Full Text 01

Silicon Drift Detector Simulations for
Energy-Dispersive X-ray Spectroscopy in

Scanning Electron Microscopy
Sebbe Blokhuizen
Department of Physics, Fysikum

Master’s thesis, 30 ECTS
Master’s Programme in Computational Physics
Supervisors:
Eddy Ardonne - Stockholm University
Aron Beekman - Sioux Technologies
Thijs Withaar - Sioux Technologies
June 25, 2023
Silicon Drift Detector Simulations for Energy-Dispersive X-ray
Spectroscopy in Scanning Electron Microscopy
Sebbe Blokhuizen
Abstract
Scanning Electron Microscopy combined with Energy Dispersive X-ray Spectroscopy (SEM-EDS) is a
widely applied elemental microanalysis method. The integration of silicon drift detectors (SDDs) has notably
enhanced EDS performance, enabling precise elemental identification due to its large sensitive area and low
output capacitance.
Accurate simulations of SDDs can provide insights that enable the design and optimization of future
models without the need for costly and time-consuming experimental iterations. Moreover, the current
model-based quantification methods for EDS applications have reached their maximum predictive accuracy.
As such, creating a more accurate simulation model could help achieve a higher level of precision in these
quantification models, which would be immensely valuable for all EDS applications.
With this objective in mind, a simulation framework for modeling SDDs in EDS was developed based
on Geant4, Allpix Squared, and COMSOL Multiphysics. The simulation encompasses the entire physics
pipeline, including characteristic X-ray emission from the target sample and its absorption in the detector.
The generated charge carriers within the detector are propagated through the internal electric field of the
SDD, and their individual charge contribution is measured to simulate EDS spectra.
The simulated model was compared to existing literature and in-house experimental measurements,
showing strong agreement in the case of a well-tuned SDD. Limitations of the simulation framework are
discussed, and further research to enhance accuracy and speed is explored.
Keywords: X-ray spectroscopy, Silicon drift detector, Scanning Electron Microscopy, Detector Simula-
tion
i
Contents
1 Introduction 1
2 Theory 4
2.1 SEM-EDS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.1.1 SEM Operating Principles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.1.2 Energy-dispersive X-ray spectroscopy . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1.3 Characteristic X-rays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1.4 Bremsstrahlung . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.1.5 X-ray Interactions With Matter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2 Semiconductor Basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2.1 Band Theory of Solids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2.2 Intrinsic Semiconductors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2.3 Extrinsic Semiconductors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.2.4 pn-Junction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.2.5 Biased pn-Junctions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.3 Semiconductor Detector Principles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.3.1 Charge Creation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.3.2 Charge Propagation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.3.3 Signal Formation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.3.4 Digitalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.3.5 Incomplete Charge Collection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.4 Silicon Drift Detector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3 Methods 32
3.1 Software Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.1.1 Geant4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.1.2 Allpix Squared . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.1.3 COMSOL Multiphysics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.2 Pre-Simulation Phase . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.2.1 Electric and Weighting Field Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.2.2 Geant4 Geometry Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.3 Main Simulation Phase . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.3.1 Initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.3.2 Core Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.4 Post-Simulation Phase . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.4.1 Data extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.5 Allpix Squared Modifications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.5.1 Discretization Error in Pair Creation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.5.2 Front Contact ICC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.5.3 ROOTOutputWriter Speed Issue . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.5.4 Additional Changes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
ii
CONTENTS CONTENTS
4 Results 48
4.1 SDD Field Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.1.1 Electric Field . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.1.2 Doping Dependence Drift Lines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.1.3 Weighting Field . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.2 ICC front contact . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.2.1 Absorption Depth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.2.2 Detector Response Modelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
4.2.3 ICC Control Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
4.3 Comparison Experimental XRF Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.3.1 Setup of the XRF-Tool . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.3.2 Materials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.3.3 Source . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.3.4 Spectrum Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
4.4 Bias Tuning of SDD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.4.1 Effects of BC Voltage on Electric Fields . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.4.2 Effects of BC Voltage on Spectra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.5 Speed-Accuracy Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.5.1 Charge Grouping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.5.2 Spatial Precision . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
5 Future Research Recommendations 69

5.1 Inaccurate SDD Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.2 Fano factor and charge creation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.3 Front Contact Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.4 Open-Source pipeline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.5 Digitization Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.6 Drift Ring Biasing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.7 Speed Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.7.1 Adjoint Monte Carlo . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.7.2 GPU Programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
6 Conclusion 73
7 Bibliography 75
A Bias Voltage Sweep - Electric Fields 80
B Bias Voltage Sweep - Spectra 83
C Charge Grouping Spectra 85
D Spatial Precision Spectra 87
iii
Chapter 1
Introduction
Background and Context

Scanning Electron Microscopy (SEM) is a powerful technique widely used in various scientific fields for
imaging and analyzing the microstructure of materials at high resolution [1]. It provides valuable insights
into the surface morphology, composition, and elemental distribution of targeted samples [2]. In SEM, an
electron beam is scanned across the sample, and the resulting interactions between the electrons and the
sample produce signals that can be detected and analyzed to generate detailed images and spectroscopic
information [3].
One important application of SEM is Energy Dispersive X-ray Spectroscopy (EDS), which enables the
elemental analysis of materials [4]. When the high-energy electron beam interacts with the sample, it induces
the emission of characteristic X-rays. These X-rays have an energy that is specific to the elements present in
the targeted sample. When an emitted X-ray finds its way into the detector and interacts with the detector
material, a cloud of electron-hole pairs is created. The total number of created electron-hole pairs is closely
related to energy of the incident X-ray, as well as the effective band gap of the absorbing medium. The
electron-hole pairs can be detected to obtain the energy of the incident X-ray.
A commonly used X-ray detector in EDS is the Silicon Drift Detector (SDD) [5]. SDDs utilize a unique
circular drift ring design to guide charge carriers through the silicon bulk towards a very small centrally
located anode, which effectively reduces capacitance and detector noise. As a result of this, SDDs have
recently become widespread in scientific experiments as well as in industrial applications for detecting radia-
tion and energetic particles [6]. They are utilized in a wide range of applications, from cancer cell detection
research [7] to satellite-based spectral timing measurements [8].
If the generated charge carriers in the SDD are only partly collected, this leads to degraded energy
resolution and distorted X-ray spectra [9]. As the detector will measure fewer charge carriers, a lower energy
count will be registered. This phenomenon is responsible for lower-energy tailing in the spectra and is referred
to as incomplete charge collection (ICC). An important cause for ICC, especially in the case of lower energy
X-rays, is the dead layer in the front of the detector. This is a thin region on the surface of the detector
material where charge carriers generated by incident radiation are not efficiently collected.
To accurately simulate the energy spectra obtained from an SDD, it is crucial to model a broad range of
the underlying physical processes. This requires a comprehensive understanding of the interactions between
X-rays, electrons, and the active material of the detector. Furthermore, the behavior of charge carriers in
response to the electric field within the detector is essential to model the effects of ICC.
Research Problem and Motivation

The primary research problem of this thesis is to develop a comprehensive computational model that ac-
curately represents the underlying physical principles governing the operation of SDDs used in EDS. The
aim is to improve the understanding of how different tuning parameters, materials, or geometries affect the
response of the detector, enabling the design and optimization of future SDD models without the need for
costly and time-consuming experimental iterations.
1
CHAPTER 1. INTRODUCTION
Moreover, the widely used Phi-Rho-Z quantification model, which relies on a polynomial fit with empir-
ically determined correction factors, has limitations in terms of accuracy with a relative standard deviation
of approximately 2-4% [10]. By adopting a more physically rigorous approach through particle physics sim-
ulations within SEM and utilizing the simulation data to obtain precise quantification predictions, we could
overcome the limitations of traditional models and achieve higher accuracy quantification in EDS.
In summary, the research problem addressed in this thesis is to develop a computational model that
accurately represents the principles underlying the operation of SDDs in EDS. This model aims to overcome
the limitations of existing empirical models, improve accuracy in quantification predictions, and advance the
understanding of SEM-EDS systems.
Related Work
Modeling the detector response curve is a procedure that often relies heavily on empirical fudge factors to
fit the simulated spectrum to an experimental ground truth. The energy deposition spectra, which are the
result of energy in the form of particles or radiation entering the sensitive detector volume, are obtained
either through a Monte Carlo approach, such as Geant4, or through a deterministic approach. The detector
response is then obtained by convolving the energy deposition spectrum with a Gaussian distribution. Such
an approach is taken in several works, [11–14]. Additional spectral features such as the lower-energy tail,
due to ICC, and the shelves, due to hot photoelectrons, are added based on empirical parameters. While
this approach is computationally cheap, it lacks physical rigor.
Others tackle the issue solely with a Monte Carlo approach, such as [15], where Geant4 was used to
model the energy depositions in the SDD. Here, the SDD is modeled as a simple silicon volume, in which
energy is deposited by the incident X-rays. While this accurately portrays the total amount of energy that
is deposited within the silicon block during the interaction with incident X-rays, it treats the detector as an
ideal charge collector. As a consequence, all energy that is deposited within the silicon block is measured.
However, actual detectors can have losses due to several reasons, such as charge losses near the front contact,
recombination within the lattice, and incorrect biasing conditions, which cause the spectra to lose resolution.
Improvements to this approach are made in recent works by Biassoni and Gugiatti [16, 17], by incorporating
an analytical model charge collection efficiency due to the dead layer of the device.
An important work by Scholze and Procop [9] describes an analytical model of EDS in silicon detectors.
The model accounts for detector statistical noise, electronic noise, escape peaks and ICC due to the dead
layer. This ICC model was incorporated into the simulation used in this thesis to model the dead layer of
the detector.
As computational power increased over the last years, more physically robust methods based on a first-
principles approach were developed. In this approach, the electron-hole pairs generated within the detector
are tracked, and their individual charge or current contributions are collected for each measured count on
the energy spectrum [18–20].
In this thesis, a similar approach is taken as described in a recent work by Saleem et al. [19]. They utilize
a simulation pipeline based on Allpix Squared and COMSOL to model multi-element germanium detectors
under irradiance of an X-ray beam of 5 to 80 keV. Germanium detectors are much thicker and have a higher
atomic number than silicon, making them more suitable for higher-energy X-rays. They obtain accurate
results when comparing the simulation to experimental results of a soil sample rich in cadmium. In this
thesis, Saleem’s methodology is explored to model SDDs for use in EDS.
Methodology
A simulation framework was set up that encompasses the entire physics pipeline, starting from the particle
beam. Particle interactions, such as the photo-electric effect and scattering, are modeled using the Monte
Carlo simulation toolkit Geant4 [21]. Geant4 has been the simulation engine of choice for high-energy physics
experiments at the LHC and is extensively utilized by ESA and NASA for radiation dose estimation [22].
Recent advances to the toolkit have facilitated its expansion into new, lower-energy applications, such as
medical physics, biochemistry, and material science.
Following the absorption of radiation in the detector sensor, the energy depositions from Geant4 are
transferred to a recently developed framework Allpix Squared [23]. Allpix Squared handles the propagation
of charge carriers within the sensor, making use of a drift-diffusion model. The charge carriers are propagated
2
CHAPTER 1. INTRODUCTION
based on an imported map of the electric field within the detector. Additionally, charge induction on the
anode is modeled by using an imported weighting potential map, utilizing the Shockley-Ramo theorem
[24, 25]. A number of modifications to the Allpix Squared framework are introduced to improve its uses for
EDS.
The three-dimensional electric and weighting potential fields of the SDD are modeled using COMSOL
Multiphysics [26]. The generated fields are converted into a regular grid and passed to the Allpix Squared
framework.
Finally, the results for each individual event are combined and extracted to obtain the simulated energy
spectra.
Initial State at Sioux Technologies

This thesis was performed under supervision of Sioux Technologies at the Mathware department in Eind-
hoven. Sioux Technologies is a high-tech solutions partner that develops, innovates and assembles high-tech
systems. They were interested in developing an accurate model for SEM-EDS systems using particle physics
simulations.
At the start of the thesis, an operational prototype simulation was provided which modeled a SEM-EDS
setup in Geant4. The simulation encompassed an electron beam shooting at a sample, and a rudimentary
detector and window model. The detector was modeled using an idealized approach, using a silicon disk. This
simulation was extended by utilizing the Allpix Squared framework to accurately model the behaviour within
the detector, and scripts were created such that the new framework could be linked to the old simulation,
as well as be run in a stand-alone manner.
Thesis Outline
The thesis is organized into several chapters that cover the theoretical background, simulation techniques,
simulation results, limitations, and future research directions.
• Chapter 2 provides an overview of the theoretical background, including the principles of semiconductor
detectors, X-ray interactions, and charge transport mechanisms.
• Chapter 3 describes in detail the simulation techniques employed in this study. Also, modifications
that were made to the existing software are presented.
• Chapter 4 presents the simulation results and discusses the effects of different parameters on the
detector’s performance. The results are compared to experimentally obtained spectra.
• Chapter 5 highlights some shortcomings of the model and suggests future research directions.
• Chapter 6 summarizes the conclusions drawn from the study.
3
Chapter 2
Theory
This chapter will present the theoretical aspects related to the research. It covers the operating principles of
Scanning Electron Microscopy (SEM) and Energy-Dispersive X-ray Spectroscopy (EDS). Additionally, the
fundamentals of semiconductor detectors will be covered. The chapter concludes by introducing the specific
detector of interest, the Silicon Drift Detector (SDD).
2.1 SEM-EDS
A Scanning Electron Microscope, or SEM, is a type of electron microscope. It operates by directing a high-
energy electron beam towards the sample surface. As the electrons interact with the atoms of the sample,
they generate multiple different signals that carry information about the composition and topography of
the sample. One of the main reasons to use a SEM as opposed to a conventional optical microscope is its
great resolving power. The maximum magnification power of conventional optical microscopes is limited to
around 1000x magnification due to the fundamental limitations of visible light. As SEMs rely on electrons
for imaging, they can obtain a resolution that is orders of magnitude greater than their optical counterparts.
Modern SEMs have reached magnifications up to 1,000,000x, with resolutions down to the nanoscale [1].
Because of this, SEMs are currently used in many disciplines, where they prove to be an invaluable tool in
the analysis of organic and inorganic materials.
2.1.1 SEM Operating Principles

In a SEM, a sample is scanned by an energetic electron beam in a raster pattern. These electrons are then
accelerated and focused through a series of electromagnetic lenses and guided towards the sample. When
the electrons reach the sample, they can interact with the atoms of the sample and emit different types of
electrons and photons, depending on the depth of the interaction.
The main two types of emitted electrons used in SEM are backscattered electrons (BSEs) and secondary
electrons (SEs). BSEs and SEs are emitted from different interaction depths, as shown in figure 2.1.
BSEs are high-energy electrons that originate from interactions occurring in deeper regions of the sample
(∼500 nm) and are the result of elastic collisions between the incident electrons and the sample. Atoms with
a higher atomic number have a higher probability of producing an elastic collision because of their greater
cross-sectional area. As a result, the total number of BSEs generated is directly proportional to the atomic
number of the sample. This provides information regarding the sample’s elemental composition, as heavy
elements will appear brighter than light elements in the resulting image. The BSE detectors are usually
arranged in a ring-shaped detector above the sample. The reasons for this are twofold; first, this design
optimizes the collection of electrons by having a large solid angle used for detection. The second reason is
to allow for the detection of electrons that are scattered from an asymmetric sample, which does not have a
uniform radial scatter pattern.
In contrast, SEs originate from the surface regions (∼50 nm) of the sample and are a result of inelastic
interactions between the primary electron beam and the sample. They have lower energies than BSEs, usually
below 50 eV. Detectors for SEs are placed at the side of the sample at an angle. The intensity of the signal
4
2.1. SEM-EDS CHAPTER 2. THEORY
relies on the number of secondary electrons that are detected. When the primary beam is perpendicular to
the surface of the sample, the interaction volume is uniform around the beam axis. As the angle of incidence
increases, the number of emitted secondary electrons increases. This leads to steeper surfaces and edges
appearing brighter compared to flat surfaces, resulting in images with detailed information regarding the
surface topography [2]. A schematic of an SEM setup is shown in figure 2.2.
Figure 2.1: Cross section of the electron inter-

action volume in a sample. The type of signal
generated depends strongly on the depth of the
interaction: SEs - 50 nm, BSEs - 500 nm, char- Figure 2.2: Schematic drawing of an SEM-EDS
acteristic X-rays - 1 µm, Bremsstrahlung - 1.5 setup.
µm.
2.1.2 Energy-dispersive X-ray spectroscopy

The previous section discussed the information that can be obtained through electron emissions. However,
as shown in figure 2.1, the sample will not only emit electrons but also X-rays. The energy of these X-
rays provides valuable quantitative and qualitative information regarding the elemental properties of the
sample. Qualitative analysis can be carried out by measuring the energy of the emitted X-rays. Additionally,
quantitative analysis can be performed by looking at the count distribution of the measured X-rays. Through
the ratio of peak heights in the spectra it can be inferred how much of a specific element is present within the
sample. The process of measuring the emitted X-rays and dispersing them based on their energy is known
as Energy-Dispersive X-ray Spectroscopy or EDS.
Bombardment of the sample with high energy electrons will cause the emission of two types of X-rays,
• Characteristic X-rays - These provide local information about the elemental composition of the sample.
• Bremsstrahlung - These are usually not of any use in EDS applications and are often considered
background noise.
These processes through which the two types of X-rays are generated will be briefly discussed.
2.1.3 Characteristic X-rays

The most important X-rays emitted from the sample for the use of EDS are Characteristic X-rays. These
X-rays have an energy that is characteristic of the targeted element.
If a sufficiently energetic electron penetrates the outer electron shell of one of the sample’s atoms, it can
interact with one of the inner-shell electrons. By transferring some of its kinetic energy to an inner-shell
5
electron, the inner-shell electron can be ejected from the shell, leaving behind a hole in the inner orbital
shell. The affected atom is thus excited into an ionized state. It can relax back into its ground state by
filling the inner-shell vacancy with an electron from an outer shell. As the electrons in the outer shell lie at a
higher energy level than the inner shell, the difference in energy has to be released to adhere to the principle
of energy conservation. This can be achieved in two ways, either by the emission of an X-ray or by emitting
an electron through the Auger effect.
In the Auger effect, the energy difference is transferred to another electron, called an Auger electron.
The emission through the Auger effect is more likely in atoms with lower Z, whereas the emission of a
characteristic X-ray is more prevalent in atoms with a higher Z, [27].
In both cases, the emitted energy relies on the specific electron energy levels. As each element has a
unique set of energy levels, X-rays are emitted at energies that are specific to each element. Figure 2.3 shows
a schematic of both processes in the case of a silicon sample.
Figure 2.3: Schematic of electron transitions in case of removal of an electron from the K-shell. Left shows
a KLL Auger transition and right shows the emission of a Kα2 characteristic X-ray. The electron binding
energies for silicon, in electron volts, are shown on the left.
In order to use emitted characteristic X-rays for analytical purposes, it is required to know which specific
energy maps to which element. For this, we have to look at the electron energy levels of the atom in question,
and the possible transitions between them. The characteristic X-rays are identified with two letters and a
number according to their transitional states, with some common examples being Kα1 , Kα2 , Kβ1 and Lα1 .
This naming scheme is called the Siegbahn notation and can be identified as follows,
• The first letter indicates which shell the emitted electron was knocked out from, i.e. K, L, M, N.
• The Greek letter indicates the shell that the vacancy-filling electron originated from. For example, for
Kα an electron from the L-shell is lowered into the K-shell, while for Kβ an electron from the M-shell
is lowered into the K-shell.
• The last number denotes the specific subshell the vacancy-filling electron originated from. For example,
for Kα1 , this is the L3 (2p3/2 )-subshell, while for Kα2 this is the L2 (2p1/2 )-subshell.
At first glance, it might seem that there are many possibilities of transitions for a specific atom, which would
mean that a wide range of measured energies could characterize this atom. However, there are usually only
a handful of characteristic energies with a high enough probability to occur to be viable candidates for EDS
characterization [28]. For example, Kα1 and Kα2 are emitted in a ratio of 2:1. This difference arises from
the fact that there are 4 electrons in the L3 -subshell, whilst there are only 2 electrons in the L2 -subshell.
The Kα1 and Kα2 X-rays are generally only a few eV apart for lower Z atoms, which is far too low to be
distinguished when measured, as a result of the fundamental statistical noise of the charge carrier creation,
as will be explained in detail in section 2.3.1. Therefore, this pair is called a doublet and is usually simply
referred to as the Kα line. Furthermore, the prevalence of vacancy-filling electrons from higher shells quickly
6
decreases. For example, the Kα line is seven to eight times more intense than the Kβ line [29]. Some of the
most common transitions are shown in figure 2.4.
Unfortunately, it is often not possible to only consider the K series of lines, due to the fact that the
energy required to knock electrons out of the K-shell quickly increases as the atomic number increases. For
example, for gold (Z = 79) the energy required is roughly 69 keV , which is much higher than the usual EDS
electron beam operating ranges of 10 to 30 keV. Thus, for elements with higher Z other characteristic lines
than the K series must be considered. Usually, the L series of lines are considered for heavy elements and
the M series for even heavier elements.
Figure 2.4: The most common electron transitions that result in the emission of K and L characteristic
X-rays.
2.1.4 Bremsstrahlung
The second process that is responsible for X-ray emission in EDS is Bremsstrahlung. The term Bremsstrahlung
comes from the German words “bremsen” and “strahlung”, which mean “to brake” and “radiation”, respec-
tively.
Bremsstrahlung is emitted whenever an incident electron penetrates the outer electron shells and interacts
inelastically with the nucleus of an atom of the target sample. The incident electron experiences a Coulomb
force due to the presence of the nucleus and as a consequence, loses some of its kinetic energy. In order to
adhere to the conservation of energy principle, an X-ray is emitted with an energy equal to the loss of kinetic
energy of the electron. The energy of the X-ray depends strongly on the interaction distance. Electrons that
interact farther away from the nucleus emit lower-energy X-rays compared to those interacting closer to the
nucleus, [30]. A schematic showing the process bremsstrahlung is shown in figure 2.5a.
Electrons can lose varying amounts of energy depending on the strength of their interaction with the
nucleus. Therefore, the energy of bremsstrahlung X-rays spans a wide range of energies and is not dependent
on the specific sample material. This is why it is often considered background radiation, for the purposes
of EDS. A typical spectrum showing the Bremsstrahlung background curve along with the peaks from the
characteristic X-rays is shown in figure 2.5b.
7
(a) (b)
Figure 2.5: (a) Bremsstrahlung produced by a high-energy electron with kinetic energy E1 . After interaction
with the nucleus, the electron is decelerated to a kinetic energy of E2 . The lost energy is converted into an
X-ray of energy E1 − E2 . (b) Schematic of a typical SEM-EDS spectrum containing characteristic X-ray
peaks superimposed on a Bremsstrahlung background.
2.1.5 X-ray Interactions With Matter

The X-rays generated as a result of the bombardment of electrons interacting with the sample can in turn
interact with the objects surrounding it. This could be for instance the sample itself, the detector, or the
protective housing of the setup. When these interactions happen within the sensitive volume of an X-ray
detector, the emitted electrons and holes generated from this can be measured to obtain the energy of the
X-ray. This process will be discussed further in section 2.3.
The are four main processes through which a photon can interact with matter. These include photoelectric
absorption, Compton scattering, Rayleigh scattering, and pair production.
The first two processes will be treated in detail in the following sections, as these are the most important
in EDS [31]. The latter two processes will be explained briefly.
In brief, in Rayleigh scattering, the photon will interact elastically with the atom. The excited atom
immediately radiates this energy as an emitting photon of the same energy but in a different direction. Since
EDS relies on the detection of X-rays with specific energies corresponding to the characteristic X-rays of
elements in the sample, the contributions from Rayleigh scattering are not relevant for elemental analysis.
Pair production refers to the creation of an electron and positron as a result of the interaction of a nucleus
and a photon of at least 1 MeV. As these energies are orders of magnitude higher than those in use during
EDS, this process is not relevant in this case. It must be noted that the pair production in this setting must
not be confused with the electron-hole pair creation that will be discussed in later chapters.
The cross sections for photoelectric absorption τ , Compton scattering σ and pair production κ are related
to the atomic number of the absorbing medium Z and the photon energy E as follows [32],
τ ∝ Z 5 /E 7/2
σ ∝ Z/E (2.1)
2
κ ∝ Z ln(2E)
In the case of a narrow mono-energetic photon beam, the decrease in intensity as a result of absorption
of the photons occurs exponentially, this is known as the Beer-Lambert law,
I = I0 e−µx (2.2)
Where I0 denotes the initial intensity, and I the intensity after passing through distance x of the absorbing
8
medium. Finally, µ denotes the linear attenuation coefficient, which is related to the total cross section by,
ρ
µ= σtot (2.3)
uA
Where u the atomic mass unit, ρ is the density and A the relative atomic mass of the absorbing medium [33].
The total cross-section σtot can be written as the sum of the individual photon interaction cross-sections, as
shown in equation 2.1,
σtot = τ + σ + κ (2.4)
In figure 2.6, the interplay between photoelectric absorption, Compton scattering, and pair production is
shown for different absorbing media with atomic number Z. It can be seen that for the energies used in EDS,
which range from a few hundred to 30 keV, the interactions are dominated by photoelectric absorption.
This is especially true when silicon is used as the detecting medium, due to its low atomic number. Figure
2.7 provides the mass attenuation coefficients for silicon across a wide range of energies. It can be seen that
in the lower energy regime, the prevalence of photoelectric absorption is significantly higher than the other
processes by several orders of magnitude.
Figure 2.6: The regions of dominant photon interactions for absorbing media of different atomic numbers
Z. The boundary where the photoelectric effect and the Compton have the same cross-section is given by
σ = τ . Similarly, the boundary for the Compton effect and pair production is given by σ = κ. The red
dashed line shows the cross sections for silicon. Adapted from [34].
Figure 2.7: Photon mass attenuation coefficients for silicon. The sudden increase in photoelectric absorption
is due to the binding energy of the silicon K shell, which lies at 1.8 keV. Data extracted from the NIST
XCOM database [35].
9
Photoelectric Absorption
Photoelectric absorption or the photoelectric effect can occur when, as a result of the absorption of an
incident photon, an electron from a tightly bound inner shell is ejected. The emitted electron is referred
to as the photoelectron. As the photon is fully absorbed during this process, the kinetic energy of the
photoelectron is given by Kpe ,
Kpe = hν − EB (2.5)
where h is Planck’s constant, ν the frequency of the incident photon and EB the binding energy of the
photoelectron. The atom is left in an excited state and can again relax back to the ground state by emitting
characteristic X-rays or Auger electrons, in a similar process as discussed in section 2.1.3. A schematic of
the photoelectric effect is shown in figure 2.8. The resultant photoelectron, characteristic X-ray or Auger
electron can, in turn, interact within the absorbing medium, generating new X-rays and electrons in a
cascading manner.
Figure 2.8: Schematic of the photoelectric effect which results in the emission of a Kα1 photon.
Compton scattering
Compton scattering or incoherent scattering results when the incident photon elastically scatters off an
electron and hereby transfers a fraction of its energy to a recoil electron. The change of direction of the
photon is related to the energy transferred to the electron. The energy of the photon after scattering can be
found by using conservation of energy (eq. 2.6) and momentum (eq. 2.7),
Eγ + Ee = Eγ′ + Ee′ (2.6)

Pγ = P′γ + P′e (2.7)
Where the subscripts γ and e denote the photon and electron, respectively. Solving these equations leads to
the Compton equation,
Eγ
Eγ′ = Eγ
(2.8)
1 + m0 c2 (1 − cos θ)
Where θ is the scattering angle of the photon. m0 is the rest mass of the electron and c is the speed of
light. The maximum energy loss occurs when the photon scatters in the opposite direction to the incident
direction. Figure 2.9 shows a schematic of Compton scattering.
10
2.2. SEMICONDUCTOR BASICS CHAPTER 2. THEORY
Figure 2.9: Schematic of Compton scattering. An incident photon collides with an electron and loses some
energy to the recoiling electron.
2.2 Semiconductor Basics

One of the most commonly used materials as a detecting medium for soft X-rays in solid-state detectors is
silicon. To understand why, some important material properties of semiconductors will be covered. Further-
more, it will be covered how the electrical properties of silicon can be altered by means of doping to facilitate
its use in silicon detectors.
2.2.1 Band Theory of Solids

Solid-state materials can be grouped into one of three classes: insulators, semiconductors, and conductors.
As the names suggest, these materials vary significantly in the level of electric conductivity. These differences
can be explained by examining their electronic band structures.
For isolated atoms, the electrons occupy discrete energy levels. However, when multiple atoms are brought
together to form a solid material, the outer electrons start to overlap and interact with each other. This
interaction involves attractive and repulsive forces between the atoms, resulting in a change in energy levels.
The energy levels will become very closely spaced and a continuous energy band is formed. Consequently,
the electrons can no longer be considered solely belonging to their individual parent atoms but rather to the
crystal as a whole.
In the case of a silicon atom, which has a total of 14 electrons, 10 of these electrons lie in deep energy levels
which have an orbital radius significantly smaller than the interatomic separation within the crystal lattice.
These deeply bound electrons are relatively localized and do not participate prominently in interatomic
interactions [36]. The remaining four electrons lie in the 3s and 3p subshells and are known as the valence
electrons. These electrons can participate in chemical bond formation.
When silicon atoms are brought together in a lattice, the allowed electron energy bands will split into
two, the valence band and the conduction band. The top of the valence band lies at an energy EV , while the
bottom of the conduction band lies at the energy EC . The forbidden energy gap between the two bands is
known as the band gap and is defined to be the difference of both energy bands, EG = EC − EV .
At absolute zero temperature, the electrons will occupy the lowest available energy states, completely
filling the valence band up to the Fermi energy. When the temperature is increased, electrons from the
valence band are able to jump to the conduction band, where they are free to move around inside the
material. For a system of identical fermions in thermodynamic equilibrium, the Fermi-Dirac distribution can
be used to find the occupation probability of a certain state with energy E,
1
f (E) = (2.9)
e(E−EF )/kB T +1
Where EF signifies the Fermi energy, kB the Boltzmann constant and T the temperature. At absolute zero,
the electrons will fill up all the available states up to the Fermi energy. For an electron with energy EF , we
find the probability of that state being occupied to be 1/2.
Insulators are characterized by a very large band gap. The valence electrons form strong bonds between
neighboring atoms and as a result, these are not easily brought into the conduction band by external stimuli
11
such as an electric field or thermal excitations. This causes these materials to have a very low electrical
conductivity.
On the other hand, conductors are characterized by the fact that they have a zero band gap. The Fermi
level lies inside one of the bands, allowing electrons in the valence band to transition to the conduction band
without needing to overcome an energy gap. As a consequence, a significant number of electrons reside in
the conduction band, causing conductors to have a very high conductivity.
Lastly, semiconductors have a very small band gap, in the order of 1 eV. EF lies in the middle of
the energy gap, separating occupied states in the valence band from unoccupied states in the conduction
band. At zero temperature, the valence band is completely occupied and the conduction band is empty,
causing semiconductors to have a low conductivity at low temperatures. However, the band gap is easily
crossed as a result of thermal excitations. Silicon belongs to the semiconductor category, with a band gap
of EG = 1.12 eV .
Figure 2.10: A band gap diagram showing the different band gap sizes EG for conductors, semiconductors
and insulators. The Fermi level is given by EF .
2.2.2 Intrinsic Semiconductors

Understanding the properties of intrinsic semiconductors and the behavior of charge carriers is crucial for
its use in semiconductor detectors. As SDDs need to detect a very small number of charge carriers, it is
necessary to eliminate the charge carriers within the detector material. In order to do this, it must first be
understood how many charge carriers are present within the semiconductor material. The following section
will present a derivation of how the concentration of free charge carriers within the semiconductor can be
obtained.
We can use the dispersion relation in the free electron model [37] to obtain a relation between the energy
of an electron E and its wave vector k,
h̄2 k 2
r
2mef f E
E= → k= (2.10)
2mef f h̄2
Where h̄ is Planck’s constant, and mef f the effective mass of the electron. At low temperatures, states below
the wave vector k are filled, while states above k are empty. Taking the derivative of k with respect to the
energy, we find,
−1/2
dk mef f 2mef f E
= (2.11)
dE h̄2 h̄2
Furthermore, the number of states N (k) lying in a shell in k-space is,
1
N (k) = 4πk 2 dk (2.12)
(2π)3
1 2
Here, (2π)3 is the number of allowed k-values per unit volume of the k-space. Furthermore, 4πk dk denotes
the spherical shell, where k is the radius and dk is the thickness of the shell [38].
12
We can rewrite N (k) to be in terms of energy, by substituting equations 2.10 and 2.11 in equation 2.12,
and we find,
3/2
1 2mef f
N (E) = E 1/2 dE (2.13)
4π 2 h̄2
Lastly, we have to account for the fact any k state can contain two different electrons of different spins,
spin-up, and spin-down. The density of state g(E) can now be obtained,
3/2
1 2mef f
g(E) = E 1/2 (2.14)
2π 2 h̄2
We are interested in finding the electron concentration in the conduction band and the hole concentration
in the valence band. As typically the Fermi energy EF is positioned near the middle of the conduction and
valence band, we can approximate the Fermi-Dirac distribution in equation 2.9 for electrons and holes
respectively as follows,
E−EF
−
fe (E) ≈ e kB T
(2.15)
EF −E
−
fh (E) = 1 − fe (E) ≈ e kB T
(2.16)
Now, the density of free electrons within the conduction band ne can be found by integrating the product
of the Fermi-Dirac distribution and the density of states over the conduction band region,
Z ∞
ne = fe (E)g(E)dE (2.17)
EC
Z ∞ 3/2
1 − 2me
E−EF
= e kB T
E 1/2 dE (2.18)
EC 2π 2 h̄2
3/2 E −E
2πme kB T − C F
=2 e kB T (2.19)
h2
EC −EF
−
= NC e kB T
(2.20)
Similarly, the density of holes in the valence band nh can be found by integrating over the valence band
region,
Z EV
nh = fh (E)g(E)dE (2.21)
−∞
Z EV 3/2
1− 2mh
EF −E
= e 2
kB T
2 E 1/2 dE (2.22)
−∞ 2π h̄
3/2 E −E
2πmh kB T − F V
=2 2
e kB T (2.23)
h
EF −EV
−
= NV e kB T
(2.24)
Here, we have used NC and NV to be the effective densities of states for the conduction band and the valence
band respectively.
In intrinsic semiconductors, the density of electrons in the conduction band and holes in the valence should
be equivalent, as these are created in pairs, ni = ne = nh . As a result of this, the intrinsic charge-carrier
density ni can be written as,
√ p −
EC −EV p E
− 2k GT
ni = ne nh = NC NV e 2kB T
= NC NV e B (2.25)
13
Where EG denotes the band gap energy. In addition, an expression of the Fermi energy EF,i in intrinsic
semiconductors can be now formulated,
ne = nh (2.26)
3/2 E −E
3/2 E −E
2πme kB T − Ck TF,i 2πmh kB T − F,i V
2 e B =2 e kB T
(2.27)
h2 h2
EC + EV 3kB T mh
⇒ EF,i = + ln (2.28)
2 4 me
We find that in intrinsic semiconductors, the Fermi Energy lies approximately in the middle of the top of the
valence and the bottom of the conduction band, with a shift based on the temperature and the difference
in effective masses of the electrons and holes. We will see in the next section that this will not be the case
when we introduce impurities to the semiconductor.
Lastly, we can express the density of electrons ne and holes nh in equations 2.20 and 2.24, respectively,
in terms of the intrinsic charge-carrier density, which will prove useful when we consider extrinsic semicon-
ductors. EF −EF,i
ne = ni e kB T (2.29)
and,
EF,i −EF
nh = ni e kB T
(2.30)
2.2.3 Extrinsic Semiconductors

In the previous section, intrinsic semiconductors were considered. These types of semiconductors are pure,
meaning they are free of any impurities. Contrary to this, there exist extrinsic semiconductors, which do
contain relatively small amounts of impurities. The process of deliberately manufacturing a semiconductor
with trace elements or chemicals is called doping.
By doping a semiconductor, the electrical properties of the material can be altered, making it suitable for
electronic applications, such as detectors. The dopant atoms introduced to the semiconductor can replace
a silicon atom in the lattice. Dopants will have the effect of either donating an additional electron to the
conduction band, or creating a hole in the valence band. This hole can then accept an electron from the
conduction band. As a result, these types of dopants are called donors and acceptors, respectively. In the
case of a silicon semiconductor material, dopants from the III and V groups are usually chosen as acceptor
and donor dopants, for example, boron (B) and phosphorus (P). These dopant atoms will contain either 3
or 5 electrons in their valence shell. The doping with donors is called n-type doping and the doping with
acceptors is called p-type Doping.
n-type doping
In n-type doping, a small number of impurity atoms, usually from group V of the periodic table, are added
to the silicon crystal. These impurities have one extra electron in their outermost shell. The most commonly
used n-type dopants are phosphorus (P), arsenic (As), and antimony (Sb).
By introducing n-type dopant atoms, a donor level just below the conduction band is created. These
donor levels act as sources of free electrons that can move through the lattice when an electric field is applied.
As a result, the concentration of dopant atoms in the silicon lattice determines the number of free electrons
that are available for conduction.
The concentration of dopant atoms is expressed in terms of doping concentration, which is the number
of dopant atoms per unit volume. The Fermi energy in the case of n-type doped material can be found up
to a high concentration by equation 2.20, where we set the electron concentration in the conduction band to
be equal to the donor concentration ND ,
−
EC −EF NC
ND = NC e kB T
→ EF = EC − kB T ln (2.31)
ND
14
From this, we can see that the Fermi energy lies slightly below the bottom of the conduction band. As
a result, these electrons are easily excited into the conduction band, where they can act as a charge carrier.
Figure 2.11 shows the lattice and band structure in the case of an n-type doped silicon lattice.
Figure 2.11: Schematic of an n-type silicon lattice doped with phosphorus (left) and its band structure
(right). The phosphorus atom contains an additional fifth electron in its valence shell. These electrons lie
just below the conduction band and are easily excited into the conduction band, where they can act as a
free charge carrier.
p-type Doping
Unlike n-type doping, p-type doping introduces positively charged carriers (holes) into the silicon lattice. In
p-type doping, a small number of impurity atoms, usually from group III of the periodic table, are added
to the semiconductor crystal. These impurities contain only three electrons in their outermost shell, making
them more likely to accept an electron from the silicon lattice. The most commonly used p-type dopants are
boron (B), aluminum (Al), and gallium (Ga).
The dopant atom introduces a hole into the lattice, creating an acceptor level just above the valence
band. These acceptor levels act as sources of holes that can move through the lattice when an electric field
is applied. The concentration of dopant atoms in the lattice determines the number of holes available for
conduction.
In a similar way to n-type doping, we can find the Fermi energy for a p-type material up to a high
concentration by equation 2.24, where we set the hole concentration in the valence band to be equal to the
acceptor concentration NA ,
−
EF −EV NV
NA = NV e kB T
→ EF = EV + kB T ln (2.32)
NA
The Fermi energy lies slightly above the top of the valence band. As a result, holes can easily drop
down into the valence band, where they can act as charge carriers. Figure 2.12 shows the lattice and band
structure in the case of a p-type doped silicon lattice.
15
Figure 2.12: Schematic of a p-type silicon lattice doped with boron (left) and its band structure (right). The
boron atom contains only three electrons in its valence shell, producing a hole. These holes lie just above
the valence band. A nearby valence electron can easily move into the vacant space, leaving behind a new
hole in turn. This allows the holes to act as free charge carriers in the valence band.
2.2.4 pn-Junction
When a p-type and n-type material are brought into contact with each other, the electrons from the n-type
material diffuse across the junction and into the p-type material, where they recombine with the holes.
Similarly, the holes in the p-type material diffuse across the junction and into the n-type material, where
they recombine with the electrons. This diffusion of majority carriers across the junction creates a region
near the junction where there are no mobile charge carriers left. This region is called the depletion region,
and it has a fixed width that depends on the doping concentrations of the p-type and n-type materials.
Built-in Voltage
Within isolation, the n-type and p-type material have different Fermi levels, as shown in equations 2.31 and
2.32. When there exists a thermal equilibrium, however, the Fermi energy must remain constant throughout
both regions. As a result, the energy bands are shifted to keep the Fermi energy constant throughout the
junction, in a phenomenon called band bending, as shown in figure 2.13.
Figure 2.13: The energy band diagram of a pn-junction under thermal equilibrium. The donator electrons
(in black) sit just below the conduction band, whilst the acceptor holes (in white) sit just above the valence
band. Under equilibrium, the drift forces due to the electric field and the diffusion forces will cancel out and
a depleted region will arise.
The built-in voltage at a pn-junction is a result of the diffusion of charge carriers across the junction due
to differences in the doping concentration on either side of the junction. The p-type region has an excess of
holes, while the n-type region has an excess of electrons. When the two regions are brought into contact,
16
the excess carriers diffuse across the junction. In so doing they leave behind positive ions at the donor
impurity sites, and negative ions at the acceptor impurity sites, which causes a net charge difference across
the junction. This exposed space charge creates an electric field that opposes further diffusion of carriers,
resulting in an equilibrium state known as the built-in potential Vbi . This built-in voltage is typically around
0.5 volts for a silicon pn-junction [4].
The magnitude of the built-in voltage depends on the doping concentrations of the two regions and is a
critical parameter in determining the behavior of the p-n junction. When an external voltage is applied to
the junction, it can change the magnitude and direction of the built-in voltage, leading to changes in the
junction behavior. For example, if the external voltage is forward-biased, it reduces the built-in voltage,
allowing the current to flow through the junction. Conversely, if the external voltage is reverse-biased, it
increases the built-in voltage, which acts as a barrier to current flow. This behaviour will be further explored
in section 2.2.5.
The built-in voltage also affects the depletion region, which is a region near the junction where the free
charge carriers are absent. The width of the depletion region is proportional to the built-in voltage and is a
critical parameter in determining the sensitivity of semiconductor detectors.
The built-in voltage is equal to the difference between the Fermi energy of the n- and p-type regions,
1
Vbi = (EF,p − EF,n ) (2.33)
e
Here we can use the result found for the density of electrons and holes as shown in equations 2.29 and 2.30,
EF −EF,n
n e = ND = n i e kB T
(2.34)
EF,p −EF
nh = NA = ni e kB T
(2.35)
Combining the two above, we obtain,

EF,p −EF,n
ND NA = n2i e kB T
(2.36)

ND NA
⇒ EF,n − EF,p = kB T ln (2.37)
n2i
Hence, by combining equation 2.33 and 2.37, we find that the built-in voltage is,

kB T ND NA
Vbi = ln (2.38)
e n2i
The magnitude of the built-in voltage depends on the doping concentrations and the temperature of the pn-
junction. If we consider a Si pn-junction with NA = 1017 cm−3 and ND = 1016 cm−3 at room temperature
at which the intrinsic carrier concentration is ni = 1010 cm−3 , we find a built-in voltage of Vbi ≈ 0.7V .
Depletion Depth
Depletion depth is an important parameter in semiconductor detectors that describes the region in which
free charge carriers are absent. The main idea of semiconductor detector devices relies on expanding the
depleted region as far as possible, such that the entire sensitive region is absent of any charge carriers. In
this manner, the generated electron-hole pairs due to the absorption of an X-ray can be distinguished, as
there reside no free carriers belonging to the silicon material.
In order to find the length of the depleted region, we must first find an expression for the electric field.
The electric field can be found using Gauss’ law, where we consider the integral form,
I Z
E ⃗ = 1
⃗ · dA ρdV (2.39)
ϵSi ϵ0 V
We will consider the following geometry, shown in figure 2.14 with a small highly doped p+ region on the
left, and a large n-type region, or bulk on the right.
17
Figure 2.14: Schematic of a pn-junction, containing a relatively thin highly doped p+ -type region on the left
and a large n-type bulk on the right side.
We will consider an integral over the depth of the pn-junction in the z-direction. The charge density as
a result of the ionized dopant atoms in this region is,
ρ = −eNA for z≤0 (2.40)

ρ = eND for z>0 (2.41)
Note that at the n-side the dopant atom donates an electron, leaving the exposed space charge positively
charged. Inversely, at the p-side the dopant atom accepts an electron, leaving it negatively charged. We will
consider an elongated cubic structure with a total cross-sectional area of A = X · Y , and dV = Adz. We can
thus rewrite equation 2.39 to,
Z
1
E(z) · A = −eNA Adz for z ≤ 0 (2.42)
ϵSi ϵ0
Z
1
E(z) · A = eND Adz for z > 0 (2.43)
ϵSi ϵ0
We can solve the integral expression by using the fact that the electric field has to be 0 outside of the
depletion region, as any field here would cause the free carriers to move and thereby eliminating the electric
field. We thus enforce E(−dp ) = E(dn ) = 0, where dp and dn are the thickness of the depletion region on
the p-side and n-side respectively. We thus find the following expression for the electric field,
eNA
E(z) = − (z + dp ) for z ≤ 0 (2.44)
ϵSi ϵ0
eND
E(z) = (z − dn ) for z>0 (2.45)
ϵSi ϵ0
We find the maximum of the electric field at the intersection between the p-type and n-type at z = 0,
eNA dp eND dn
Emax = − =− → NA dp = ND dn (2.46)
ϵSi ϵ0 ϵSi ϵ0
This expresses the fact that the total positive charge in the n-type depletion region balances the total negative
charge in the p-type depletion region. Now, the potential along the junction can be found by integrating,
Z
eNA eNA
Vp (z) = (z + dp )dz = (z + dp )2 for z ≤ 0 (2.47)
ϵSi ϵ0 2ϵSi ϵ0
Z
eND eND z eND 2
Vn (z) = (dn − z)dz = (dn − )z + d for z > 0 (2.48)
ϵSi ϵ0 ϵSi ϵ0 2 2ϵSi ϵ0 p
Here, we used the fact that at the boundary the potential must be identical, Vp (0) = Vn (0). Furthermore,
since we are interested in the potential difference across the junction, we can arbitrarily set it to zero at one
18
side. Here, we defined the potential at the edge of the depletion region on the p-side as zero, Vp (−dp ) = 0.
Figure 2.15 shows the behavior of the charge, electric field strength, and potential,
Figure 2.15: A schematic of an unbiased pn-junction where NA > ND . The exposed space charge accumulates
at the junction, causing a sharp charge distribution Q(z). As a consequence, an internal electric field E(z)
pointing from the N towards the P side is formed. The electrostatic potential ϕ(z) varies sharply across the
depleted region and the potential difference forms the built-in voltage Vbi of the junction.
We find that the maximum of the potential lies at the edge of the depletion region on the n-side, namely,
e
Vmax = Vn (dn ) = (ND d2n + NA d2p ) = Vbi (2.49)
2ϵSi ϵ0
This potential is equal to the built-in voltage we found in equation 2.38. We can use this result, and the
result from equation 2.46, to find expressions for the depth of the depletion region on either side,
s
2ϵSi ϵ0 Vbi NA
dn = (2.50)
e ND (NA + ND )
s
2ϵSi ϵ0 Vbi ND
dp = (2.51)
e NA (NA + ND )
The width of the total depletion region is equal to the sum of the depletion width on either side, hence
we find the total depletion width of the pn-junction,
s
2ϵSi ϵ0 Vbi 1 1
d = dn + dp = + (2.52)
e NA ND
Often, in detectors, we have a highly doped p+ region and a lowly doped n-region. In this case, we can
assume NA ≫ ND , and we can approximate the total depletion depth as,
r
2ϵSi ϵ0 Vbi
d≈ (2.53)
e ND
19
As an example, we will consider a Si pn-junction with a strongly doped p+ region, NA = 1018 cm−3 and a
lightly doped n ND = 1012 cm−3 at room temperature at which the intrinsic carrier concentration. From
equation 2.38, we find a built-in voltage of Vbi = 0.58 V . Thus we find a total depleted region of d = 27.4 µm.
Note that almost the entirety of the total depleted region lies in the lowly doped n-type region, in fact, the
depleted region in the p+ region extends only dp = 2.74e-5 µm.
From this example, we can see that the depleted region is very small, which means that the total active
detector volume would be very small. Of course, we would like the active volume to be as large as possible
to obtain the highest quantum efficiency. Therefore, we would like to expand the total depleted region. This
can be done by introducing an external bias on the junction, which will be covered in the next section.
2.2.5 Biased pn-Junctions

Applying external bias to pn-junctions is a crucial step in the creation of semiconductor detectors. By
applying an external voltage to a pn-junction, the magnitude and direction of the built-in voltage can be
controlled, leading to changes in the junction behavior.
Up to this point, we have assumed that the pn-junction meets a thermal equilibrium, which causes the
Fermi energy through the entire junction to be constant. When we start to introduce an external potential,
this thermal equilibrium does not hold anymore. Therefore the equations used hold only approximately in
the case of external biasing.
The depletion depth of a pn-junction can be found using,
r
2ϵSi ϵ0
d≈ (Vbi − Vbias ) (2.54)
eND
Forward Biasing
In the case of a forward biased junction, a positive external bias voltage is applied, Vbias > 0. This biasing
setup can be seen in figure 2.16,
Figure 2.16: A schematic of a forward biased pn-junction. By connecting the positive terminal to the p-type
side and the negative terminal to the n-type side and applying a voltage, the depletion region shrinks.
When a forward bias is applied to a pn-junction, the potential difference across the junction is reduced,
and as a result, the depletion region will shrink. When the external potential overcomes the built-in voltage,
the depletion region will disappear and the carriers will be able to freely move and conduct a current. This
effectively makes the forward-biased pn-junction a voltage-controlled electrical switch. For this reason, they
are a critical component in electronic devices such as diodes, transistors, and solar cells.
Reverse Biasing
In the case of a forward biased junction, a negative external bias voltage is applied, Vbias < 0. This biasing
setup can be seen in figure 2.17,
20
2.3. SEMICONDUCTOR DETECTOR PRINCIPLES CHAPTER 2. THEORY
Figure 2.17: A schematic of a reverse biased pn-junction. By connecting the positive terminal to the n-type
side and the negative terminal to the p-type side and applying a voltage, the depletion region expands.
When a reverse bias is applied to a pn-junction, the potential difference across the junction is increased,
and as a result the depleted region increases. In a semiconductor detector, we would like the active volume
to be fully depleted. This means that the depletion region spans the entire sensor bulk. The reverse bias
voltage required to fully deplete the volume is known as the depletion voltage. This voltage is usually much
higher than the built-in voltage, therefore Vbi is often neglected, and the depletion voltage in a p+ n-junction
can be found from equation 2.53, such that,
eND d2
Vdep ≈ (2.55)
2ϵSi ϵ0
For example, to fully deplete a silicon sensor volume with a thickness of 500 µm and doping concentration
ND = 1012 cm−3 , we need an external voltage of Vdep = −193 V .
2.3 Semiconductor Detector Principles

The X-ray detection process in silicon semiconductor detectors consists of several stages, which can be
summarized into four main steps:
• Charge creation: When X-rays interact with the detector bulk, they ionize the silicon atoms, generating
electron-hole pairs in the process. The energy of the incident X-ray is proportional to the number of
electron-hole pairs generated.
• Charge propagation: The cloud of electron-hole pairs created in the first stage then move through the
silicon bulk under the influence of the electric field in the depletion region of the pn-junction. The
electrons drift towards the anode, whilst the holes drift towards the cathodes.
• Signal formation: As the charge carriers approach the electrodes, they induce a current on them.
The total charge induced on the anode over the time of propagation is proportional to the number of
electron-hole pairs.
• Digitization: The final stage involves converting the induced charge to a voltage signal by a FET
preamplifier, followed by pulse processing and dispersion of the measured energies into a histogram, or
spectrum.
Each of these stages will be treated in detail in the following sections.
2.3.1 Charge Creation

Whenever an X-ray enters the detector volume, it can interact with the silicon atoms of the bulk in the
sensor. As a result of this interaction, electron-hole pairs will be generated.
21
It was shown in section 2.1.5 that the X-ray interactions in EDS fall almost exclusively within the
photoelectric absorption regime. Whenever a photon is absorbed through photoelectric absorption, an
energetic photoelectron is created with an energy equal to the absorbed X-ray minus the binding energy of the
electron. The resulting photoelectron and Auger electron, if emitted, have much shorter attenuation lengths
in the silicon bulk compared to X-rays, namely usually in the order of tens of nanometers, as determined
experimentally by Solokha et al. [39]. When these electrons are sufficiently energetic, they quickly interact
again in a cascading manner, creating further secondary electron-hole pairs along the way. This process
continues until the secondary particles do not have enough energy to generate new electron-hole pairs.
After the cascading interactions are finished occurred, a cloud of electrons is excited into the conduction
band, and an equal number of holes is introduced in the valence band. The total number of charge carriers
created is dependent on the number of interactions that occur and thus is dependent on the incident photon
energy.
Charge Creation Energy

Once the cascading interactions are completed, it is necessary to determine the number of generated electrons
and holes, denoted as N . Naively, it could be expected that this is simply related to the band gap of the
absorber and the incident photon energy E,
E
N= (2.56)
EG
Where EG is the band gap of silicon, which is around EG = 1.12 eV . However, experimentally, much
fewer electron-hole pairs are created than this. An average charge creation energy ω is observed, which is 2
to 3 times larger than the band gap, such that,
E
N= (2.57)
ω
The culprit is the fact that not all absorbed energy goes into producing electron-hole pairs. A significant
portion of the absorbed energy goes into the creation of phonons, or lattice vibrations. A good approximation
is shown in [40], relating the average charge creation energy to the band gap of the absorber,
ω ≈ 2.8EG + 0.6 eV (2.58)
For silicon, this results in an average charge creation energy of ωSi ≈ 3.7 eV . This means that roughly 70%
of the ionization energy goes into phonon excitations, whereas only 30% percent goes into the creation of
electron-hole pairs [41]. Using the average charge creation energy ω and the energy of the incident photon
E, the total number of generated electron-hole pairs can be calculated using equation 2.57.
Fano Factor
Imperfect energy resolution in a detector can arise from several factors such as drift in the detector’s operating
characteristics, random noise from the detector and instrumentation system, and statistical noise from the
discrete nature of the measured signal. The statistical noise represents a minimum amount of fluctuation
that will always be present in the detector signal and therefore sets an upper limit on detector performance.
The statistical noise arises from the fact that the number of electron-hole pairs generated within the
detector as a result of the absorption of an X-ray is not a continuous variable. Instead, it is represented by
a discrete number of electron-hole pairs. It was found that the total number of created electron-hole pairs
fluctuates, even when the energy of the absorbed X-ray is kept constant.
Fano [42] first introduced the degree of deviation from Poisson statistics in the context of ionizing radiation
in a gas. It quantifies the discrepancy between the observed variance and the expected variance based on
Poisson statistics.
According to Poisson statistics, in a system where events occur randomly and independently at an average
rate, the variance should be equal to the mean, σ 2 = µ. However, in the case of generating electron-hole
pairs in a semiconductor detector, the experimentally observed variance differs from the Poisson expectation.
22
The Fano factor, denoted as F, is defined as the ratio of the observed variance to the expected variance based
on Poisson statistics,
σ2
F = (2.59)
µ
In the case of charge carrier creation in semiconductors, this becomes,
2
σN
F = (2.60)
N̄
2
where σN is variance and N̄ the mean of the number of charges created. A Fano factor of F = 1 indicates
that the system follows Poisson statistics, as the observed variance matches the expected variance. In silicon
semiconductors, the Fano factor is much smaller, usually measured to be in the order of F ≈ 0.1, which is
known as underdispersion.
Underdispersion arises from the involvement of multiple excitation mechanisms which contribute to the
reduction of statistical spread. In addition to the creation of electron-hole pairs, lattice vibrations are also
generated, which have much smaller excitation energies in the order of meV. As a result, a greater number
of excitations are involved than what is evident from the created charges alone, leading to a decrease in
statistical fluctuations [43].
By rewriting equation 2.60, the upper limit of resolution in detectors in terms of Full Width at Half
Maximum (FWHM) can be found,
√
F W HM ≈ 2.35 ωEF (2.61)
using the fact that F W HM ≈ 2.35σ.
Figure 2.18 shows the FWHM for silicon, where the average charge creation energy and Fano factor were
taken as ω = 3.64 eV and F = 0.115.
Figure 2.18: The intrinsic resolution of silicon as a result of the Fano noise in FWHM for different photon
energies. This is the highest resolution physically attainable.
We can see that the FWHM increases as we increase the incident photon energy, this means that lower-
energy photons can be detected with much higher resolution than higher-energy photons.
When discussing detector resolution, it typically refers to the resolution at the Mn-Kα line, approximately
at 5.9 keV. This particular energy is selected because Fe-55, which is commonly employed as a reference
material for testing detector resolution, undergoes electron capture decay to Mn-55, resulting in the emission
of manganese X-rays. From equation 2.61, it can be shown that the resolution at this energy is around 118 eV
FWHM.
23
Measured values for the average charge creation energy and the Fano factors for soft X-rays were tabulated
for different kinds of silicon detectors by G.W. Fraser et al. [44]. Here, values for the average charge creation
energy are reported to be between 3.6 and 3.8 eV, and values for the Fano factor are around 0.1. Furthermore,
experimental results show that the average charge creation depends both on the incident photon energy and
on the temperature of the detector, as shown by Mazziotta [45].
2.3.2 Charge Propagation

Once electron-hole pairs are created within the detector bulk, they move through the silicon by two processes,
drift and diffusion [30].
Diffusion
After the formation of the pair, both the electron and hole will start to diffuse as a result of random thermal
motion. This motion leads to a gradual expansion of the generated charge cloud over time. The cross-section
of the cloud can be approximated by a Gaussian with the following standard deviation [4],
√
σ = 2Dt (2.62)
Where D is the diffusion coefficient and t the elapsed time. The diffusion coefficient can be predicted by the
Einstein relation,
kB T
D=µ (2.63)
e
where µ is the charge carrier mobility, kB the Boltzmann constant, T the temperature and lastly e the
elementary charge.
Drift
The charges will be separated as a result of the electric field that is present due to the reverse biasing.
Besides the diffusion, the charge carriers will undergo net drift velocity in the direction of the electric field.
The electrons will have a net movement in the negative direction of the electric field. As holes move by
transferring electrons from nearby atoms in the lattice, these will experience a net movement in the positive
direction of the electric field.
As a result, the electrons will be pulled towards the anode, while the holes will be pulled towards the
cathode, following the electric drift lines in the detector bulk. The drift velocities of the electrons and holes
are defined respectively as [46],
ve = µe E (2.64)
vh = µh E (2.65)
where E is the electric field vector, and µ the mobility of the charge carrier.
Jacoboni-Canali Model
The most commonly used mobility model is the Jacoboni-Canali model [47, 48]. This model consists of
analytical expressions which are in good agreement with the experimental data of charge transport in silicon.
The mobility for electrons and holes are given by,
vm 1
µ(E) = (2.66)
Ec (1 + (E/Ec )β )1/β
where E is the magnitude of the local electric field, while vm , Ec , and β are phenomenological fitting
parameters, which best describe the temperature dependence of the charge carrier drift velocities in high-
purity silicon. The parameters are given below for the case of electrons and holes respectively,
24
vm,e = 1.53 · 109 · T −0.87 cm/s vm,h = 1.62 · 108 · T −0.52 cm/s
Ec,e = 1.01 · T 1.55 V/cm Ec,h = 1.24 · T 1.68 V/cm (2.67)
−2 0.66 0.17
βe = 2.57 · 10 ·T βh = 0.46 · T
where T is the temperature of the silicon bulk.
2.3.3 Signal Formation

Charge collection is not a simple case of counting the number of charge carriers which end at the anode. In
order to understand the principle of charge induction on the metal electrode, the Shockley-Ramo theorem can
be used, which was discovered by Ramo and Shockley independently from each other in the late 30s [24, 25].
We shall first treat a simple model containing a system of four electrodes. This system is called system
A, and is shown on the left in figure 2.19. We consider an infinitesimally small point-like electrode, labeled
by n = 0, which has a charge of QA 0 = q. It is surrounded by three grounded electrodes labeled with n from
1 through 3. As these three electrodes are grounded, we have V1A = V2A = V3A = 0. In this scenario, the
charge induced on electrode 1 due to the presence of the charged point-like electrode is to be determined.
An important relation that can be used to solve this problem is Green’s reciprocity theorem [49]. This
theorem relates two possible sets of charges Qn and potentials Vn of a system of conductors,
N
X N
X
QA B
n Vn = QB A
n Vn (2.68)
n=1 n=1
In order to understand situation A, we will consider a hypothetical situation B, where we keep the
geometry of our system the same as in situation A. However, in scenario B the potential at electrode 1 is
kept at Vw , whilst the remaining electrodes are kept grounded, V2B = V3B = 0. The charge of the point-like
electrode is also set to zero, QB
0 = 0. Using equation 2.68, we can find the charge induced on electrode 1,
V0B VB
qV0B + QA B
1 V1 = 0 → QA
1 = −q B
= −q 0 (2.69)
V1 Vw
Figure 2.19: Two scenarios containing three finite electrodes and a point-like electrode. Scenario A describes
the three electrodes as grounded, while the charge of the point-like electrode is set to q. Scenario B describes
two grounded electrodes and one electrode at Vw , where the charge of the point-like electrode is set to zero.
Since the point-like electrode is an infinitesimally small uncharged electrode in situation B, this situation
is equal to there being no electrode at all. Therefore, the potential V0B is equal to the potential at a point
⃗x in the potential field, resulting from electrode 1 being kept at Vw . We call this potential field ψ(⃗x), the
weighting potential.
25
Thus, we have now found a way to find the induced charge of a general electrode, as a result of its
weighting potential,
q
Q=− ψ(⃗x) (2.70)
Vw
However, we are interested in looking at the induced current as a result of a charge moving through this
weighting potential. The case of a charge moving along the trajectory ⃗x(t) is shown in figure 2.20.
Figure 2.20: A charge moving along trajectory ⃗x(t) from ⃗x0 to ⃗x1 . While traveling it induces a charge on
the nearby electrodes.
The instantaneous induced charge as a result of the presence of a point particle was shown in equation
2.70. Therefore, the time-dependent induced charge given by,
q
Q(t) = − ψ(⃗x(t)) (2.71)
Vw
The time-dependent induced current can be found by taking the derivative,
∂Q
I(t) = − (2.72)
∂t
q ∂ψ
= (⃗x(t)) (2.73)
Vw ∂t
q ∂⃗x(t)
= ∇ψ(⃗x(t)) · (2.74)
Vw ∂t
Suppose the charge moves along the trajectory ⃗x(t) from point ⃗x0 = ⃗x(t0 ) to point ⃗x1 = ⃗x(t1 ). The total
charge induced on the electrode accumulated over time can then be found by integration,
Z t1
Qind
tot = I(t)dt (2.75)
t0
Z t1
q
= ∇ψ(⃗x(t)) · ⃗x′ (t)dt (2.76)
Vw t0
q
= [ψ(x⃗1 ) − ψ(x⃗0 )] (2.77)
Vw
Where we used the gradient theorem to evaluate the integral expression.
We thus find that the induced charge is solely dependent on the endpoints of the trajectory, and not on
the specific path taken. This result is known as the Shockley-Ramo theorem, and this will be used in the
simulation to obtain the total charge induced by a charge carrier traveling through the silicon bulk.
26
2.3.4 Digitalization
Once the charge has accumulated on the anode, it is converted into a digital signal where each event is
binned according to the measured energy.
FET Preamplifier
The charge that is collected at the anode is first transformed into a voltage signal by a FET preamplifier.
During the measurement, the charge builds up on a feedback capacitor, originating from two different sources.
Firstly, as a result of small imperfections in the silicon bulk, a small continuous leakage current can flow
through the detector. This causes a steadily increasing voltage slope in the output of the FET. The second
source is the generation of electron-hole pairs which induce a current on the anode. These cause a sharp
step-like increase of the voltage for each X-ray absorption event. The capacitor is periodically discharged in
order to prevent saturation of the preamplifier. A schematic of the detector setup and an example of the
output of the FET is shown in figure 2.21.
Figure 2.21: (left) Schematic of an SDD connected to a FET pre-amplifier and a feedback capacitor. (right)
Typical output voltage ramp showing several photon absorption events. Each voltage step corresponds to
an absorbed photon and the step height determines the energy of the photon.
Pulse Procession
The height of the voltage jumps in 2.21 is directly related to the amount of charge that has accumulated on
the anode and is thus indicative of the energy of the X-ray absorption event. The aim of the pulse processor
is to remove the noise and to accurately measure the height of each voltage step. It does so by averaging
the voltage signal before and after the jump for a period of time called the shaper time. By utilizing a long
shaper time, the noise is averaged out and the highest resolution can be obtained. The drawback, however,
is a low throughput which results in longer acquisition times. Conversely, by utilizing a small shaper time,
the noise will be reduced less but a higher throughput is achievable. The resulting step measurement is then
sent to the multi-channel analyzer (MCA), where it is stored in a histogram according to the specific energy
measured.
2.3.5 Incomplete Charge Collection

If not all the electron-hole pairs produced by an incident X-ray are effectively propagated through the
detector, the measured charge signal by the FET will be lower than anticipated, leading to an underestimated
energy measurement of the incident X-ray, which leads to tailing on the lower-energy side of the peak in the
measured spectra. This phenomenon is referred to as incomplete charge collection (ICC).
Front Contact ICC

ICC at the entrance side is one of the leading causes of lower energy tailing, especially for lower energy
X-rays. This is because the electrical field in this region is poorly defined and the interface between the two
27
materials creates charge reflection phenomena [50]. This layer causes a fraction of electron-hole pairs created
by photon interactions to not be collected.
Scholze ICC Model

Front contact ICC is usually modeled by an exponential decrease of charge collection efficiency at the entrance
side [9,51]. Other models extend this by including a complete dead layer where the charge collection efficiency
is 0, followed by a similar exponential CCE model [17,18]. For this thesis, the model by Scholze and Procop [9]
was used to account for the front contact ICC. The model will briefly be discussed here. The detector is
modeled as a thick silicon bulk with a front contact. Directly beneath the front contact begins the sensitive
volume of the detector. The sensitive volume is split into two regions.
Firstly the ICC region, here any charge generated will not be completely collected. The ICC region spans
from the interface of the front contact (z = 0) up to z = z0 . In the case of sufficiently high X-ray energies,
the charge carriers are generated deeper than z0 and are completely collected. In the case of photons being
absorbed at z < z0 the charge will only be partly collected and as a result, a lower energy count will be
recorded by the MCA. This model is shown in figure 2.22.
Photoelectrons and Auger electrons which were emitted by an absorbed X-ray near the interface may
traverse the interface, leading to them not being propagated through the active volume. The measured
energy will then be the energy of the incident photon minus the kinetic energy of the electron which left the
active volume.
In the case of photon absorption within the front layer of the detector, the opposite process occurs. The
charges created in the front layer can pass the interface and end up in the active volume. The measured
energy will then be between zero and the kinetic energy of the electron. These processes contribute to the
generation of a spectral background that extends from zero to the full energy of the incident photon, known
as the shelf. These spectral characteristics can be seen in figure 2.24
Figure 2.22: Front contact model by Scholze et al. [9]. Auger and photoelectrons can cross the silicon-front
contact interface, causing shelves. Characteristic X-rays can leave through front contact, causing escape
peaks. Charge carriers too close to the interface can be lost, causing lower-energy tailing.
The charge collection efficiency (CCE) is modeled by Scholze as a monotonically increasing function of
the absorption depth z. It is assumed that the lowest CCE occurs directly at the interface C0 = C(z = 0)
and increases up to the depth z0 , where full absorption occurs; C(z0 ) = 1. The CCE is then effectively
28
modeled using the following exponential,

a
z
C(z) = C0 + (1 − C0 ) for z < z0 (2.78)
z0
C(z) = 1 for z ≥ z0 (2.79)
here z0 is the length of the ICC layer, whilst a and C0 are free-fitting parameters. Scholze reports the
following values for an SDD with accurate experimental correlation,
C0 = 0.9 (2.80)
a = 0.5 (2.81)
z0 = 50 nm (2.82)
Figure 2.23 shows the CCE curve in the case of the parameters for an SDD.
Figure 2.23: Charge collection efficiency curve for an SDD. Charge carriers near the interface z = 0 have a
higher probability to be lost.
Escape peaks
One of the most prominent detector artifacts is the presence of escape peaks. Escape peaks are caused by
the loss of a Si-Kα X-ray during the conversion of photon energy to charge.
When an X-ray of energy E enters the detector, it can knock out a silicon K-shell electron from its shell.
After relaxation of the Si atom, Si-Kα of 1.74keV can be emitted. In most cases, this X-ray is absorbed
back into the silicon bulk, and the total absorbed energy is still E.
However, when an atom near the front of the detector volume is affected, it is possible for this Si-Kα to
exit the detector volume. The attenuation length of a Si-Kα X-ray is in the order of micrometers. This is
many times higher than the attenuation length of the Auger and photoelectrons which are generated in the
detector bulk. Therefore, if the Kα X-ray leaves the detector, whilst all remaining charge is collected, a total
energy of E − ESi−Kα is measured. This results in a new peak in the spectrum that lies 1.74 keV lower than
the main peaks of the analyzed materials. Usually, only the escape peaks of Si-Kα are considered, as the
probability of generating and escaping Si-Kβ x-rays is so low that it can be ignored as noise [4]. A schematic
showing the shape of a typical EDS spectrum is shown in figure 2.24.
29
2.4. SILICON DRIFT DETECTOR CHAPTER 2. THEORY
Figure 2.24: Schematic of a typical EDS spectrum measured by a silicon detector. It shows the main Kα
and Kβ Gaussian peaks, as well as the influence of incomplete charge collection (ICC) and the silicon escape
peak.
2.4 Silicon Drift Detector

The noise level of detectors can be reduced in two ways, either by lowering the leakage current or by lowering
the capacitance. Even when a diode is reverse-biased, there is still a small amount of current that can
flow through the depletion region. This is known as the leakage current. The amount of leakage current is
determined by the minority current carriers, which are electrons in the p-type material and holes in the n-
type material. Through improvements in fabrication techniques of the silicon, it was shown that the leakage
current could be reduced drastically, such that it did not play an important role anymore [52, 53].
In the search for decreasing the capacitance of detectors, Gatti and Rehak [54] proposed an idea for a
novel detector design based on the process of sideward depletion. The concept involved using a small n+
ohmic contact that is positively biased compared to the p+ contacts on both surfaces of a high-resistivity
semiconductor bulk such as n-type silicon, to fully deplete it. This depletion creates a potential valley in the
center of the bulk, through which the electrons are guided toward the anode. This was done by dividing the
diode into a number of thin strips and applying a raising voltage for each subsequent strip.
This device is shown in figure 2.25. The main advantage of this design, compared to a standard PN diode
of similar size, is the small size of the anode. The anode is, contrary to PN diodes, almost independent of
the total area of the active volume, allowing for a very low capacitance to be reached [5].
30
2.4. SILICON DRIFT DETECTOR CHAPTER 2. THEORY
Figure 2.25: Early silicon drift detector design by Gatti and Rehak. Principles of semiconductor drift
detectors. The charge carriers created by incident X-rays are transported parallel to the detector surface
onto a small capacitance anode. Adapted from [55].
Further advances introduce the circular SDD, which instead of linear p+ strips, makes use of a circular
design, reducing the detector’s capacitance even further. The SDD consists of a cylindrical silicon lightly-
doped n− -type wafer that makes up the bulk of the active volume. Near the entrance window, there is a
highly doped p+ back contact (BC). On the opposite side lies a series of p+ concentric drift rings of increasing
sizes. The innermost of these drift rings is referred to as R1, whereas the outermost ring is referred to as
RX. In order to shield the integrated FET from incoming charge, a deep p-type guard is placed around it,
known as the Inner Guard Ring or IGR. The anode is responsible for the collection of charges and consists
of a highly doped n+ ring placed in between R1 and the IGR. Figure 2.26 shows a schematic of an SDD.
Figure 2.26: Cross-section of a cylindrical silicon drift detector with an integrated FET. The gate of the
transistor is connected to the collecting anode, which collects charges transported by the field strips. The
back contact serves as the entrance window for ionizing radiation. From [56].
31
Chapter 3
Methods
This chapter will describe the software packages and simulation steps that are required to simulate the EDS
spectra for SDDs. First, a quick overview of each software and its uses will be provided, after which the
details regarding its uses within the complete simulation chain are covered.
The simulation chain will be divided into three distinct phases. These are the pre-simulation phase, the
main simulation phase, and the post-simulation phase.
• The pre-simulation phase contains the creation of internal electric and weighting fields, as well as the
geometry required to define the virtual setup.
• The main simulation phase includes the steps which are taken during the simulation, such as the
simulation of the passage of particles through the virtual setup. It also includes charge carrier creation
and propagation, and signal induction.
• The post-simulation phase describes how the data is extracted and spectrum data is obtained from the
simulation.
3.1 Software Overview

The simulation consists of multiple different software packages which will each handle a specific task within
the complete simulation chain. The main software packages used in this thesis are the following,
• Geant4 [21]
• Allpix2 [23]
• COMSOL Multiphysics [26]
Figure 3.1 shows a schematic of the entire simulation pipeline and portrays in which step of the pipeline
each software package is involved. Each software and its uses will be briefly introduced in the following
sections.
32
3.1. SOFTWARE OVERVIEW CHAPTER 3. METHODS
Figure 3.1: Schematic of the simulation pipeline.
3.1.1 Geant4
Geant4 [21] is a powerful software toolkit used for simulating the passage of particles through matter, using
a Monte Carlo approach. The name “Geant4” stands for “GEometry ANd Tracking 4”, reflecting its ability
to simulate complex geometries of materials and track the paths of particles as they interact with the matter,
including energy loss, scattering, and absorption processes. The toolkit is written in C++ and is designed
to be modular, allowing users to choose the specific functionality they need for their simulations.
Penelope
In Geant4, a physics list is a collection of physics processes and models that describe the interactions of
particles with matter. The physics list that has been used throughout this thesis was the Penelope low-
energy electromagnetic model [57]. The Penelope code system is a widely used tool for simulating the
coupled transport of electrons and photons in the energy range of 50 eV to 1 GeV [58]. The Penelope
code has been applied to diverse fields, and its reliability has been verified through benchmark comparisons
with experimental measurement data from various sources, as shown by Sempau et al. [59]. The Penelope
codebase was written in Fortran 77, with some extensions of Fortran 90. As a result, it lacked the ability to
run parallel simulations and to be linked easily to other simulation packages. To overcome these limitations,
a C++ version of the original Fortran subroutines was developed, which could in turn be integrated with
Geant4 [60]. This allows access to the Penelope physics as part of the Geant4 toolkit, by means of assigning
it as a physics list. Some important lower energy processes which are handled through the Penelope model
are the photoelectric effect, Compton Scattering, and Bremsstrahlung.
33
3.2. PRE-SIMULATION PHASE CHAPTER 3. METHODS
3.1.2 Allpix Squared

Allpix Squared [23] is a relatively new modular open-source simulation framework for silicon detectors. This
modularity arises from the implementation of independent modules which communicate through a messaging
system. The modeling of charge carriers within the detector plays a critical component in accurately modeling
the device. In Geant4, materials are treated as a continuum of atoms, instead of crystal lattices. Secondary
particles are only produced above the mean ionization potential of the material, which is much higher than
the energy of electron-hole pair creation.
Allpix Squared attempts to resolve this by collecting the energy deposits in the semiconductor material
from Geant4 and converting them into clouds of electron-hole pairs, propagating them through a provided
internal electric field of the detector, and finally modeling the signal formation as a result of the propagating
charge carrier cloud.
A number of modifications, including bug-fixes as well as additional features were added to the Allpix
Squared framework. These modifications will be discussed in section 3.5.
3.1.3 COMSOL Multiphysics

COMSOL Multiphysics [26], hereafter COMSOL, is a well-established simulation software package that
handles finite element analysis for various physics and engineering applications. In our case, we are especially
interested in its capabilities in electrostatic field modeling. An electric field is required in order to perform
the propagation of the charge carriers. Additionally, a weighting potential field is required to perform charge
induction on the anode as a result of the Shockley-Ramo theorem, as explained in section 2.3.3. Both the
electric field and weighting field are modeled in COMSOL and are imported into Allpix Squared.
3.2 Pre-Simulation Phase

In the pre-simulation phase, the data is generated which has to be imported into Allpix Squared, before the
main simulation can start. This includes the simulation of the potential field and its resulting electric field,
and the weighting field in COMSOL. This also includes the geometry creation of the detector and setup,
which are supplied to Geant4 and Allpix Squared.
3.2.1 Electric and Weighting Field Modeling

The electric field and weighting fields are modeled using COMSOL Multiphysics. The simulated detector
was modeled after an in-house SDD by PNDetector, namely the MOD-SDD-P-030r-pnW-xxx-SLM. This
SDD has an active area of 30 mm2 and a thickness of 450 microns.
The manufacturer did not provide specific details about the inner geometry of the detector, as such
information is considered confidential. For instance, the exact layout of the drift rings, such as the total
number of rings, and the size and thickness of the rings was unknown. Furthermore, specific properties
regarding the type of silicon, such as doping concentration was unknown. As a result, many estimated
guesses had to be made to recreate a digital layout of the geometry. These guesses were based on available
literature of SDDs, such as [5, 61, 62].
First, the creation of the electric field will be discussed, followed by a description of how the model is
adjusted to obtain the weighting potential field.
Geometry
The geometry of the SDD closely follows the illustration in figure 2.26. A thickness of Z = 450 µm was
chosen for the axial symmetric model. The final drift ring was placed at a radius of r = 3.09, and an extra
450-micron layer was included to accommodate the guard rings at the edges, resulting in a total radius of
R = 3.54 mm.
A number of simplifications were made to the model in order to simulate the electric field. The pn-junction
is treated as an ideal diode, which has a number of consequences. Firstly, the n− -bulk is assumed to be
uniformly doped with a doping concentration ND . Secondly, the n− -bulk is assumed to be fully depleted,
34
meaning every dopant atom in the bulk will be ionized. As a result, every dopant atom will contribute to
the total space charge in the region and the charge density ρ in the n− -bulk will be,
ρ = ND e (3.1)
where ND is the doping concentration and e the elementary charge.

Furthermore, the boundary conditions of the SDD were simplified. At the boundary, there exist transition
regions where the charge density and effective doping concentration change rapidly over the range of a single
micron [18]. These transition regions appear when traversing from the n− -bulk to n+ anode and from the
bulk to the p+ drift rings. The transition regions are neglected for the simulation of the fields. As a result, the
charge density will be a constant value according to equation 3.1, throughout the entire simulated volume.
The boundary of the detector geometry is split into several segments, a schematic of which is shown in
figure 3.2. The top of the device contains the drift rings and therefore is split into a circle surrounded by
a set of concentric rings. The inner circle indicates the IGR and is surrounded by the anode ring and R1
ring. Concentric rings of increasing sizes up to RX make up the drift rings of the device. On the entrance
side of the device, there is a large circular Back Contact (BC), through which the photons will enter the
detector. On either side of the device, guard rings are placed at the edge, which are used to gradually bring
the potential field to zero.
Figure 3.2: COMSOL geometry schematic of a silicon drift detector.
Boundary Conditions
The boundary conditions were treated as follows. At each of the biasing contacts, the potential was kept
at a predefined value, using Dirichlet boundary conditions. These values were taken from the experimental
setup, which was tuned to the optimal biases voltages to obtain the highest resolution. These bias voltages
are given in table 3.1,
Contact Bias voltage [V]

BC -111
IGR -18
R1 -15
RX -130
Table 3.1: Optimal bias voltages for the silicon drift detector. These are obtained through a tuning algorithm
that minimizes the FWHM of a main spectrum peak.
The potential of the intermediate drift rings Ri are set to be linearly decreasing values between R1 and
RX, such that,
(RX − R1 )
Ri = R1 + · (i − 1) for i ∈ {1, 2, . . . , N } (3.2)
N −1
35
Where, R1 = R1 and RN = RX. The total number of drift rings was set to N = 50.
Five guard rings were placed at the outermost edge that linearly decrease the potential from BC (at the
entrance side) or RX (at the drift ring side) to zero. In between the rings, the boundary conditions were
given COMSOL’s “zero charge” condition. As a result, the normal part of the electric field vector at these
boundaries is zero, meaning electric drift lines are not allowed to cross them. The aim of this boundary
condition is to model the insulating layer of silicon oxide that lies at these boundaries.
The edge of the device at r = R was kept at a potential of 0 V. Similarly, the anode is also kept at a zero
potential.
Mesh Builder
COMSOL offers automatic mesh generation that can create a mesh based on the defined geometry and user
settings.
Built-in quality assessment of the generated mesh is performed by COMSOL to prevent numerical inac-
curacies or convergence issues. Therefore, the mesh is checked for element shape, aspect ratio, and skewness.
The mesh is built using the extremely fine element size preset, which results in a generated mesh according
to the values given in table 3.2. A section of the axisymmetrical mesh that was used is shown in figure 3.3.
Maximum Minimum Maximum Number of

Curvature Resolution of
element size element size element triangular
factor narrow regions
[mm] [mm] growth rate elements
0.00708 0.00142 1.1 0.25 1 92436
Table 3.2: Mesh Parameters for COMSOL Simulations
Figure 3.3: Visualization of the COMSOL Mesh.
Electrostatic Solver
The potential field in the sensitive detector volume is calculated by solving the Poisson equation in the
previously described geometry, using COMSOL’s electrostatics module. The Poisson equation,
ρ ND e
∇2 V = − = (3.3)
ϵ ϵ0 ϵSi
is solved, where e is the elementary charge, ϵ0 the vacuum permittivity and ϵSi the relative permittivity
of silicon, which is set to ϵSi = 11.68. The doping concentration in the bulk ND was not revealed by
the manufacturer. Therefore, an estimation was performed based on the depletion voltage provided by the
manufacturer and equation 2.55, as well as visual inspection of the drift lines during optimal performing
conditions. The doping concentration was chosen to be ND = 7.5e11 cm−3 .
The electric field can be found by taking the negative gradient of the potential, such that,
E = −∇V (3.4)
36
The solution is solved on the mesh shown in figure 3.3 using the stationary solver. After generating
the solution for either the electric field or weighting field on the two-dimensional axisymmetrical slice, a
three-dimensional grid is obtained by revolving the slice around the axis r = 0.
The field reader modules of Allpix Squared expect the fields to be in a regularly-spaced three-dimensional
grid. Therefore, the solution was interpolated at regular intervals in to include the entire detector,
2R
xi = −R + · (i − 1) for i ∈ {1, 2, . . . , N } (3.5)
N −1
2R
yi = −R + · (i − 1) for i ∈ {1, 2, . . . , N } (3.6)
N −1
Z
zi = · (i − 1) for i ∈ {1, 2, . . . , N } (3.7)
N −1
Where R is the radius, and Z the thickness of the detector. N refers to the number of points used for the
grid. N = 300 was used for the electric field, providing grid size in the z-direction of 1.5 µm and in the x-
and y-direction of 23.6 µm. For values higher than N > 300, memory issues would occur due to having to
load the entire grid into memory before exporting. As the design of the SDD is cylindrical, the grid points
that lie at x2 + y 2 > R2 will be undefined. Therefore, these values were clamped to a zero value.
A custom python script was written to perform the file conversion from the COMSOL output into the
regular grid INIT format that is required by Allpix Squared’s field reader modules.
Weighting Field
The weighting potential field is constructed in a similar manner as the electric field, with a few alterations.
As explained in section 2.3.3, the weighting field is the result of the Shockley-Ramo theorem that describes
how much of a signal is induced on the anode, when a charge moves through space. It can be obtained by
setting the boundary conditions for the cathodes to a zero potential. This includes the drift rings, BC, IGR,
and guard rings. The anode is given set to unit potential. In addition, the space charge is set to zero, thus
effectively solving the Laplace equation,
∇2 V = 0 (3.8)
The geometry, mesh, and solver were kept the same as for the electric field creation. Although, since fewer
values needed to be exported, the number of grid points in the mesh could be increased to N = 500, before
running into memory issues.
3.2.2 Geant4 Geometry Modeling

The GeometryBuilderGeant4 module in Allpix Squared is responsible for constructing the Geant4 geometry
of the simulation. The Geant4 geometry includes information about the shape, orientation, and size of the
simulated objects, but also on their material and elemental properties.
The GeometryBuilderGeant4 module handles the construction of both the active and passive volumes
in the simulated environment. The active volume describes the silicon sensor volume within the detector
in which the electron-hole pairs are formed, propagated, and measured. The passive volumes describe the
geometry configurations of all remaining parts of the setup. This could for instance be the detector collimator
and shielding, but also the sample, sample holders, and setup housing. Principally, passive volumes contain
all the geometries, excluding the silicon sensor, that an emitted particle can interact with.
GDML files The GeometryBuilderGeant4 module supports loading of arbitrary Geometry Description
Markup Language files [63], or GDML files for short. GDML is a language based on XML that was designed
specifically for describing detector geometries. Since GDML is pure XML, it can be used universally and is
particularly useful as a format for exchanging geometries between different applications. GDML files contain
geometry trees that correspond to the hierarchy of volumes that make up a detector geometry.
A GDML file is structured with several different components. The first section includes numerical values
for various constants, positions, and rotations that will be used later in constructing the geometry. The
second section includes definitions for all materials in the setup. These can be made from a single element
37
3.3. MAIN SIMULATION PHASE CHAPTER 3. METHODS
or mixtures, which can be composed based on a fractional mass or atomic ratio. The third section comprises
all the solid definitions used in the geometry description. Simple solids such as boxes, spheres, and cones,
as well as more complicated solids like hyperbolic tubes and twisted solids, are supported. Composite solids
made using boolean operations such as union, subtraction, and intersection are also supported. The fourth
section contains the actual implementation of the geometry tree, along with the assignment of solids and
materials. The hierarchy of volumes is defined by specifying daughter volumes positioned inside a volume.
The last section contains information about the top volume of the geometry tree, such as the world volume.
3.3 Main Simulation Phase

During the main simulation phase, the externally created geometries and fields will be imported into the
Allpix Squared framework. Allpix Squared acts as a wrapper for Geant4, and particles are created through
a particle source. The creation and propagation of charge carriers is performed, and the Shockley-Ramo
theorem is used to find the induced signal. All modules are described in detail in the Allpix Squared User
Manual [64], however the most important modules and their inner workings will be briefly discussed here.
3.3.1 Initialization
In the first step of the Allpix Squared simulation, all included modules are initialized. The electric field
and weighting field INIT files are imported using the ElectricFieldReader and WeightingPotentialReader
modules, respectively. The geometry is loaded through a detector configuration file, which describes the
geometries and materials of the detector and setup. This can be done manually through the configuration
file, or by loading an external GDML file.
Particle Source
In order to start the simulation, energy has to be injected into the system by means of a particle source.
This particle source may be in the form of an electron or photon beam or other, such as a radioactive decay
process. The former are useful when simulating SEM-EDS, whilst the latter is useful when measuring the
detector response under irradiation of a radioactive source, like Iron-55 (Fe-55). When dealing with a simple
source, like a point or beam source with a single energy, the source can directly be defined within the Allpix
Squared main configuration file. However, to create more complex sources, an external file containing Geant4
General Particle Source (GPS) macros can be provided and passed to the Allpix Squared framework. This
allows for the creation of sources with user-defined energy and angular distributions.
3.3.2 Core Simulation

The Core simulation will start after all modules have been initialized. It consists of many events, which
are independently treated. During each event, a single particle emission from the particle source will be
treated. The energy that finds its way into the detector will be converted to electron-hole pairs and be
propagated through the provided electric field. Finally, the signal that is induced on the anode as a result
of the propagating charge clouds is measured and a measured energy spectrum can be obtained.
Geant4 Deposition
First, a particle will be emitted from the particle source that was defined in the main configuration file. This
particle is stepped through the world volume using the geometry that was provided in the initialization.
The step size is dependent on a number of different factors. Each particle has access to a list of available
interactions based on the material it resides in. At every step, all processes available in the volume are
invoked to get a proposed physical interaction length, based on attenuation factors for that process. The
process which requires the shortest attenuation length limits the total size of the step. Furthermore, each
step is limited at the boundaries of volumes. Depending on which process a particle undergoes, secondary
particles are generated, for example in a photoelectric absorption event, a secondary photoelectron is created.
Figure 3.4a shows a flowchart for the stepping procedure used in Geant4.
38
The particles in Geant4 are tracked down to the value of the production cut, which is given at the start
of the simulation. The production cut is the distance a secondary must minimally travel to be produced.
Specifically, secondaries that do not have enough energy to travel at least a distance equal to the production
cut are not produced, and instead, this energy is “deposited” into the material. This process continues until
all particles have reached the zero energy limit.
If none of the secondary particles leave the active volume of the detector, the sum of all energy deposits
will add up to the total incident X-ray energy. Each energy deposit, along with additional information like
deposition time and position are collected by Allpix Squared for further processing.
Figure 3.4b shows a schematic of the stepping procedure through multiple volumes. The particle’s initial
position is marked by the red circle, and as it propagates through the simulation volume, it undergoes
various physics interactions and collisions with geometry boundaries, represented by the orange circles.
These interactions give rise to the generation of secondary particles, depicted as blue circles. The stepping
algorithm propagates the particles until they reach the zero energy limit, thereby completing the simulation
for a single event. This is an example of a single event of Geant4.
(a) Flowchart of the Geant4 stepping algorithm. (b) Schematic of a single event in Geant4.
Figure 3.4: Stepping of Geant4 particles through the geometry.
Charge Carrier Creation

Geant4 does not treat processes at very low energy, such as electron-hole pair creations in silicon. Therefore,
the energy depositions are loaded into the Allpix Squared framework. As shown in section 2.3.1, the absorbed
energy does not simply map one-to-one to a number of created charges. The number of electron-hole pairs
created by a single energy deposition is calculated using the mean pair creation energy. Additionally, the
fluctuations are modeled using the Fano factor, assuming Gaussian statistics. A Gaussian peak with the
39
following mean µ and variance σ 2 is sampled to obtain the mean number of charges created for each energy
deposition.
Edep
µ= (3.9)
ω
σ2 = F µ (3.10)
Here, the ω and F were set to their default values for silicon in Allpix Squared, ω = 3.64 eV and F = 0.115.
The number of charge carriers created for each energy deposit in the active volume along with deposition
times and positions are stored in an Allpix Squared object called DepositedCharge. Objects can be used to
transfer data between modules using Allpix Squared’s messaging system.
Charge Propagation
The GenericPropagation module is responsible for the simulation of the propagation of electrons and holes
through the active volume of the detector. The module listens for the DepositedCharge object from the
DepositionGeant4 module and retrieves the charge carriers and their deposit locations. These will then form
the initial positions for propagation. Furthermore, the charges can be grouped and propagated together to
save computation time.
The process of charge carrier propagation involves both a drift and diffusion step. The charge group will
be propagated using these two steps until one of two conditions is reached: Either the total propagation time
is surpassed, or the charge is considered to not be in motion anymore. The latter occurs when the charge
has taken a step outside of the detector volume, or has recombined with the lattice. This process will be
repeated for every deposited charge until every charge has finished propagating.
Drift The drift step is computed using the charge carrier mobility, which is automatically selected based
on the charge carrier type. This ensures both electrons and holes are handled correctly. The electric field of
the detector has to be sampled during each step of the integration. The velocity of the charge carrier is set
according to the drift velocity equation,
v = µE (3.11)
Here, the mobility µ is determined by the Jacoboni-Canali theorem as explained in section 2.3.2.
The position of the charge carrier is then integrated over time by means of a fourth-order Runge-Kutta-
Fehlberg method [65] along the sampled electric field. The integration process enables the equations to
be solved iteratively, with the flexibility to customize the step size. Additionally, the step size is adjusted
during the integration based on the error of the Runge-Kutta step, by utilizing a tableau with error estimation
capabilities.
Diffusion As explained in section 2.3.2, the charges also experience diffusion when they propagate across
the sensor volume. In order to properly simulate this behavior, each charge carrier takes another step after
each Runge-Kutta step. This step is randomly sampled from a Gaussian distribution with the mean µ = 0
and standard deviation σ based on the Einstein relation,
r
2kb T
σ= µ∆t (3.12)
e0
Using charge carrier mobility µ, temperature T , and time step size ∆t. Effectively, a random walk is
performed for a single step based on the Einstein relation at the end of each Runge-Kutta step. The
information of the charge propagation is then stored in the PropagatedCharge object.
Charge Induction
The InducedTransfer module is responsible for measuring the total induced charge on the anode as a result of
the charge propagation. The module listens for the PropagatedCharge object from the GenericPropagation
module and retrieves the final position of the charge carrier after the propagation. As shown in the section
2.3.3, the Shockley-Ramo theorem can be used to determine the total induced charge on an electrode, as a
40
3.4. POST-SIMULATION PHASE CHAPTER 3. METHODS
result of a moving charged particle. It was found that the total induced charge on the electrode was solely
dependent on the endpoints of the trajectory of the charged particle, and not on the specific path it takes.
For a group of NG electrons, the total charge of a group is equal to q = −NG e, with elementary charge e.
Since the total charge, the start and end positions of the charge group, and the weighting field ψ(x) are
known, the total induced charge can be found by the Shockley-Ramo theorem. The total induced charge on
the anode as a result of a single charge group being propagated from x0 to x1 can be calculated as follows,
Qind
G = −NG e0 [ψ(x1 ) − ψ(x0 )] (3.13)
where ψ(x) is the weighting field at position x.
This induction is performed for every set of electrons or holes in a single event. The total charge that is
found during a single event, is the summation of all the induced charge by the individual sets of carriers,
X
Qind
tot = Qind
G (3.14)
G
The total induced charge is then stored in the PixelCharge object. This object contains the data required
to generate the final energy spectra.
3.4 Post-Simulation Phase

After the main simulation phase is finished, Allpix Squared will generate output in the form of ROOT files.
These files can be read out to obtain the data required for generating energy spectra.
3.4.1 Data extraction

ROOT
ROOT [66] is a data analysis framework developed by CERN. Allpix Squared relies strongly on ROOT for
its internal and external handling of simulation data. For example, internally ROOT is used for handling
histogram creation and coordinate transformations.
When outputting data simulation data in Allpix Squared, this is usually done using the ROOTOb-
jectWriter module, which stores the requested objects on disk using in a ROOT TTree file format. The data
is stored on separate trees of each object and a TBranch is created for each unique message, which contains
a combination of the detector, the module, and the message output name. For each event, the data of the
objects is stored under a TLeaf reference in the appropriate TBranch. Internal references between objects
within a single event are stored internally as ROOT TRefs, which allows for the retrieval of related objects.
For example, a single PixelCharge caught by the detector can be related back to the initial energy deposition
which caused the charge to be created, stored in the DepositedCharge object. In this manner, the entire
history of the simulated charge carriers can be tracked.
Furthermore, a specific object can be loaded back into the framework through the ROOTObjectReader
module, which converts the object tree data from a given ROOT file back into a message. This is useful
for instance when comparing a single simulation with multiple different electric fields. In this case, only
a single simulation has to be performed to obtain the DepositedCharge objects for the specific source and
geometry of the setup. Subsequently, the energy deposits can be retrieved using the ROOTObjectReader
and further simulation such as propagation and charge induction can be performed, without rerunning the
charge deposition step.
ROOT Macros
ROOT Macros were used to extract or analyze the data included in a ROOT file. ROOT macros are created
in C++ and are compiled just-in-time through the ROOT CLI. The macros contain functions to gather the
data pertaining to a specific object tree and detector branch. The properties of the tree, such as charge
positions or gathered charge are then plotted internally using ROOT histograms. Alternatively, the data is
extracted with a ROOT macro to a human-readable output and analyzed using Python scripts.
41
3.5. ALLPIX SQUARED MODIFICATIONS CHAPTER 3. METHODS
ROOT also supports a python interface called PyROOT to directly access the data stored in TTrees.
This method proved to be much slower than the C++ counterpart, especially in the case of millions of
deposited charges. As a consequence, PyROOT was not used.
3.5 Allpix Squared Modifications

Allpix Squared is a fairly recently released simulation framework and is still under very active development,
with many improvements being committed to its master branch every day at the time of writing. Further-
more, its open-source nature allows the scientific community to implement new modules and features, or
solve bugs encountered during simulation. Some bugs were found during the simulations that significantly
impacted both the quality of the spectra, as well as the performance of the simulation speed. A number of
modifications, including bug fixes and implementations of new features, were made to the Allpix Squared
framework during this thesis. These will be discussed in this section.
3.5.1 Discretization Error in Pair Creation

Allpix Squared transforms the energy deposits generated by Geant4 into electron-hole pairs. The total charge
conversion procedure happens in the following order,
• Energy deposition in Geant4

• Gaussian sampling
• Type conversion
• Charge summation
First, the energy deposit during a hit in the active volume from Geant4 is loaded by Allpix Squared. After
this, a Gaussian distribution based on the Fano factor and mean charge creation energy is sampled and a
floating point value is returned. This floating point value is then converted to an integer that represents
the total number of charge carriers created during that specific energy deposit. Finally, for each event, the
number of charge carriers of all hits is summed together to obtain the energy deposit spectrum.
A fundamental problem was found in the implementation of the DepositionGeant4 module, namely at the
type conversion step explained above. The type conversion was declared explicitly through the static cast
operator, which truncates the number of charges for every single energy deposit. As a consequence, the
number of charges sampled was always rounded down to the nearest integer.
For example, if the Gaussian sampling would return 9.9, the number of charge carriers created would be
truncated to the integer 9. It was found that this introduced a bias towards the lower energies, caused by
the accumulation of truncation errors during a single event. This had a significant impact on lower energies
in use during EDS, as this caused peak shifts of the main peaks to lower energies. An improvement has
been made by instead rounding the mean charge energies to the nearest integer, instead of truncating. The
following analysis was carried out to assess the performance of the improvement.
Firstly, a photon beam of energy 10 keV was fired at a cube of silicon with an edge length of 10 cm.
This ensured that it was thick enough to completely absorb the incident photons. As the mean charge
creation energy was set to its default value of 3.64 eV, it is expected that at full absorption, a total of
10, 000/3.64 = 2747.25 charge would be created on average during each event.
In order to verify this, 100,000 events were simulated with a Fano factor of 0. This choice of Fano factor
eliminates intrinsic variations in the charge creation process. While this particular configuration does not
directly replicate a physical scenario, it offers insight into the discretization process.
The original Allpix Squared algorithm can be compared to the improved modification in figure 3.5. The
original Allpix discretization has a significant amount of lower energy tailing as a result of the truncation,
whereas the improved version is symmetrical around the expected value of 2747.25 mean charge creations.
42
Figure 3.5: The total number of charge deposits for 100,000 events for 10 keV incident photons. Comparison
of the original Allpix Squared discretization vs. the improved version.
This can be explained by means of the following example. A photon is absorbed in the silicon and
Geant4 deposits energy in 6 separate deposition events, as shown in table 3.3. The mean charge creation is
then calculated by dividing the energy deposit by the mean charge creation energy, which is 3.64 eV in this
case. Next, the old and improved type conversion from floating point to integer is performed in the last two
columns and indicates the total number of charges created by the energy deposit.
Energy Depositions Mean charge Number of charges created

[keV] creation Original Allpix Improved
Dep. 1 0.02614 7.18 7 7
Dep. 2 0.09491 26.07 26 26
Dep. 3 0.0956 26.26 26 26
Dep. 4 1.61185 442.82 442 443
Dep. 5 0.98038 269.34 269 269
Dep. 6 7.19112 1975.58 1975 1976
Sum 10.0000 2747.25 2745 2747
Table 3.3: Discretization process for an event with six energy depositions. Comparison of the original Allpix
Squared discretization vs. the improved version.
The loss of just a couple of charges by the truncation might seem like a trivial difference, however
with increasing energies in the low energy regime, more energy deposit events occur and as a result, more
truncation errors accumulate. Eventually, this can lead to a significant error in the output spectrum.
The verify this, a range of simulations was performed where 300,000 photons of increasing energies were
measured on a block of silicon, in this case with the physical Fano factor of 0.115. The output energy deposit
spectra were analyzed by means of a Gaussian fit on the main peak. The center of the Gaussian fit was
compared to the expected peak energy and from this, the peak shift was recorded for each simulation.
The result is found in figure 3.6. It can be seen that the original implementation has an increasing peak
shift as a result of truncation error when incident photon energy is increased, whereas the improved method
hovers around zero, meaning the simulated energy peak lies at the expected theoretical energy.
43
Figure 3.6: Peak shift of the main incident peak for varying incident photon energies. Comparison of the
original Allpix Squared discretization vs. the improved version.
This problem was brought up to the Allpix Squared developer team by means of a merge request (https:
//gitlab.cern.ch/allpix-squared/allpix-squared/-/merge_requests/982), and was promptly accepted.
3.5.2 Front Contact ICC

Incomplete charge collection as a result of the front contact is an important effect, as explained in section
2.3.5, that causes lower energy tailing in the peaks. It was found that Allpix Squared handles the charges at
the front contact in an insufficiently accurate manner to represent this phenomenon. This is a consequence
of how Allpix Squared handles charge propagation at the boundary of the sensitive volume. The Allpix
Squared propagation is performed by the GenericPropagation module and it relies on the following steps.
First, the charge group is loaded according and their deposition time and position are loaded from the
DepositedCharge object. The Runge-Kutta step will be done based on the position of the charge and the
local electric field vector. Next, the diffusion step will be performed and the charge will take a step in a
random direction. It was found that in the case of low-energy incident photons, due to the shallow absorption
depth, charges would have a high probability to take a diffusion step backward out of the sensor volume,
causing this charge to be lost. Therefore, the ICC would be too pronounced using this model.
In order to verify this behavior, the simulated results are compared with the results of a PIN Si(Li)
detector by Scholze et al. [9]. A PIN detector was chosen as these usually have a thicker dead layer than
SDDs. Scholze provides the response function at a monochromatic photon excitation of 4 keV for a detector
with a gold front contact of 18.1 nm thick. The measured experimental spectrum is shown in figure 3.7. The
spectrum shows two peaks, the first is the main excitation peak of 4 keV and the second is a smaller escape
peak. In between the peaks lie shelves, which are primarily caused by the Au-MNN (for E < 1.5 keV) and
Si-KLL (for 2.5 < E < 4 keV) Auger electrons. At energies below 0.5 keV, there is a sharp peak, which
is caused by electronic noise. The main peak shows signs of ICC, which tapers off into the Si-KLL shelf.
Lastly, the Au-Mα is just visible, showing up as lower energy tailing of the escape peak at 2.1 keV.
The response function was modeled using Allpix Squared, using a 3 mm thick PIN detector layout with
a simple linear internal field. 3 million photons with an energy of 4 keV were shot perpendicular to the
detector surface. The simulation results are compared to the measurement and Scholze’s model in figure 3.7.
The main properties of the spectrum are displayed well, namely the main peak at 4 keV, the escape peak,
and the Au-Mα . The shelves originating from Au-MNN and Si-KLL Auger electrons are also well represented.
44
However, it is clearly visible that the ICC behavior is too excessive, causing higher lower-energy tailing than
the experimental data obtained by Scholze.
Figure 3.7: Detector response function for a PIN detector under irradiation of 4 keV X-rays. No front contact
ICC is modeled.
Allpix Squared’s model was improved upon by incorporating Scholze’s front contact ICC model [9] into the
GenericPropagation module, by utilizing the charge collection efficiency (CCE) curves explained in section
2.3.5. According to the CCE curve displayed in figure 2.23, charges are stochastically recombined depending
on the deposited depth within the silicon.
Figure 3.8 shows an example showing the new improved ICC model. Here, the ICC profiles were slightly
modified from the fitting parameters used by Scholze, namely z0 = 150 nm and a = 0.4. This brings the
ICC tailing down to the experimentally measured ICC tail. However, as an expense, this also brings down
the level of the shelves slightly.
Figure 3.8: Detector response function for a PIN detector under irradiation of 4 keV X-rays. Front contact
ICC is modeled using Scholze’s front contact model.
45
3.5.3 ROOTOutputWriter Speed Issue

A speed-related issue was found related to the ROOTOutputWriter when multiple output objects were
written to disk. When each object was written to disk in separate simulations, the total time spent in the
ROOTOutputWriter module was drastically lower than when writing all the objects to disk during the same
run.
Tests were done comparing the simulated times for 50,000 events containing a 4 keV photon beam at
a PIN detector model. Three different objects were written to disk, namely the DepositedCharge, Propa-
gatedCharge, and PixelCharge objects. First, three simulations were run to output each object separately
and a fourth simulation was run where all three objects were written in a single simulation run. Table 3.4
shows the total time spent in the ROOTOutputWriter module for each simulation. The expected behavior
is that value in the rightmost column would be roughly the sum of the previous three columns, however,
a greatly higher value is observed. In the case where multiple objects are written to disk, the majority of
the total simulation time is spent in the ROOTOutputWriter module, whilst for the other three cases this
ranges between 0.2-2% of the total simulation time.
DepositedCharge,
DepositedCharge PropagatedCharge PixelCharge PropagatedCharge,
PixelCharge
Time [s] 5.09 32.91 3.038 1664.22
Table 3.4: Total time spent in the ROOTObjectWriter module when writing different objects to disk.
This issue was brought up to the Allpix Squared developer team and the problem was identified. The
underlying cause was related to how Allpix Squared stores object references. In order to preserve a trackable
history for each charge, each object contains references to their parent objects, by means of TRefs. These
TRefs act similarly to C-style pointers, allowing for direct access to the referenced objects without additional
bookkeeping. Moreover, these references persist when written to a ROOT file and are subsequently read
back into memory. ROOT implements this functionality through a central lookup table, which maintains
information about the referenced objects and their memory locations.
However, this approach has certain limitations, particularly in the case of multithreaded simulations. The
lookup table is a global object, necessitating the use of mutexes for access. When multiple threads generate or
utilize TRef references, they must share this mutex and consequently experience significant waiting time for
lock release. As more TRefs are generated throughout the simulation, the size of the central reference table
increases. The lookup table is initialized with a fixed size, and once the number of TRef objects surpasses
this pre-allocated space, new memory must be acquired. This results in the reallocation of memory for the
entire expanded size of the table. Given the potential presence of millions of entries, this operation quickly
becomes computationally expensive, significantly slowing down the simulation process. This issue was then
resolved by the Allpix Squared developers in collaboration with the ROOT developers and merged into the
master branch (https://gitlab.cern.ch/allpix-squared/allpix-squared/-/merge_requests/952).
3.5.4 Additional Changes

A number of smaller tweaks were also made to the Allpix Squared framework. A few of these tweaks are
mentioned here in brief.
Custom Shapes
The GeometryBuilderGeant4 module supports building a geometry from the ground up by utilizing a small
number of primitive solids to construct the geometry. Available primitives include classes for cubes, cylin-
ders, and spheres. These primitives proved to be too limited to accurately describe the geometry of the
experimental setup in section 4.3, therefore additional passive model classes were created to be used by the
Allpix Squared framework. These models included conical shapes, sliced cylinders, and shapes with holes.
Figure 3.9 displays some of the custom geometry models that were loaded into Allpix Squared.
46
Figure 3.9: Custom geometry models created for Allpix Squared.
Custom Materials
Allpix Squared comes with a number of predefined materials in a MaterialManager class. The number of
defined elements and compound materials was limited. Therefore, several custom materials were added to
the MaterialManager class to accurately describe materials used in the experimental setup. Added materials
included AISI 303 stainless steel, nickel and aluminum alloys, and silicon dioxide, to name a few. The
elementary makeup of the compound materials, as well as their densities, were added.
GDML loading bug fix

Allpix Squared supports the loading and exporting of GDML files, as explained in section 3.2.2. An issue
was found when exporting a GDML file of the detector setup, and subsequently loading this GDML file
for a new simulation. This issue arose from a mismatch in the expected values of the export and import
functions. The GDMLOutputWriter module exported the GDML file containing the names along with the
corresponding memory addresses of the solids. The importer in turn expected names of the solid without
memory addresses, causing a fatal exception to occur. This issue was resolved and a merge request (https://
gitlab.cern.ch/allpix-squared/allpix-squared/-/merge_requests/984) was sent to the developers,
who merged the fix to the master branch.
47
Chapter 4
Results
This chapter will describe the results obtained through the simulated model, their analyses, and their im-
plications. The chapter begins with an examination of the electric field in the detector, discussing its
characteristics and the impact of doping dependence on drift lines. The section then explores the modeling
of the ICC front contact, including the absorption depth and detector response. Furthermore, the model is
compared to experimental XRF data, detailing the setup, materials, and source used. The impact of bias
tuning on the detector’s electric field and spectral response is discussed in detail. Lastly, the chapter focuses
on speed-accuracy considerations, specifically analyzing charge grouping and spatial precision as parameters
affecting simulation efficiency.
4.1 SDD Field Modeling

In this section, the simulation results of the electric and weighting fields are presented, obtained through the
modeling of the geometry, mesh, and boundary conditions in COMSOL, as described in Section 3.2.1. The
characteristics of the fields are identified, as well as the role that the bias voltages and doping concentration
play in establishing a well-defined field.
4.1.1 Electric Field

First, the results of the obtained potential and corresponding electric drift field are discussed. The optimal
bias voltages, as described in table 3.1, were used for the boundary conditions. The simulated potential field
and electric drift lines are presented in figure 4.1.
Within the active region (r < 2.7 mm) the bias conditions generate a “gutter-like” potential field, the
minimum of which runs radially along the entire device towards the anode. The electric field lines at the top
and bottom of the device, quickly find their way into the potential valley. As there is a strong electric field
at the top and bottom of the device, the electrons created here will drift along the drift lines and quickly
settle at the bottom of the potential valley. Once there, they will slowly propagate toward the anode, where
the charge will be induced and a signal will be generated.
The holes will propagate in the opposite direction as the electrons along the drift lines. As a result, the
holes will quickly drift towards the nearest cathode at either z = 0 mm or z = 0.45 mm.
On the far right, the effect of the guard rings is visible. The potential field is gradually brought to zero
at the edge of the cylindrical wafer. The far-right portion does not belong to the active volume, the charges
that end here do not induce any signal. In order to prevent photons from entering the edge of the wafer,
this region is collimated to ensure photons only land in the active volume.
The potential fields in the three-dimensional render are shown for the side containing the drift rings, and
the entrance side in figures 4.2. Furthermore, the magnitude of the resulting electric field is shown in figure
4.3.
48
4.1. SDD FIELD MODELING CHAPTER 4. RESULTS
Figure 4.1: Potential field and electric drift lines for an SDD with optimal biasing conditions.
(a) Side containing drift rings. (b) Entrance side.
Figure 4.2: Three-dimensional renders of the potential field in the SDD at optimal bias conditions.
In the following analysis, the charge carrier drift paths through the electric field under two different
scenarios are investigated under varying photon energies. The electric field shown in figure 4.1 is utilized
in Allpix Squared’s ElectricFieldReader module to propagate the charges through the GenericPropagation
module. Within the detector, ten photons with energies of either 2 keV or 10 keV are emitted from a
uniform distribution within a circular region centered at the middle of the detector with a radius of 2 mm.
Throughout their motion within the detector bulk, the trajectories of both electrons and holes are tracked.
Figure 4.4a illustrates the drift paths of electrons and holes for 2 keV photons, and figure 4.4b depicts the
paths for 10 keV photons. The electrons are represented by blue lines, while the holes are represented by
red lines.
In the case of the 2 keV photons, all photons are absorbed very close to the front contact of the detector.
In this region, there exists a strong electric field, as can be seen from figure 4.3. As a result, the electrons
quickly drift toward the bottom of the potential valley. Once in the valley, there exists a small but nonzero
electric field in the direction of the anode, causing a slow drift. As a result of this slow drift speed, the
diffusion becomes much more apparent in this section of the drift lines. In this case, the holes immediately
drift towards the BC and are barely visible due to the low absorption depth.
Conversely, the 10 keV photons penetrate much deeper into the detector. This can be seen by the drift
paths of the holes, which are much longer compared to the previous case. One of the ten photons even passed
49
through the detector completely, without interacting within the detector volume, leaving only 9 electron-hole
drift lines to be displayed. Also, in this case, the holes immediately drift toward the BC.
Figure 4.3: Magnitude of the electric field for the SDD in the case of optimal biasing conditions.
(a) Eγ = 2 keV (b) Eγ = 10 keV
Figure 4.4: Drift paths of electron (in blue) and holes (in red) under different incident photon energies Eγ .
4.1.2 Doping Dependence Drift Lines

The tuning of bias voltages in SDDs serves the purpose of ensuring that the drift lines align. It was found
that the doping concentration in the bulk of the sensor was a major factor in determining the shape and
destination of these electric drift lines. As no information regarding the doping concentration was supplied
by the manufacturer, this value had to be estimated using the depletion voltage of the detector, using
equation 2.55. It was found that by slightly altering the doping concentration, the potential field could
change drastically. Figure 4.5 illustrates this phenomenon.
First, the case is treated where the doping concentration is deemed too high for the current set of bias
voltages, at ND = 1.5e12 cm−3 . This case is shown in figure 4.5a. As a result of the additional exposed
space charge, the main potential valley is shifted down the z-axis, towards the center of the device. This
causes the drift lines to be unable to reach the anode, and instead, they settle at the center of the device,
below the IGR. This has a great impact on the output spectra, as the electrons created in the sensor will
follow the drift lines toward the center of the device. Here, they are trapped, unable to induce any significant
charge on the anode. Therefore, the total spectra would show a completely distorted spectrum, using this
set of bias voltages and doping concentration.
50
Conversely, the doping concentration can be too low for the set bias voltages, as such is the case in
figure 4.5b, where a doping concentration of ND = 6e11 cm−3 was used. In the case where the doping
concentration is too low, the potential valley is unable to dip deep enough to ensure alignment with the
anode. This can be observed in the drift lines as a shift of the potential valley up the z-axis. In this case,
most of the created electrons will end up in the potential valley and drift towards one of the intermediate
drift rings and recombine, resulting in little to no induced signal. Only the electrons deposited in the very
center of the device, at r < 0.2 mm, can make their way to the anode. As a result, only a tiny amount of
charge is induced on the anode.
In conclusion, it was found that the doping concentration in the bulk of the SDD plays a critical role in
shaping the electric drift lines and their destination. By making slight adjustments to the doping concen-
tration, the potential field undergoes significant changes. This was demonstrated by the vertical shift of the
primary potential valley along the z-axis, leading to partial or complete misalignment of charges with the
anode in extreme cases. These observations highlight the crucial role of doping concentration in achieving
optimal charge collection in SDDs. Proper tuning of bias voltages and doping concentration is essential to
ensure efficient charge transport and accurate spectral measurements.
(a) ND = 1.5e12 cm−3
(b) ND = 6e11 cm−3
Figure 4.5: Potential fields and drift lines for the SDD, using varying doping concentrations ND .
4.1.3 Weighting Field

The weighting field was obtained, as explained in section 2.3.3, by solving the Laplace equation for the same
geometry used in the electric field model. The boundary conditions were set such that the cathodes are
grounded and the anode is a unit potential. The resulting field is the weighting field and is displayed in
figure 4.6. In this figure, only the inner region (r < 0.8 mm) is shown because the rest of the region has a
weighting field of 0.
The resulting weighting field is localized around a small region near the anode. It exhibits a rapid decrease,
approaching zero as one moves away from the anode. As a consequence, only the charges that move within
this region are able to induce a charge on the anode. Outside of this region, the motion of the charge carriers
induces very little charge. This phenomenon is known as the near-field effect, or small-electrode effect [67].
This has implications for the contribution of the induced signal by the two types of charge carriers,
electrons and holes. As the X-rays in use for EDS operation are of relatively low energy, the penetration
depth through the depleted silicon is quite small. As a result, most of the electron-hole pairs are produced
close to the entrance window. In this region, the weighting potential is essentially zero, meaning any
movement of a charge through this region does not induce a signal on the anode.
51
4.2. ICC FRONT CONTACT CHAPTER 4. RESULTS
As the holes will propagate towards the nearest cathodes following the drift lines, few will actually move
through the near-field region. In fact, only the holes which are generated within a non-zero region of the
weighting field near the anode, are able to induce a charge on it, before they are moved away. This is also
visible in the drift paths in figure 4.4. As a consequence, holes will only marginally contribute to the total
induced charge on the anode, and nearly all the induced charge is the result of electrons finding their way
to the anode.
Figure 4.6: Weighting field of the SDD.
4.2 ICC front contact

Incomplete charge collection is a significant factor contributing to the loss of charge in the detector, as
elaborated in section 2.3.5. The presence of a dead layer at the entrance side of the detector gives rise to a
region where some charge carriers are lost due to trapping and recombination, leading to an underestimation
of the photon energy. The occurrence of ICC is highly influenced by the energy of the incident photons and
their respective absorption depths within the silicon sensor. In this regard, an analysis was conducted to
examine how different energy X-rays and their corresponding absorption depths influence the occurrence of
ICC.
4.2.1 Absorption Depth

The absorption depth for different energy X-rays was analyzed. 50,000 X-rays with energy Eγ are fired
at a cube of silicon with sides of 10 mm. The photons were fired using an infinitely thin beam aimed at
the middle of the block, at coordinates X = Y = 0. The characteristic Kα X-rays for aluminum, copper,
and cadmium were analyzed. Figure 4.7 shows absorption profiles for three different X-rays energies. Two
important observations can be made from the absorption profiles.
Firstly, when a higher energy X-ray enters the silicon block, the spread of the energy depositions in the X-
and Y-direction is greatly increased. Higher-energy X-rays introduce more energy into the system, allowing
for more cascading interactions and a higher number of energy deposits Ndep to occur. As more interactions
occur within the sensor, the interaction volume expands radially outward from the point of incidence.
Secondly, as the incident photon energy increases, the penetration or absorption depth also increases
significantly. For instance, in the case of an Al-Kα beam, the deposited photons can only penetrate the
silicon to a maximum depth of 100 microns. On the other hand, Cd-Kα X-rays are capable of penetrating
the entire 10 millimeters of the silicon block. Moreover, 0.4% of X-rays were able to completely traverse the
silicon block without interacting with it.
52
(a) Aluminum Kα1 , Eγ = 1.486 keV , Ndep = 202, 562 (b) Copper Kα1 , Eγ = 8.048 keV , Ndep = 386, 523
(c) Cadmium Kα1 , Eγ = 23.174 keV , Ndep = 827, 960
Figure 4.7: Absorption profiles for a beam of characteristic X-rays in a 10 mm thick block of silicon.
To determine the absorption depth of photons at a specific energy, a set of simulations was performed.
A monochromatic photon beam consisting of 50,000 photons was directed at the center of a silicon cube
with an edge length of 10 cm. The incident photon energy ranged from 300 eV to 22.5 keV. The resulting
positions of energy deposits were analyzed, and the average penetration depth within the silicon block was
calculated. The obtained average absorption length was compared to the linear attenuation coefficient for
silicon obtained from the NIST X-ray Attenuation Database [68], as depicted in Figure 4.8.
The simulated average absorption closely aligns with the NIST database within the energy range relevant
to EDS. Notably, there is a sharp rise in the attenuation coefficient of silicon around 1.8 keV, corresponding
to the energy threshold for photoelectric absorption with silicon’s K-shell electrons, which possess a binding
energy of 1.839 keV.
53
Figure 4.8: Average absorption depth in silicon for different incident photon energies.
4.2.2 Detector Response Modelling

When a flux of mono-energetic X-ray photons interacts within an SDD, only a fraction of the detected counts
is recorded in the channel corresponding to the incident photon energy. The majority of counts are recorded
in adjacent channels. This resulting spectrum is referred to as the detector response to the incoming mono-
energetic photon flux. Understanding the detector response is essential for accurate energy calibration, peak
identification, and quantitative analysis of the EDS spectra.
During the project, it was not possible to obtain experimental results for the monochromatic detector
response spectra for an SDD. Therefore, the response spectra from Scholze et al. [9] were used as a reference.
Scholze provides the detector response curve for an SDD using monochromatized radiation of 2.622 keV
(Cl-Kα1 ). The front layer consisted of an Al/SiO2 front layer, the thickness of which was determined from
efficiency measurements to be 150 nm. The spectrum was accumulated with an input count rate of 5 · 4 s−1
to a total of 5 · 106 pulses.
The experimental setup of Scholze was recreated. A 2.622 keV photon beam was fired perpendicular to
the detector surface, centered at the middle of the sensor. The beam was composed of a uniformly distributed
photon flux across the area of a circle with a radius of 2 mm. The detector was protected by a 12.5 µm
beryllium window. Furthermore, a 500 µm zirconium collimator was used, to prevent photons from reaching
the outer edges of the silicon bulk. The front contact of the model consisted of two layers, an aluminum and
SiO2 layer, which was placed in front of the active detector volume. For the SiO2 layer, a thickness of 100
nm was used, while the aluminum layer had a thickness of 50 nm. The geometry of the detector under the
bombardment of photons is shown in figure 4.9.
54
Figure 4.9: The SDD geometry under the irradiation of a 2.622 keV photon beam, in green. The beryllium
window is highlighted in blue, the zirconium collimator window is shown in red. The sensor of the detector
is enclosed within a steel housing, represented in grey.
A total of 2 million events were simulated. The propagation was performed under a temperature of
-50 °C, using a total integration time of 2 us, which was ample time for charges to completely propagate
throughout the silicon bulk towards the anode. For the front contact ICC model, Scholze’s model was used,
as explained in section 2.3.5. The parameters used in the charge collection efficiency curve were set to be
the same as reported by Scholze’s SDD model, namely,
z0 = 50 nm (4.1)
a = 0.5 (4.2)
C0 = 0.9 (4.3)
The total energy per event was obtained from the PixelCharge object, by summing up the total induced
charge contribution on the charge. The comparison between the pulse probability of the simulated device,
and the model prediction and measurement by Scholze is shown in figure 4.10.
Figure 4.10: Response spectra for an SDD irradiated by monochromatic 2.622 keV photons. The measured
and modeled data by Scholze [9] are in red and black, respectively, and the simulated data is in blue.
The main peak at 2.622 keV and the escape peak at 882 eV are well represented in the simulation.
Furthermore, the characteristic peaks resulting from interactions of the incident photon in the front contact
55
are visible, at 1.486 keV for Al-Kα and at 1.739 keV for Si-Kα . The shelves are arguably better represented
in the simulated model compared to Scholze’s predictive model, which slightly over-represents the shelf
contribution.
The experimental data shows a significant dip below the 100 eV region, which can be attributed to
electronic noise. Furthermore, energies above 2.7 keV are measured in the experimental data. These are the
result of pile-up events, where the charge from multiple photon events is detected as a single event, leading
to the sum of both events to be measured. As in our simulated model each emitted photon event is treated
independently in the simulation, no pile-up effects can occur, resulting in the absence of this behavior in the
simulated spectrum.
The simulated data is noisier than the experimental data. The reason for this is twofold. Firstly, some
noise was lost during the digitization process of the experimental data for the figure. Secondly, the total
number of events in the simulation was considerably lower than the number of pulses obtained by Scholze.
Increasing the number of events in the simulation to match the total number of pulses in the experimental
data is expected to yield noise levels more similar to the experimental data.
4.2.3 ICC Control Parameters

The ICC region is modeled using the charge collection efficiency (CCE) curve based on the model by Scholtze
et al, as described in section 2.3.5. By tuning the parameters of the CCE curve, the amount of tailing as a
result of the front contact ICC can be modified.
For example, by adjusting the parameter z0 , the depth of the dead layer can be controlled. When z0 is
increased, the total number of energy depositions that occur within the dead layer is increased. Therefore,
the probability of more charge groups being incompletely collected is increased. This effect can clearly be
seen in figure 4.11, where z0 is increased from 0 nm to 75 nm. The other variables in the simulation were
kept identical.
Figure 4.11: Response spectra for an SDD irradiated by monochromatic 2.622 keV photons for increasing
the dead layer thicknesses z0 .
Increasing the value of z0 primarily impacts the occurrence of ICC and, consequently, the extent of
lower-energy tailing in the spectrum. The main peak exhibits a noticeable change in its lower-energy tailing.
When z0 is set to zero, representing an SDD without front contact ICC, the full-energy peak intersects the
shelf without any lower-energy tailing. Moreover, a slight variation in tailing can also be observed at the
escape peak.
By utilizing this model for the front contact ICC, a detector response curve of a specific detector can be
56
4.3. COMPARISON EXPERIMENTAL XRF DATA CHAPTER 4. RESULTS
modeled. The parameters used to obtain the charge collection efficiency (CCE) curve allow control over the
amount of lower-energy tailing observed in the detector’s response.
This could also prove useful for modeling older detectors that may have accumulated radiation damage
in the silicon lattice of the front layer. These damage sites can act as trapping centers. Drifting charge
carriers may get trapped in the damage site and be lost for read-out, resulting in ICC. As a result, the
charge efficiency in the front layer is decreased. The framework could therefore also also model radiation
damage in the front layer. Given the scope of this thesis, this idea was not further explored.
4.3 Comparison Experimental XRF Data

To investigate the performance of the SDD simulation in a complex setup, the simulation results were
compared to an in-house X-ray fluorescence (XRF) setup. The experimental setup named XRF-Tool was
made available by Sioux Technologies. The XRF technique is very similar to SEM-EDS. The difference
arises from a difference in the excitation method, SEM-EDS uses electrons to excite the sample, while in
XRF X-rays are used to obtain elemental spectra of the sample. The key components of the setup are as
follows,
• Source - The X-ray beam responsible for exciting the sample.
• Sample - The target sample exposed to the X-rays from the source.
• Detector - The SDD used to measure the emitted X-rays from the sample.
• Housing - The thick protective housing enclosing the setup.
4.3.1 Setup of the XRF-Tool

The XRF-Tool utilizes a Mini-X2 X-Ray Tube by Amptek to produce X-rays. The Mini-X2 is aimed directly
at a cylindrical nickel sample positioned at a 45-degree angle relative to both the detector and the Mini-X2,
in order to optimize the number of X-rays reaching the detector. The sample is attached to the protective
housing by means of a copper sample holder, which keeps the sample at the same height as the SDD. To
shield the copper sample holder from X-rays, the sample is covered with a conical aluminum cover. The
detector is positioned at a distance of 10 mm from the sample and consists of a 12.5 µm beryllium window,
and a 500 µm zirconium collimator. The entire setup is contained in a thick 15 mm stainless steel housing.
The structure through which the detector is attached to the housing is neglected in the simulation setup.
The Geant4 model used during the simulation is shown in figure 4.12.
57
Figure 4.12: GDML model of the XRF-Tool.
4.3.2 Materials
The XRF-tool contains multiple different materials, for each of the separate parts in the setup. The specific
material codes for these are shown in table 4.1,
Material Material code Function

Nickel 2.4068 - N02201 Sample
Aluminum EN AW-6082 T6 - 3.2315.72 Sample cover
Copper C10100-H2 Sample holder
Stainless steel AISI 303 Protective housing
Table 4.1: Materials used in the XRF-Tool.
The tabulated materials are not completely pure and a number of different trace elements can be found
in each. These trace elements are in high enough abundance that they can appear as small peaks on the
output spectra. The weight percentage for each trace element in the chemical composition varies for each
batch of the manufacturer, therefore upper limits are provided in the datasheet for the materials. As the
specific composition of the materials and the amount of trace elements were unknown in the experimental
setup, these provided upper limits were used in the material definitions for Geant4. Table 4.2 shows the
exact chemical compositions of each material used by Geant4 during the simulations. As the copper sample
holder was completely out of the line of sight for the X-ray source and because copper C10100-H2 is already
at 99.99% purity, 100% pure copper was used for this material.
58
Nickel Aluminum Stainless steel

Ni 98.775 Al 95.35 Fe 67.50
C 0.02 Si 1.30 Ni 10.00
Fe 0.40 Fe 0.50 Cr 19.00
Mn 0.30 Cu 0.10 C 0.150
Si 0.10 Mn 1.0 Mn 2.000
Cu 0.25 Mg 1.2 P 0.200
Mg 0.05 Cr 0.25 S 0.150
Ti 0.10 Zn 0.20 Si 1.000
S 0.005 Ti 0.10
Table 4.2: Chemical composition in weight percentage for the materials of the XRF-Tool used in Geant4.
4.3.3 Source
The X-ray source was an Amptek Mini-X2 X-ray tube with a gold anode. The output spectrum of the
Mini-X2 consists of two components, a bremsstrahlung continuum, and characteristic X-rays originating
from the gold anode. The output spectra of the Mini-X2 at multiple tube voltages are shown in figure 4.13a,
as retrieved from the manufacturer documentation [69].
(a) Output spectrum at different voltages for a gold target. (b) X-ray cone angle.
Figure 4.13: The output spectrum and cone angle of the Amptek Mini-X2 X-ray tube.
In the experimental setup, the Mini-X2 operated at 10kV, predominantly emitting X-rays through
bremsstrahlung radiation. The Mini-X2 was collimated using a 2 mm brass collimator with an aluminum
insert. This collimation reduced the X-ray cone angle from 120° degrees to 5° degrees, as depicted in Figure
4.13b.
The Mini-X2 source was simulated in Geant4 using GPS macros, employing a circular planar source
with a user-defined energy distribution extracted from figure 4.13a. The 5-degree cone angle resulted in
some X-rays landing on the aluminum sample cover, as shown in the simulation render of figure 4.14. As a
result, many characteristic X-rays of aluminum are emitted, some of which find their way into the detector.
Therefore, the output spectrum will exhibit a noticeable peak corresponding to aluminum.
59
Figure 4.14: Simulation of the XRF-Tool under irradiation of 500 X-rays. The X-rays are in green and
electrons are in red.
The simulation was run for 109 events, for each of which a single X-ray is emitted from the Mini-X2.
However, only 0.03% events (3 million) resulted in energy being deposited within the active volume of the
detector. In the remaining events, all energy is deposited in either the sample structure or the protective
housing.
4.3.4 Spectrum Comparison

The experimental result was obtained for an acquisition time of 60 seconds and a shaper time of 10µs,
resulting in a total of 1.3 · 106 counts. The counts were binned in an energy histogram with a bin size of 5
eV. The comparison between the simulation and experimental spectrum for the XRF-Tool is shown in figure
4.15.
Figure 4.15: Comparison of the output spectra for the XRF-Tool for a nickel sample.
60
4.4. BIAS TUNING OF SDD CHAPTER 4. RESULTS
The simulation successfully captures the full-energy nickel Kα and Kβ peaks coming from the sample. An
aluminum Kα peak originating from the aluminum sample cover is well represented. Moreover, additional
background peaks originating from the stainless steel housing and trace impurities such as Cr, Mn, and Fe
are also represented. The background bremsstrahlung curve resulting from the Mini-X2 source is visible in
the energy range of 5 keV to 10 keV. Furthermore, the escape peaks for nickel are not visible, as they are
obscured below the noise floor originating from the bremsstrahlung curve.
The exact heights of the elemental peaks depend strongly on the specific chemical compositions of the
materials, which were unknown during the simulation. As a result, some discrepancies in exact peak heights
are observed. For example, the simulated peak height of aluminum is slightly lower than the experimentally
obtained result. This could be caused by the fact that the aluminum ring was purer than the chemical
composition used in the simulations. This can be attributed to several factors, including discrepancies in
the modeled geometry and the Mini-X2 spectrum. Differences in geometry, such as steeper conal angles,
affect the number of X-rays entering the detector volume, while variations in the Mini-X2 energy spectrum
influence the emission of X-rays within the aluminum ring. Matching the chemical compositions in the
simulation with those in the physical setup is expected to minimize these differences in peak ratios.
Additionally, a hump between 4 keV and 5 keV is observed in the experimental results, likely caused
by the Mini-X2 source. Due to the unavailability of an accurate measurement of the output spectrum from
the Mini-X2, the provided spectrum by the manufacturer in figure 4.13a was used to configure the Geant4
source. The spectrum provided by the manufacturer for the Mini-X2 source is coarse, especially for the
case at 10 kV. As a result, there are uncertainties regarding the exact output spectrum of the Mini-X2.
It is expected that having precise knowledge of the Mini-X2 spectrum would improve the alignment of the
background curve in the simulation.
In conclusion, the simulation of the XRF-Tool successfully captured the main nickel peaks and detected
an aluminum peak from the sample cover. The background peaks from the housing and trace impurities
were also observed. However, discrepancies in peak heights and background radiation were noticed and are
expected to be the result of inaccurate modeling of unknown parameters, such as material properties and
Mini-X2 output spectra.
4.4 Bias Tuning of SDD

Bias tuning is performed on SDDs to optimize their performance and enhance their sensitivity and resolution.
The bias voltage creates a strong electric field within the detector, which allows for efficient collection of the
charge generated by the incident X-rays. By adjusting the parameters of the bias voltages, it is possible to
control the electric field strength and shape within the detector. By tuning of the bias voltages, the drift
path towards the anode can be optimized to obtain the highest resolution. Experimental measurements
showed that by altering a bias parameter from its optimal value, the FWHM of the peaks would increase,
and in the worst case a complete breakdown of the signal would occur.
The experimental results showed that the signal from the SDD was sensitive to variations in the BC
voltage. When the BC voltage was increased beyond its optimal value, the spectra exhibited additional
broad peaks at lower energies, commonly referred to as camel humps. The previous section described the
performance of the simulations in the case of an optimally tuned SDD. The goal of the current section was
to assess whether the simulation framework could also accurately reproduce the behavior of an improperly
tuned SDD.
Experimental measurements were conducted using a Fe-55 source. The BC voltage was increased from
-110V up to -150V in increments of 5V. The spectra had an acquisition time of 60 seconds each, using a
shaper time of 10 µs. The remaining bias voltages were maintained at their optimal values, as outlined in
table 3.1.
The simulation setup was modeled in a similar fashion as the XRF-Tool described in section 4.3. In this
case, however, The nickel sample, as well as the copper sample holder and the aluminum rings were replaced
by a Fe-55 decay source. The bias voltages in the COMSOL model were adjusted to match the experimental
bias voltages.
In the case where BC = −110 V , a clean spectrum is obtained, as depicted in figure 4.16. The main
two peaks from the Fe-55 source pertain to the Mn-Kα and Mn-Kβ characteristic X-rays. These peaks, as
61
well as their respective escape peaks, are well-represented by the simulation. The measured spectrum shows
two additional peaks. These were identified to be chlorine (2.6 keV) and calcium (3.7 keV). It is expected
that these originate from the housing that the Fe-55 source resides in. The Fe-55 housing is made from
PVC, which contains chlorine. Calcium carbonate is often added to PVC to increase its strength, which
could explain the calcium peak. The simulated model does not model the plastic housing, which explains
the absence of these two peaks in the simulation.
(a) Potential field and electric drift lines, using BC = −110 V.
(b) Simulated and measured Fe-55 spectra.
Figure 4.16: Simulated results of a Fe-55 source for an SDD with BC = −110 V.
The discussion begins by examining the impact that adjusting the BC voltage has on the simulated
electric fields. Subsequently, the effects on the measured spectra are addressed. Finally, a comparison is
made between the simulated spectra and the experimental measurements.
4.4.1 Effects of BC Voltage on Electric Fields

The potential fields and electric drift lines for the range of BC voltages are shown in appendix A. Increasing
the BC parameter causes the main potential valley to shift closer to the top of the device. When the potential
valley is pushed too high, the drift lines do not end their trajectory at the anode. This has three possible
outcomes depending on the extent of the BC voltage increase.
Firstly, if the BC is only turned slightly above the optimal result, the majority of the electrons are still
able to induce a full charge on the anode. However, as the electrons drift towards the anode, the spread of the
charge cloud grows larger due to thermal diffusion. The charges that propagate near the edge of the charge
cloud could end up closer to the R1 drift ring than the anode. As a result, they do not end their trajectory
in the region where the weighting field is equal to 1 and only induce part of their charge. Therefore, these
charges do not fully contribute to the induced charge, leading to some lower energy tailing caused by ICC.
62
Secondly, when the BC voltage is further increased, the drift lines end between the anode and the first
drift rings. Depending on an electron’s position within the charge cloud, it may either induce its full charge
or none at all. This phenomenon contributes to the observed shifted peaks known as camel humps, caused
by ICC and the mismatched drift lines. Higher BC voltages lead to more charges ending their trajectory in
the ICC region, causing the camel humps to shift to lower energies.
Lastly, if the BC voltage is raised excessively, the drift lines completely miss the anode, and nearly all
electrons end up in the region with a zero weighting field, resulting in little to no measured charge. This
scenario can be seen in figure 4.17b.
(a) BC = −110 V
(b) BC = −150 V
Figure 4.17: Potential field and electric drift lines of the SDD for two different BC voltages. In (a) the drift
lines reach the anode. In (b) the BC voltage is too high and the drift lines cannot reach the anode.
4.4.2 Effects of BC Voltage on Spectra

The EDS spectra of the resulting BC sweep can be found in appendix B. The spectra are normalized with
respect to the case where BC = −110 V. In this manner, the change in peak height could also be analyzed.
The experimental results demonstrate that a spectrum breakdown occurs when the BC potential reaches
-120 V. At this voltage, an extra peak emerges below the full-energy Mn-Kα peak. This additional peak
arises due to the ICC of charge clouds belonging to fully absorbed Mn-Kα or Mn-Kβ events. As the BC
potential is further increased, the additional peak gradually shifts towards lower energies, leaving behind an
ICC plateau between the peak and the full energy peak.
When the BC potential exceeds -130 V, also the full energy Mn-Kα and Mn-Kβ peaks will significantly
shift towards the lower energies. For BC voltages above -135 V, the full-energy peaks have completely shifted,
leading to only a minimal number of full-energy events being detected.
In figure 4.18 the experimental and simulated spectra are compared for a BC voltage of -115 V and -120
V. It was found that the simulated device was much more sensitive to the increase in BC voltage than the
experimental setup. In the simulated spectra, an increase from -110 V to -115 V resulted in a significant
occurrence of ICC. Even though the potential valley in the case of BC = −115 V aligns correctly with the
anode. It is expected that the cloud of charge carriers has significantly grown in size due to diffusion and
that a portion of the charges does not end up at the anode. These charges instead recombine near the drift
rings, inducing their charge only partly.
In the case of -120 V, the drift lines are sufficiently shifted towards the top of the device, such that the
majority of charges are not able to reach the anode. Instead, they land in a broad region where the weighting
field is between zero and one, causing partial induction on the anode. This is characterized by a broadening
and shifting of the full-energy Mn-Kα and Mn-Kβ peaks.
63
4.5. SPEED-ACCURACY CONSIDERATIONS CHAPTER 4. RESULTS
(a) BC = −115 V (b) BC = −120 V
Figure 4.18: Simulated and experimental results of a Fe-55 source for an SDD with different BC voltages.
When increasing the BC further to BC = −125 V, most charge carriers follow the potential valley land
in the region where the weighting field is zero and thus induce no charge. Two full absorption peaks remain
visible at the characteristic Mn-Kα and Mn-Kβ energies. These are the result of photon events that are
absorbed in the very center of the device (r = 0). These charges propagate upwards towards the anode and
are repelled by the IGR, causing them to still be completely collected.
Above BC = −125 V, the simulated spectra do not change significantly, as the electric drift lines lie too
far away from the anode to induce any charge on the anode.
In conclusion, bias tuning is crucial for optimizing the performance of SDDs. Experimental measurements
showed that the SDD’s response was sensitive to variations in the BC voltage. Increasing the BC voltage
beyond its optimal value resulted in additional broad peaks at lower energies, due to ICC. The simulations
accurately represented the behavior of an optimally tuned SDD. However, under sub-optimal biasing condi-
tions, the simulation’s accuracy depended on precise knowledge of the SDD’s internal geometry, which was
unknown as this information was not provided by the manufacturer. Understanding the internal structure
of the SDD proved to be essential for modeling the detector’s behavior in non-optimal biasing scenarios.
4.5 Speed-Accuracy Considerations

Allpix Squared provides several parameters that can be adjusted to improve simulation times, at the expense
of a loss in accuracy. As the majority of simulation time is spent in the propagation of charges, two
parameters in the GenericPropagation module were investigated in their ability to speed up the simulation
times. Specifically, the investigation focused on the grouping of charge carriers and the spatial precision of
the integration steps. This section examines the trade-off between accuracy losses and computational gains
associated with these parameters and offers guidance on their optimal usage.
4.5.1 Charge Grouping

Allpix Squared allows charges to be propagated together in groups. These groups are independent from
each other and no interaction between groups is modeled. The charge per step parameter sets the maximum
group size of the propagated charges. The total number of created charge carriers at a specific energy deposit
is divided into groups of this size and a group with the remaining charge carriers.
For example, consider the case that an energy deposit contains the full absorption of an Mn-Kα X-ray.
On average, 1621 electron-hole pairs will be generated. By setting the charge per step value to 1, all charges
will be independently propagated, and thus a total of 1621 propagation will have to be simulated for each
event. However, by setting it to 500, only a total of four groups will have to be propagated; three groups
64
containing 500 charges and one containing the remaining 121 charges. As a consequence, the latter scenario
will have a significantly shorter computation time.
The accuracy loss and computational speed up for different values of charge per step were analyzed for
the same setup described in the previous section, using a BC = −110 V. The simulation contains 50 million
events from a Fe-55 source, which were loaded from an external simulation using the ROOTObjectReader.
Figure 4.19 shows the computation times for each separate module for a range of simulations where the
charge per step parameter was varied between 500 and 1. These results highlight that the GenericPropa-
gation module dominates the total computational time, particularly when a larger number of charges are
propagated. As the number of propagated charges increases, the time spent in the GenericPropagation
module significantly rises.
Figure 4.19: Computation time for the Allpix Squared modules for 50 million events of a Fe-55 source for
different charge grouping sizes.
Furthermore, an increase in computation time can be seen in the induced transfer module. This can be
explained by the fact that the total number of propagated charges is increased. Therefore, more induction
calculations have to be performed to find the total induced charge sum on the anode.
Interestingly, by increasing the number of charge groups, the computation time in the ROOTObjec-
tReader decreases. It is expected that this is a consequence of how multithreading is implemented in Allpix
Squared. The thread pool in Allpix Squared consists of two separate queues. The first queue is an unsorted
queue that follows a first-in-first-out order, and it is constantly filled with new events to be processed by the
main thread. The second queue is a priority-ordered queue designed to temporarily buffer events that have
to wait to be handled in a certain sequence.
The ROOTObjectReader module requires the ROOT file to be read in a sequential order and thus does
not allow random read events. If a read event is not ready yet, all subsequent events are buffered until
the event is done and the sequence can continue. Since a significant portion of the 50 million events may
not result in deposited charges, the ROOTObjectReader module needs to read and skip over many events.
If the computational time of the remaining modules is relatively low, such as for charge per step = 500,
the event buffer will gradually fill up. When the number of buffered events reaches the maximum allowed
value, the simulation will be slightly bottle-necked by the reading speed, and thus relatively more time will
be spent in the ROOTObjectReader module. However, if the remaining modules have significantly longer
computational times, this effect is not observed since the total number of buffered events does not reach its
maximum capacity.
The time spent in the GenericPropagation module is shown in figure 4.20 on a log-log plot. The com-
putation time roughly follows a power law, and as a result, shows as a straight line in the log-log plot. The
result was linearly fitted using the least-squares method, and a slope or exponential of 0.86 was found.
65
Figure 4.20: Computation time of the GenericPropagation module for different charge per step values. The
computation time closely follows the relation y = 103.31 · x0.86 .
Appendix C shows the complete set of spectra obtained for all different values of the charge per step
parameter. The main peaks, such as the Kα , Kβ and escape peak appear to be well-represented irrespective
of the value used. For values below 25, the shelf between the escape peak and the main peak is well-
represented. For values above 25, however, the shelves lose accuracy significantly. Instead of appearing as a
straight shelf, spurious peaks are introduced into the spectrum.
At lower energies, especially below 2 keV, narrow periodic artificial peaks are observed. These peaks
lie at multiples of the charge per step times the average charge creation, which was set to 3.64 eV in this
case. They reduce in magnitude for increasing energy and disappear into the shelf completely at energies
above 3 keV. This effect is caused by the fact that most charge deposits will be split into groups which
have a size that is a multiple of charge per step times the average charge creation. Therefore, the energy
values related to these group sizes have a much higher chance of being measured, than the energy levels in
between. Especially in the case of lower energies, where only a limited number of charge groups are created,
this effect becomes especially apparent. This effect appears to be an inherent byproduct of charge grouping
and therefore only completely disappears in the case of a charge per step value of 1.
Concluding, if an accurate representation of the shelves is required, but the lower energy region is not of
importance, a charge per step value less than 25 should be used during the simulations. This choice provides
a speedup of approximately 14 times compared to propagating each charge individually. Furthermore, if the
lower energy region is of importance, a charge per step value of 1 ought to be used in order to not obtain
any artificial periodic lower energy peaks.
4.5.2 Spatial Precision

The second parameter which was analyzed was the spatial precision of the integration. This parameter
describes the spatial precision that the Runge-Kutta integrator will aim for. The time step of the Runge-
Kutta propagation is adjusted to achieve the desired spatial precision, taking into account the uncertainty
calculated from the fifth-order error method. As a result, the maximum allowed time step was increased to
1 ms, such that this did not pose any limitations for the maximum allowed step size.
Figure 4.21 shows the computation times for each separate module for a range of simulations where the
spatial precision parameter was varied between 5000 nm and 1 nm. Each charge was individually propa-
gated for all simulations. Again, the GenericPropagation module dominates the total computational time
throughout all simulations. As the spatial precision is increased, the time spent in the GenericPropagation
module rises, resulting in a significant increase in the total simulation time.
66
Figure 4.21: Computation time for the Allpix Squared modules for 50 million events of a Fe-55 source under
different spatial precision values.
The time spent in the ROOTObjectReader module decreases for increased spatial precision, expectedly,
for the same reason as explained in previous section. The computation times for the other modules were not
significantly affected by changing the spatial precision.
The time spent in the GenericPropagation module is shown in figure 4.22 on a log-log plot. The com-
putation time roughly follows a second order polynomial in the log-log plot. The computational time was
2
fitted using the least-squares method to the function 10(Ax −Bx+C) , with the following constants,
A ≈ 0.06211 (4.4)
B ≈ −0.4971 (4.5)
C ≈ 6.02675 (4.6)
Figure 4.22: Computation time of the GenericPropagation module for different spatial precision values. The
2
computation time closely follows the relation y = 10(0.062x −0.497x+6.027) .
Appendix D shows the complete set of spectra obtained for all different values of the spatial precision. For
spatial precision values below 5 nm, the spectrum retains it shape, broadly speaking. The full absorption
peaks for Mn-Kα and Mn-Kβ exhibit a slight peak shift to lower energies. As the spatial precision is
67
increased, the peak shift increases as well. For spatial precision values above 100 nm, a complete breakdown
of the spectrum occurs.
Figure 4.23 displays the shift of the Mn-Kα peak for various spatial precision values. The peak shift was
determined by extracting the center of a Gaussian fit applied to the Mn-Kα peak. Results corresponding
to spatial precision values exceeding 100 nm were excluded from the analysis as they no longer accurately
represented a Gaussian shape.
Figure 4.23: Peak shift of Mn-Kα peak for different spatial precision values. The peak shift closely follows
the relation y = 100.59 · x0.93 .
In conclusion, to maintain an accurate spectrum without significant distortion, it is recommended to use

a spatial precision of up to 10 nm. Below this value, the spatial precision can be adjusted based on the
acceptable error margins for the peak shifts. For example, if a maximum peak shift of 10 eV is desired, the
spatial precision should be kept below 2.5 nm.
68
Chapter 5
Future Research Recommendations
This section highlights the limitations of the current model and suggests steps for improvement.
5.1 Inaccurate SDD Geometry

The lack of accurate SDD geometry information posed a significant challenge in the thesis project. The
manufacturer provided only limited specifications, such as the thickness of the sensor and active area, without
any details on the concentration of impurities within the silicon. It was found that some of these parameters
play a crucial role in determining the electric field drift lines within the SDD, and as a consequence play a
major role in accurately generating EDS spectra.
To address this issue, alternative approaches could be considered. One possibility is to collaborate with
the manufacturer or research institutions that specialize in SDD fabrication to obtain more comprehensive
and accurate information about the device’s geometry. This collaboration could involve conducting joint
experiments or accessing proprietary data that would allow for a more precise characterization of the SDD
structure.
Another approach could be to explore reverse engineering techniques. For instance, by analyzing photos
of the SDD, information about the number and sizes of the drift rings could be obtained. Furthermore, by
analyzing the SDD’s response to different inputs and comparing it with simulated data, it might be possible
to infer certain aspects of the device, such as the doping concentration.
Overall, the lack of detailed information regarding the SDD’s inner geometry presents a significant chal-
lenge in accurately modeling its electric field. Addressing this limitation is crucial for improving the reliability
and accuracy of the simulation.
5.2 Fano factor and charge creation

The charge carriers were generated using the energy deposits in the detector from Geant4, based on the
Fano factor F and average charge creation energy ω. Research suggests there is both a temperature and an
energy dependence for these values [44, 45].
During the simulations, the default constant values for F and ω were used. Currently, Allpix Squared
does not support the inclusion of energy and temperature dependence for these parameters. As they have a
significant effect on the FWHM of the measured peaks, improving the model by incorporating energy and
temperature dependence might prove necessary.
5.3 Front Contact Model

The current front contact model relies on an empirical approach based on Scholze’s ICC model [9]. This
approach utilizes empirical parameters to determine the charge collection efficiency of charges near the front
69
5.4. OPEN-SOURCE PIPELINE CHAPTER 5. FUTURE RESEARCH RECOMMENDATIONS
contact layer. However, to achieve more robust and accurate modeling of energy dispersive spectroscopy
(EDS), it is essential to explore alternative approaches.
A more robust model would be to apply Monte Carlo simulation based on physical interaction within in
the front contact. The dead layer is a result of charge carriers landing in the shallow BC implant [61], and
in the silicon bulk region which is not completely depleted. As the supplied electric field was assumed to
be fully depleted, this phenomenon could not be modeled. Incorporating a more detailed electric field and
doping field modeling could address these limitations.
To achieve this, a nonregular implementation of the electric field might be necessary. Currently, Allpix
Squared only supports regularly spaced electric fields. The depth of the dead layer is usually in the order
of several nanometers, which would require a very regular fine grid size for the electric field throughout the
entire detector bulk. This could be improved by modeling irregular grid sizes in Allpix Squared, which could
be refined such that there is higher precision near the front contact.
Improving the front contact model could better improve the accuracy of modeling EDS, especially in the
case of low-energy X-rays, where front contact becomes a considerable factor.
5.4 Open-Source pipeline

During the later stages of the thesis, advances were made in replacing the COMSOL implementation with
an open-source FEM solver. The reasons for this are twofold.
Firstly, this implementation allows the entire pipeline to be run without any external dependencies in
a single environment, such as a containerized application. This approach provides numerous benefits. By
utilizing a self-contained environment where all the required libraries, frameworks, and dependencies are
packaged together, conflicts and compatibility issues that may arise when running applications on different
systems are eliminated. Secondly, open-source alternatives provide transparent code, eliminate licensing
costs, and offer flexibility for customization.
To this end, a python package for solving partial differential equations called scikit-fem [70] was used. A
working principle for modeling of electric and weighting fields of the SDD was achieved, and a small dummy
experiment was developed. This served as a proof of concept, demonstrating the ability to use the scikit-fem
package to effectively simulate the desired fields. However, due to the time constraints of this thesis, this
pipeline was not further developed.
5.5 Digitization Effects

Digitization of the signal introduces additional artifacts in the measured energy spectrum. For instance,
electronics noise, such as noise dependent on the input capacitance and shaper time, decreases the resolution
of the detector setup.
Furthermore, if two X-rays get absorbed within a time frame shorter than the shaper time, the measured
energy will be counted as a single occurrence. This process depends on the pulse processor and specific
pile-up rejection algorithms.
The effects of digitization have not been included yet within the simulation framework. Allpix Squared
contains the DefaultDigitization module, which translates the collected charges at the anode into a digitized
signal. This module incorporates Gaussian noise from the readout electronics and supports a customizable
threshold. It also can also simulate the linear response of a Charge-to-Digital-Converter and Time-to-Digital-
Converter with adjustable resolution.
Allpix Squared currently does not include any modules for modeling pile-up effects, though this can be
done in offline analysis, [19]. These options were not further explored within this thesis, but could further
increase the accuracy of the simulated spectra.
5.6 Drift Ring Biasing

For generating the electric field, the bias voltage across the drift rings was assumed to be linear. This is
generally well within accordance in the case that a voltage divider is used between each drift ring, [71]. In
70
5.7. SPEED CONSIDERATIONS CHAPTER 5. FUTURE RESEARCH RECOMMENDATIONS
the ideal case of a silicon drift detector, where the hole current is zero, the voltage drop between any two
adjacent drift cathodes is constant. However, in real conditions, there is always some hole dark current that
enters the drift cathodes. As a result, the potential on the drift rings can exhibit a parabolic shape, instead
of a linear shape [72].
This deviation in the external biasing voltage leads to changes in the internal electric field, which, in
turn, affects the drift paths of the charge carriers within the detector. Therefore, accurately representing
the potential distribution across each drift ring becomes crucial.
To assess the actual voltage distribution, the potential across each drift ring can be measured using
a probe station [61]. By comparing the measured potential with the expected linear voltage drop, any
deviations or nonlinearity in the drift rings can be identified and included in the simulated model.
5.7 Speed Considerations

A couple of options for speed improvements for the Allpix Squared framework were examined, with a main
focus on improving the propagation integration of the charge carriers. Here, the energy deposits from an
external Geant4 simulation were loaded in to prevent unnecessary additional simulation time from rerunning
the energy deposit steps. However, the Monte Carlo portion of the pipeline in Geant4 also takes up a consid-
erable amount of time. For instance, the 109 events pertaining to the XRF-tool energy deposit simulations
took 29 hours to complete. Some thoughts on possible speed-ups are treated in this section.
5.7.1 Adjoint Monte Carlo

In a traditional forward MC simulation, particles are tracked from their source through the system until
they reach the detector. This can be computationally expensive, especially when the detector is small, as
most particles do not reach the detector. Therefore, their tracking and resulting computation time do not
contribute to the final result.
The adjoint MC method addresses this issue by starting the simulation from the detector and tracking
the particles in reverse towards the particle source, applying the reverse reaction processes [73]. As a result,
this technique is used primarily when the sensitive volume is small compared to the entire geometry of the
system of the external particle source.
Unfortunately, in the case of EDS applications, the main particle source will the be primary electron
beam, which has a diameter in the order of nanometers, [74]. As a result, it is expected that the reverse
Monte Carlo might not be an appropriate method in order to speed up the simulations. In fact, as the point
source in EDS is smaller than the detector area, it might perform worse than regular forward MC.
5.7.2 GPU Programming

GPUs have a high number of cores, which enables them to perform calculations at a significantly faster
rate than CPUs, particularly for problems that can be effectively parallelized. Geant4 involves running
straightforward calculations on a large number of particles. Importantly, each particle’s behavior is treated
independently of others. This type of problem is well-suited for parallelization and is expected to yield
substantial performance improvements when executed on GPUs.
Currently, Geant4 lacks native support for GPU acceleration in general use cases. However, there exist
several packages that incorporate GPU acceleration for lower energy physics simulations. GGEMS [75]
focuses on fast simulation platforms for imaging applications and particle therapy. While GGEMS supports
X-ray processes such as Compton scattering and the photoelectric effect, electron interactions are not yet
modeled but are planned for future implementation.
Another notable package, MPEXS is a CUDA-based application for medical physics using Geant4-derived
physics models, also demonstrated efficient use of GPU resources for regular voxelised geometries. MPEXS
demonstrates a significant enhancement in simulation time for standard electromagnetic physics, achieving a
250-fold speed increase compared to a single-core CPU without sacrificing accuracy. Additionally, MPEXS
employs the Livermore model to handle low-energy X-ray interactions down to 100 eV.
Two ongoing R&D efforts, namely Celeritas [76] and AdePT [77], focus on developing CUDA-based im-
plementations of electromagnetic physics on GPUs. The main objectives of these projects include addressing
71
5.7. SPEED CONSIDERATIONS CHAPTER 5. FUTURE RESEARCH RECOMMENDATIONS
significant performance bottlenecks of electromagnetic shower simulations, specifically aimed at simulations

for the Large Hadron Collider [78]. If these could be extended to the lower energy regime, this might prove
very useful for the speed-up of EDS simulations.
In conclusion, the use of GPUs for accelerating calculations in Geant4 shows great potential for improving
simulation times. While native GPU support is currently lacking in Geant4, several packages have demon-
strated successful GPU acceleration for specific physics simulations. It is expected that Future incorporation
of GPU acceleration could drastically improve the simulation times for simulating EDS in Geant4.
72
Chapter 6
Conclusion
This thesis focused on the modeling of silicon drift detectors (SDDs) in energy-dispersive X-ray spectroscopy
(EDS) and aimed to accurately simulate EDS spectra.
The proposed simulation framework combined the use of multiple software packages. Geant4 was used
for tracking electron and photon interactions, Allpix Squared for charge carrier propagation and induction,
and COMSOL for obtaining the electric field and weighting field for the SDD.
Through the modeling of the potential and electric field using COMSOL, it was found that the internal
geometry and doping concentration of the SDD had a significant impact on the potential field and resulting
electric drift lines, and consequently on the detector response. A strong interplay between the doping
concentration and bias voltages was observed. When the doping concentration is increased, the main drift line
of the silicon sensor volume shifts towards the center; conversely, reducing the doping concentration causes
the main drift line to shift towards the drift rings of the device. This behavior is similar to the observations
made when altering the bias voltages. Specifically, a high back contact (BC) voltage pushes the main drift
line towards the drift rings, whereas a lower BC voltage pulls it down towards the center. Consequently, it
is necessary to closely tune the biasing potentials according to the specific doping concentration present in
the silicon bulk.
Additionally, the weighting field was simulated in order to model Shockley-Ramo induction. The near-
field effect on charge collection on the anode was observed, causing only a small region near the anode to
contribute to the total charge induction. As a result, the vast majority of induced charge is the result of
electrons moving near the anode. Since few of the holes propagate within this near-field region, little to no
charge is induced by them on the anode, causing holes only contribute marginally to the output spectra.
Geant4 was employed to simulate electron and photon interactions within the setup materials. Geant4
demonstrates excellent accuracy in handling the depth of energy depositions, as evidenced by its close
agreement with the NIST X-ray Attenuation Database in terms of absorption length. Furthermore, it
is observed that increasing the incident X-ray energy results in a significant expansion of the interaction
volume within the silicon sensor, with Cd-Kα X-rays extending over several millimeters laterally.
Allpix Squared was used to convert the energy deposits from Geant4 into charge carriers, using the
average charge creation energy and Fano factor for silicon. These charge carriers are propagated through the
silicon using the Jacoboni mobility model, and charge is induced on the anode by Shockley-Ramo induction.
Multiple modifications were implemented in the Allpix Squared framework. To enhance the simulation’s
accuracy near the front contact of the SDD, a charge collection efficiency model was incorporated to account
for charge loss in the dead layer of the detector. Additionally, an issue regarding charge conversion was
resolved that caused unwanted peak shifts in the spectra.
The resulting EDS spectra were compared to detector response curves from literature and an experimental
in-house XRF setup provided by Sioux Technologies, both demonstrated close agreement in the case of an
SDD with well-tuned bias voltages.
Next, the influence of altering the bias potential, specifically the back contact (BC), on the output spectra
was considered in the case of irradiation by a Fe-55 source, and compared to an experimental setup. It was
found that the behavior of the spectra was very dependent on the internal geometry model of the SDD,
especially in the cases where a shift of drift lines caused most of the charge carriers to not get collected. As
73
CHAPTER 6. CONCLUSION
the internal geometry of the SDD was unknown, satisfactory agreement with the experimental setup was
only obtained in the case where the BC voltage was close to the optimal case. However, it is expected that
the framework could handle these situations correctly if the precise geometry and corresponding electric field
are known.
Furthermore, various computational speed-up options were investigated, and their impact on the resulting
spectra was analyzed. Specifically, the charge grouping and spatial precision of propagated charge carriers
was investigated. It was found that charges should not be grouped in sets larger than 25 in order to preserve
the shelves in the spectra. Furthermore, if accurate values in the lower energy regime below 1 keV are
required, charges should not be grouped to avoid artificial periodic peaks. Additionally, it is advised that a
spatial precision below 10 nm ought to be used to preserve the main peak position.
Overall, the proposed simulation chain provides a powerful tool that can be used to accurately model EDS
simulations in the case of SDDs. In this light, it has the potential to enhance the design and optimization
of SDDs, improve existing quantification models, and advance the understanding of SEM-EDS systems.
74
Chapter 7
Bibliography
[1] Azad Mohammed and Avin Abdullah. Scanning electron microscopy (sem): A review. In Proceedings of
the 2018 International Conference on Hydraulics and Pneumatics—HERVEX, Băile Govora, Romania,
volume 2018, pages 7–9, 2018.
[2] Karen D Vernon-Parry. Scanning electron microscopy: an introduction. III-Vs Review, 13(4):40–44,
2000.
[3] David B Williams, C Barry Carter, David B Williams, and C Barry Carter. The transmission electron
microscope. Springer, 1996.
[4] Barrie Glyn Lowe and Robert Anthony Sareen. Semiconductor X-ray detectors. CRC Press, 2013.
[5] L Strüder, P Lechner, and P Leutenegger. Silicon drift detector–the key to new experiments. The
Science of Nature, 85(11):539–543, 1998.
[6] Giuseppe Bertuccio, Mahdi Ahangarianabhari, Claudio Graziani, Daniele Macera, Yongbiao Shi,
A Rachevski, I Rashevskaya, A Vacchi, G Zampa, N Zampa, et al. A silicon drift detector-cmos front-
end system for high resolution x-ray spectroscopy up to room temperature. Journal of Instrumentation,
10(01):P01002, 2015.
[7] P Zambianchi and JK Zambianchi. Monte carlo simulation of micro-x-ray fluorescence (µ-xrf) based
detection of immunomagnetically labeled tumor cells. Physics in Medicine & Biology, 68(11):115017,
2023.
[8] Marco Feroci, Giovanni Ambrosi, Filippo Ambrosino, Matias Antonelli, Andrea Argan, Viktor Babinec,
Marco Barbera, Joerg Bayer, Pierluigi Bellutti, Bruna Bertucci, et al. The large area detector onboard
the extp mission. In Space Telescopes and Instrumentation 2022: Ultraviolet to Gamma Ray, volume
12181, pages 466–483. SPIE, 2022.
[9] F Scholze and M Procop. Modelling the response function of energy dispersive x-ray spectrometers with
silicon detectors. X-Ray Spectrometry: An International Journal, 38(4):312–321, 2009.
[10] Yu G Lavrent’Ev, VN Korolyuk, and LV Usova. Second generation of correction methods in electron
probe x-ray microanalysis: Approximation models for emission depth distribution functions. Journal of
Analytical Chemistry, 59(7):600–616, 2004.
[11] Jorge E Fernández, Viviana Scot, and Lorenzo Sabbatucci. A modeling tool for detector resolution and
incomplete charge collection. X-Ray Spectrometry, 44(3):177–182, 2015.
[12] Jorge E Fernandez and Viviana Scot. Simulation of the detector response function with the code
mcshape. Radiation Physics and Chemistry, 78(10):882–887, 2009.
[13] Viviana Scot and Jorge E Fernandez. The monte carlo code mcshape: Main features and recent devel-
opments. Spectrochimica Acta Part B: Atomic Spectroscopy, 108:53–60, 2015.
75
CHAPTER 7. BIBLIOGRAPHY
[14] G Calzolai, S Tapinassi, M Chiari, M Giannoni, S Nava, G Pazzi, and F Lucarelli. Silicon drift detector
response function for pixe spectra fitting. Nuclear Instruments and Methods in Physics Research Section
B: Beam Interactions with Materials and Atoms, 417:51–55, 2018.
[15] Siming Guo, Jinjie Wu, Jian Zhang, Zihao Fan, Weina Che, Mengshi Li, Dongjie Hou, and Ji Wang.
Study on silicon drift detector efficiency using monte–carlo simulation and experimental methods. The
Journal of Engineering, 2019(23):8543–8545, 2019.
[16] Matteo Biassoni, Matteo Gugiatti, Silvia Capelli, Marco Carminati, Oliviero Cremonesi, Carlo Fiorini,
Pietro King, Peter Lechner, Susanne Mertens, Lorenzo Pagnanini, et al. Electron spectrometry with
silicon drift detectors: a geant4 based method for detector response reconstruction. The European
Physical Journal Plus, 136:1–13, 2021.
[17] Matteo Gugiatti, Matteo Biassoni, Marco Carminati, Oliviero Cremonesi, Carlo Fiorini, Pietro King,
Peter Lechner, Susanne Mertens, Lorenzo Pagnanini, Maura Pavan, et al. Characterisation of a silicon
drift detector for high-resolution electron spectroscopy. Nuclear Instruments and Methods in Physics Re-
search Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 979:164474, 2020.
[18] LQ Qi, G Li, YP Xu, N Yu, W Li, YS Wang, H Wang, MY Ge, ZW Li, J Zhang, et al. Simulation
of the silicon drift detector for the spectroscopy focusing array onboard the extp. Nuclear Instruments
and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated
Equipment, 1025:166105, 2022.
[19] Tasneem Saleem, FJ Iguaz, and F Orsini. Allpix squared simulations of multi-element germanium
detectors for synchrotron applications. Journal of Instrumentation, 17(02):P02013, 2022.
[20] Manuele Bettelli, Nicola Sarzi Amadè, Davide Calestani, Bruno Garavelli, Pietro Pozzi, Daniele Macera,
Luca Zanotti, Carlo Andrea Gonano, Matthew C Veale, and Andrea Zappettini. A first principle method
to simulate the spectral response of cdznte-based x-and gamma-ray detectors. Nuclear Instruments
Equipment, 960:163663, 2020.
[21] Sea Agostinelli, John Allison, K al Amako, John Apostolakis, H Araujo, Pedro Arce, Makoto Asai,
D Axen, Swagato Banerjee, GJNI Barrand, et al. Geant4—a simulation toolkit. Nuclear instruments and
methods in physics research section A: Accelerators, Spectrometers, Detectors and Associated Equipment,
506(3):250–303, 2003.
[22] John Allison, Katsuya Amako, John Apostolakis, Pedro Arce, Makoto Asai, Tsukasa Aso, Enrico Bagli,
A Bagulya, S Banerjee, GJNI Barrand, et al. Recent developments in geant4. Nuclear instruments and
835:186–225, 2016.
[23] Simon Spannagel, Koen Wolters, Daniel Hynds, N Alipour Tehrani, Mathieu Benoit, Dominik
Dannheim, Neal Gauvin, Andreas Nürnberg, Paul Schütze, and Mateus Vicente. Allpix2: A modu-
lar simulation framework for silicon detectors. Nuclear Instruments and Methods in Physics Research
Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 901:164–172, 2018.
[24] William Shockley. Currents to conductors induced by a moving point charge. Journal of applied physics,
9(10):635–636, 1938.
[25] Simon Ramo. Currents induced by electron motion. Proceedings of the IRE, 27(9):584–585, 1939.
[26] COMSOL Multiphysics. Introduction to comsol multiphysics®. COMSOL Multiphysics, Burlington,
MA, accessed Feb, 9(2018):32, 1998.
[27] Alexander Barth. Magnetic and structural properties of the Gd/Ni-Bilayer-System. PhD thesis, Uni-
versität Konstanz, Konstanz, 2007.
76
[28] Norman Allen Dyson and Norman Allen Dyson. X-rays in Atomic and Nuclear Physics. Cambridge
university press, 1990.
[29] Peter J Goodhew and John Humphreys. Electron microscopy and analysis. CRC press, 2000.
[30] Glenn F Knoll. Radiation detection and measurement. John Wiley & Sons, 2010.
[31] Rita Joana da Cruz Roque. X-Ray Imaging Using 100 µm Thick Gas Electron Multipliers Operating in
Kr-CO 2 Mixtures. PhD thesis, Universidade de Coimbra (Portugal), 2019.
[32] Elena Aprile, Aleksey E Bolotnikov, Alexander I Bolozdynya, and Tadayoshi Doke. Noble gas detectors.
John Wiley & Sons, 2006.
[33] John H Hubbell and Stephen M Seltzer. Tables of x-ray mass attenuation coefficients and mass energy-
absorption coefficients 1 kev to 20 mev for elements z= 1 to 92 and 48 additional substances of dosi-
metric interest. Technical report, National Inst. of Standards and Technology-PL, Gaithersburg, MD
(United . . . , 1995.
[34] Robley Dunglison Evans and RD Evans. The atomic nucleus, volume 582. McGraw-Hill New York,
1955.
[35] Martin J Berger. Xcom: photon cross sections database.
http://physics.nist.gov/PhysRefData/Xcom/Text/XCOM.html, 8:3587, 1998.
[36] Simon M Sze, Yiming Li, and Kwok K Ng. Physics of semiconductor devices. John wiley & sons, 2021.
[37] M Ali Omar. Elementary solid state physics: principles and applications. Pearson Education India,
1975.
[38] Olaf Krasel. Charge collection in irradiated silicon detectors: a study of the operation conditions of
silicon sensors in the ATLAS pixel detector. PhD thesis, Dortmund, Technische Universität, Diss.,
2004, 2004.
[39] Vladyslav Solokha, Tien-Lin Lee, Axel Wilson, Kurt Hingerl, and Jörg Zegenhagen. Effective attenua-
tion length of kev photoelectrons in silicon measured by transmission through thin membranes. Journal
of Electron Spectroscopy and Related Phenomena, 225:28–35, 2018.
[40] Alan Owens and Anthony Peacock. Compound semiconductor radiation detectors. Nuclear Instruments
Equipment, 531(1-2):18–37, 2004.
[41] Helmuth Spieler. Semiconductor detectors. SLUO (SLAC User’s Organization) Lectureson Detector
Techniques, October, 23, 1998.
[42] Ugo Fano. Ionization yield of radiations. ii. the fluctuations of the number of ions. Physical Review,
72(1):26, 1947.
[43] Helmuth Spieler. Semiconductor detector systems, volume 12. Oxford university press, 2005.
[44] GW Fraser, AF Abbey, A Holland, K McCarthy, A Owens, and A Wells. The x-ray energy response of
silicon part a. theory. Nuclear Instruments and Methods in Physics Research Section A: Accelerators,
Spectrometers, Detectors and Associated Equipment, 350(1-2):368–378, 1994.
[45] MN Mazziotta. Electron–hole pair creation energy and fano factor temperature dependence in silicon.
Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors
and Associated Equipment, 584(2-3):436–439, 2008.
[46] Alexander Chao, Christian W Fabjan, and Frank Zimmermann. Particle Acceleration and Detection.
Springer, 2009.
77
[47] Canali Jacoboni, C Canali, G Ottaviani, and A Alberigi Quaranta. A review of some charge transport
properties of silicon. Solid-State Electronics, 20(2):77–89, 1977.
[48] ATLAS collaboration et al. Modeling the mobility and lorentz angle for the atlas pixel detector. Technical
report, ATL-INDET-PUB-2018-001, CERN, Geneva, Switzerland, 2018.
[49] Walter Blum, Werner Riegler, and Luigi Rolandi. Particle detection with drift chambers. Springer
Science & Business Media, 2008.
[50] F Scholze and G Ulm. Characterization of a windowless si (li) detector in the photon energy range 0.1–
5 kev. Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers,
Detectors and Associated Equipment, 339(1-2):49–54, 1994.
[51] Johann Plagnard, Christophe Bobin, and Marie-Christine Lépy. Accurate efficiency calibration of a
low-energy hpge detector using a monochromatic x-ray source. X-Ray Spectrometry: An International
Journal, 36(3):191–198, 2007.
[52] Josef Kemmer. Fabrication of low noise silicon radiation detectors by the planar process. Nuclear
Instruments and Methods, 169(3):499–502, 1980.
[53] J Kemmer, G Lutz, E Belau, U Prechtel, and W Welser. Low capacity drift diode. Nuclear Instruments
Equipment, 253(3):378–381, 1987.
[54] Emilio Gatti and Pavel Rehak. Semiconductor drift chamber—an application of a novel charge transport
scheme. Nuclear Instruments and Methods in Physics Research, 225(3):608–614, 1984.
[55] Pavel Rehak, Jack Walton, Emillo Gatti, Antonio Longoni, Marco Sanpietro, Josef Kemmer, Hans Dietl,
Peter Holl, Robert Klanner, Gerhard Lutz, et al. Progress in semiconductor drift detectors. Nuclear
Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and
Associated Equipment, 248(2-3):367–378, 1986.
[56] Catalina Curceanu, Aidin Amirkhani, Ata Baniahmad, Massimiliano Bazzi, Giovanni Bellotti, Carolina
Berucci, Damir Bosnar, Mario Bragadireanu, Michael Cargnelli, Alberto Clozza, et al. X-ray detectors
for kaonic atoms research at daϕne. Condensed Matter, 4(2):42, 2019.
[57] J Sempau, JM Fernandez-Varea, E Acosta, and F Salvat. Penelope—a code system for monte carlo
simulation of electron and photon transport. In Workshop Proceedings. Issy-les-Moulineaux, France,
pages 5–7, 2001.
[58] Makoto Asai, Miguel A Cortés-Giraldo, Vicent Giménez-Alventosa, Vicent Giménez Gómez, and
Francesc Salvat. The penelope physics models and transport mechanics. implementation into geant4.
Frontiers in Physics, page 660, 2021.
[59] J Sempau, JM Fernández-Varea, E Acosta, and F Salvat. Experimental benchmarks of the monte carlo
code penelope. Nuclear Instruments and Methods in Physics Research Section B: Beam Interactions
with Materials and Atoms, 207(2):107–123, 2003.
[60] John Allison, Katsuya Amako, John Apostolakis, Pedro Arce, Makoto Asai, Tsukasa Aso, Enrico Bagli,
A Bagulya, S Banerjee, GJNI Barrand, et al. Recent developments in geant4. Nuclear instruments and
835:186–225, 2016.
[61] N Golshani. Development of silicon drift detectors using boron layer technology. Delft University of
Technology, 2015.
[62] P Lechner, C Fiorini, R Hartmann, J Kemmer, N Krause, P Leutenegger, A Longoni, H Soltau,
D Stötter, R Stötter, et al. Silicon drift detectors for high count rate x-ray spectroscopy at room
temperature. Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrom-
eters, Detectors and Associated Equipment, 458(1-2):281–287, 2001.
78
[63] Witold Pokorski, R Chytracek, J McCormick, and G Santin. Geometry description markup language
and its application-specific bindings. IEEE Trans. Nucl. Sci, 53, 2006.
[64] Paul Schütze, Simon Spannagel, and Koen Wolters. Allpix2 User Manual v2.4.1, 2023.
[65] Erwin Fehlberg. Classical fourth-and lower order runge-kutta formulas with stepsize control and their
application to heat transfer problems. Computing, 6:61–71, 1970.
[66] Ilka Antcheva, Maarten Ballintijn, Bertrand Bellenot, Marek Biskup, Rene Brun, Nenad Buncic,
Ph Canal, Diego Casadei, Olivier Couet, Valery Fine, et al. Root—a c++ framework for petabyte
data storage, statistical analysis and visualization. Computer Physics Communications, 180(12):2499–
2512, 2009.
[67] JD Eskin, HH Barrett, and HB Barber. Signals induced in semiconductor gamma-ray imaging detectors.
Journal of applied physics, 85(2):647–659, 1999.
[68] Christopher Chantler, Karen Olsen, Robert Dragoset, A Kishore, Svetlana Kotochigova, and D Zucker.
X-ray form factor, attenuation and scattering tables (version 2.0), 2003-01-01 2003.
[69] AMPTEK INC. Mini-X2 X-Ray Source. Available at https://www.amptek.com/products/

mini-x2-x-ray-tube.
[70] Tom Gustafsson and Geordie Drummond Mcbain. scikit-fem: A python package for finite element
assembly. Journal of Open Source Software, 5(52):2369, 2020.
[71] Sai Liu, Yuming Xue, Rui Jia, Ke Tao, Shuai Jiang, Yiqing Wu, Haoyu Sun, Qing Guo, Liming Zhang,
and Shaojun Feng. Design and preparation of integrated voltage divider for silicon drift detector by ion
implantation. Journal of Materials Science: Materials in Electronics, 30:10152–10161, 2019.
[72] A Rashevsky. Alice silicon drift detector: qualification procedure of the mass production. In IEEE
Symposium Conference Record Nuclear Science 2004., volume 2, pages 917–921. IEEE, 2004.
[73] L Desorgher, F Lei, and G Santin. Implementation of the reverse/adjoint monte carlo method into
geant4. Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers,
Detectors and Associated Equipment, 621(1-3):247–257, 2010.
[74] Yu V Larionov. Measurement of the sem-beam diameter using a relief structure: Influence of contami-
nation. Journal of Surface Investigation. X-ray, Synchrotron and Neutron Techniques, 8(6):1137–1145,
2014.
[75] Julien Bert, Yannick Lemaréchal, Didier Benoit, Marie-Paule Garcia, and Dimitris Visvikis. Ggems:
Gpu geant4-based monte carlo simulation platform. Contributionsa la simulation Monte-Carlo pour
l’optimisation du traitement en radiothérapie, page 69, 2016.
[76] Seth R Johnson, Stefano C Tognini, Philippe Canal, Thomas Evans, Soon Yung Jun, Guilherme Lima,
Amanda Lund, and Vincent R Pascuzzi. Novel features and gpu performance analysis for em particle
transport in the celeritas code. In EPJ Web of Conferences, volume 251, page 03030. EDP Sciences,
2021.
[77] G Amadio, J Apostolakis, P Buncic, G Cosmo, D Dosaru, A Gheata, S Hageboeck, J Hahnfeld,
M Hodgkinson, B Morgan, et al. Offloading electromagnetic shower transport to gpus. In Journal
of Physics: Conference Series, volume 2438, page 012055. IOP Publishing, 2023.
[78] John Apostolakis, Marilena Bandieramonte, Sunanda Banerjee, Nazar Bartosik, Gloria Corti, Gabriele
Cosmo, V Daniel Elvira, Thomas Evans, Andrei Gheata, Simone Pagan Griso, et al. Detector simulation
challenges for future accelerator experiments. Frontiers in Physics, 10:913510, 2022.
79
Appendix A
Bias Voltage Sweep - Electric Fields
This section contains the results of the potential field and drift line simulations for a silicon drift detector,
where the back contact (BC) voltage was increased outside of the operating regime.
BC = −110 V
BC = −115 V
BC = −120 V
80
APPENDIX A. BIAS VOLTAGE SWEEP - ELECTRIC FIELDS
BC = −125 V
BC = −130 V
BC = −135 V
BC = −140 V
BC = −145 V
81
APPENDIX A. BIAS VOLTAGE SWEEP - ELECTRIC FIELDS
BC = −150 V
82
Appendix B
Bias Voltage Sweep - Spectra
This section contains the EDS spectra for a silicon drift detector, irradiated by a Fe-55 source, where the
back contact (BC) voltage was increased outside of the operating regime.
BC = −110 V BC = −115 V
BC = −120 V BC = −125 V
83
APPENDIX B. BIAS VOLTAGE SWEEP - SPECTRA
BC = −130 V BC = −135 V
BC = −140 V BC = −145 V
BC = −150 V
84
Appendix C
Charge Grouping Spectra
charge carriers are grouped in sets during propagation.
charge per step = 1 charge per step = 5
85
APPENDIX C. CHARGE GROUPING SPECTRA
86
Appendix D
Spatial Precision Spectra
spatial precision for the propagation of the charge carriers is varied.
spatial precision = 1 nm spatial precision = 2 nm
87
APPENDIX D. SPATIAL PRECISION SPECTRA
spatial precision = 1000 nm
88

Full Text 01

Uploaded by

Copyright:

Available Formats

Full Text 01

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Full Text 01

Uploaded by

Copyright:

Available Formats

Silicon Drift Detector Simulations for

Energy-Dispersive X-ray Spectroscopy in

Department of Physics, Fysikum

5 Future Research Recommendations 69

A Bias Voltage Sweep - Electric Fields 80

B Bias Voltage Sweep - Spectra 83

C Charge Grouping Spectra 85

D Spatial Precision Spectra 87

Background and Context

Research Problem and Motivation

Initial State at Sioux Technologies

2.1.1 SEM Operating Principles

Figure 2.1: Cross section of the electron inter-

2.1.2 Energy-dispersive X-ray spectroscopy

2.1.3 Characteristic X-rays

2.1.5 X-ray Interactions With Matter

Eγ + Ee = Eγ′ + Ee′ (2.6)

2.2 Semiconductor Basics

2.2.1 Band Theory of Solids

2.2.2 Intrinsic Semiconductors

2.2.3 Extrinsic Semiconductors

Combining the two above, we obtain,

ρ = −eNA for z≤0 (2.40)

2.2.5 Biased pn-Junctions

2.3 Semiconductor Detector Principles

Each of these stages will be treated in detail in the following sections.

2.3.1 Charge Creation

Charge Creation Energy

ω ≈ 2.8EG + 0.6 eV (2.58)

2.3.2 Charge Propagation

where T is the temperature of the silicon bulk.

2.3.3 Signal Formation

2.3.5 Incomplete Charge Collection

Front Contact ICC

Scholze ICC Model

modeled using the following exponential,

2.4 Silicon Drift Detector

3.1 Software Overview

Figure 3.1: Schematic of the simulation pipeline.

3.1.2 Allpix Squared

3.1.3 COMSOL Multiphysics

3.2 Pre-Simulation Phase

3.2.1 Electric and Weighting Field Modeling

where ND is the doping concentration and e the elementary charge.

Figure 3.2: COMSOL geometry schematic of a silicon drift detector.

Contact Bias voltage [V]

Maximum Minimum Maximum Number of

Table 3.2: Mesh Parameters for COMSOL Simulations

Figure 3.3: Visualization of the COMSOL Mesh.

3.2.2 Geant4 Geometry Modeling

3.3 Main Simulation Phase

3.3.2 Core Simulation

Figure 3.4: Stepping of Geant4 particles through the geometry.

Charge Carrier Creation

3.4 Post-Simulation Phase

3.4.1 Data extraction

3.5 Allpix Squared Modifications

3.5.1 Discretization Error in Pair Creation

• Energy deposition in Geant4

Energy Depositions Mean charge Number of charges created