-
tions have been presented and analyzed with regard to their accuracy of the simulation of conduc-
tors, in particular of litz wires. In this section, these methods are combined leading to a new
multi-scale approach for the three dimensional simulation of frequency depended power losses
in litz wire systems in the proximity to magnetic materials. The specific benefits of the different
numerical and analytical methods are merged to reach a closed workflow for arbitrary geometry
setups and high frequency litz wires with several hundreds of strands.
To get an overview over the task of different approaches and their interfaces, the entire workflow
is depicted in figure 3.1. A description of the overall concept is presented in this section, while

44
 3.2. SlicerPro approach

FEM

CAD
Import of standard geometry data
with simplified solid conductor

FEM
Excitation of the conductor by homoge-
nous current density and calculation of the
field distribution by a magnetostatic solver

SlicerPro

Extraction of Hext
Export of simulated magnetic field data at cuts along
the length of the conductor and extraction of Hext

Frequency
range

Analytic PEEC Measured Factors

• Calculation of skin and • Numerical calculation of • Usage of measured fac-
proximity losses by ana- power losses in realistic tors (λskin and λprox )
lytic methods [31][32] litz wire conductors [8] characterizing a litz wire
[31]

Figure 3.1: Workflow of the coupled simulation method combining the FEM on system level
with three solving approaches on litz wire level. Based on a user defined frequency and the
external magnetic field, analytical, numerical and measurement based methods enable the
calculation of the frequency-dependent resistance of the entire power electronic system.

45
CHAPTER 3. Multi-scale algorithm for litz wire systems

details about the substeps are provided in the following chapters.

The simulation process starts with the import of the geometry file in standard FEM programs.
Most programs support common CAD formats or even offer a direct interface for parametric ge-
ometry transfer. The geometric resolution of the inner structure of the conductor is not required,
because the surrounding magnetic field is equal as long as a circular solid conductor of the same
diameter is used with a homogenous current density (see chapter 2.3.3). This artificial electrical
boundary condition has to be defined on FEM level to match with the reality in litz wire systems.
In the next step, the magnetic field distribution in the entire simulation domain is calculated. In
case of constant material data within the regarded frequency range, simulations can be simplified
and speeded-up by using magnetostatic solvers. Based on these FEM results, the magnetic field
distribution on cuts along the length of the conductor is transferred to the SlicerPro program.
Within this separate program, the external magnetic field Hext is extracted and the data are pro-
vided to three different approaches calculating the power losses on litz wire level in a user defined
frequency range:

1. Within the analytical method limiting cases such as a perfectly twisted litz wire or a con-
ductor with parallel strands are calculated based on homogenization approaches (chapter
2.3.2). In addition, an advanced approach is implemented, combining the complex geo-
metric structure of realistic litz wires with the homogenization on varying geometry levels
[86].

2. The PEEC method enables calculations of power losses for realistic litz wires with arbi-
trary substructures. Varying pitch-lengths and bundle structures can be investigated by this
numerical approach. Due to the long calculation time, precalculated look-up tables are gen-
erated for a wide range of frequencies and external magnetic fields.

3. Based on highly accurate measurements of given litz wires, two extracted factors (λskin and
λprox ) enable the calculation of skin- and proximity losses [31].

Finally, the frequency-dependent losses are visualized and can be provided to thermal FEM simu-
lations, for instance.
The crucial point of the workflow is the initial replacement of the complex litz wire structure by
a simple solid conductor. As underlined in chapter 2.3.3, both conductor types have the same be-
havior on system level and therefore cause the same magnetic field distribution in the simulation
domain. This fact enables the splitting of the solving process: On one hand side, a simplified FEM
simulation in which the field distribution is carried out based on a solid conductor (yellow box in
figure 3.1). On the other hand, calculations of power losses on litz wire level are performed based
on the external magnetic field (grey box in figure 3.1).
This uni-directional coupling of the FEM with the three approaches on litz wire level yields the
frequency-dependent losses of the entire power electronic system. A feedback of the new spatial
current density distribution from the litz wire level to the system level can be neglected, as it has
been shown, that strand-level currents do not create bundle-level fields [73], [29].
In the next chapters, the workflow is described in detail according to figure 3.1.

3.3. Preparations on CAD level

In this chapter, the requirements of the geometry data in relation to the entire workflow are dis-
cussed.
The CAD data need to depict the regarded system with respect to its electromagnetic behavior.
Consequently, all non conductive and non magnetic parts are neglected to reduce computational

46
 3.4. Field simulation with FEM

resources. Another important issue concerns the support and simplification of the following mesh-
ing process. Apart from common quality standards in the design of geometries [87], the main aim
is a high resolution of the magnetic field along the length of the conductor:
• Geometric details within the litz wire conductor such as strand diameter or bundle structure
are neglected, because these details have no impact on the magnetic field distribution on
system level (as it has been shown in chapter 2.3.3). Consequently, litz wires are resolved
by a solid conductor of the same shape. However, it should be noted that wire diameters
without wrapping have to be considered.
In case of very basic meshing algorithms, the circular conductor needs to be discretised by
a polygon already in the CAD tool. Thus, a uniform distribution of nodes on the surface of
the conductor can be reached.

• In the center of the conductor, the internal magnetic field vanishes. To reach a high accuracy
of the (external) magnetic field at this defined position, the meshing algorithm has to be
forced to create grid points there. The easiest approach bases on an additional polyline
in the middle of the conductor. This geometric object is included in the meshing process,
resulting in nodes and edges along the polyline.

• Ferrite materials have significant influence on the magnetic field distribution. To reduce
deviations due to the singularity of the magnetic field at the edges of the solids, the edges
need to be chamfered.
These geometric adaptions are important to reach accurate results and do not require deeper CAD
skills nor time. In fact, most common FEM software tools also enable these minor corrections.
Moreover, a lot of tools even offer a direct interface between CAD and FEM programs. Based on
that, the influence of variations on geometry level (i.e. different number of windings or varying
ferrite geometries) can be investigated very comfortable.

3.4. Field simulation with FEM

In this section, the specifications on FEM programs are discussed with regard to the mesh genera-
tion and the electrical boundary conditions. Based on these settings, the magnetic field distribution
in the entire simulation domain is calculated.
In this thesis, tools of Ansys (Ansys Emag 16.0 and Maxwell3D Version 2015) are used. Both
programs differ significantly with regard to their solving and output capabilities and therefore rep-
resent the range of common commercial software. However, this should not be seen as a limitation
of the workflow - most electromagnetic software packages, even open-source, meet the require-
ments for using SlicerPro.
In the domain of meshing, there are big differences in the functionality of software packages, es-
pecially according to the resulting mesh quality. In a lot of cases tightly packed circular structures
need additional user efforts to avoid degenerated elements and to reach a fast convergence of the
meshing algorithm.
In a lot of simulations it has proved useful to define a fix number of discretisations of the circular
surface and along the length of the conductor. This is similar to the number of edges approximat-
ing a round conductor in 2D (chapter 2.4.2). With regard to the discretisation along the length
of the conductor, the structure should be divided into domains with lower or higher relevance:
Regions with curved windings or sections with strong inhomogeneous magnetic field need a finer
mesh to resolve the physical effects more precisely.
In figure 3.2, the meshing of the conductor is adapted to the geometric condition. The mesh of
the conductor cross section is swept along the length of the conductor with increasing number

47
CHAPTER 3. Multi-scale algorithm for litz wire systems

(a) (b)

Figure 3.2: Manually generated mesh for a choke in Ansys Emag (a). In relevant areas, i.e. near
the edges of the ferrite structure, the number of subdivisions is increased (see closed-up in (b)).

of subdivisions nearby edges of ferrite structures or curved windings (figure 3.2 b). The under-
lying meshing strategy requires user experience with the used algorithm and a defined schedule
for a chronological order of the meshing. Due to the requirement of a homogeneous mesh, one
meshed part already defines the mesh (and so the corresponding distribution of the elements) on
the boundary of all adjacent parts. Consequently, the winding structure has to be meshed first to
reach the required discretisation.
Another important issue concerns the geometric shape of the elements used. In tubular geome-
tries (i.e. pipes or conductors) with minor changes along the length of the structure, the usage
of polyhedrons instead of tetrahedrons reduces the amount of elements significantly without loss
of physical accuracy. For such a mesh structure, the number of subdivisions corresponds to the
number of evaluation cuts on which the power losses are calculated in the next step of the work
flow (visualized in figure 3.4).
Now, first FEM programs offer adaptive meshing in 3D. Based on an initial mesh the electromag-
netic system is solved and an error evaluation for the solved fields starts [88]. Depending on the
deviation between the total energy in the entire simulation domain and the definition at the bound-
aries, an iterative process of local refinement starts. In all steps, the elements with the biggest
energy error are split and the physical system is solved again. Finally, the refinement process is
stopped: Either the energy error falls below the user-defined accuracy threshold or the maximum
number of refinement steps is reached.
In Ansys Maxwell3D, this method is available for the magnetostatic and eddy-current solver.
Such meshing strategies enable non-expert users to run simulations even for complex geometries.
In case of complex geometries with unknown field intensities, the adaptive meshing is a comfort-
able approach - in particular in combination with HP computers and multi-core licenses. However,
the implementation of this meshing strategy only works with tetrahedral elements and results in a
very fine and unstructured mesh.
Although automatic meshing becomes more and more common, it has to be verified, that resulting
elements fulfill common quality standards for FEM simulations [87]. In figure 3.3 a) the mesh
after two adaptive refinement steps is visualized. The required accuracy of less than 1 % energy

48
 3.4. Field simulation with FEM

(a) (b)

Figure 3.3: Geometry of figure 3.2 meshed by using the adaptive refinement method in Ansys
Maxwell3D for the entire structure (a) and a close-up (b).

error per element was reached and similar to the manual meshing of figure 3.2, the curved parts are
meshed with a higher resolution. However, the element quality (ratio of the longest to the shortest
element edge) is worse, especially in straight parts of the circular structure (figure 3.2 b).
In the domain of conductors, manual sweeping operations are often more efficient, because the
circular structure can be resolved more accurately, especially in case of very close windings. How-
ever, it is not possible to estimate the minimum amount of discretisation elements along the length
of a conductor with regard to the final accuracy of the solution. Consequently, the amount of ele-
ments is mostly overestimated to avoid later remeshing.
In table 3.1 simulations with two geometric setups are compared with regard to adaptive and man-
ual meshing. In both cases, the calculation time of the manual meshing is significantly shorter,
even if the additional preparation time (appr. 10 min) is taken into account. While the adaptive
meshing requires two to three steps of remeshing and recalculation, the electromagnetic system
needs to be solved only once in the manual case. The manual approach overestimates the num-
ber of elements for the first setup (N = 12 solo) and fits good in the second one, compared to the
adaptive case with an energy error of less than 1 %. In case of N = 12 star, resulting field and loss
distribution is more complex and requires more computational resources. The amount of elements
and memory strongly depends on the geometric setup and the amount of refinement steps.
All in all, manual meshing enables the generation of a mesh much faster and with higher quality,
but requires a lot of experience of the user.
The simulation model has to take into account the physical conditions in real litz wire conduc-
tors by the boundary conditions. In addition to the standard constraints for 3D simulations, the
homogenous distribution of the current in the conductor has to be specified. In the software tool
Maxwell3D, the stranded option in the excitation setup meets this requirement.
In contrast, Ansys Emag requires an additional scripting (which is provided in appendix A.1) to
define this current density distribution. Initially, an arbitrary potential difference between both
endings of the wire is defined. Based on this boundary condition, the system is solved, resulting in
vectorized quantities for the field and current density distribution in the entire domain. In a second
step, only the absolute values of the current density inside the conductor are overwritten by the

49
CHAPTER 3. Multi-scale algorithm for litz wire systems

Table 3.1: Comparison of magnetostatic simulations using manual (Ansys Emag) and adaptive
meshing (Ansys Maxwell3D) for a coil with 12 windings. The first geometric setup consists of an
air coil (N = 12 solo), while the second setup has an additional ferrite star in the middle (N = 12
star).

N = 12 solo N = 12 star

Elements RAM Time Elements RAM Time

[1e6] [GB] [min] [1e6] [GB] [min]
Ansys Emag 2.57 41.6 55 2.75 43.9 69
Maxwell3D 2.10 31.6 145 2.74 50.3 209
(error < 1%) 2 refinement steps 3 refinement steps

defined current density value. The orientation of the vectorized quantity keeps the same. Con-
sequently, all nodes inside the conductor have the same current density and a realistic vectorized
orientation. Afterwards, all other results are deleted and the final solving process starts with the
homogeneous current density distribution.
In most applications in the domain of power electronic systems, materials have constant conductiv-
ity and permeability in the considered frequency range. Consequently, the simulation model can
be reduced to a magnetostatic problem. From a numerical point of view, magnetostatic solvers
have benefits with regard to the underlying system of equations (see chapter 2.2). The system
is much less complex compared to the one solved by eddy current solvers and consequently the
solving process is much faster and the results have a minor numerical error.
Based on that, in all following simulations the magnetostatic solver is used to calculate the mag-
netic field distribution. Even in simulations of transformers or coupled systems this approach is
suitable [89].

3.5. Extraction of magnetic field data

After solving the electromagnetic system with the FEM, field values are stored in all nodes and
provided for further post processing. For the following step in our workflow the magnetic field in
the domain of the conductor is required. The complexity of extracting these data strongly depends
on the export interface provided by the simulation tool. The usage of raw data (field values at the
evaluation nodes) should be preferred compared to interpolated values.
In case of Ansys Emag the magnetic field values are stored on the nodes of the elements. Because
of the special structure of the swept mesh (figure 3.4 a), nodes of the conductor can be associated
with cuts along the length of the conductor. The number of subdivisions in the meshing process
defines the number of cuts. The mesh specification for the first cross section defines the number of
nodes inside each cut. With the help of a short scripting (see appendix A.2) the data for the total
magnetic field (h_tot) for all relevant nodes are exported (table 3.2). First, nodes of the starting
cut, which corresponds to the conductor surface at the boundary of the simulation domain are
exported. Afterward the scripting proceeds with all adjacent cuts. In the output file new cuts are
initialized by the header text line (“X Y Z ...”). The chosen export format for the numbers, has to

50
 3.5. Extraction of magnetic field data

(a) (b) (c)

Figure 3.4: (a) shows the structured mesh of a conductor in Ansys Emag, with nodes on the
surface (b) and corresponding magnetic field values (c).

Table 3.2: Output file for the total magnetic field h_tot in cuts along the length of the conductor
created by an APDL scripting based on an Ansys Emag simulation. The first three columns
provide the position within the coordinate system (in [m]), followed by the corresponding values
A
of the magnetic field components (in [ m ]). The provided data correspond to the first two nodes
each for two cuts of a conductor.

POS_X POS_Y POS_Z HX HY HZ

5.848e-02 0.000e+00 1.500e-01 -4.693e+00 -1.081e+02 -5.038e-01
5.553e-02 0.000e+00 1.500e-01 4.615e+00 9.750e+01 7.972e-01
..
.
..
.
POS_X POS_Y POS_Z HX HY HZ
5.681e-02 -2.644e-04 1.489e-01 -2.131e+01 3.393e+00 -6.049e-02
5.836e-02 -4.434e-04 1.489e-01 -4.175e+01 -1.036e+02 -1.128e+00
..
.
..
.

take into account, that the magnetic field values and even the position of the nodes change in the
dimension of several magnitudes. While the first three columns provide the position of the nodes,
the last three columns list corresponding values of the magnetic field - separated in its shares for
x, y, and z direction.
To reach a high spatial resolution of the magnetic field distribution inside the conductor also the
internal node data of cuts are exported. Consequently, the total amount of nodes in one cut (appr.
100 using a discretisation with 20 edges such as in figure 3.4 a) and therefore the evaluation time
is very high: In the coil setups used in table 3.1 (12 windings and 3000 cuts) 45 min are required.
To reduce the time for data export, one can restrict the export to the surface nodes only (figure 3.4
b). The resulting data can be used in a very convenient extraction approach [9] described in the
following chapter.
In contrast to Ansys Emag, Ansys Maxwell3D (Version 15.0) does not provide any direct access
to raw data (such as vectorized quantities calculated in one node). Nevertheless, a comfortable
approach for exporting magnetic field data was developed with the help of a polyline (also known
as a polygonal chain or piecewise linear curve) at the center of the conductor (figure 3.5 b).
Based on this additional geometric part, the coordinates of points along the polyline are known and
the magnetic field can be evaluated on them. More precisely - the magnetic field data are internally
determined by a linear interpolation between results on the nodes of the unstructured mesh. Based

51
CHAPTER 3. Multi-scale algorithm for litz wire systems

(a) (b) (c)

Figure 3.5: (a) shows the unstructured mesh of a conductor in Maxwell3D, with a polyline
(piecewise linear curve) in the center of the conductor (b). With the help of scripting a cut-based
output (such as table 3.2) can be generated based on extrapolated data (c).

on the standard export functionality in Maxwell3D, the magnetic field along the polyline can be
saved. The output file is presented in table 3.3 and consists of two initial header lines followed
by the simulation data. The first three columns consist of the position of the evaluation point
(corresponding to the length along the polyline) and the following one contains the corresponding
magnitude of the magnetic field.

Scalar data LineValue(Line(Coil), Mag_H)

NumElems 10000
5.350e-02 0.000e+00 5.700e-02 2.771e+01
5.350e-02 0.000e+00 5.673e-02 2.800e+01
5.350e-02 0.000e+00 5.646e-02 2.841e+01
5.350e-02 0.000e+00 5.620e-02 2.930e+01
5.350e-02 0.000e+00 5.593e-02 3.007e+01
5.350e-02 0.000e+00 5.567e-02 3.056e+01
5.350e-02 0.000e+00 5.540e-02 3.130e+01
.. .. .. ..
. . . .

Table 3.3: Output file of Maxwell3D using an additional polyline in the middle of the conductor.
Both initial rows consist of the type of output parameters and the amount of elements in the file.
In the following the simulation data are provided: (x, y, z) coordinates (in [m]) and the
A
corresponding magnitude of the total magnetic field (in [ m ]).

Due to the location of the polyline in the center of the cross section of the conductor, the
internal magnetic field vanishes at these points. Consequently, the total magnetic field at these
points is equal to the external magnetic field. The export of the vectorized external magnetic field
would be possible, but useless due to symmetry reasons. All following evaluations on litz wire
level base on the external magnetic field of single cuts. Consequently, the direction of the external
magnetic field can be neglected, because the calculation of proximity losses only requires the
magnitude of the external magnetic field.
However, in this workflow the external magnetic field values base on interpolation at only one
point of the cut. Consequently, the numerical error is higher compared to the one in evaluation
nodes and moreover, a spatial resolution of external magnetic fields within a cut is not possible. In
case of very tight winding structures, the external magnetic field is strongly inhomogeneous even
at cross sections with only some millimeter diameter (figure 2.4).
With the help of an additional python scripting, cut planes and corresponding points are calculated
along the polyline. The coordinates of these points (figure 3.3 b) are similar to the ones from the

52
 3.6. Calculation of external magnetic field Hext

structured mesh of Ansys Emag. Finally, the vectorized total magnetic field is evaluated on all
these points by Maxwell3D (figure 3.5 c).
Both FEM programs enable the extraction of spatial magnetic field data of the conductor. All
following steps in the workflow of SlicerPro base on the standardized export files containing the
normed data on a polyline (table 3.3) or the vectorized data on cut level (table 3.2).

3.6. Calculation of external magnetic field Hext

After the termination of the FEM simulation, the magnetic field data of the conductor are provided
and the external magnetic field Hext needs to be extracted. In the last chapter two different output
formats are presented: normalized data on a polyline (table 3.3) and vectorized data on cut level
(table 3.2).
~ ext are considered based on the different
In the following, varying strategies for the calculation of H
raw data. The accuracy of the resulting external magnetic field values is discussed subsequently
in the verification section with regard to the required effort.

Case A - Normalized data on a polyline (“Hext,poly ”)

In case of a circular conductor, the magnetic field caused by the conductor itself vanishes in the
center of the conductor. Consequently the magnetic field in the center of the conductor is equal to
the external magnetic field. Based on an additional polyline on CAD level (see chapter 3.3) the
external magnetic field can be extracted directly in evaluation points along this line (Maxwell3D
approach in the last chapter).
However, the usage of single values along the length of the polyline bears the risk of numerical
errors. With the help of additional cleaning approaches such as discussed in [55], [90], [91] the
quality of the results can be improved.
In the following, results based on the normalized data on a polyline are denoted as Hext,poly .

Case B - Vectorized data on cut level

In the following, several approaches for the extraction of the external magnetic field are presented
with regard to vectorized data on cut level.
1. Linear average on the cross section boundary (“Hext,av1 ”)
This approach is limited to the nodes on the boundary of the cuts (therefore data in table
3.2 need to be filtered). The simulated magnetic field consists of different components,
which are shown in figure 3.6 for the simplified case of four nodes on one cut. The total
magnetic field (H ~ tot based on the FEM simulation) is visualized by green arrows and is the
superposition of the internal magnetic field H~ int (black arrows) and external magnetic field
~
Hext (blue arrows). Due to the tangential orientation of H~ int , only the external magnetic field
is relevant in terms of the proximity losses.
An easy approach to extract the external magnetic field H ~ ext uses the vectorized average
of the magnetic field (component-by-component for HX, HY, HZ) of one cut. Assuming
an equidistant distribution of all nodes n on the surface of the conductor, the share of the
internal magnetic field vanishes in this average vector [55]:
1 n ~ 1 n ~ n
~ ext,i ) = H
~ ext .
∑ Htot,i = ( ∑ Hint,i + ∑ H (3.6.1)
n i=1 n i=1 i=1
| {z }
equal to 0

The resulting vector describes the linearly averaged external magnetic field in one cut based

53
CHAPTER 3. Multi-scale algorithm for litz wire systems

Figure 3.6: In the left figure the FEM mesh of a circular conductor (discretized by 20 edges) is
visualized. For reasons of clarity, resulting magnetic field distribution is simplified to four points
on the right sketch. The total magnetic field H~ tot is provided by the FEM on all nodes and results
from the superposition of the internal (black) and external (blue) magnetic field components. By
calculating the vectorized average over all H ~ tot,i at the surface, the internal magnetic field
components vanish and the averaged external magnetic field remains.

on the boundary nodes. Due to the symmetry of the conductor, only the magnitude has to
be considered as relevant for the loss calculation to follow. We define therefore

~ ext k
Hext,av1 := kH

for this averaging approach.

2. Quadratic spatial average (“Hext,av1 ”)

The external magnetic fields have a quadratic influence on proximity losses:
l 2
Pprox = Hext · Ds (3.6.2)
κ
Hence, a quadratic averaging strategy would fit better in terms of loss calculation. In contrast
to the former averaging strategies, this method requires the explicit determination of H ~ ext .
~ ext is calculated by
To this end, H

~ ext = H
H ~ tot − H
~ int (3.6.3)

for all nodes of a cut to reach a higher resolution of the spatial distribution. Therefore, in
~ int is required.
contrast to the linear averaging, the calculation of H
The internal magnetic field in a 2D cut can be calculated analytically. Assuming a node with
a distance ρ to the middle of the conductor, the internal magnetic field

~ int (ρ) =~eϕ Iρ

H (3.6.4)
2πrc2

has a tangential orientation and only depends on the excitation current I. Due to the sweep-
ing strategy in the meshing process (in AnsysEmag), all nodes in one cut plane have in the
same relative position along the length of the conductor, and therefore the same internal
magnetic field magnitude. By considering the orientation of different cuts to each other,
the vectorized quantities of H~ int can be determined with linear transformations (projection

54
 3.7. Calculation of external magnetic field Hext

method). In the next step, the external magnetic field for all n nodes is normalized, squared
and averaged:
1 n ~ ~ int,i k2 = Hext
2
∑ k |Htot,i {z
n i=1
−H
}
. (3.6.5)
~ ext,i
H

In the following, the external magnetic field calculated by this quadratic spatial averaged
method is named Hext,av2 . To reach reliable results, the nodes (and therefore the elements)
need to be distributed equally on the cross section. In case of coarse meshing inside the
conductor additional weighted factors are required to avoid numerical errors.
3. Spatial 3D data
Up to now, only data of separate 2D cuts are considered. Assuming a very fine resolution
of the conductor in the length direction, data of adjacent cuts can be prepared like in the
previous approach and combined to a 3D data set of the magnetic field. This spatial vol-
umetric resolution of the external magnetic field yields the most realistic field distribution.
However, because the methods on litz wire level (such as PEEC) use piecewise linear con-
ductor geometries, these 3D data need to be straightened. Depending on the curvature of the
conductor in the simulation model, this may result in distortions and additional numerical
errors. Consequently, this approach is not yet regarded in further investigations.

Verification
The different strategies for determining the external magnetic field differ strongly with regard to
the computational effort and the required data from the FEM simulation. In case of Hext,poly nearly
no additional preparation of the provided simulation data (table 3.3) is needed, while Hext,av2 re-
quires several thousand of matrix-vector operations. Therefore, the costs and benefits of varying
methods need to be checked.
In figure 3.7 all three strategies are applied on two geometry setups. Based on the FEM result of a
choke (figure 3.3) with 0 mm and 1 mm air gap Hext,poly , Hext,av1 and Hext,av2 are calculated along
the length of the conductor. The symmetric shape of all resulting curves (with a symmetry axis
of length ≈ 0.95 m) corresponds to the symmetry in the geometric setup. Due to the same (but
mirrored) curves on both sides of the symmetry axis, the FEM demonstrates its high accuracy with
regard to the field distribution in complex geometries (even though it is not a rating of the absolute
values).
The agreement between the results of the different methods is good, except for the center of the
regarded range, which corresponds to the connection of both ferrites or the air gap, respectively.
In case of the geometry setup without air gap the external magnetic field of Hext,poly and Hext,av1
A
even vanishes, while Hext,av2 reaches approximately 50 m . The field distribution in the corre-
sponding cuts has a very high inhomogeneity in relation to the absolute values of the magnetic
field strength. Such effects can only be resolved correctly (with regard to the proximity losses)
by spatial and quadratic interpolation. The same applies to the deviation in the peaks of the 1 mm
air gap geometry. Due to high local variations of the magnetic field in proximity to an air gap,
minor changes in the position of the regarded evaluation points result in large differences between
the three methods. However, the amount of cuts with a significantly high deviation between the
different interpolation strategies is small. Regarding the quadratic average of the magnetic field
over all 3000 cuts (which is proportional to the averaged losses in the system), the results of all
three methods differ only by 1 % to 3 %.
Consequently, due to the longer export time and the higher complexity to gain Hext,av2 , the polyline
methods (Hext,poly ) and the linear average on the cross section boundary (Hext,av1 ) should be given
preference.

55
CHAPTER 3. Multi-scale algorithm for litz wire systems

Figure 3.7: External magnetic field out of the three determination strategies versus the length of
the conductor using an excitation current of 1 A. The choke of figure 3.3 is used as benchmark
geometry with an air gap of 0 mm (top) and 1 mm (bottom).

3.7. Litz wire simulation

The main challenge in the design of multi-scale approaches is to split the solving workflow in a
way that the specific benefits of varying methods can be combined. Moreover, the data exchange
needs to be standardized to enable the usage of different methods. In the presented workflow the
system level is connected with the litz wire level by the Hext data. The origin of the data just as the
interpolation or the cleaning strategy is irrelevant, as long as the data are provided in a standardized
format consisting of the length (in [m]) and corresponding magnetic field strength (such as used
for the visualization in figure 3.7). Consequently, methods and programs on one level can be
replaced without influencing the other level. This was already demonstrated on system level by
the usage of two different FEM solvers - Maxwell3D and AnsysEmag. Each one has its specific
pros and cons (table 3.1) and provides results in the same output format.
On litz wire level, in addition to the provided Hext data, a frequency value or range is required
to enable the calculation of power losses. Initially the frequency range is defined from 1 kHz up
to 1 MHz, but can be adapted depending on specific applications. As already shown in chapter
3.4, the resolution of the magnetic field data (and therefore the distance between adjacent cuts)
depends on the amount of discretizations in the meshing process (AnsysEmag) or of the polyline

56
(Maxwell3D). In the presented examples the distance between neighboring cuts was less than
1 mm. Considering the evaluation process on litz wire level, only one cut each is used for the loss
calculation. Consequently, the calculated power loss requires a rescaling to take the distance to
neighboring cuts on system level into account.
While on system level only one numerical methodology (FEM) is available (due to the complexity
with regard to the geometry and the materials), several suitable solving approaches can be used on
litz wire level. To cover common engineering requirements as good as possible, three solvers with
different alignments and focus have been implemented:

• Analytical method
Based on rough data of the litz wire (conductor diameter, amount and diameter of the
strands) the limiting range of power losses (corresponding to ideal twisted and parallel
stranded conductors) for any suitable internal structures is calculated by using homogeniza-
tion. With the help of more detailed information about the inner structure (amount and pitch
length of varying geometry levels) the homogenization can be applied on this substructure
yielding higher accuracy in terms of realistic litz wires.

• PEEC method
With the help of detailed information on the internal structure of the litz wire conductor, a
numerical model of the conductor is built up and solved, considering all physical effects.
The method is very costly with regard to the computational resources and the calculation
time.

• Measured factors
A litz wire conductor is measured in a special test-setup and two factors are extracted. Based
on these factors, analytical methods are enhanced. In contrast to both previous methods, this
approach requires a prototype of the conductor.

In SlicerPro the analytical, numerical, and measured approaches are implemented in a modular
structure (figure 3.1). All approaches can be chosen separately, but require varying input informa-
tion. These information and further details about the implementation and interfaces are provided
in the following subchapters for all three approaches.

3.7.1. Analytic approach

In many analytical approaches simplifications with regard to the geometric setup or the physical
accuracy are required. In chapter 2.3.2 an analytical method for litz wires was discussed based on
a homogenized conductor. The resulting equations are very convenient, but depict an unrealistic
system either of a perfectly twisted or of a parallel stranded conductor. However, common litz
wires show a behavior somewhere between both extreme cases.
In this chapter an analytical approach for realistic litz wire conductors is presented based on the
work of Sullivan [92], [32]. Unfortunately, technical terms in these publications are not consistent
with the ones presented in chapter 2.3.2 for the analytical method of Albach [31]. To get a better
overview on the new approach and the partitioning of the total losses, table 2.2 is enhanced: Based
on all seven theoretical loss shares, corresponding formulas for the analytical model of Albach
(limiting cases) and Sullivan (more realistic litz wires) are merged in table 3.4.
In general, the litz wire power losses in the method of Sullivan are split into four different
shares: Losses caused by skin- and proximity effect, each on strand and bundle level.
The main strategy of this approach is visualized in table 3.5 by a sketch of the current density
distribution within the conductor: Skin and proximity losses are calculated by the superposition of
the corresponding effect on different geometry levels. While the loss calculation on strand level
 Albach [31] Parameter Sullivan [32]
4
 2

p3 p2
4l 4 + 4π2 r 32 − r2
RDC = RDC Rdc,twisted r3 p3
σcu πdS2 N = (3.7.1)
Rdc,untwisted 6π2
parallel stranded (par) vs. ideal twisted (id) conductor Pskin_strand 2
Pskin = RDC Irms F0 F1 F2 . . .
r2 I1 (xS )
  
I0 (xS )
Pskin,id = RDC I 2 1 Re xS + N2 S
| {z } | {z }
2 I1 (xS ) rC2 I0 (xS ) Pskin_bundle strand bundle
| {z }
Pprox_int_strand
| {z }
skin_strand prox_int_strand Superposition of internal and external field:
 
I0 (xC ) ~ tot = H
H ~ int + H
~ ext
Pskin,par = RDC I 2 21 Re xC
I1 (xC )
Geometry level i=0 (strand level):
with normalized skin parameter
Rl
√ Pprox,strand = ∑∀N G0 |Htot (z)|2 dz with
CHAPTER 3. Multi-scale algorithm for litz wire systems

xS = (1 + j) π f µ0 σcu rS Pprox_int_bundle 0
√ πd 4 ω2 µ02 σ q
xC = (1 + j) π f µ0 σlitz rC G0 = 0 and |Htot (z)| = Htot,x 2 (z) + H 2 (z)
tot,y
128
Geometry level i≥1 (bundle level):
Pprox,bundle i =
Hext = const.

Rl   2
Pprox_ext_strand Gi ~ tot,x (z) + sin(kz)H
~ tot,y (z) dz +
l cos(kz)H
Pprox,id = l H 2 · NDid with
σcu ext 0
  Rl   2
xS I1 (xS ) Gi ~ tot,y (z) + sin(kz)H
cos(kz)H ~ tot,x (z) dz
Did = 2πRe l
I0 (xS )  
0
2π xlitz I1 (xlitz ) Pprox_ext_bundle
Pprox,par = l 2
σlitz Hext · NDpar with Dpar = Re πd14 ω2 µ02 σlitz,i
N I0 (xlitz ) with Gi = (3.7.2)
128
Table 3.4: Varying composition of loss shares in litz wire conductors according to Albach [31] and Sullivan [32]. While in the left approach the
separation of the table coincides with the separation of the method, the loss shares which correspond to equations in the right approach are highlighted by
green and blue.

58
 3.7. Litz wire simulation

Table 3.5: Sketch of the current density distribution caused by the skin and proximity effect on
different geometry levels with a visualization of corresponding direction (into or out of the plane
of the drawing) by red and blue. In the analytical approach of Sullivan the frequency dependent
power losses are calculated based on the superposition of both effects on all geometry levels.

strand level (sub)bundle level bundle level

i=0 i=1 i=2

skin effect

proximity effect

(i=0) is similar to standard approaches (such as equation (2.3.14) or (2.3.16)), power losses on
bundle level (i ≥1) are determined based on a solid conductor of the same geometric shape like
the bundle assuming a homogenous material.
Considering the implementation of Sullivan following has to be taken into account:
• Influence of twisting on the RDC
Standard analytical approaches for calculating the (untwisted) direct current resistance of
RDC, twisted
litz wires (such as equation 2.3.18) are enhanced by the factor RDC, untwisted
. Based on equation
(3.7.1) the varying lengths of strands due to twisting are taken into account by using the
outer radius r of the bundle and corresponding pitch length p.
• Skin losses on varying geometry levels
The total skin effect losses only consist of losses caused by the skin effect - internal prox-
imity losses (like in the approach of Albach) are not included.
The influence of the skin effect on different geometry levels (such as strand level, first bun-
dle level, second bundle level, etc.) is calculated by adapting the skin effect factor Fs . This
factor is already initialized in equation (2.3.15) and corresponds to the lowest geometry level
F0 within the Sullivan approach. For all higher geometry levels i (such as strands combined
in a bundle and small bundles combined in a larger bundle) a corresponding skin factor Fi is
calculated.
This strategy is similar to the homogenization approach of Albach for the calculation of the
skin losses in a parallel stranded conductor, which was introduced in chapter 2.3.2. How-
ever, the corresponding equation (2.3.30) provides the losses for the entire conductor, while
in the new approach of Sullivan the skin effect factor is calculated with a restriction to the
partial geometries (such as bundles). Consequently, the factor is given by
 
1 I0 (xCi )
Fi = Re xCi (3.7.3)
2 I1 (xCi )
with the normalized skin parameter xCi of the regarded i’s geometry level. In addition to the

59
CHAPTER 3. Multi-scale algorithm for litz wire systems

(a) (b)

(c) (d)

Figure 3.8: Litz wire conductor consisting of 3 bundles à 7 strands exposed to a homogenous
external magnetic field (a). One twisted bundle of the litz wire (b) can be transformed into a
straight bundle with a twisted external magnetic field along the length of the conductor (c). This
bundle is approximated by a conductor of the same geometric shape with homogenized material
(d).

homogenized averaged conductivity (equation (2.3.28)) the greater length of the strands due
to the twisting needs to be taken into account [93].
Based on that, the total skin losses are calculated by the superposition of losses caused by
the skin effect on different geometry levels (upper figures in table 3.5):
2
Pskin = RDC Irms · F0 · F1 · F2 . . . (3.7.4)
| {z }| {z }
strand level bundle level

• Proximity losses on varying geometry levels

The calculation of the proximity losses based on [32] exhibits fundamental differences to
previous methods.
Firstly, the magnetic field used in all calculations bases on the superposition of the internal
and external magnetic field, given by
~ tot = H
H ~ int + H
~ ext . (3.7.5)
The magnetic field (in a single strand just as in the homogenized solid conductors) is as-
sumed to be constant on the cross section and equal to the field in the center position.
Consequently, every single conductor is exposed to a varying field
q
2 (z) + H 2 (z),
|Htot (z)| = Htot,x (3.7.6)
tot,y

which only depends on its length in z direction.

The proximity losses on strand level (i=0) consist of the losses of all N strands of the litz

60
 3.7. Litz wire simulation

wire. Based on the individual curve l of each strand within the conductor, these losses are
calculated by

Zl
πd04 ω2 µ20 σcu
Pprox,strand = ∑ G0 |Htot (z)|2 dz with G0 = . (3.7.7)
∀N 128
0

The usage of G0 is very common in a lot of publications and yields a very accurate approx-
imation of σ1Cu Did , the exact analytical result used in the Albach approach.
With regard to the calculation of the proximity losses on bundle level, the approach of Sul-
livan is much more complex compared to the Albach approach. Due to the complexity of
the proximity effect within the conductor, each bundle is considered as a cylindrical con-
ductor of the same shape with a homogenized electrical conductivity σlitz (equation 2.3.28
and [94]). Just as the previous approach for the skin losses, the method of Sullivan bases on
the same homogenization approach like Albach. However, this simplification is applied at
each geometry level separately, instead of only once on the entire conductor.
Moreover, in terms of the mathematical integration, the point of view is changed: In gen-
eral, the magnetic field is assumed to be constant and strands in the conductor are twisted,
corresponding to a change of the x-y coordinates along the z-length of the conductor (figure
3.8 a, b).
In this method, the x-y position of the strands is assumed to be fixed and the magnetic field
changes its position along the length of the conductor. This approach corresponds to a co-
ordinate transformation to a local coordinate system along the center of a parallel stranded
conductor (figure 3.8 c). Moreover, the length of this line conductor has to be adapted, be-
cause of the larger length due to the twisting on bundle level (figure 3.8 b).
By neglecting the inner structure, the bundle can be approximated by a solid line conductor
with the same bundle diameter and homogenized material properties (figure 3.8 d). In the
resulting Pprox,bundle this coordinate transformation is taken into account by scaling the total
magnetic field with trigonometric functions along the length of the bundle (equation (3.7.2)
in table 3.4).

First results in [32], [95], [96] demonstrate the great potential of this sophisticated analytical
approach. However, all litz wires used in these publications only consist of several dozen strands
in a simple twisted structure. The transfer of this approach to high frequency litz wires is very
challenging, because the determination of the input parameters is not unique:

• The homogenized quantities have to be calculated based on the geometry data of the litz
wire, such as strand diameter, pitch lengths, and amount of strands of varying bundle levels.
Even though each strand can be assigned to the varying bundles, the same is difficult for the
surroundings. This separation is not unique, but crucial for the resulting parameter.

• The packaging of strands in bundles often results in a non-circular shape of the bundles.
However, the analytical approaches are only valid for round geometries.

• In case of non-circular bundle structures, the position of the center position, which is used
for the integration of the proximity losses in not defined.

Despite all these challenges, a first version of this analytical approach is implemented in the
SlicerPro workflow. However, these results are not yet reliable for HF litz wires and further re-
search activities are currently in progress. Consequently, this analytical approach is not validated
in the verifications provided in chapter 4.

61
CHAPTER 3. Multi-scale algorithm for litz wire systems

3.7.2. PEEC approach

The PEEC method and especially the implementation of Fastlitz [8] provides an accurate resolu-
tion of both, the geometry of litz wires and the superposition of the corresponding electromagnetic
fields.
In chapter 2.5 the theoretical background of this numerical method was presented and its realiza-
tion was discussed in chapter 2.5.1 based on Fastlitz for one sample of a conductor. Moreover,
the quality of the simulation results has been documented in various publications [97], [72]. How-
ever, all these examples use litz wires with only a minor amount of strands. In the following
sections, requirements and vulnerabilities are discussed with regard to high frequency (HF) litz
wires consisting of some hundreds of strands in different bundle levels.

Geometry
Initially, the user has to define a variety of input parameters for the design of the litz wire geometry
(table 3.6). Based on that, a mathematical algorithm (implemented in twist.m and maskHelix.m in
the software package Fastlitz) generates curves for the strand and bundle structure.
However, real litz wires are tightly coated, which results in a high copper fill factor (equation
2.3.19). Considering over 150 litz wires of the manufacturer Pack [7], 2/3 of the conductors have
a fill factor larger than 50 % (figure 3.9). In contrast, almost 90 % of the simulated litz wires in
Fastlitz yield a fill factor of less than 50 % (for more than 1500 test wires). These results indicate
that a realistic resolution of tightly coated litz wires will be problematic for the twisting algorithm.
In figure 3.10 a real litz wire conductor with 420 x 0.1 mm strands (a) and corresponding simula-
tion models with different twisting structures (b) - (e) are presented. The experimental conductor
has a diameter of 2.95 mm and consists of 14 bundles à 30 strands, which are all isolated against
each other by an insulation layer with a thickness of 0.01 mm. Due to the coating of the conductor,
the resulting copper fill factor reaches 48 %. Even though this value is rather below the average
of experimental conductors, the simulation models b) - e) only yield copper fill factors between
30 % and 40 %. In the four models generated by the twisting algorithm of Fastlitz, the 420 strands
of the litz wire are separated in 10, 12, 14, and 15 bundles, which have a much larger distance
to each other compared to the distances between the strands. All modelled geometries are very
fluffy and the diameter of the regarded 14 bundle structure differs by 10 % between experiment
and simulation.
Although all simulated geometries meet the main requirement (no intersection of strands along the
length of the conductor), the defined distances between adjacent strands (parameter insu) are only
fulfilled between strands in one bundle. Even the closest distance of adjacent strands consisting of
varying bundles, is largely 2 - 3 times bigger than the defined parameter.

Table 3.6: Example of geometric input parameters used for Fastlitz. Based on these definitions, a
litz wire conductor consisting of 5 bundles à 20 strands with a diameter of 0.1 mm each is
created. The total simulation length of 30 mm is discretized by 60 intervals and the pitch length
on strand and bundle level is 30 mm.

len = 0.03; % [m] length of wire

pitch = [0.03 0.03]; % pitch length in [m] on each level
ns = [20 5]; % number of strands at each level
srad = 1e-5/2; % [m] strand radius
insu = 0.1; % (diam + insulation) / diam - 1
drad = 3; % discretization param
odd = true; % filament at center?
ndiv=60; % number of divisions

62
 3.7. Litz wire simulation

Figure 3.9: Probability distribution of the copper fill factor for real litz wires (from the
manufacturer Pack [7]) and simulated ones.

To sum up, Fastlitz primarily focuses on the implementation of the PEEC method, the geometric
construction is considered to be of minor importance. Even though all design parameters of a real
HF litz wire are taken into account, the conductor diameter in the simulation exceeds the experi-
mental one in all cases. The exact values differ strongly depending on the amount and structure
of the bundles. In case of litz wire structures with two or even more bundle levels, the deviation
increases up to 25 %. Due to the fact that the skin and the proximity effect differ depending on the
geometric structure, these deviations between the experimental conductor and the litz wire model
lead to an increasing error of the simulation.
A minor restriction additionally results from the vectorized definition of the strand and bundle
structure [82], such as ns = [20 5] in table 3.6 for a conductor consisting of 5 bundles à 20 strands.
The entire substructure of the litz wire conductor is defined from global bundle level down to local
strand level (by ns and pitch parameter). Based on these definitions, all properties of structures in
one geometry level are equal. Consequently, litz wires consisting of different amount of strands in
different bundles cannot be resolved with Fastlitz.

Discretization
Based on the formulation of the PEEC method, only conductive structures have to be considered.
The discretization along the length of the conductor and the partition of the cross section (figure
2.15) are defined by ndiv and drad (equal to k in figure 2.15), respectively. The amount of both dis-
cretizations is crucial with respect to the accuracy of the solution and the required computational
resources. Common litz wire systems are too complex for reference simulations with the FEM
or analytical approaches. Consequently, setups with a higher amount of discretizations (ndiv = 90
and drad = 8) are used as a benchmark with regard to the power losses. This approach does not
provide absolute values for the estimated error, but it enables meaningful conclusions about the
accuracy.
In table 3.7 the normalized deviation of the power losses to the reference setup of two different
litz wire conductors (225 and 420 strands à 0.1 mm diameter) is provided for varying discretiza-
tions of ndiv (30, 45, 60 per length of 30 mm) and drad (2, 4, 6). All setups are calculated at a
A
frequency of 100 kHz and a homogenous external magnetic field of 100 m , orthogonally directed
to the length of the conductor.

63
CHAPTER 3. Multi-scale algorithm for litz wire systems

2 .9 5 m m
(a)

(b) (c) (d) (e)

Figure 3.10: Microscopic picture of the cross section of a real litz wire conductor [7] with
420 x 0.1 mm strands (a) and corresponding geometry in Fastlitz consisting of 10, 12, 14, and 15
bundles (b-e).

It can be seen that the amount of discretizations of the cross section (drad) has a significantly
higher influences compared to the discretization along the length of the conductor (ndiv). At least
drad = 4 is needed to reach a deviation of approximately 1 % in both setups. By taking the calcu-
lation time (provided on the right part of table 3.7) into account, a discretization of ndiv = 30 and
drad = 4 is the most efficient configuration for the regarded setups. The results of both litz wires
are sufficiently precise to reach less than 1 % deviation from the reference simulation in less than
one, respectively, two hours of calculation time on an HPC computer (16 cores Intel(R) Xeon(R)
CPU E5-2667 v2 @ 3.30 GHz). A significantly higher accuracy would require an increase of both
parameters and would require a much higher computational effort (factor 4 to 6).
However, additional refinements are necessary in case of higher frequencies or stronger external
magnetic fields. To quantify the accuracy more generally, varying discretizations of drad and ndiv
A A A
are considered for minor, medium, and high external magnetic fields (10 m , 100 m , and 1000 m )
up to 1 MHz for a 420 x 0.1 mm litz wire. The evaluation is split in a variation of drad from 2
to 6 (figure 3.11 a) and ndiv from 30 to 60 (figure 3.11 b), respectively, to detect the frequency
dependent influence of both refinements each.
In both analyses the reference setups contain additional refinements of the variation parameter
(drad = 8 in (a) and ndiv = 90 in (b)), while the parameter which is not changed stays at its mini-
mum value (ndiv = 30 and drad = 2, respectively) to reduce the calculation time.
In the frequency range up to 200 kHz the deviations due to the drad variation are significantly
larger than the deviations due to the variation of ndiv. This difference corresponds to a discrep-

64
 3.7. Litz wire simulation

0 n d iv = 3 0 , d ra d s w e e p
1 0

d e riv a tio n to re fe re n c e (d ra d = 8 , n d iv = 3 0 )
-1
1 0

-2
1 0

-3
1 0 H - fie ld , D is c re tis a tio n
ex t
1 0 A /m , d ra d = 2
1 0 0 A /m , d ra d = 2
-4
1 0 1 0 0 0 A /m , d ra d = 2
1 0 A /m , d ra d = 4
1 0 0 A /m , d ra d = 4
-5 1 0 0 0 A /m , d ra d = 4
1 0
1 0 A /m , d ra d = 6
1 0 0 A /m , d ra d = 6
1 0 0 0 A /m , d ra d = 6
-6
1 0
1 k 1 0 k 1 0 0 k 1 M
f [H z ]
(a)

drad = 2 , ndiv sweep

100
Hext - field, Discretisation
10 A/m, ndiv=30
derivation to reference (drad=2, ndiv=90)

100 A/m, ndiv=30

10-1
1000 A/m, ndiv=30
10 A/m, ndiv=45
100 A/m, ndiv=45
10-2 1000 A/m, ndiv=45
10 A/m, ndiv=60
100 A/m, ndiv=60
1000 A/m, ndiv=60
10-3

10-4

10-5

10-6
1k 10k 100k 1M
f [Hz]
(b)

Figure 3.11: Frequency dependent deviation of the power losses for different discretizations of
drad (a) and ndiv (b) for varying external magnetic fields (colorized). Simulations with drad = 8,
ndiv = 30 (a) and drad = 2, ndiv = 90 (b) are used as reference setups for a 420 x 0.1 mm litz wire
exposed to varying external magnetic fields.

65
CHAPTER 3. Multi-scale algorithm for litz wire systems

Table 3.7: Simulation of two litz wires (225 x 0.1 mm and 420 x 0.1 mm) for different
A
discretizations along the length (ndiv) and of the cross section (drad) at 100 kHz and 100 m . On
the left, the normalized deviation versus a highly accurate setup (ndiv = 90 and drad = 8) is given
and the calculation time is provided on the right.

error time
( norm. deviation in [%] [min]
to drad = 8, ndiv = 90)
ndiv \ drad 2 4 6 2 4 6
30 2.5 0.6 0.9 8 44 170
litz wire
225

45 2.5 0.6 0.4 12 60 162

60 2.6 0.5 0.3 14 63 257
30 2.9 0.9 1.4 16 100 477
litz wire
420

45 2.7 0.7 0.4 24 129 325

60 2.8 0.6 0.3 29 136 584

ancy of several orders of magnitude. Consequently, ndiv has only minor importance with regard
to the accuracy up to 200 kHz - above, both discretization parameters have similar influence. Due
to changes of sign within the regarded frequency range, the normalized deviation shows several
outliers in both analysis.
Regarding figure 3.11 a), it can be seen that increasing discretizations of drad are particularly ef-
A
fective if the conductor is exposed to high external magnetic fields. Up to 100 m , a discretization
of drad = 4 is sufficient to limit the deviation for frequencies up to half a megahertz to approxi-
mately 1 %. In case of high external magnetic fields, the deviation exceeds 1 % and converges to
5 % at 1 MHz, even for drad = 6.
In case of ndiv, the deviation increases strongly for frequencies of several hundred kilohertz for
setups with medium and high external magnetic fields. Additionally, a higher amount of dis-
cretizations has not such a large impact compared to drad. In case of 1 MHz, the deviation of the
results with the highest discretization of ndiv exceeds the counterpart of drad for medium and
high external magnetic fields.
In conclusion, setups using ndiv = 30 and drad = 4 yield very accurate results up to 200 kHz and for
external magnetic fields of some hundreds ampere per meter. In simulations with higher frequen-
cies and stronger external magnetic fields, both discretization parameters need to be increased.
However, calculations such as the one used for the reference setup (drad = 8 and ndiv = 90) re-
quire more than 15 hours and 250 GB RAM for a single 30 mm sample of a 420 x 0.1 mm litz wire
conductor.

Length of the simulated conductor

The previous sections have demonstrated that the PEEC implementation Fastlitz requires several
hours on HP computers to calculate frequency dependent power losses in case of complex litz
wires. Although the amount of strands corresponds with real HF litz wires, the inner structure
is indeed not yet completely resolved: Up to now only conductors with one fix pitch length (all
entries of the pitch vector in table 3.6 are equal) have been investigated. In realistic litz wires, the
pitch lengths of strand and bundle level significantly differ just because of varying manufacturing
processes. To avoid errors in the calculation of proximity losses, the underlying equation (2.5.21)
requires closed loops between all strands in the conductor. Consequently, the length of the simu-
lation model has to be enlarged to reach an equal position of the strands at the cross sections at the

66
 3.7. Litz wire simulation

4 5 .5

5 b u n d le s à 2 0 s tra n d s
4 5 .0 4 b u n d le s à 5 0 s tra n d s
a t 1 0 0 k H z a n d H ex t= 1 0 0 A /m
n o rm a liz e d p o w e r lo s s e s [m W /m m ]

C lo s e d lo o p s
4 4 .5
o n b o th
g e o m e try le v e ls
4 4 .0
p itc h le n g th
o n b u n d le le v e l
4 3 .5
p itc h le n g th
o n s tra n d le v e l
4 3 .0

S k e tc h o f th e
4 2 .5 litz w ire g e o m e try

4 2 .0
0 2 0 4 0 66 00 8 0 1 0 0 1 2 0
le n g th o f th e s im u la tio n m o d e l [ m m ]

Figure 3.12: Normalized power losses versus length of the simulation domain for two different
litz wires. Both conductors have a pitch length of 20 mm and 30 mm on strand and bundle level,
respectively.

beginning and the end of the model.

In figure 3.12 two litz wire structures consisting of 5 bundles à 20 (a) and 4 bundles à 50 strands
(b) are investigated with regard to varying lengths of the simulation domain. Both litz wires have
a fix pitch length of 20 mm on strand level and 30 mm on bundle level. The power losses shown
are normalized to the value for 1 mm length of the conductor at 100 kHz with an external magnetic
A
field of 100 m .
The normalized power losses vary strongly with the length of the simulation domain. However, for
closed conduction loops on both geometry levels (length of 60 mm and 120 mm), the results are
equal. This is the case for both litz wire geometries. The corresponding length of the simulation
domain satisfies the least common multiples at 60 mm and its multiple at 120 mm. With increasing
length of the conductor, the influence of the unclosed loops decreases and the normalized power
losses converge against the values calculated for the simulation model with the length of the least
common multiple.
Consequently, to resolve power losses in litz wires correctly, the least common multiple of varying
pitch lengths has to be used as the length of the simulation model. Depending on the ratio of the
pitch lengths, the required computational resources can be many times larger compared to the one
needed for the standard 30 mm sample.
Regarding the workflow in SlicerPro, each input configuration consists of a frequency and an ex-
ternal magnetic field value in one cut corresponding to a defined length (such as 1 mm). Within
the PEEC method, power losses are calculated for a litz wire conductor with a much larger length
due to the need to have closed loops between the strands.

67
CHAPTER 3. Multi-scale algorithm for litz wire systems

(a) (b)

Figure 3.13: Interpolation and corresponding error of the skin effect dependent resistance (a) and
A
proximity losses (b) at 100 m and 1 A up to 1 MHz.

Look-up table
Litz wires are used in a wide field of applications. However, frequencies up to 1 MHz and external
magnetic fields up to 1 kAm meet the requirements of most applications, at least for the present time.
Consequently, the characterization of a litz wire conductor with regard to its power losses in these
ranges can be applied to a large variety of systems relevant for different applications.
In figures 3.13 a) and b), the skin effect dependent resistance (RAC and Rskin ) and the proximity
A
losses for 100 m are visualized up to 1 MHz. The reference curves (solid lines) base on very ac-
curate simulations of a 420 x 0.1 mm litz wire with 7 bundles à 60 strands, evaluated on more than
50 supporting points (equally distributed on a logarithmic scale). The dashed lines correspond to
approximations with significantly less amount of supporting points. For the skin effect dependent
resistance (figure 3.13 a), the distribution of the points is limited to one starting point at 1 kHz
and 5 within each following decade. Considering the entire frequency range, the shape of the
approximation curves is very similar to the ones of the reference simulation, respectively. A de-
tailed analysis of the frequency dependent deviation between the approximated and the reference
curve depicts a maximal error of approximately 1 %, even though the curves consists of only 10
evaluation points.
The proximity losses in litz wire conductors have an approximately polynomial increase versus
the frequency [94]. Consequently, the solid curve in figure 3.13 b) corresponds to a quite straight
line within a logarithmic scaling, with a gradient equal to the original exponent. Based on a lin-
ear interpolation on the transformed values, the amount of evaluations can be reduced strongly in
comparison to figure 3.13 a) : only four supporting points are required to reproduce the reference
results in the entire frequency range with a maximum deviation of 1.5 %. The largest deviation is
observed in the frequency range with the highest curvature. However, this range depends strongly
on the geometry of the litz wire and is difficult to anticipate. Consequently, the relevant frequency
range of power electronic systems should be resolved at least by one or two additional supporting
points in the range from 100 kHz to 1 MHz. Due to the proportionality of the proximity losses to
the square of the external magnetic field, the resulting curve can be adapted to any field of appli-
cations.
Both interpolation strategies enable a significant reduction of the total amount of PEEC simula-
tions without loss of accuracy. The Fastlitz approach was adapted and enhanced to take these
findings into account. All output file consists of three parts: frequency dependent skin losses

68
 3.7. Litz wire simulation

(without Hext ), frequency dependent proximity losses (optional with different Hext values) and de-
tailed information about the simulation setup and calculation time. In appendix A.3 an example
of such a standard output file of Fastlitz is provided.
Based on more than 300 PEEC simulations of different strand and bundle configurations, one
characteristic output file each has been generated. The data for these litz wires are provided in
a database. Each of the look-up table can be used as input for the SlicerPro approach, and ap-
plied on different Hext distributions (corresponding to different geometries) without requiring high
computational resources.

3.7.3. Measured factors

In both previous chapters, mathematical models have been presented to calculate the frequency
dependent power losses of litz wire conductors. Due to limitations of the computational resources
and/or lack of detailed information about the inner structure, these approaches may have a large
deviation to real applications. If the litz wire is available for experiments, specialized measurement
setups can be used to extract characteristic data for the loss calculation. Based on that, common
theoretical approaches can be adapted and enhanced to reach more realistic results.
Due to the superposition of varying physical effects, the measurement process has to be split to
allow the quantification of the different loss types [31]. In the following, DC-, skin-, and proximity
losses (corresponding to Albach in table 2.2) are measured individually:

• RDC - Direct current resistance

The RDC value of a piece of litz wire conductor can be measured using a micro-ohmmeter.
The quality of the solder joints can be verified by reference values such as the ones provided
by the data sheet from the manufacturer. Moreover, analytical calculations based on equa-
tion (2.3.18) or (3.7.1) enable reliable results.

• Rskin - Skin resistance

To enable the measurement of pure skin losses, proximity losses and parasitic effects have
to be avoided. Similar to a coaxial waveguide, the conductor under test is inserted into a bar
made out of electrically conductive material. With the help of an impedance analyzer, the
electrical system of the litz wire and the bar is measured for a wide frequency range. Based
on that, the skin resistance is extracted for the electromagnetically isolated conductor.

• Pprox - Proximity losses

In contrast to previous losses, the external proximity loss is independent of the excitation
current and only depends on the time varying external magnetic field. With the help of
a test setup (figure 3.14) the litz wire conductor (apart from the solder joints) is exposed
to an external magnetic field. To reach a homogeneous field distribution along the length
of the conductor, the litz wire is located in the middle of the air gap between a system of
excitation coils, which are enforced and focused by several U-shaped ferrite cores. The
external proximity loss of the litz wire is calculated based on the difference between the
impedance of the excitation coil with and without the conductor in the test setup.

Based on [31], the frequency-dependent losses of any real litz wire conductor are limited by the
losses of an ideally twisted (lower bound) and a parallel stranded (upper bound) conductor. As-
suming an accurate measurement over a wide frequency range, two parameters λskin and λprox can
be calculated to approximate the measured curve based on both limiting bounds for the skin re-
sistance and the proximity losses, respectively. In the following formulation, the skin resistance
Rskin,λ consists of the measured results of RDC and Rskin , and the measured proximity losses are

69
CHAPTER 3. Multi-scale algorithm for litz wire systems

Figure 3.14: Test setup for the measurement of external proximity losses. The litz wire under test
(green) is exposed to a homogenous external magnetic field caused by the excitation coils (red).

1 0 0 0
λs k i n / λp r o x
0 .0 / 0 .0
1 .0 / 1 .0
0 .5 / 0 .5
1 0 0 0 .3 / 0 .5
0 .7 / 0 .5
0 .5 / 0 .3
R
F

0 .5 / 0 .7

1 0

1
1 k 1 0 k 1 0 0 k 1 M
f [H z ]

Figure 3.15: Normalized resistance FR = RRAC calculated for varying combinations of λskin and
DC
A
λprox for a 245 x 0.1 mm litz wire exposed to a constant external magnetic field of 100 m .

defined as Pprox, λ .
The ideal twisted and the parallel stranded conductor have been introduced in chapter 2.3.2 just
like the calculation of their losses. Combining equations (2.3.27) and (2.3.30) for the ideal and
parallel case, λskin can be calculated by

Rskin,λ = λskin Rskin,id + (1 − λskin )Rskin,par , (3.7.8)

for the measured resistance value Rskin,λ .

Based on equation (2.3.31) and (2.3.32), λprox can be determined by

Pprox,λ = λprox Pprox,id + (1 − λprox )Pprox,par . (3.7.9)

In case of measurements consisting of a setup with a fix frequency and one external magnetic
field, both factors λskin and λprox can be determined exactly. However, regarding a frequency range

70
 3.8. Realization and coding

(such as 1 kHz to 1 MHz) all resistance and loss values are vectorized quantities. In such a case, the
determination of λskin and λ prox results in an optimization problem. Depending on the operating
point, both parameters are fitted in a way, that Rskin,λ and Pprox,λ agree the best with the measured
curves [31].
In figure 3.15 several variations of λskin (dotted line) and λprox (dashed line) are visualized for a
A
245 x 0.1 mm litz wire conductor exposed to a constant external magnetic field of 100 m . It can be
observed that the curves of the normalized resistances FR are much more sensitive to changes of
the proximity factor compared to the skin factor. However, in common HF litz wires λskin varies
between 0.5 and 0.9, while λprox only differs between 0.98 - 0.99.
Within a research project [98] more than hundred litz wire conductors have been measured and
both factors were extracted for frequencies up to 30 MHz. Based on these data, the power losses of
litz wire conductors can be evaluated easily for a wide range of frequencies and external magnetic
fields.
The SlicerPro workflow enables the usage of these measurement-based factors as a third possibil-
ity to calculate the frequency dependent power losses. In addition to λprox and λskin the amount
and diameter of the strands within the litz wire conductor needs to be provided by the user.

3.8. Realization and coding

The workflow described in previous chapters is not only an academic approach for solving a multi-
scale problem in the domain of electric engineering, but rather a helpful tool in the design process
of power electronic systems. To enable the shift from a useful theoretical approach to a reliable
support in practical terms, engineers need to make use of the workflow. Therefore, the imple-
mentation has to provide a comfortable calculation of the results and additionally has to allow
adaptions and future enhancements.
In order to meet these requirements, following priorities considering the implementation and cod-
ing have been taken into account:

• Modular structure
The main idea of the modular structure consists of a separation of the complex workflow
in individual components which are implemented in one separate subroutine each. This
strategy requires the definition of interfaces and standardized exchange formats within the
program and also for the in- and output. Based on that, all subroutines in the SlicerPro
implementation can be replaced and enhanced, and therefore enable an easy extension of
new bugs or features.
The entire program is written in the scripting language Python and the manipulation of all
program settings is controlled by only one input script. This file can be filled out manually
or automatized by a graphical user interface (GUI), and is executable directly in the batch-
mode.
• Visualization
To enable a fast comparison between varying setups, the visualization of the simulation
results is mandatory. In the current implementation the power losses for a user-defined fre-
quency range and the spatial distribution of the external magnetic field data is provided.
Moreover, the import of measured curves for comparing simulation results with the experi-
ments is possible, as well as scaling, numbering and labeling of all curves in the plots.
• Advanced mathematical calculations
The computational costs of the varying simulation approaches differ strongly. Even tough

71
CHAPTER 3. Multi-scale algorithm for litz wire systems

Table 3.8: Workflow and relevant functions of the SlicerPro implementation.

Start 0

Import batch_commands_in 1
List of relevant functions
For each input file,
0 00 batch_slicerpro.py
for each litz file

Extraction of H ext 2 11 parse_input_commands

22 import_h_extern
Normalized data 3 Vectorized data 4
33 maxwell_two_col
linear quadratic
44 fill_cut_list
Find boundary points 6 Calculate h_explicit 5
55 get_h_extern_explicit
[Projection] 7
66 extract_rim_values
Lin. average of boundary
8
points 77 project_values

88 calc_h_average
Hext data in standardized
format 99
Calculate cut_distance
peec_solver

Solver 0
10
10 lambda_solver

11
11 analytic_solver
Analytic 11 PEEC 9 Measured factors 10
12
12 export_h
Use analytic approach 11 Import litz file 9 Extract lambda parameter 10
13
13 gui_slicer
Interpolation 9 Calculate skin and proximity
10
losses

Calculate
0
P(f) [W] or R(f) [Ω ]

Export data 12 Plot in GUI 13

geometrical projections, evaluations of modified Bessel functions and integrals are no big
deal for common computers, the huge amount of cuts and evaluation points requires ad-
vanced mathematical functions. Consequently, these functions use the numpy and scipy
packages which are comparable to Matlab with regard to the mathematical functionality
and calculation time.

• Platform-independent
To work directly with the input data of all kinds of simulation software, the implementation
has to be portable for both operating systems, Windows and Linux.

The described implementation of the SlicerPro workflow is realized in the free Python distribution
Anaconda. In table 3.8 the corresponding workflow is depicted and additionally correlated with
the main functions in the code. More detailed information about in- and output parameters, as well
as the variables and interfaces are provided in the documentation [99].
To simplify the handling of the program and to take into account the usage habits of engineers,
a graphical user interface (GUI) was designed. In figure 3.16 a) the standard interface for the cal-
culation is visualized. The option menu (for the PEEC method) provides the selection of different

72
 3.9. Conclusion

(a) (b)

Figure 3.16: Screenshot of the SlicerPro GUI providing the visualization of the
frequency-dependent resistance (a) and the external magnetic field distribution (b).

FEM data at the top left, and litz wire input files at the bottom left. Resulting frequency-dependent
curves are plotted on the right side while the solver options can be adapted by the corresponding
button. Additionally, a second tab provides the local magnetic field distribution (figure 3.16 b)
bottom), and the spatial distribution of the magnetic field and power losses on the cut and system
level, respectively.

3.9. Conclusion
The investigations of different solving approaches for Maxwell’s equations have revealed that a
single method cannot cover the complexity of litz wire systems. Consequently, the solving ap-
proach is split by taking benefit from the separation of global and local effects in power electronic
systems: The magnetic field distribution surrounding a litz wire conductor is equal to the distribu-
tion caused by a solid conductor with a homogeneously distributed excitation current. However,
the latter case does not require the resolution of the complex inner litz structure. Based on that, an
initial FEM simulation is used to calculate the magnetic field distribution in the entire simulation
domain by a solid conductor. Subsequently, the external magnetic field along the length of the
conductor is extracted from these data and provided for a local evaluation of the power losses on
litz wire level by analytic (Sullivan), numeric (PEEC), and measurement-based (lambda factors)
methods.
The verifications in this chapter demonstrated the accuracy, the computational effort, and the ap-
plicability of the individual solving steps for current and prospective power electronic systems. In
this context, the following issues can be highlighted:

• The FEM approach used for the calculation of the magnetic field distribution on system
level is the most advanced method for complex systems considering materials with electric
and magnetic permeability, geometries with arbitrary shape, and the direct import of CAD
data.

73
CHAPTER 3. Multi-scale algorithm for litz wire systems

• The extraction and evaluation of the external magnetic field along the length of the conductor
(based on the aforementioned magnetic field distribution) is realized by three approaches:
evaluation of the center nodes of a polyline, and a linear and a quadratic interpolation on
cut based data. Even though the computational effort and the amount of raw-data differ
significantly, the quality of the results is very similar. Consequently, the most efficient
approach using the center nodes of the polyline is applied in all following examples.

• The advanced adaptive meshing algorithms in common FEM implementations enable a high
accuracy of the simulation results without many user interactions. However, the computa-
tional time in this case is at least 2 - 3 times larger than in case of a manually (expert level)
prepared mesh.

• On litz wire level, analytical methods using the homogenization of the inner structure of the
conductor are state of the art for the approximation of ideal twisted and parallel stranded litz
wires (best and worst case). An enhanced approach of Sullivan uses the homogenization on
varying geometry levels of litz wires (sub-bundle, bundle, and conductor). However, due
to their complexity, common high frequency (HF) litz wires cannot be addressed by this
method, especially in case of non-circular bundles or conductors.

• The numeric PEEC method and corresponding implementation Fastlitz is specialized to

calculate the power losses of litz wire samples. Based on that approach, a look-up table is
generated consisting of the frequency dependent resistances (only based on the skin effect)
and proximity losses.

• The resolution of tightly packed litz wires in Fastlitz is problematic due to weaknesses in
the twisting algorithms. In most cases, the diameter of the simulation model is significantly
larger than the real litz wire.

• The lambda factors λskin and λprox are calculated by a fitting between the theoretical best and
worst case of a litz wire conductor (ideal twisted and parallel stranded litz wire) based on
measurements on experimental conductor samples. The measurement effort of this method
is manageable and the results are widely accepted and recognized.

The coupling of the FEM data with the three varying methods on litz wire level is implemented
in SlicerPro. This Python program provides a GUI based import and calculation, as well as a
visualization and export of the results.

74
 Experimental verification
4
The accuracy of the SlicerPro approach is checked by the results of several representative experi-
ments covering the wide field of power electronic applications.

75
CHAPTER 4. Experimental verification

4.1. Requirements on the experiments

The workflow described in previous chapters yields frequency-dependent power losses of litz wire
systems. Even though the individual numerical methods have been verified in chapter 2.4.2 (FEM)
and 2.5.2 (PEEC), the results of SlicerPro need to be benchmarked with real systems by measure-
ments.
The choice of the experiments and the findings gained from them is crucial to rate the significance
and reliability of any new simulation approach. The primary field of applications of the SlicerPro
implementation focuses on inductive components and systems with a high power density. There-
fore, the experiments have to be representative with regard to close winding distances, complex
ferrite structures, and HF litz wires. A close agreement of experimental and simulation geome-
try is important, especially due to the complexity of the electromagnetic effects which have to be
resolved. With this in mind, the evaluation has not only consist of the comparisons of absolute
values, but also of relative correlations. Especially in the early design period when the geometric
design is not yet fix, engineers require reliable predictions about the relative change of parameters
due to variations in the setups.
In general, experiments and measurements are expensive, especially in case of prototypes. All
parts and devices are individually constructed and installed, and specialized measurement equip-
ments are required. Moreover, experts for all fields of the realization and evaluation are needed
to achieve a high accuracy in every sub steps. Consequently, one of the most important issues
during the design of the experimental setups is the limitation of the number of prototypes within
industrial processes, as well as in research projects. Accordingly, test setups for the SlicerPro
workflow have to be designed in a special manner that components of one setup can be reused to
gain further information in a new setup.
It should be emphasized that all of the following verifications and evaluations only serve to demon-
strate the scope of applications of the new simulation approach and its accuracy with experimental
results. The improvement of power electronic systems based on results of the SlicerPro approach
is not in the scope of this thesis. However, first experiments provided in [89], [100] demonstrate
this potential.
In the following subchapter, the experimental setups are defined and described with regard to their
significance in the evaluation procedure. Afterward, the requirements on the construction of the
prototypes are specified and the quality of measurement results is discussed. In chapter 4.3.3 and
following, the experiments are documented and compared with the simulation results.

4.2. Definition of experimental setups

The coupled simulation approach presented in this thesis is a very general method to deal with the
multi-scale problem of litz wire conductors in electronic systems. Thus, the range of applications
for the SlicerPro implementation is very diverse. Within this thesis the main objective consists of
inductive components in the domain of power electronics, such as coils, chokes and transformers.
In the following experiments the coupled simulation approach is checked with regard to its support
in the early design period of power electronic systems. At the beginning, engineers require infor-
mation about the sensitivity and the corresponding optimization potential of the aforementioned
components. Apart from variations of the geometry and the material properties, the influence of
different litz wire types is of prime importance. Commonly, several prototypes are built up to
obtain the relevant information. To enable an additional hedge of these results or even the replace-
ment of this costly work, the new simulation workflow has to be verified close to reality.
In cooperation with application engineers at our institute following sensitivities are chosen con-

76
 4.3. Measurement settings and error estimation

sidering priority issues during the design of prototypes:

• Influence of varying number of strands and different bundle structures of the litz wire

• Resolution of complex power electronic systems with a high amount and tightly packed
windings

• Calculation of power losses for a wide frequency range

• Influence of different ferrite structures

• Solution quality of the two simulation approaches on litz wire level

The relevance of most of the items listed have already been studied in various publications with
regard to a large variety of components and devices. However, the elaborations in this thesis are
limited to the question, whether the SlicerPro approach yields reliable results for these sensitiv-
ities. Hence, different experimental setups with varying focus are required to gain meaningful
results. Nevertheless, pure academic examples or setups with dominating effects are not suitable
according to the aspired fields of applications of the simulation approach.
To achieve a high relevance of the experiments for real applications most of the experimental
setups analyzed in the following base on devices used in projects of the power electronic depart-
ment of our institute. Moreover, costs and efforts are significantly reduced due to the reuse of litz
wires, wrapping tools and adapted measuring facilities.
However, all the aforementioned sensitivities can not be tested in one single experimental setup.
On one hand, there are setups which can not be adapted due to their type of construction, on
the other hand the complexity of the experiments and the simulations would exceed financial and
human resources. Moreover, the resulting size of all the measured data would complicate the
analysis. For reasons of clearness, four different experimental setups are designed and tested with
regard to only several sensitivities (out of the listing above).
In table 4.1 all experiments used for the verification are listed and shortly described. The corre-
sponding sensitivities are specified in the right columns for the respective setup.

4.3. Measurement settings and error estimation

In a variety of publications [101], [31], [102] various recommendations are provided to reach reli-
able and reproducible measurement results of power electronic systems. To enable the comparison
to the simulation, the construction and measurement process needs to be documented very accu-
rately. Moreover, the virtual model has to be adapted to reach a close match with the experiment
with regard to the geometry, material properties, and physical and electrical boundary conditions.
In terms of litz wire systems, the following issues should additionally be taken into account during
the preparation:

• The position of the conductor has to agree as accurate as possible with the CAD data. How-
ever, conductors which were wrapped on small cable drums for a long time tend to be
warped later in the application. To avoid imprecise windings, it has approved as useful to
create a wrapping tool. Based on the CAD data, a simplified “negative” volume of the initial
winding geometry is constructed (figure 4.1 a-b) and manufactured by using a 3D printer
(figure 4.1 c). With this tool, the conductor can be adapted easily to the required winding
shape and fixed with instant glue.

77
CHAPTER 4. Experimental verification

Table 4.1: Experimental setups and corresponding sensitivities used for the verification of the
SlicerPro approach.

• Choke with and without ferrite core

Setup 1

• N = 39 windings of a 245 x 0.1 mm litz wire

- X X X X
• Coupled simulation approach based on the PEEC
method and lambda approach

• Air coil with N = 12 windings

Setup 2

• 420 x 0.1 mm litz wire consisting of 7 and 14 bundles X - X - -

• PEEC method on litz wire level

• Coil with N = 12 windings using four different ferrite

Setup 3

setups
- - X X -
• 420 x 0.1 mm litz wire
• Lambda approach on litz wire level

• Lateral displacement of two coils in a transformer

system
Setup 4

• 800 x 0.071 mm litz wire - X X X -

• Coupled simulation approach based on the lambda
approach

Sensitivities with regard to

. . . number of strands and complexity of the bundle structures
. . . complexity of the winding structure / power electronic system
. . . frequency (sweep between 1 kHz - 1 MHz)
. . . ferrite structures
. . . varying methods on litz wire level (PEEC approach and lambda factors)

78
 4.3. Measurement settings and error estimation

(a) (b)

(c)

Figure 4.1: Based on the CAD data of the coil (a), the corresponding wrapping tool is
constructed (b) and manufactured (c).

The wrapping tool can be reused and several identical coils can be created easily. However,
especially in case of small winding distances, the number of copies is limited because of the
fragile structure of the tool.

• The solder joint is important to achieve an equal distribution of the total current among
all strands. With the help of a soldering bath (instead of manual soldering) strands are
homogeneously embedded in solder and have equal resistance to the electrical contact [102].
Moreover, a cable lug needs to be shrunk on the joints to enable reproducible results at the
analyzer.

Regarding the measurement requirements for the verification, the considered frequency range
needs to cover the operating frequency of existing and future power electronic systems. In most ar-
eas of current applications, devices work within a limited range up to maximal 200 kHz. However,
due to the progress in the field of semiconductors increasing frequencies become more widespread
in power electronic systems. Especially in the domain of inductive power transfer (IPT) systems
frequencies up to 1 MHz are aspired to achieve a higher efficiency. To take these applications into
account, the measurement equipment has to provide a precise resolution from 1 kHz to 1 MHz.
All results presented in this thesis base on no-load measurements with a 4294A Precision Impedance
Analyzer of Agilent Technologies. To obtain highly accurate results, following preparatory work
is required:

• At the beginning of the measurement, one has to ensure that the analyzer has reached its
operating temperature. Following, the calibration procedure is started to gauge the system
based on calibrated attenuators [103].

79
CHAPTER 4. Experimental verification

• Common measurement equipment enables the usage of different connection types. Screw-
ing connections allow an enhanced fastening of the experimental joints and should be pre-
ferred compared to standard clamps.

• Within the measurement mode of the analyzer, the type of equivalent circuit which is used
in the experiment has to be defined. In case of inductive components a standard RLC equiv-
alent circuit is chosen for the impedance spectroscopy. To reduce outliers, several dozens
of measurements are combined and averaged to one final result. Afterward, an automatic
export script transfers all relevant results to Excel and an additional scripting evaluates the
frequency-dependent resistance and corresponding measurement error [104].

(a) (b)

Figure 4.2: Three identically manufactured setups of a circular transformer (a) consisting of two
coils (b) made of a 270 x 0.04 mm litz wire.

A meaningful test of the quality standards and principles mentioned above is enabled by the
analysis of three identical experimental setups (figure 4.2 a). Two coils each are manufactured in
a ferrite shielded transformer used for inductive power transfer to rotating equipment [105]. The
system is very compact and consists of 10 and 9.5 windings at the primary and the secondary
side, respectively. Moreover, due to the usage of a 270 x 0.04 mm litz wire the power density is
extraordinary high.
All three experimental setups are built up and measured at the impedance analyzer. The resulting
frequency-dependent no-load resistance is plotted for all setups in figure 4.3. In addition, the max-
imum deviation between these three curves is provided. Especially in the lower kilohertz range all
curves show an oscillation due to the switching of the measurement range of the analyzer. Apart
from that, the shape of the resistance curves matches over the entire frequency range up to 1 MHz.
Nevertheless, the deviation between the maximal and minimal resistance of all three setups is ap-
proximately 10 % in average. Regarding the measurement error of the analyzer (visualized by the
dashed blue curve) in the regarded frequency range, this effect is negligible. Consequently, the
majority of the deviation of these identically manufactured experimental setups is mainly caused
by minor differences in the position of the windings or by small deviations in the material prop-
erties. Due to the very high compactness of the system smallest differences in the experimental
setups yield a strong deviation of the frequency-dependent resistance. Therefore, such tight setups
correspond to the worst case scenario for the benchmark of power electronic systems.
To reduce the effect of tiny deviations in the manufacturing process, the following experimental
systems are not designed as close and compact as it would be possible. Especially in case of a

80
 4.3. Measurement settings and error estimation

Figure 4.3: Measurement of the no-load resistance of all three coil systems (figure 4.2 a) up to
1 MHz. The frequency-dependent deviation between the maximal and minimal resistance of all
systems and the maximal measurement error are visualized by the blue curves (with the scaling
on the right side).

minor amount of windings, the spacing between adjacent conductors is limited to at least 0.5 mm.
Taking this into account, identically manufactured components, such as the coil in figure 4.9 e)
yield a deviation of maximal 3 % over the entire frequency range up to 1 MHz by using an addi-
tional gauging based on the DC resistance.

81
CHAPTER 4. Experimental verification

4.3.1. Simulation vs. experiment - Choke

In the first verification setup, the simulation workflow presented in this thesis is applied on com-
mon basic components of power electronic applications: Power inductors and high performance
chokes. Especially in the domain of non-isolated DC-DC converters, coils with a few dozen wind-
ings based on litz wire conductors are widely used. A precise determination of the power losses
during operation enables the definition of a suitable converter topology and provides helpful input
considering the thermal management and cooling of the entire system.
In this verification setup and all following elaborations, the considerations are limited to the RAC
values, as the inductance can be calculated easily by standard simulation approaches.
To reuse the costly prototype in varying configurations, the setup A in this experiment consists
of a pure air coil, which is enhanced by several E-shape ferrite cores in setup B. Even though
the operating frequency in these applications is usually limited to several kilohertz, the following
measurements and simulations are performed up to 1 MHz to allow the proof of the simulation
concept.

(a) (b)

Figure 4.4: Geometry of the air coil (a) and enhancement by four 3C95 ferrite E-cores (b),
corresponding to the experimental setup A and B, respectively.

Experiment

The experimental coil for setup A was designed as a rectangular air coil with N = 39 windings in
a two layer configuration of 20 and 19 windings, respectively, one above the other. With the help
of a standard wrapping-tool for E65 ferrite cores, windings with a 245 x 0.1 mm litz wire were
tightly stacked with 0.5 mm distance in average (figure 4.4 a). To reduce the manufacturing based
deviation of the experimental geometry from the CAD data, two air coils are built up in the same
manner. Based on the aforementioned standardized procedure, both coils were measured and the
resulting DC resistances differ by less than 1 %.
In the case of setup B, the coils of setup A were enhanced by four 3C95 ferrite E-cores each (figure
4.4 b), two at the top and two turned at the bottom. With the help of these ferrites, the inductance
of the choke is significantly increased as well as the frequency dependent resistance measured at
the impedance analyzer. To guarantee comparable results even at high frequencies, the positions
of the ferrite parts and the distances to the coil are fixed by screw-type terminals.
The measured curves were generated by averaging the frequency dependent results of both coils
used in setup A and B.

82
 4.3. Measurement settings and error estimation

(a) (b)

(c)

Figure 4.5: Simulation model of the choke without (a) and with (b) ferrite cores, and
corresponding external magnetic field distribution Hext calculated by the FEM simulation along
the length of the conductor (c) for an excitation current of 1 A.

Simulation
The design of the CAD model requires much effort and a lot of experience due to the tight and
equally spaced packaging of the windings. Based on more than 5000 points, a polyline was calcu-
lated corresponding to the center position of the conductor. In a next step, this line was transformed
to the air coil model and additionally enhanced by the ferrite parts for the modified setup (figure
4.5). Even though, the accurate positioning is much easier in the simulation model than in the
experiment, the edges of the four ferrite bodies have to be chamfered to avoid artifacts such as
singularities in the FEM simulation.
The meshing in the FEM simulation was performed by using the adaptive meshing option in Ansys
Maxwell3D (with a 1 % limitation of the energy error). In case of the pure air coil, the magnetic
field distribution was calculated for an excitation current of 1 A using the magnetostatic solver.
Hext (figure 4.5 c) was extracted along the polyline and used as an input for the two simulation
methods on conductor level: On one hand, the pure numeric approach using the PEEC method
with an exact copy of the inner structure of the 245 x 0.1 mm litz wire. Based on the twisting
model of the conductor, the frequency dependent behavior is calculated in Fastlitz and provided
in the characteristic look-up table (appendix A.3). On the other hand, the second approach uses

83
CHAPTER 4. Experimental verification

(a) (b)

Figure 4.6: Frequency dependent resistance of setup A (pure air coil) and B (choke with ferrite
cores) based on measurements (red) and the coupled simulation approach using the PEEC
method (blue scatter) and the lambda factors (black dots).

the measurement based lambda factors. Both factors λskin = 0.49 and λprox = 0.989 are provided by
[31] for the litz wire used in this experiment.
Within the SlicerPro implementation, the FEM results on system level (Hext data) are combined
with both litz wire approaches on conductor level (look-up table of the litz wire for the PEEC
method and lambda factors). The calculated resistance of the system is only based on effects
caused by the frequency and the external magnetic fields in the conductor. In case of setup A, the
coupled simulation approach already covers all relevant effects. However, in case of setup B, addi-
tional losses caused by hysteresis and eddy currents occure within the magnetic parts. Commonly,
these effects are not included in the simulation model of the SlicerPro approach, even though they
are measured at the impedance analyzer. Unfortunately, there is no reliable simulation model for
the calculation of these effects in complex 3D geometries. In particular, they require a detailed
characterization (considering the magnetic field strength, shape, temperature, and frequency) of
the ferrite material.
However, a rough approximation of the ferrite losses can be calculated with the help of the Stein-
metz coefficients [33], [106]. Based on the (very imprecise) material data of the 3C95 material,
the factors k, α, and β are determined and implemented in Maxwell3D. Based on that, the core
losses can be calculated within the magnetic material by using the eddy current solver. The corre-
sponding losses are converted and added to the frequency dependent resistance.
In figure 4.6 the measured data and the simulation results are shown for the setup A and B up to
1 MHz.

Comparison
Due to the closed E-cores, most of the magnetic flux density of the coil is focused in the ferrite
core. However, this change does not change the external magnetic field distribution along the
length of the conductor, as long as the E-cores have no air-gap (in such cases, the nearby windings
would be exposed to extremely high fields, similar to figure 3.7).
Considering the frequency dependent resistance in figure 4.6 a), the curves of both simulation ap-
proaches differ only slightly in case of setup A and accurately agree with the measurements up to
1 MHz. Even though the resistance of the air coil increases by a factor of 100 over the regarded
frequency range, the simulation enables an accurate prediction of the frequency dependent resis-

84
 4.3. Measurement settings and error estimation

tance.
As described in other sources [6], the resistance of the coil increases by using ferrite cores. Due
to the closed E-cores in setup B, the flux densities inside the ferrite material are very high and
yield core losses exceeding the winding losses by several magnitudes. Moreover, due to the ferrite
cores, the inductive behavior of the entire device changes to a capacitive behavior at approx-
imately 100 kHz. Consequently, the measured results of the resistance are only reliable up to
this frequency. Unfortunately, in the data sheet of 3C95 the relevant information for determining
the core losses (based on the Steinmetz parameters) are only provided for 100 kHz, 200 kHz and
400 kHz. Due to a high sensitivity of these parameters (with regard to the frequency and flux den-
sity), an extrapolation is very inaccurate and therefore the simulation results poorly agree with the
measurements.

Conclusion
By using the polyline approach (chapter 3.6) the resolution of complexe winding structures and
the extraction of the associated external magnetic field does not require any skripting or user in-
teraction. The entire FEM process requires less than half an hour for the calculation of the Hext
data on a standard simulation computer (Intel I7, 4 Cores). However, in case of resolving the
ferrite losses with the artificial conductivity approach, this time is required for one eddy current
simulation each, corresponding to a defined frequency.
The results of the PEEC and lambda approach differ only slightly for setup A and B in the re-
garded frequency range. However, both methods require varying raw data (PEEC: detailed inner
structures, Lambda approach: measured factors) to resolve the litz wire behavior. In general, a
quantification of the effort for getting these data in a suifficant quality is difficult and differs con-
sidering the litz wire type, the availability of measurement equipement, and the computational
ressources.
In these simulation setups, the numeric modeling was very efficient, because all relevant infor-
mation of the inner litz wire structure was provided by the manufacturer Elektrisola. Moreover,
the division of the strands in 7 bundles can be resolved realistically in the twisting algorithm of
Fastlitz (figure 3.9): One bundle stays in the center position and the others turn around along the
length. Due to the minor amount of strands and a good ratio of the varying pitch lengths (detailed
description in figure 3.12), the generation of the characteristic look-up table for the 245 x 0.1 mm
litz wire takes only 10 min for the preparation and less than 3 h calculation time.
The good agreement of the simulation results with the measurements in setup A demonstrate the
potential of the coupled simulation workflow in case of optimal input data and a pure litz wire sys-
tem. Up to now, the usage of magnetic material within the regarded system reduces the quality of
the simulation results significantly. However, it should be taken into account, that the chosen fer-
rite setup corresponds to one of the worst-case scenarios, because the saturation within the ferrite
core is extremly high. The resolution of these parasitic effects requires further research, especially
with regard to a reduced amount of measured input data for the loss calculation.
Finally, it has to be noted that due to the fact, that all litz wires (with the same diameter) cause
the same amount of losses in the surrounding, the losses calculated by the simulation workflow
remain unaffected and can be used for finding the optimal litz wire for the system.

85
CHAPTER 4. Experimental verification

4.3.2. Simulation vs. experiment - Bundle structure

In this part of the verification, the analysis concentrates on studying varying bundle structures in
litz wires by the coupled simulation appraoch using the FEM and PEEC method.
Commonly, the skin and proximity effect are the two dominating effects in litz wire conductors
influencing the frequency dependent power losses. While the skin losses are significantly reduced
by a smaller strand diameter, the decrease of the proximity losses is more complex.
In case of an ideal twisted litz wire, the loop currents induced by external magnetic fields are
canceled out and vanish for a sufficiently long conductor (equation 2.3.31). However, real compo-
nents are not perfect: In a lot of litz wires, one bundle stays in the center position (cable core) and
adjacent bundles rotate on a fix radius along the length of the conductor. The position of each in-
dividual strand influences the size of spanned loops between the strands (figure 2.4) and therefore
the proximity losses of the conductor.
In the following, two bundle structures are analyzed in the frequency range up to 1 MHz for an
air coil consisting of N = 12 windings. These elaborations mainly focus on the verification and
benchmarking of the coupled simulation approach.

(a) (b)

Figure 4.7: Measurements performed in a top (a) and bottom (b) configuration.

Experiment
The portfolio of litz wire manufacturers consists of a huge variety of litz wire types, but in general
each of them is produced with only one corresponding bundle structure. Due to large purchase
quantities in case of an individual conductor design, the first challenge of this experiment was the
identification of two litz wires with the same amount and diameter of strands and similar conduc-
tor cross section, but with a major difference in the bundle configuration.
Two circular 420 x 0.1 mm litz wires meet these requirements (table 4.2): The conductor manu-
factured by Pack consists of 14 bundles à 30 strands (wire A) and the configuration of Elektrisola

86
 4.3. Measurement settings and error estimation

Table 4.2: List of all properties (based on table 3.6) used for the Fastlitz simulation of both
420 x 0.1 mm litz wires. Deviations in the simulation model compared to the experimental wires
are highlighted.

Pack Elektrisola
(Wire A) (Wire B)
14 bundles à 30 strands 7 bundles à 60 strands

number of [30, 14] [60, 7]

strands
pitch length [m] [0.026, 0.039] [0.042, 0.056]
[strands, bun- [0.026, 0.037] [0.042, 0.056]
dle]
length of the 0.078 0.168
simulation
model [m]
drad, discr. of 4 4
cross section
(fig. 3.11)
ndiv, discr. of 78 168
length (fig.
3.11)
diameter [mm] 3.26 3.05
(exp.) 2.95 3.00

uses 7 bundles à 60 strands each (wire B). According to the data sheets, the variation of both litz
wire diameters is 0.05 mm and therefore smaller than the manufacturing tolerances. Thus, there is
no difference in the external behavior of both conductors considering the magnetic field distribu-
tion on system level.
Due to the fact, that the estimated deviation between the resistances of both bundle structures is
assumed to be very small, a larger amount of experiments is required: Three helical planar coils
with N = 12 windings were built up with the help of a wrapping tool (figure 4.1 a-c) for wire A and
B, respectively. All six experimental coils are analyzed in a top and bottom configuration (figure
4.7) and each measured curve bases on 10 separate measurements.
Finally, the comparison quantity RRwire_A
wire_B
− 1 is calculated for all pairs of wire A and wire B in the
regarded frequency range. The resulting variation is provided in figure 4.8, with the averaged
frequency dependent deviation between both conductors, as well as the minimum and maximum
values of all measurements. The spectrum between minimum and maximum can be interpreted as
the (experimental) system-related error which needs to be taken into account when the measure-
ments are compared to the simulation.

87
CHAPTER 4. Experimental verification

Figure 4.8: Frequency dependent deviation between the Pack (wire A) and Elektrisola (wire B)
litz wire based on measurements and the coupled simulation approach. The curves of the
measurements (average, minimum and maximum values) base on the results of three
experimental coils for wire A and B, respectively.

Simulation

The geometry of the coils is simple and can be generated with an extruded polyline defined by an
analytical function in the CAD tool. The calculation of the magnetic fields and the extraction of
the external magnetic fields is performed just like in the previous setup.
In table 4.2 all information about the design of the inner structure of both experimental conductors
is provided, including their implementation in Fastlitz. In case of the Pack litz wire, the pitch
length on bundle level has to be changed slightly in the simulation model to reach a smaller least
common multiple and therefore a manageable size for the calculation.
Unfortunately, in case of the Elektrisola conductor the pitch lengths on strand and bundle level
have a bad ratio, resulting in a simulation length of approximately 16 mm. This size is more than
twice the length of the Pack model and increases standard PEEC simulations by a factor of 5.
Using a discretization of the cross section with drad = 5, more than 400 GB RAM are required
and the corresponding system of equations reaches almost 10 trillion unknowns. Moreover the
calculation fails due to convergence problems of the GMRES method at high frequencies. The
solver already requires more than 300 iterations at 100 kHz to reach a residual of 1e-3. In standard
PEEC calculations the target residual is set to 1e-7 and is commonly reached after 5 - 50 GMRES
iterations. Consequently, the PEEC method has been restricted to a Fastlitz setup using drad = 4.
To avoid deviations due to varying discretization levels in the simulation results of both conductors,
this parameter was also used for the Pack litz wire.
Even though both experimental litz wires have the same conductor diameter, these values differ
in the simulation models. This is due to weaknesses in the twisting algorithm (chapter 3.7.2) and
leads to a deviation of approximately 10 % between the Pack and Electrisola litz wire models.
Based on both characteristic litz wire look-up tables, the coupled simulation approach is used to
determine the frequency dependent ratio of the resistance of wire A and wire B.

88
 4.3. Measurement settings and error estimation

Comparison
The measurement results in the considered frequency range (figure 4.8) can be divided in four
sections:
In the first range from 1 kHz - 20 kHz both conductors have the same resistance, followed by a sec-
tion with a slight tendency for a better performance of wire B up to 80 kHz. However, this deviation
is smaller than the experimental error and therefore not reliable. In section 3 (80 - 500 kHz) the
bundle structure used in wire A yields a significantly better performance of up to 10 % compared
to wire B. In the following section up to 1 MHz, the measurements demonstrate a fundamental
change of the ratio resulting in a clear benefit of wire B.
The segmentation of the experimental curves in the aforementioned four sections can be quali-
tatively reproduced by the simulation. The frequency range between the deviation peaks as well
as the sign changes are similar. Moreover, at higher frequencies the size of the deviation can be
predicted accurately by the coupled simulation approach.
However, especially in section 1 and 2 the absolute values of the measurements and the simulation
differ. In the range of several kilohertz the simulated curve shows a monotonous (frequency depen-
dent) increase of the deviation, which does not occur in the measured curves. This is remarkable,
because the skin and the proximity effect are negligible for such high frequency litz wires in this
frequency range. However, due to the fact, that the simulation model of wire A has a significantly
larger diameter of the cross section, the influence of the skin effect on bundle and conductor level
(chapter 3.7.1) differs in comparison to the experiment. Unfortunately, this results in a varying
superposition of the electromagnetic effects, which is more (at lower frequencies) or less (above
100 kHz) significant depending on which effect causes the dominating share of the losses.

Conclusion
Both litz wires used in this verification cannot be distinguished from the outside and only differ
by their inner bundle structure. However, these minor differences have a significant impact on the
frequency dependent resistance on the system level.
Considering the exact bundle structure in the PEEC method, the coupled simulation approach
allows one to qualitatively resolve the shape of the curves measured at experimental setups up to
1 MHz. However, to reach a higher accuracy of the simulation results two problems have to be
overcome and will be subject of further studies: The agreement of the simulation models with real
litz wire structures has to be improved and the amount of computational resources for the solving
process of the PEEC method has to be reduced.
It should be noted that progress on these issues will not only provide an improvement of the
workflow methodology, but its application in optimization loops will allow a significant increase
of the efficiency of power electronic systems.

89
CHAPTER 4. Experimental verification

4.3.3. Simulation vs. experiment - Ferrite variation

The experimental device used in this verification has already demonstrated its technical relevance
in an inductive power transfer (IPT) system [107]. The coils used in this system consist of N = 12
windings each and enable a power transfer of 3.6 kW from the charging station to the vehicle. To
reach a high efficency of the system and to reduce the magnetic fields in the surrounding, ferrite
material is used to guide the field.
However, as anywhere in the domain of the automotive, the manufacturing costs and thus the
amount of ferrite material are always scrutinized. Consequently, the following verifications con-
sider RAC up to 1 MHz for an air-coil and three different ferrite setups.
Based on the findings of the previous chapter and the standard operating frequencies of such appli-
cations (100 - 150 kHz), the 420 x 0.1 mm litz wire with a 14 bundles à 30 strands configuration is
used in this experiment. Due to the fact that the resolution of the twisting structure of this conduc-
tor type is insufficient in the PEEC method, the coupled simulation approach bases on the lambda
approach on litz wire level.

(a) (b) (c) (d)

(e) (f)

Figure 4.9: The experimental coil (e) is measured in four different configurations (a)-(d) with
varying ferrite structures. The magnetic field distribution on the surface of the conductor is
visualized in (f) for the star-shaped setup and an excitation current of 1 A based on the FEM
results.

Experiment
The helical planar coil used in this experiment consists of N = 12 windings with approximately
0.5 mm radial spacing (figure 4.9 e). The inductive component is measured in four different setups

90
 4.3. Measurement settings and error estimation

5 0 0
fu ll
s ta r
4 0 0
c ro ss
a ir c o il

H [A /m ]
3 0 0

2 0 0

1 0 0

0
0 .0 0 .5 1 .0 1 .5 2 .0 2 .5 3 .0
le n g th [m ]

Figure 4.10: Averaged external magnetic field along the length of the conductor (starting in the
top left of the coil) for four different structures (figure 4.9 a-d).

(figure 4.9 a-d) based on a 420 x 0.1 mm litz wire internally divided in 14 bundles [7]: the air-coil
by itself and the coil located on a cross-shaped, star-shaped and fully covered ferrite structure. In
all ferrite setups, the magnetic material 3C90 [108] is used and the stripes are fixed to agree with
the positions in the simulation model at the best.
All measurements have been performed under the standardized test conditions and results are
provided in figure 4.11.

Simulation
The FEM model was built up in AnsysEmag with an excitation current of 1 A, homogeneously dis-
tributed on the cross-section of the solid conductor. The magnetic permeability of the ferrite was
set to 2000 for the entire frequency range and the power losses in the ferrite parts are neglected.
The meshing algorithm was forced to generate 3000 cuts along the length of the conductor and
the system was simulated by the magnetostatic solver for all four setup. In figure 4.9 f) the total
magnetic field Htot is visualized for the cross-shaped ferrite structure on the surface of the conduc-
tor. Based on these data, the external magnetic field was evaluated by the linear averaging method
(chapter 3.6). In figure 4.10 the resulting external magnetic field distribution along the length of
the conductor is provided for all four setups. The periodic structure of the peaks corresponds to the
ferrite distribution and the crossing of the coil windings with the return conductor. The diameter of
the winding within the coil becomes smaller with increasing length and consequently the distance
between adjacent peaks decreases while the absolute value of the field rises.
Based on some samples of the litz wire, the conductor is measured in the test setup described in
chapter 3.7.3 and the corresponding factors λskin = 0.58 and λprox = 0.99 are extracted. Based on
these input data, the frequency dependent resistance is calculated with the SlicerPro approach.

Comparison
The measured (solid line) and simulated (dashed line) data provided in figure 4.11 show increasing
resistances with increasing frequency as well as with increasing coverage with ferrite. The shape
of the measured curves can be reproduced by the simulation, even thus the absolute values of the

91
CHAPTER 4. Experimental verification

Figure 4.11: Measured and simulated (SlicerPro approach with lambda factors) resistance of all
four geoemtry setups (figure 4.9 a-d) up to 1 MHz.

coupled simulation approach are largely smaller.

Considering the pure air coil, the measured and simulated curves perfectly match in the regarded
frequency range. However, in case of the ferrite setups the differences between the simulation
results and the experiments grow with increasing frequency. It could be observed that the deviation
is larger in case of an increasing coverage of the ferrite material.

Conclusion
In case of simple coil geometries, variations of the coil shape or ferrite distribution can be inveti-
gated within several minutes using common FEM programs. Based on the resulting Hext data, the
SlicerPro approach enables a fast comparison of all setups within a few seconds. However, the
quality of the results differs significantly and reveals some similarities to the choke analyzed in
chapter 4.3.2.
Also in this case, the frequency dependent resistance of the air coil is resolved very accurately by
the coupled simulation approach, even up to 1 MHz. Moreover, this verification also demonstrates,
that the error of the simulation results is much larger for the ferrite setups (compared to the one
without magnetic material). However, even though there are no ferrite losses considered in this
verification, the simulated curves have a similar shape as the measured ones. In this setup, the
ferrite material is fragmented in an increasing amount of unconnected parts. In comparison to the
compact E-cores setup in chapter 4.3.2), the alternating magnetic fields (caused by the excitation
coil) cannot generate large closed eddy current loops in the material. Consequently, the losses
in the ferrites are less important and the simulation results are more reliable. Especially in the
relevant frequency range of IPT systems up to 200 kHz the coupled simulation approach provides
useful information for the design of the power electronic system.

92
 4.3. Measurement settings and error estimation

4.3.4. Simulation vs. experiment - Coil system

During the design of power electronic systems the consideration and the testing of varying opera-
tion modes is a time-consuming and cost intensive process.
In this verification, the displacement in a coupled coil system for the inductive power transfer is
investigated. Even though advanced systems such as the one in figure 4.12 a) already provide a
good performance over a wide range of positions, this is won at the price of 9 coils (in a two layer
configuration) each on the primary and secondary side.
Standard IPT systems only consist of one coil on each side which are designed and optimized for a
defined distance and position. Such a simple coil system (figure 4.12 b) is used as the verification
setup considering the idle mode. This operation mode corresponds to the worst case scenario,
because due to the no-load on the secondary side, the magnetic fields, and therefore the losses,
on the primary side are maximal. This is, because in the standard case (assuming power transfer),
the fields on the primary side are reduced due to fields caused by the induced currents on the sec-
ondary side.
In this verification setup, the frequency dependent resistance of a primary coil is investigated up to
1 MHz, while the secondary coil is displaced between 0 mm and 55 mm from the center position.
Due to the fact that real IPT systems work at 140 kHz, the principal focus is set on this frequency.

(a) (b)

Figure 4.12: Inductive power transfer systems (a) with a primary (left) and a secondary side
(integrated in the license plate) consisting of 7 coils each. Simplified verification setup (b) with
one coil each on both sides [89].

Experiment
The experiments presented in this section are the most challenging in this thesis and base on the
setups and investigations of [89].
Both coils have the same geometrical shape as the ones used in previous sections: N = 12 helical
planar windings and a coil diameter of 110 mm. The conductor cross section is also very similar,
even though a high frequency litz wire with 800 x 0.071 mm strands [7] is used.
The air gap between primary and the secondary coil is 10 mm in the vertical direction (y-axis in
figure 4.13 a) and four ferrite plates each are located on the rear side to shield the magnetic field.
The lateral displacement in x-direction is realized by a specialized fixture (figure 4.13 b) enabling
an accurate movement of the secondary coil from 0 mm to 55 mm. The frequency dependent re-
sistance of the primary coil is measured up to 1 MHz with a step size of 5 mm. In figure 4.14 the

93
CHAPTER 4. Experimental verification

(a)

(b)

Figure 4.13: Model of the coil system with 0 mm (solid) and 55 mm (wire frame) lateral
displacement and corresponding Hext along the length of the conductor (b).

results for the standard operation frequency of 140 kHz is visualized. The measured curve of the
resistance factor FR depicts a relatively slight reduction due to the displacement. However, the
unsteady curve shape indicates a measurement error of up to 5 %, especially because of varying
positions of the connection wire for the different measurements. In figure 4.15 a) FR is provided
for the entire frequency range up to 1 MHz versus the displacement by a color plot. The hori-
zontal shape of the isolines of FR demonstrate a significantly higher influence due to increasing
frequencies compared to the displacement with regard to the resistance.

Simulation
The verification setup is built up with the exact shape of the windings and the ferrites in a FEM
simulation model (figure 4.13 a). The air gap in the simulation is set to 12 mm to take into account
the realistic conditions in the experiments including the thickness of the litz wire wrapping and a
slight warping of the coils.
The lateral movement of the secondary coil (and associated ferrites) is defined as a sweeping
parameter in x-direction. Due to the iterative refinement of the mesh in Ansys Maxwell3D, the
electrostatic simulations require only minor user interactions and take less than 1 h for 12 varia-
tions. The resulting external magnetic field along the length of the conductor is provided in figure
4.13 b) for displacements up to 55 mm. With an increasing displacement the external magnetic
field distribution gets increasingly inhomogeneous and even reaches an oscillation by a factor of 5

94
 4.3. Measurement settings and error estimation

Figure 4.14: Measured and simulated resistance factor FR at varying displacements at 140 kHz.

within the windings (due to the high magnetic fields at the ferrite edges of the secondary side).
On conductor level, the PEEC approach is used to predict the losses of the 800 x 0.071 mm litz
wire. Due to the lack of information about the inner structure, the simulation model was built up
based on measured data of the bundle structure and the pitch lengths (on varying geometry levels)
using an unwrapped sample of the conductor. However, the pitch lengths determined in this pro-
cess represent just approximate values, because the measured system (litz wire) was changed due
to the removing of the wrapping. Moreover, the discretization of the cross section of the strands
in the Fastlitz approach has to be limited (drad = 3), because the large amount of strands requires
too many computational resources.
The resulting PEEC model of the litz wire is applied to the FEM data on system level. With the
help of the SlicerPro software the frequency dependent resistance is calculated for all displace-
ments up to 1 MHz (figure 4.15 b).

Comparison

First, the simulation results and the measurements are considered at the operating frequency of the
IPT system at 140 kHz (figure 4.14). Both curves of the resistance factor FR versus the displace-
ment have a very similar shape, but with smaller values in the simulation. The simulation data
provide a very smooth curve, while the measured curve has a more uncontentious shape which
indicates a system related error of the resistance of 5 - 10 %. This deviation represents minor in-
accuracies in the positioning of the coils and their connections as well as at the calibration of the
analyzer between varying measurements.
Considering the resistance up to 1 MHz (figure 4.15 a, b), the deviation between simulation and
measurement is growing strongly with increasing frequency. Due to the usage of a linear scaling
of the frequency in the color plot, the error of the simulation at higher frequencies is emphasized
even though the deviation is similar to figure 4.11. In all cases the absolute values predicted by
the simulation are smaller - up to factor 4 at 1 MHz.

95
CHAPTER 4. Experimental verification

(a) (b)

Figure 4.15: Resistance factor FR versus displacement (x-axis) and frequency (y-axis) based on
measurements (a) and the coupled simulation approach using FEM and PEEC method (b).

Conclusion
With the help of common FEM programs, the complexity of power electronic systems can be re-
solved in the simulation. Based on a parametric CAD model, modifications of the geometry can
be built up and electromagnetic simulations can be calculated for these models in parallel.
The results of the coupled simulation approach have a sufficiently high accuracy in the relevant
frequency range (up to 150 kHz) of IPT systems. Especially in the early design period of power
electronic systems these data can be used to predict the system behavior for (minor) variations of
the frequency and the displacement using a circuit simulator.
At higher frequencies the quality of the simulation results is significantly reduced due to the weak-
nesses in the resolution of the litz wire geometry and especially of the ferrite losses.

4.4. Conclusion
The goal of the experimental verification is to demonstrate the functionality of the coupled simu-
lation approach with regard to the calculation of winding losses in litz wire systems. The setups
studied have to cover realistic problems of the design of power electronic systems, but should not
exceed a reasonable amount of prototypes and measurements. To reach a high quality of the ex-
perimental results for tightly packed windings as well as for setups with ferrite material, the design
and measurement process is based on standardized procedures.
Four experimental setups were presented, each starting with a description of the background and
its field of application in the power electronic domain, a presentation of the experimental and simu-
lation setup, the comparison of simulation versus experiment, and a final conclusion. The findings
gained by these experiments and further investigations [89], [9] lead to the following results for
reliability the coupled simulation approach:

• The FEM part of the simulation approach calculating the external magnetic field is nu-
merically stable, well documented and provides a high usability even for non-expert users.
Tightly packed windings and complex 3D systems do not limit the accuracy of the sim-
ulation results. The computational resources increase with increasing complexity of the
geometry, but commonly require less than one hour calculation time on standard simulation
computers (Intel I7, 4 Cores).

96
 4.4. Conclusion

h ig h

a m o u n t o f fe rrite s / s a tu ra tio n
s a tu ra tio n
in a c c u ra te re s u lts

qu
a li
ta t
p a rtia l
iv e
fe rrite s p re
d ic
tio
ns
a c c u ra te re s u lts

a ir c o il

1 k 1 0 k 1 0 0 k 1 M
f [H z ]

Figure 4.16: Sketch of the quality of the results calculated by the coupled simulation approach
depending on the freqency and the ferrite coverage in the power electronic system.

• The results of the coupled simulation approach using the PEEC method agree well with the
ones using the lambda factors and match with the measurements of air coils up to 1 MHz.
However, the PEEC model of the litz wire requires a fine discretization of the cross section
and length of the simulation model and needs to reproduce the exact inner twisting structure
of the conductor. Up to now, the latter is a crucial limitation of this numerical method.

• In addition to the winding losses, also the ferrite losses are measured by the impedance
analyzer in case of magnetic material within an experimental setup. The resolution of these
effects in the simulation requires proper material data (such as Steinmetz coefficients) and
a separate FEM simulation for each frequency (which is very time consuming). Due to the
weakness of resolving the ferrite losses in common simulation approaches, the deviation
between simulation and experimental results increases with increasing ferrite coverage and
frequency.

• Reliable qualititative predictions of winding losses can be generated for varying litz wire
types using the same ferrite structure (or even in case of a similar amount of ferrite material).

The SlicerPro approach has demonstrated its potential for calculating the winding losses with re-
gard to complex power electronic systems and high frequency litz wires. The two implementations
using the PEEC method and the lambda factors have only a minor deviation in the regarded fre-
quency range, but the accuracy of the coupled simulation approach suffers due to parasitic effects
in the ferrite parts (figure 4.16).

97
CHAPTER 4. Experimental verification

98
 Summary and outlook
5
In the domain of power electronics inductive systems are very complex with regard to the geome-
try and the superposition of various physical effects. For example, the prediction of the frequency
dependent losses of air coils, chokes or even entire transformers is nearly impossible based on
pure analytic formula, especially in case of litz wires. Even though analytical and measurement
based methods are considered in this work, the thesis aims at developing a design process without
approximation approaches and experimental prototypes. Consequently, the focus of these elab-
orations is on two numeric methods: the finite element method (FEM) and the partial element
equivalent circuit (PEEC) approach. Both methods are well-established in the engineering do-
main, but differ significantly considering the simulation of litz wire systems.
The fundamental difference is based on varying mathematical formulations of the Maxwell’s equa-
tions describing the electric system. While the FEM uses a differential formulation, the PEEC ap-
proach bases on the integral formulation for evaluating the differential equations in the simulation
domain.
In case of complex geometries, such as some hundreds of twisted strands in a litz wire, the FEM
approach requires very large computational resources. The entire simulation domain including the
free space has to be discretized, and the electromagnetic field and the current density distribution
need to be calculated in each element. In contrast, the PEEC method is restricted to conductive
materials and transfers the electromagnetic field problem into an electric circuit problem. The re-
sulting circuit typically consists of several thousand of resistor (R), inductor (L) and capacitor (C)
elements and can be solved by evaluating the corresponding linear system of equations. Moreover,
this approach can focus on dominating effects: In case of litz wire systems, second order effects
such as the capacitive coupling between strands can be neglected by leave out the associated ele-
ments on circuit level, resulting in a significant speed up of the calculation time.
In addition to these basic considerations on theory level, the accuracy of the numeric approaches
was investigated and benchmarked with the exact solutions. Based on the analytic results for a
single solid line conductor the requirements on the FEM mesh were specified to guaranty a high
quality of the resolution of the dominating effects in the electric system. When fulfilling these
requirements, the numeric error of skin and proximity losses is less than 1 % within the frequency
range from 1 kHz to 1 MHz. Afterwards, the FEM and the PEEC method were compared for a
more realistic test structure: A 3D setup of a twisted conductor of 30 mm length consisting of 19
strands. The resulting frequency dependent loss curves of both numeric methods agreed well up

99
CHAPTER 5. Summary and outlook

to 1 MHz and therefore demonstrate the reliability of the two approaches. However, due to the
aforementioned differences in the solving process, the PEEC method enables the reduction of the
computational time and the memory requirements by several orders of magnitude.
Nevertheless, even though the computation of all relevant effects in litz wires got feasible with the
PEEC approach, the simulation of an entire inductive system is not yet possible with reasonable
computational resources: The conductor length in common applications exceeds the one in the test
setup by more than a factor of 100 and additionally the effects caused by materials with magnetic
permeability are insufficiently resolved in the PEEC method.
Due to the fact that the entire litz wire system is not solvable by one individual method, the simu-
lation problem itself has to be split: The simulation of the internal effects causing litz wire losses
has to be separated from the calculation of the external magnetic field distribution in the entire
simulation domain.
The approach realized in this thesis bases on the fact that the resistance of each strand within a
well-twisted litz wire is almost the same in case of a good solder joint. Consequently, the exci-
tation current is evenly distributed on all strands. Even in case of a local inhomogeneous current
density distribution within one strand (due to the skin and the proximity effect), the current density
distribution in the entire cross section of the conductor is approximately homogeneous. Based on
that, there is no difference between the external behavior of a litz wire and a solid conductor with
a homogenous current density distribution. However, the latter case does not require the resolu-
tion of the complex substructure of the conductor. Consequently, this task is predestined for the
FEM, because the required homogenous current density distribution can be realized by specialized
excitation conditions in the simulation. Moreover, this numeric method is well-established in the
industry and therefore provides a direct geometry import and sophisticated autonomous meshing
algorithms. Based on those results, the magnetic field distribution is calculated in the entire simu-
lation domain and evaluated in cuts along the length of the conductor. The corresponding external
magnetic field is extracted and provided as a boundary condition to three different approaches
calculating the frequency dependent power losses on litz wire level:

• An analytic approach using a homogenization of the stranded conductors.

• The numeric PEEC approach with a realistic geometrical litz wire structure.

• The measured lambda factors interpolating between the losses of ideal twisted and parallel
stranded conductors.

All three approaches are described in this thesis in detail starting with the mathematical basis, the
implementation in the algorithms, and the pros and cons with regard to realistic litz wire conduc-
tors.
The presented coupled workflow is implemented in a Python program called SlicerPro enabling
the calculation of the frequency dependent power losses for varying litz wires with the three cal-
culation methods. Based on the cut based magnetic field data provided by the FEM simulation,
the program calculates and visualizes the results in several seconds. Additionally, all approaches
presented in this thesis concerning the preparation and the interpolation of the input data is imple-
mented and accessible by the graphical user interface.
The workflow was verified with regard to its functionality and accuracy in case of realistic prob-
lems in the design period of power electronic systems. Four representative experimental setups
are manufactured and measured, and additionally “replicated” and simulated on the computer. In
the simulation domain, the coupled simulation approach is used with the PEEC method and the
measurement-based lambda factors on litz wire level. The results demonstrate the fundamental
suitability of these two configurations of the workflow for calculating litz wire losses in realistic
power electronic systems. For air coil setups the simulation enables the prediction of the absolute

100
 5.0.

values of the power losses up to 1 MHz with less than 5 % deviation from the measured data in av-
erage, even for highly complex winding structures. However, the accuracy of the results decreases
by using ferrite material in the simulation setup. This is due to the fact, that core losses in the fer-
rites need to be treated, but corresponding simulation models are not yet reliable. This inaccuracy
primarily relates to comparisons of absolute values. Nevertheless, the simulation results of setups
containing some unconnected ferrite strips can be still used for qualitative studies in a lot of cases.
Especially in the early design process of power electronic systems, variations of the geometry are
needed to identify the best setup with regard to efficiency, losses or costs.
Nevertheless, for extreme scenarios of closed E-cores with a high magnetic flux density, the para-
sitic effects in the core dominate the measurements and thus the simulation results differ strongly
from the measured data.
Considering the two approaches on litz wire level, both methods significantly differ with regard to
the information required for the calculation of the frequency depending losses. The measurement
based lambda approach requires a sample of the experimental conductor and suitable measurement
equipment. The numeric PEEC approach needs detailed data about the inner litz wire structure
and a twisting algorithm to build up the simulation model. However, even though first manu-
facturers provide all relevant information about their conductors, common simulation approaches
fail to design realistically packed litz wires. In most cases, the copper fill factor is too low and
therefore the resulting diameter of the simulated conductor is larger than that of the real litz wire.
Consequently, the PEEC method is not yet reliable for tightly packed conductors and therefore the
lambda approach is recommend to be used in such cases.
The experiments and verifications presented in this thesis have demonstrated that the coupled
simulation approach provides a reliable workflow for real applications in the field of power elec-
tronics. Due to the implementation in the SlicerPro program a large variety of geometries, and
winding and litz wire structures can be simulated in several minutes. This improvement closes the
methodological gap with regard to the calculation of litz wire losses in power electronic systems.
However, the verification of the simulations by comparison with experimental results must be ex-
tended to gain more insight into the limits of different simulation methods for real problems. As
a consequence, the design process of new power electronic components should be accompanied
by the simulation in parallel. Based on that, improvements of the experimental system and the
simulation methods can be gained the most efficiently. However, in the next step the simulation
algorithms need to be enhanced and adapted to allow the application to future power electronic
requirements: The design and calculation of non-circular conductor cross sections, more real-
istic packaging of litz wires and the resolution of the physical effects in the domain of several
megahertz. Due to the modular structure of the SlicerPro program such future adaptions can be
implemented in a straight-forward way and directly compared to the standard methods.
Another aspect of further research activities should focus on the optimization of the litz wire twist-
ing structure with regard to application requirements. Depending on the power electronic system
and associated load cases an individual litz wire can be designed taking into account manufactur-
ing limitations and costs. Such a virtual prototyping significantly reduces costs and time of the
development process and paves the way to customized power electronic systems. Moreover, the
efficient prediction of litz wire losses enables the optimization of the overall system - consisting of
circuit, core and winding losses. Based on parametric geometries, a variety of setups can be built
up with different core materials and litz wire structures. With the help of these results, engineers
will gain a much deeper and more fundamental knowledge and understanding of power electronic
systems. Moreover, the combination of virtual prototyping with optimization strategies, e.g. using
genetic algorithms, allows a further extension of the simulation capabilities in the field of system
and product development.

101
CHAPTER 5. Summary and outlook

102
 A
Appendix

1.1. Definition of a homogenous current density distribution in a con-

ductor

Source Code A.1: apdl_stranded_current.apdl

1 /com , NEEDED : arg1 = current in the coil ; dia = diameter of the conductor
2 /com , NEEDED : named selection for coil ( all elements )
3 /com , NEEDED : named selection for c%i% ( volume ) and f%i% ( start face )
4 /com , for each coil segment
5 PI =3.14159265359
6 r= dia /2/1000 ! diameter of conductor [m]
7 sd = arg1 /( r*r* Pi ) ! current density [A/m ^2]
8 sd = sd * sqrt (2) ! amplitude
9
10 / prep7
11 ! local coordinate system ( CS ) for each element in the coil
12 cmsel ,s , coil
13 *get , enum1 , elem ,, count
14 ee =0
15 *do ,i ,1 , enum1
16 ee = elnext ( ee )
17 n1 = nelem (ee ,1)
18 n2 = nelem (ee ,2)
19 n5 = nelem (ee ,5)
20 cs ,100+ i ,0 ,n1 ,n2 , n5
21 emod ,ee , esys ,100+ i
22 * enddo
23 alls
24 [...] !
25
26 ! static calculation
27 / solu
28 solve
29

103
 APPENDIX

30 ! Current density direction is used for gauging the local CS

31 ! ’sd ’ is defined everywhere as the absolute value
32 cmsel ,s , coil
33 bfe ,all ,js ,2 , sd
34 alls
35

36 ! Harmonic calculation starts

37 / solu
38 frq = arg2
39 anty , harmic
40 harf , frq
41 alls

1.2. Export of the spatial magnetic field distribution (Ansys EMag)

B - Export of the spatial magnetic field distribution (Ansys EMag)

Source Code A.2: apdl_post_hdaten.apdl

1  /com , NEEDED : named selection for coil named coil ( elements )
2 /com , NEEDED : named selection for of_coil ( face for post ) and
3 /com , of_coil_start ( start edges ) for total coil
4
5 / post1
6 set , last ! use the last results
7
8 / delete , hdaten , txt ! delete existing file
9
10 /com , speichern aller H Werte an allen Knoten
11 /com , Vorbereitung der Daten
12 alls ! select all ( parts )
13 csys ,0 ! use coordinate system 0
14
15 cmsel ,s , coil ! select component coil
16 nsle ,s ! select nodes
17

18 /com , Evaluate the number of nodes in the selection (-> nmax )

19 *get , nmax , node ,,num , max
20
21 /com , Prepation of an Array for all nodes (x ,y ,z coordinates , H - field )
22 *dim , mypost , array , nmax ,6
23 * vget , mypost (1 ,1) , node ,,h ,x ,
24 * vget , mypost (1 ,2) , node ,,h ,y ,
25 * vget , mypost (1 ,3) , node ,,h ,z ,
26 * vget , mypost (1 ,4) , node ,,loc ,x
27 * vget , mypost (1 ,5) , node ,,loc ,y
28 * vget , mypost (1 ,6) , node ,,loc ,z
29 *dim , mask , array , nmax ,
30
31 /com , Macro for the export of the data
32 * create , exp_data , mac
33 * vget , mask (1 ,1) , node ,, nsel ! defines the mask - flags on the selected node
34 * cfopen , hdaten ,txt ,, append ! open file - add data
35 * vwrite ,’HX ’,’HY ’,’HZ ’,’ POS_X ’,’ POS_Y ’,’ POS_Z ’,
36 (12 A ) ,(12 A ) ,(12 A ) ,(12 A ) ,(12 A ) ,(12 A),
37 * vmask , mask (1 ,1) ! activate mask for further order ( data of the simulation )

104
 APPENDIX

38 * vwrite , mypost (1 ,1) , mypost (1 ,2) , mypost (1 ,3) , mypost (1 ,4) , mypost (1 ,5) , mypost (1 ,6)
39 (6 ES14 .7)
40 * cfclose , hdaten , txt ! close file
41 * end
42
43 /com , 1. Segment
44 cmsel ,s , of_coil ! selection of_coil ( boundary face of the coil )
45 esln ,s
46 cmsel ,r , coil ! get elements at the face consting to the coil
47 nsle ,r , corner ! restrict to nodes
48 cm , post_node , node ! post_nodes = all nodes at the boundary face
49 cmsel ,r , of_coil_start ! start face of the coil
50 cm , n_akt , node ! n_akt - nodes used in the current segment
51 exp_data ! evaluation of the results
52
53 nn =1
54 /com , all further segments
55 * dowhile , nn
56 esln ,s ! provides adjacent elements to current selection
57 cmsel ,r , coil ! restriction to coil elements
58 nsle ,s
59 cmsel ,r , post_node ! restrict to post_nodes
60 cmsel ,u , n_akt ! ! unselect nodes of the last segment
61 exp_data
62 *get ,nn , node ,, count
63 cmsel ,a , n_akt
64 cm , n_akt , node ! update n_akt
65 * enddo
66

67 /com , Export of the Data

68 / COPY , hdaten ,txt ,, hdaten_ % arg1 %,txt , ’/ user_data / rosskopf /’

1.3. Standard output file of the PEEC method for litz wires
#1 R without H_ext: |Frq[Hz]|Imag(Z)|real(Z)[Ohm]|
1.00E+01,1.14E-05,1.61E-03,
3.16E+01,3.60E-05,1.61E-03,
1.00E+02,1.14E-04,1.61E-03,
3.16E+02,3.60E-04,1.61E-03,
1.00E+03,1.14E-03,1.61E-03,
1.00E+04,1.14E-02,1.64E-03,
1.21E+04,1.38E-02,1.65E-03,
1.47E+04,1.67E-02,1.66E-03,
1.78E+04,2.02E-02,1.69E-03,
2.15E+04,2.45E-02,1.72E-03,
2.61E+04,2.97E-02,1.76E-03,
3.16E+04,3.59E-02,1.81E-03,
3.83E+04,4.35E-02,1.88E-03,
4.64E+04,5.26E-02,1.96E-03,
5.62E+04,6.37E-02,2.05E-03,
6.81E+04,7.71E-02,2.15E-03,
8.25E+04,9.33E-02,2.26E-03,

105
APPENDIX

1.00E+05,1.13E-01,2.39E-03,
1.21E+05,1.37E-01,2.54E-03,
1.47E+05,1.65E-01,2.72E-03,
1.78E+05,2.00E-01,2.94E-03,
2.15E+05,2.43E-01,3.23E-03,
2.61E+05,2.94E-01,3.60E-03,
3.16E+05,3.55E-01,4.05E-03,
3.83E+05,4.30E-01,4.60E-03,
4.64E+05,5.21E-01,5.26E-03,
5.62E+05,6.30E-01,6.03E-03,
6.81E+05,7.63E-01,6.96E-03,
8.25E+05,9.24E-01,8.10E-03,
1.00E+06,1.12E+00,9.50E-03,
#2 P_loss H_ext: |Frq[Hz]|H_ext|
0.00E+00,1,
1.00E+02,3.65E-15,
1.00E+04,3.65E-11,
1.00E+05,3.68E-09,
2.15E+05,1.75E-08,
1.00E+06,4.44E-07,
#3 Sim_infos
len:0.18
pitch:3.000000e-02,3.60E-02
number of strands:35,7
radius[strand]:5e-05
insulation:0.1
discretization[cross-section]:4
discretization[length]:30
max[x,y,z]:1.800549e-01,1.22E-03,1.23E-03
Date:20-Jul-2016 04:51:51

106
 List of publications
B
• C. Joffe, S. Ditze, and A. Rosskopf, “A novel positioning tolerant inductive power transfer
system,” in Electric Drives Production Conference (EDPC), 2013 3rd International, pp. 1–7,
Oct 2013

• A. Roßkopf, E. Bär, and C. Joffe, “Influence of inner skin- and proximity effects on con-
duction in litz wires,” IEEE Transactions on Power Electronics, vol. 29, pp. 5454–5461, Oct
2014

• A. Roßkopf, C. Joffe, and E. Bär, “Calculation of ohmic losses in litz wires by coupling
analytical and numerical methods,” in Electric Drives Production Conference (EDPC), 2014
4th International, pp. 1–6, Sept 2014

• C. Joffe, A. Roßkopf, S. Ehrlich, C. Dobmeier, and M. März, “Design and optimization of a

multi-coil system for inductive charging with small air gap,” in 2016 IEEE Applied Power
Electronics Conference and Exposition (APEC), pp. 1741–1747, March 2016

• A. Rosskopf, C. Joffe, E. Baer, and C. Bonse, “Calculation of power losses in litz wire
systems by coupling FEM and PEEC method,” Power Electronics, IEEE Transactions on,
vol. PP, no. 99, pp. 1–1, 2015

• S. Ditze, A. Endruschat, T. Schriefer, A. Rosskopf, and T. Heckel, “Inductive power transfer

system with a rotary transformer for contactless energy transfer on rotating applications,”
2016

Best Paper

• A. Roßkopf, “Dimension reduction methods for solving multiscale problems in electric sys-
tems,” Ansys CADFEM Users Meeting, 2013

• A. Roßkopf, “Wicklungsverluste in leistungselektronischen Systemen mit Hochfrequenzl-

itzen,” Ansys CADFEM Users Meeting, 2015

107
CHAPTER 5. List of publications

108
 Bibliography

[1] Prof. Dr. Martin März, “Lösungen für lokale Gleichstromnetze: Fraunhofer IISB zeigt
Neuheiten auf der PCIM Europe 2014,” in Pressemitteilung, IISB.

[2] T. Heckel, C. Rettner, and M. März, “Fundamental efficiency limits in power electronic
systems,” in 2015 IEEE International Telecommunications Energy Conference (INTELEC),
pp. 1–6, Oct 2015.

[3] S. Selberherr, Analysis and simulation of semiconductor devices. Springer Science & Busi-
ness Media, 2012.

[4] D. A. Tziouvaras, P. McLaren, G. Alexander, D. Dawson, J. Esztergalyos, C. Fromen,

M. Glinkowski, I. Hasenwinkle, M. Kezunovic, L. Kojovic, et al., “Mathematical mod-
els for current, voltage, and coupling capacitor voltage transformers,” IEEE Transactions
on Power Delivery, vol. 15, no. 1, pp. 62–72, 2000.

[5] H. Kaimori, A. Kameari, and K. Fujiwara, “Fem computation of magnetic field and iron loss
in laminated iron core using homogenization method,” IEEE Transactions on Magnetics,
vol. 43, no. 4, pp. 1405–1408, 2007.

[6] C. J. Dunlop, “Modeling magnetic core loss for sinusoidal waveforms,” tech. rep., Mas-
sachusetts Institute of Technology, Master’s thesis, 2008.

[7] Rudolf Pack GmbH & Co., “Technical Data RUPALIT® Litz wire.” http://www.
pack-feindraehte.de/en/products/litzwire/litzentabelle.pdf.

[8] R. Zhang, J. White, and J. Kassakian, “Fast simulation of complicated 3-d structures above
lossy magnetic media,” Magnetics, IEEE Transactions on, vol. 50, pp. 1–16, Oct 2014.

[9] A. Rosskopf, C. Joffe, E. Baer, and C. Bonse, “Calculation of power losses in litz wire
systems by coupling FEM and PEEC method,” Power Electronics, IEEE Transactions on,
vol. PP, no. 99, pp. 1–1, 2015.

[10] G. S. Ohm, Die galvanische Kette, mathematisch bearbeitet. T. H. Riemann, Berlin, 1827.

[11] J. P. Joule, “Annals of electricity, magnetism, and chemistry 8,” Philosophical Transactions
of the Royal Society of London, pp. 219–224, 1842.

[12] J. C. Maxwell, “A dynamical theory of the electromagnetic field,” Philosophical Transac-

tions of the Royal Society of London, pp. 459–512, 1865.

109
Bibliography

[13] J. Mühlethaler, J. Kolar, and A. Ecklebe, “Loss modeling of inductive components em-
ployed in power electronic systems,” in Power Electronics and ECCE Asia (ICPE ECCE),
2011 IEEE 8th International Conference on, pp. 945–952, May 2011.
[14] J. Lammeraner and M. Stafl, Eddy Currents. Iliffe Books Ltd., 1966.
[15] A. Kirsch and F. Hettlich, “The mathematical theory of maxwell’s equations,” Lecture
notes, 2009.
[16] H. C. M. u. L. J. T. M. J. Turner, Ray W. Clough, “Stiffness and deflection analysis of
complex structures,” Journal of the Aeronautical Sciences (Institute of the Aeronautical
Sciences), vol. 23, no. 9, 1956.
[17] A. E. Ruehli, “Inductance calculations in a complex integrated circuit environment,” IBM
J. Res. Dev., vol. 16, pp. 470–481, Sept. 1972.
[18] S. Lupi, F. Dughiero, E. Baake, and J. Lavers, “State of the art of numerical modeling for
induction processes,” COMPEL-The international journal for computation and mathematics
in electrical and electronic engineering, vol. 27, no. 2, pp. 335–349, 2008.
[19] M. N. Sadiku, Numerical Techniques in Electromagnetics. CRC Press LLC, 2 ed., 2001.
[20] F. W. Grover, Inductance calculations: working formulas and tables. Courier Corporation,
2004.
[21] E. B. Rosa, The self and mutual inductances of linear conductors. US Department of Com-
merce and Labor, Bureau of Standards, 1908.
[22] P. Lucht, “Transmission lines and maxwell’s equations,” 2014.
[23] S. Ramo and J. Whinnery, Felder und Wellen in der modernen Funktechnik. VEB Verlag
Technik, 1960.
[24] J. Mühlethaler, Modeling and Multi-Objective Optimization of Inductive Power Compo-
nents. PhD thesis, ETH Zürich, 2012. DISS. ETH NO. 20217.
[25] J. Biela, “Wirbelstromverluste in Wicklungen induktiver Bauelemente,” ETH Zürich, V1.73
/ Okt. 2012.
[26] M. Albach, “Vorlesung: Elektromagnetische Felder II,”
[27] J. Ferreira, “Appropriate modelling of conductive losses in the design of magnetic compo-
nents,” in Power Electronics Specialists Conference, 1990. PESC ’90 Record., 21st Annual
IEEE, pp. 780–785.
[28] J. Ferreira, “Analytical computation of ac resistance of round and rectangular litz wire wind-
ings,” in IEE Proceedings B (Electric Power Applications), vol. 139, pp. 21–25, IET, 1992.
[29] J. Ferreira, “Improved analytical modeling of conductive losses in magnetic components,”
Power Electronics, IEEE Transactions on, vol. 9, pp. 127–131, Jan 1994.
[30] M. Albach, M. Döbrönti, and H. Roßmann, “Wicklungsverluste in Spulen und Trafos aus
HF-Litze,” elektronik industrie, no. 10, pp. 32–34, 2010.
[31] H. Rossmanith, M. Döbrönti, M. Albach, and D. Exner, “Measurement and characterization
of high frequency losses in nonideal litz wires,” Power Electronics, IEEE Transactions on,
vol. 26, pp. 3386–3394, Nov 2011.

110
 Bibliography

[32] C. Sullivan and R. Zhang, “Analytical model for effects of twisting on litz-wire losses,” in
Control and Modeling for Power Electronics (COMPEL), 2014 IEEE 15th Workshop on,
pp. 1–10, June 2014.

[33] M. Albach, “Vorlesung: Induktive Komponenten,”

[34] A. Roßkopf, E. Bär, and C. Joffe, “Influence of inner skin- and proximity effects on con-
duction in litz wires,” IEEE Transactions on Power Electronics, vol. 29, pp. 5454–5461,
Oct 2014.

[35] P. G. Ciarlet, The finite element method for elliptic problems, vol. 40. Siam, 2002.

[36] ANSYS, Mechanical Release, Help 15.0 , “Documentation: help/ans/elem/hlp/e/-

solid236.html.”

[37] S. Brenner and R. Scott, The mathematical theory of finite element methods, vol. 15.
Springer Science & Business Media, 2007.

[38] A. Ern and J.-L. Guermond, Theory and practice of finite elements, vol. 159. Springer
Science & Business Media, 2013.

[39] A. Quarteroni and A. Valli, Numerical approximation of partial differential equations,

vol. 23. Springer Science & Business Media, 2008.

[40] S. Larsson and V. Thomée, Partial differential equations with numerical methods, vol. 45.
Springer Science & Business Media, 2008.

[41] R. Clough, “The finite element method in plane stress analysis,” in 2nd Conference on
Electronic Computation, A.S.C.E. Structural Division, September 1960.

[42] O. C. Zienkiewicz and R. L. Taylor, The Finite Element Method Set. Butterworth-
Heinemann, 6 ed., 2005.

[43] E. Schmidt, “Notentialformulierungen für Finite Elemente Analysen von Elektrischen

Maschinen und Transformatoren,” Magnetics, IEEE Transactions on, vol. 44, pp. 718–721,
June 2008.

[44] S. Humphries, “Finite-Element Methods for Electromagnetics.” Website. http://www.

fieldp.com/femethods.html.

[45] J. A. Stratton, Electromagnetic theory. John Wiley & Sons, 2007.

[46] O. Biro and K. Preis, “On the use of the magnetic vector potential in the finite-element
analysis of three-dimensional eddy currents,” Magnetics, IEEE Transactions on, vol. 25,
pp. 3145–3159, July 1989.

[47] J. Weiss and V. Garg, “One step finite element formulation of skin effect problems in mul-
ticonductor systems with rotational symmetry,” Magnetics, IEEE Transactions on, vol. 21,
pp. 2313–2316, Nov 1985.

[48] Ansys Mechanical Release, Help 15.0, “Documentation: help/ans/thry/thy/emg1.html.”

[49] R. Acevedo and G. Loaiza, “A Fully-Discrete Finite Element Approximation for the Eddy
Currents Problem,” Ingenieray Ciencia, vol. 9, pp. 111 – 145, 06 2013.

[50] J. D. Jackson, Classical Electrodynamics. John Wiley & Sons, 3 ed., 1999.

111
Bibliography

[51] P. Zhou, Z. Badics, D. Lin, and Z. Cendes, “Nonlinear T-Ω Formulation Including Motion
for Multiply Connected 3-D Problems,” Magnetics, IEEE Transactions on, vol. 44, pp. 718–
721, June 2008.

[52] M. Repetto and F. Trevisan, “Global formulation of 3D magnetostatics using flux and
gauged potentials,” International Journal for Numerical Methods in Engineering, vol. 60,
no. 4, pp. 755–772, 2004.

[53] I. Munteanu, “Tree-cotree condensation properties,” tech. rep., CST – Computer Simulation
Technology, 2002.

[54] G. Poole, Y.-C. Liu, and J. Mandel, “Advancing analysis capabilities in ansys through solver
technology,” Electronic Transactions on Numerical Analysis, vol. 15, pp. 106–121, 2003.

[55] A. Roßkopf, C. Joffe, and E. Bär, “Calculation of ohmic losses in litz wires by coupling
analytical and numerical methods,” in Electric Drives Production Conference (EDPC), 2014
4th International, pp. 1–6, Sept 2014.

[56] M. Costabel, “Principles of boundary element methods,” Computer Physics Reports, vol. 6,
no. 1-6, pp. 243–274, 1987.

[57] A. Sutradhar, G. Paulino, and L. J. Gray, Symmetric Galerkin boundary element method.
Springer Science & Business Media, 2008.

[58] A. E. Ruehli and P. A. Brennan, “Efficient capacitance calculations for three-dimensional

multiconductor systems,” Microwave Theory and Techniques, IEEE Transactions on,
vol. 21, pp. 76–82, Feb 1973.

[59] A. E. Ruehli, “Equivalent circuit models for three-dimensional multiconductor systems,”

Microwave Theory and Techniques, IEEE Transactions on, vol. 22, pp. 216–221, Mar 1974.

[60] n. MultiPEEC (3.5.2), title= User’s Guide 2012.

[61] C. Wollenberg and A. Gurisch, “Analysis of 3-d interconnect structures with peec using
spice,” Electromagnetic Compatibility, IEEE Transactions on, vol. 41, pp. 412–417, Nov
1999.

[62] K. Nabors and J. White, “Fastcap: a multipole accelerated 3-d capacitance extraction pro-
gram,” Computer-Aided Design of Integrated Circuits and Systems, IEEE Transactions on,
vol. 10, pp. 1447–1459, Nov 1991.

[63] M. Kamon, M. Tsuk, and J. White, “FASTHENRY: a multipole-accelerated 3-D inductance

extraction program,” Microwave Theory and Techniques, IEEE Transactions on, vol. 42,
pp. 1750–1758, Sept 1994.

[64] M. Kamon, M. J. Tsuk, C. Smithhisler, and J. White, “Efficient techniques for inductance
extraction of complex 3-d geometries,” in Computer-Aided Design, 1992. ICCAD-92. Di-
gest of Technical Papers., 1992 IEEE/ACM International Conference on, pp. 438–442,
IEEE, 1992.

[65] A. Ruehli, G. Antonini, J. Esch, J. Ekman, A. Mayo, and A. Orlandi, “Nonorthogonal peec
formulation for time- and frequency-domain em and circuit modeling,” Electromagnetic
Compatibility, IEEE Transactions on, vol. 45, pp. 167–176, May 2003.

112
 Bibliography

[66] V. Jandhyala, Y. Wang, D. Gope, and R. Shi, “Coupled electromagnetic-circuit simulation

of arbitrarily-shaped conducting structures using triangular meshes,” in Quality Electronic
Design, 2002. Proceedings. International Symposium on, pp. 38–42, 2002.

[67] G. Antonini, J. Delsing, J. Ekman, A. Orlandi, and A. Ruehli, “Peec development road map
2007,” tech. rep., Department of Electrical Engineering, University of L’Aquila, 2007.

[68] A. Ruehli and H. Heeb, “Circuit models for three-dimensional geometries including di-
electrics,” Microwave Theory and Techniques, IEEE Transactions on, vol. 40, pp. 1507–
1516, Jul 1992.

[69] V. Okhmatovski and A. Cangellaris, “A new technique for the derivation of closed-form
electromagnetic green’s functions for unbounded planar layered media,” Antennas and
Propagation, IEEE Transactions on, vol. 50, pp. 1005–1016, Jul 2002.

[70] F. Ferranti, M. S. Nakhla, G. Antonini, T. Dhaene, L. Knockaert, and A. E. Ruehli, “Mul-

tipoint full-wave model order reduction for delayed peec models with large delays,” IEEE
TRANSACTIONS ON ELECTROMAGNETIC COMPATIBILITY, vol. 53, no. 4, pp. 959–
967, 2011.

[71] J. Ekman, Electromagnetic Modeling Using the Partial Element Equivalent Circuit Method.
PhD thesis, EISLAB Dept. of Computer Science and Electrical Engineering, University of
Technology Luleå Sweden, 2003.

[72] C. Sullivan and R. Zhang, “Simplified design method for litz wire,” in Applied Power Elec-
tronics Conference and Exposition (APEC), 2014 Twenty-Ninth Annual IEEE, pp. 2667–
2674, March 2014.

[73] J. Acero, I. Lope, J. Burdio, C. Carretero, and R. Alonso, “Loss analysis of multistranded
twisted wires by using 3d-fea simulation,” in Control and Modeling for Power Electronics
(COMPEL), 2014 IEEE 15th Workshop on, pp. 1–6, June 2014.

[74] J.-R. Li and M. Kamon, “PEEC Model of a Spiral Inductor Generated by Fasthenry,” in Di-
mension Reduction of Large-Scale Systems (P. Benner, D. C. Sorensen, and V. Mehrmann,
eds.), vol. 45 of Lecture Notes in Computational Science and Engineering, pp. 373–377,
Springer Berlin Heidelberg, 2005.

[75] J. E. Garrett, Advancements of the Partial Element Equivalent Circuit Formulation. PhD
thesis, The University of Kentucky, 1997.

[76] R. Unbehauen, Grundlagen der Elektrotechnik: Allgemeine Grundlagen, Lineare Netzw-

erke, Stationäres Verhalten. Springer-Lehrbuch, Springer Berlin Heidelberg, 2013.

[77] S. M. Lorenz-Peter Schmidt, Gerd Schaller, Grundlagen der Elektrotechnik 3. Netzwerke.

Addison-Wesley Verlag, 2006.

[78] E. Horneber, Simulation elektrischer Schaltungen auf dem Rechner. Fachberichte Simula-
tion, Springer Berlin Heidelberg, 2013.

[79] M. Kamon, “Nonuniformly discretized reference planes in fasthenry 3.0.” Website, 1996.
http://www.layouteditor.net/wiki/FastHenry.

[80] J. W. Cooley and J. W. Tukey, “An algorithm for the machine calculation of complex fourier
series,” Mathematics of computation, vol. 19, no. 90, pp. 297–301, 1965.

113
Bibliography

[81] J. Phillips and J. White, “A precorrected-fft method for capacitance extraction of compli-
cated 3-d structures,” in Proceedings of the 1994 IEEE/ACM international conference on
Computer-aided design, pp. 268–271, IEEE Computer Society Press, 1994.

[82] Zhang, R.Y., “FastLitz - Realistic litz wire characterization,” March 2014. http://web.
mit.edu/ryz/www/zip/fastlitzv1a.zip.

[83] T. E. Tezduyar and A. Sameh, “Parallel finite element computations in fluid mechanics,”
Computer methods in applied mechanics and engineering, vol. 195, no. 13, pp. 1872–1884,
2006.

[84] R. Sabariego, P. Sergeant, J. Gyselinck, P. Dular, L. Dupré, and J. Melkebeek, “Fast mul-
tipole accelerated finite element-boundary element analysis of shielded induction heaters,”
IEEE transactions on magnetics, vol. 42, no. 4, pp. 1407–1410, 2006.

[85] D. Kürschner, Methodischer Entwurf toleranzbehafteter induktiver Energieübertra-

gungssysteme. Berichte aus der Elektrotechnik, Shaker, 2010.

[86] X. Nan and C. Sullivan, “An improved calculation of proximity-effect loss in high-
frequency windings of round conductors,” in Power Electronics Specialist Conference,
2003. PESC ’03. 2003 IEEE 34th Annual, vol. 2, pp. 853–860 vol.2, June 2003.

[87] N. C. Eccles, J. P. Steinbrenner, and J. Abelanet, “Solid modeling and fault tolerant mesh-
ingtwo complimentary strategies,” in 17th AIAA Computational Fluid Dynamics Confer-
ence, pp. 6–9, 2005.

[88] ANSYS Maxwell_v16 , “Documentation: Maxwell/v16/l01/introduction.pdf.”

[89] C. Joffe, A. Roßkopf, S. Ehrlich, C. Dobmeier, and M. März, “Design and optimization of a
multi-coil system for inductive charging with small air gap,” in 2016 IEEE Applied Power
Electronics Conference and Exposition (APEC), pp. 1741–1747, March 2016.

[90] S. Ramaswamy, R. Rastogi, and K. Shim, “Efficient algorithms for mining outliers from
large data sets,” vol. 29, no. 2, pp. 427–438, 2000.

[91] Z. He, S. Deng, and X. Xu, “An optimization model for outlier detection in categorical
data,” in Advances in Intelligent Computing, pp. 400–409, Springer, 2005.

[92] C. R. Sullivan, “Optimal choice for number of strands in a litz-wire transformer winding,”
IEEE Transactions on Power Electronics, vol. 14, no. 2, pp. 283–291, 1999.

[93] C. Bonse, “Analystischer ansatz zur verlustberechnung von komplexen litzenstrukturen für
hochfrequente leistungselektronik,” Masterthesis, 2017.

[94] D. C. Meeker, “An improved continuum skin and proximity effect model for hexago-
nally packed wires,” Journal of Computational and Applied Mathematics, vol. 236, no. 18,
pp. 4635–4644, 2012.

[95] J. Deng, W. Li, T. D. Nguyen, S. Li, and C. C. Mi, “Compact and efficient bipolar coupler
for wireless power chargers: Design and analysis,” IEEE Transactions on Power Electron-
ics, vol. 30, no. 11, pp. 6130–6140, 2015.

[96] V. Prasanth, S. Bandyopadhyay, P. Bauer, and J. A. Ferreira, “Analysis and comparison

of multi-coil inductive power transfer systems,” in Power Electronics and Motion Control
Conference (PEMC), 2016 IEEE International, pp. 993–999, IEEE, 2016.

114
 Bibliography

[97] R. Zhang, J. White, J. Kassakian, and C. Sullivan, “Realistic litz wire characterization us-
ing fast numerical simulations,” in Applied Power Electronics Conference and Exposition
(APEC), 2014 Twenty-Ninth Annual IEEE, pp. 738–745, March 2014.

[98] H. Roßmanith, R. Pack, and A. Stadler, Charakterisierung und Optimierung von Litze in
Drosseln und Transformatoren zur besseren Ausnutzung der Energie : Abschlussbericht
zum Verbundvorhaben. 2012.

[99] C. Bonse and A. Roßkopf, “Documentation: Slicerpro help,” 2015.

[100] C. Joffe, Entwurfsmethodik wirkungsgradoptimierter induktiver Energieübertragungssys-

teme mit erhöhter Spulenpositionstoleranz. PhD thesis, Friedrich-Alexander-Universität
Erlangen-Nürnberg, 2017.

[101] S. Prabhakaran and C. R. Sullivan, “Impedance-analyzer measurements of high-frequency

power passives: Techniques for high power and low impedance,” in Industry Applications
Conference, 2002. 37th IAS Annual Meeting. Conference Record of the, vol. 2, pp. 1360–
1367, IEEE, 2002.

[102] New England Wire Technologies, “Termination Methods for Litz Wire.” Website, 2007.
https://www.newenglandwire.com/~/media/Files/Litz_Wire_Termination_
guide_4th_Edition.pdf.

[103] Agilent Technologies, “Service Manual - 4294A Precision Impedance Analyzer.” Website.
http://www3.nd.edu/~nano/facilities/at_man_Agilent4294aServiceManual.
pdf.

[104] C. Rettner, “Forschungspraktikum Erstellung eines Fehlerrechnungsprogramms für den

Impedanceanalyzer Agilent 4294A in Excel,”

[105] S. Ditze, A. Endruschat, T. Schriefer, A. Rosskopf, and T. Heckel, “Inductive power transfer
system with a rotary transformer for contactless energy transfer on rotating applications,”
2016.

[106] A. Stadler, Messtechnische Bestimmung und Simulation der Kernverluste in weichmag-

netischen Materialien.

[107] C. Joffe, S. Ditze, and A. Rosskopf, “A novel positioning tolerant inductive power transfer
system,” in Electric Drives Production Conference (EDPC), 2013 3rd International, pp. 1–7,
Oct 2013.

[108] Ferroxcube, “3C90 - Material Specification.” Website, 2004. http://www.ferroxcube.

com/FerroxcubeCorporateReception/datasheet/3c90.pdf.

[109] A. Roßkopf, “Dimension reduction methods for solving multiscale problems in electric
systems,” Ansys CADFEM Users Meeting, 2013.

[110] A. Roßkopf, “Wicklungsverluste in leistungselektronischen Systemen mit Hochfrequenzl-

itzen,” Ansys CADFEM Users Meeting, 2015.

115