Steinbuch Centre for Computing

Scientific Computing & Mathematics

Bachelor Thesis and Master Thesis in

Computer Science or Mathematics
Development of massive parallel algorithms for domain applications

Background
We develop the NAStJA framework, that provides
an easy way to enable massively parallel
simulations for a wide range of multi-physics
applications based on stencil algorithms.
It includes the phase-field method and phase-field
crystal model for material science. Based on a
cellular Potts model, NAStJA simulates tissue
growth, e.g., cancer, for the life science.

Possible tasks
● Enhance the parallel scalability and performance of the NAStJA framework.
● Introduce new algorithm and extend the models.
● Porting of algorithms to GPUs.
● Perform simulation studies and analyze the model behavior.
● Please ask for further open topics.

Prerequisite
● Basic knowledge of programming, preferably C++.
● Interest in numerical simulations and applied sciences.

About us
At the SCC, we do research as an interdisciplinary team with other international
research groups at the interface between mathematics, computer science, physics
and biology. We offer:
● Intensive supervision in german or english.
● Access to high-performance computing systems.
● Productive and dynamic atmosphere in a team of employees.
● Career perspectives as a young scientist.
● Excellent home working possibilities and online networking opportunities.

Contact
Dr. Marco Berghoff
NAStJA development team
SimLab NanoMicro
[email protected]
https://nastja.gitlab.io

KIT – The Research University in the Helmholtz Association

 Bachelor Thesis or Master Thesis
Title: Neural Differential Equations

Speciality: Students from the Department of Mathematics

Motivation: Deep learning is reshaping the discipline of scientific computing with the
rapid development of computing power. As illustrated in the universal approximation
theorem, neural network can be regarded as an effective function approximator in the
latent space. Compared to classical convolutional networks based on discrete iterating
sequence, e.g. ResNet, the neural ordinary differential equation provides a novel way
to model continuous dynamics with all the trainable parameters naturally coupled.

Fig: An example on applying neural network in solving Boltzmann equation

Subject: Basically the neural ODE is adopted for classification and regression in
machine learning community. The work on this thesis serves as a heuristic attempt to
widen its possibility in scientific computation and numerical simulation. More
specifically, classical physical modeling can be hybridized with neural network, which
is expected to form a differentiable formulation of desirable accuracy and efficiency.

Tasks: The work on the thesis will include following steps:

• To develop a differentiable programming within deep learning frameworks
• To implement neural ordinary and partial differential equations
• To conduct hybrid mechanical and neural modeling and develop universal
differential equations
• To compare the performance and efficiency of the newly designed differentiable
model and classical methods in both machine learning and scientific computing
tasks

Contact: Tianbai Xiao, PhD (SCC/SCM, CSMM Group)

Prof. Martin Frank (SCC, Department of Mathematics)
Email: [email protected]
 Master’s Thesis
Filters For Uncertainty Quantification

Courses: Mathematics, Computer Science, Physics

Topic
A vast amount of physics and engineering applications depend on uncertain measurement data and
modeling parameters. The increase of computing power allows researchers to include these un-
certainties into simulations in order to check the validity of their results. A commonly used tool to
compute quantities such as the expectation value or the variance of the solution is the stochastic
Galerkin method. This method yields satisfactory solution approximations in smooth regimes, howe-
ver exhibits heavy oscillations at shocks.

In this project, we are interested in computing the expectation value and the variance of the state of
a gas whose initial state is uncertain. Due to the formation of shocks one needs to find strategies to
mitigate oscillations of the stochastic Galerkin method. A promising idea is using filters, since they
show non-oscillatory solution approximations while maintaining important solution characteristics.
However, several urgent questions remain unanswered: Which kind of filter should be chosen for
which problem? How should one choose the filter strength? Are there strategies which allow an
automated choice? Can one analytically quantify the effect certain filters have on the solution? If yes,
how can this be used to choose an adequate filter strength?

Task
The task of this thesis project is to (partially) answer these questions. The individual steps are:
• Analytically and numerically investigate the filtered stochastic Galerkin system
• Implement different filters and study their effect on the solution
• Find a dimensionless representation of the filter strength

Contact
Prof. Dr. Martin Frank, Jonas Kusch
Faculty of Mathematics
Steinbuch Centre for Computing
www.scc.kit.edu
[email protected],[email protected]
 Bachelor or Master Thesis
Title: Uncertainty-aware dose optimization for radiation therapy

Course of Study: Mathematics, Computer Science, Physics or similar

Motivation: In radiation therapy, parameters like the position or energy of beams are
optimized with respect to the fulfillment of a radiation plan. This plan prescribes certain
upper and lower bounds for the amount of dose in tumors, organs and regular tissue.
For fixed parameters, the dose distribution is then computed by solving the transport
equation. However, the assumption that parameters can be fixed deterministically is
not a realistic one. Set-up errors in the placement of the patient, movements or
uncertainties in the tissue density occur frequently and affect the outcome – especially
in the case of proton therapy and heterogeneous regions like the lung. In this case
simply including a margin to account for uncertainties is not sufficient and can lead to
great deviations from the radiation plan.

air

bone

Irradiation of lung tumor with several beams Beam hitting a bone/air interface

Subject: Therefore, this thesis is concerned with finding and testing efficient strategies
of including uncertainties into the dose optimization process, e.g. with methods from
stochastic/probabilistic optimization. Here the focus will lie foremost on uncertainties
concerning the beam parameters, such as positions and range/energy of the beams.

Tasks: The work on the thesis could include the following steps:
 Mathematically analyze the problem and develop/determine different
optimization strategies
 Implement a simplified framework for dose computation & set-up a
representative small-scale example (e.g. 2D CT scan with heterogeneities)
 Test the ideas in the simplified example, analyze the results with respect to
quality, speed and generalizability
 Evaluate and maybe improve the approaches based on results, mathematical
properties and applicability to the specific problem at hand

Contact: Pia Stammer, MSc (SCC/SCM, DKFZ)

Prof. Martin Frank (SCC, Department of Mathematics)
Email: [email protected]
Phone: +49 721 608-25492
 Bachelor Thesis or Master Thesis
Title: Symmetrized Polynomial Propagators for Hamiltonian Systems

Field of study: Mathematics

Description: Dynamics of Hamiltonian systems is simulated using numerical time inte-

gration methods specific for this type of ordinary differential equations. Typical applica-
tions are molecular dynamics and celestial mechanics simulations. These applications
require simultaneously several properties from time integrators: sufficient order of accu-
racy, long-term stability, low computational complexity and high scalability and short
time to solution. The small and fixed time steps necessary to fulfil some of the require-
ments limit the success of currently used integrators. Expanding the time evolution op-
erator in series of Chebyshev or Newton polynomials provides promising schemes, al-
lowing larger time steps and parallelization. On one hand, these schemes are insuffi-
ciently studied in the case of classical dynamics, and on the other hand, they do not
preserve the intrinsic time reversal symmetry of the solution required for long-term sta-
bility.

Subject: For a small test system (with three particles), you will investigate one-step
implicit time-symmetric schemes based on the original polynomial schemes focusing on
the local truncation error (LTE), the global error and their relation to the time step size
and polynomial order.

Tasks:
• Investigate how the LTE of the two original schemes and of their symmetrized ver-
sions behave with variation of expansion order and time step.
• Show analytically and numerically that the Chebyshev propagator is time-symmetric
in the limit of infinite expansion order and compare to the Newtonian propagator.
• Measure the global energy error (energy drift) for the original and the symmetrized
schemes again with increasing expansion order and time step.
• Decompose the LTE of the symmetrized scheme into components accounting for the
asymmetry and for the series truncation error, and study their behavior in the limit of
very short time step size using sufficiently high decimal precision.
• Suggest a modification of the symmetrized scheme allowing adaptive time step size
using a suitable criterion for the LTE.

Contact: Dr. Ivan Kondov, Prof. Dr. Martin Frank

Email: [email protected]
Phone: 0721 608-28644
 Bachelor Thesis or Master Thesis
Title: Tracking Virtual Flame Particles for Studying Flame Dynamics
Motivation: Objective of this theoretical work is to gain an in-depth understanding of
the flame dynamics during turbulent combustion processes, which are of great
importance for the stable, efficient and clean burning of fuels. The topic covers mainly
the effect of the interaction between the turbulent flow and combustion reaction. The
flame is stretched due to the unsteady, non-uniform flow in terms of curvature and
straining, which locally leads to a modified balance between underlying transport and
reaction processes. In extreme case, the flame may be extinguished by a high stretch
rate on the flame surface. As this interaction process occurs at a wide range of length
and time scales, most existing models take simply the unstretched, laminar flame
speed for the interpretation of local flame characteristics, e.g. in Flamelet models. This
assumption is not valid especially under conditions of highly turbulent and unsteady
flows.

Subject: Previous studies on flame dynamics are mostly limited to global flame
features, such as time mean total flame surface or flame angle, which neglects the
effect of time history of the turbulent flow. A novel method based on tracking virtual
massless particles along the flame surface will be used in this thesis work for a
detailed analysis of the correlation of modified transport processes by turbulent flow
stretch. Direct numerical simulation on simple 2D oscillating jet flames will be carried
out using the OpenFOAM code at varying fluctuating time scales. The behaviors of
local and global flame speed in response to unsteadiness of the incoming flow will be
analyzed and at best, recast to a correlation law by means of the flame’s relaxation
time. The computational grids as well as tools required for DNS and particle tracking
are already developed.

Tasks: Detailed work steps include:

• Literature research and familiarization with
required tools
• Running DNS for different inflow conditions
• Detailed analyses of the simulation results
• Summary and presentation of the work

Contact:
Thorsten Zirwes
(SCC/SCM, SimLab Energy and EBI/VBT)
Email: [email protected]
Phone: +49 721 608-29278
 Bachelor Thesis or Master Thesis
Title: Benchmarking of Transport Equations

Speciality: Students from the Department of Mathematics

Motivation: Transport equations are of great importance in many fields. Practical

applications include radiation therapy, optical imaging in biological media and heat
transfer as well as non-destructive material analysis. In practice, transport equations
are solved numerically and the verification of solvers is of the utmost importance.

Subject: The subject of this thesis is the method of manufactured solutions. It allows
to compute the boundary conditions and source terms that impose certain analytical
solutions for partial differential equations. One can then study the convergence
behaviour of numerical solutions to the analytical ones to verify numerical solvers for
said partial differential equations. In this work, the method of manufactured solutions
shall be used to derive analytical solutions for the Boltzmann transport equation
respectively one of its approximations, the so-called SPN-equations. Of special
importance for this application is the handling non-trivial boundary conditions for these
equations.

Tasks: The work on the thesis will include following steps:

• Definition of suitable test cases.
• Development of an algorithm that generates manufactured solutions for these
test cases with the appropriate boundary conditions.
• Implementation of this algorithm in a software tool.
• For a master thesis, the above points can be investigated in greater detail.
Alternatively, an extension to other approximations of the Boltzmann equation
or adjoint transport is possible.

Contact: Prof. Dr. Martin Frank, Alexander Jesser

Faculty of Mathematics
Steinbuch Centre for Computing
www.scc.kit.edu
Email: [email protected], [email protected]
 Bachelor Thesis or Master Thesis
Title: Generating Turbulent Velocity Fields with Controlled Properties

Speciality: Students from the Department of Mathematics

Motivation: Turbulent flows are known to intensify the processes of mixing in nature
and in technology devices, they also accelerate the speed of chemical reactions in
apparatus of process engineering or in combustion technology. Optimisation of
processes in modern devices requires the detailed study of fluid flow,
thermodynamics, transport of species and energy in great detail. For these purposes
studies with very fine numerical grids are required and one of it is the Direct Numerical
Simulation (DNS) of turbulence. The first step in such simulations is to generate an
initial turbulent field and to maintain it throughout the complete simulation time.

Small-scale turbulence Large-scale turbulence

Subject: The work on the thesis is connected with the generation and the maintaining
of turbulent fields with special, controlled properties. This is not a trivial task when the
forcing is generated in physical space by means of additional source-terms. Two of the
most important parameters which are of primary interest, are the kinetic energy of
turbulence and the integral length scale. Many others parameters like e.g.
Kolmogorov’s scale, RMS of turbulence u’ need to be evaluated too.

Tasks: The work on the thesis will include following steps:

• To setup an OpenFOAM model in a cuboidal geometry with a fine enough grid
resolution and periodic boundary conditions
• To implement and test different kinds of linear forcing in the momentum
equations (Navier-Stokes equations) according to recent publications in the
literature
• To carry out several DNS which generate turbulence with different properties
• To comparing the methods for generation of turbulence and the obtained
turbulent fields with appropriate tools in Python which will be developed during
the thesis

Contact: Jordan Denev, PhD (SCC/SCM, SimLab Energy)

Prof. Martin Frank (SCC, Department of Mathematics)
Email: [email protected]
Phone: 0721 608-25771
 Steinbuch Centre for Computing
Scientific Computing & Mathematics

Bachelor Thesis and Master Thesis in

Computer Science or Mathematics
Porting the radiation transport moment model to GPUs

Background
The radiation transport equation is a mesoscopic equation in high dimensional phase
space. Moment methods approximate it via a system of partial differential equations in
traditional space-time. One challenge is the high computational intensity due to large
vector sizes (1600 components for P39) in each spatial grid point.
We develop the NAStJA framework, that provides an easy way to enable massively
parallel simulations for a wide range of multi-physics applications based on stencil
algorithms. It includes a CPU implementation for the radiation transport. For real-time
applications the usage of a large cluster is not practical. Here, GPUs offer an
alternative with their compact compute capacity. NAStJA already provides an interface
to solve simple models on the GPU.

Tasks
● Extend NAStJA's GPU interface for vector spaces.
● Porting the model on GPUs.
● Performance analysis on single and multi-GPUs.

Prerequisite
● Basic knowledge of programming, preferably C++, CUDA.
● Interest in numerical simulations and applied sciences.

About us
At the SCC, we do research as an interdisciplinary team with other international
research groups at the interface between mathematics, computer science, physics
and biology. We offer:
● Intensive supervision in german or english.
● Access to high-performance computing systems.
● Productive and dynamic atmosphere in a team of employees.
● Career perspectives as a young scientist.
● Excellent home working possibilities and online networking opportunities.

Contact
Dr. Marco Berghoff
NAStJA development team
SimLab NanoMicro
[email protected]
https://nastja.gitlab.io

KIT – The Research University in the Helmholtz Association

 Master Thesis
Title: Solving the simplified PN equations using GPU acceleration

Area of expertise: Students from the Department of Mathematics

Motivation: Based on an industry project, the SCC is developing a measurement

facility used for non-destructive sample analysis based on the concepts of prompt-
gamma neutron activation. The measurement facility evaluates characteristic signals
of recorded gamma spectra and maps them the sample’s elemental masses. In order
to compute this mapping, the neutron flux in the sample needs to be known. This flux
can in general be described by the linear Boltzmann equation, but can be
approximated for scattering dominated domains using the multi-group simplified PN
equations. This system of equations consists of several coupled diffusion equations.
Due to the high number of required energy groups, a fast numerical code is crucial in
order to determine the neutron flux several times during the evaluation of a sample.

Checkerboard test case Neutron flux within the sample

Subject: The goal of the thesis is to develop a CUDA-capable neutron flux solver for
simple 3-D geometries using the simplified P1 approximation. The equations should be
discretized using a finite element approach and utilize the deal.ii framework and its
built-in CUDA capabilities.

Tasks: The work on the thesis will include following steps:

• Derive a weak form of the simplified P1 approximation.
• Develop a working 2-D solver and verify the solution on academic test cases.
• Add multi-group capabilities.
• Extend the solver to 3-D compare the solution to existing implementations (will
be provided by the supervisor).
• Analyze the scaling capabilities of the implementation.

Contact: Jannick Wolters, M.Sc. (SCC/SCM)

Prof. Martin Frank (SCC, Department of Mathematics)
Email: [email protected]
 Master’s Thesis
Non conservative methods for linear Boltzmann equations

Courses: Mathematics, Computer Science, Physics

Topic
Kinetic equations play a crucial role in many fields of application, such as radiation cancer therapy,
fluid dynamic of rarefied gases, or the simulation of astrophysical phenomena. The example of the
ray transport equation is given below.

∂t ψ(t, x, v) + v · ∇x ψ(t, x, v) = (Sψ)(t, x, v).

The difficulty in solving a kinetic equation stems from the high dimensional space of input varia-
bles. Namely, the kinetic density ψ is dependent on time t, space x and velocity v. Multiple solution
approaches are based on discretization strategies of the velocity variable v. The MN method uses
a Galerkin discretization in v that relies on the so-called minimal entropy principle. The method is
computationally expensive, since one needs to solve a constrained optimization problem in each
grid cell. The research question for this thesis is to construct a numerical method based on the MN
approximation, that does not rely on the solution of this optimization problem. A promising approach
uses non-conservative methods.

Task
The project has the following main tasks.

• Understanding the MN approximation and the minimal entropy approach.

• Implementation of these methods into an existing numerical framework (C++ skills are advan-
tageous) or developement of an own solver.

• Literature research on non conservative methods for transport equations.

• Comparison of the results to the traditional method and to own approaches.

Contact
Prof. Dr. Martin Frank, Steffen Schotthöfer,
Jonas Kusch
Faculty of Mathematics
Steinbuch Centre for Computing
www.scc.kit.edu
[email protected]
 Bachelor’s Thesis
Neural network solvers for differential equations

Courses: Mathematics, Computer Science

Topic
Since the advent of neural networks, they have been successfully applied to many task in the field of
computer vision, natural language processing and other life sciences. Quite recently, techniques from
neural networks have been used to solve ordinary and partial differential equations. Let us consider
the ODE

x0 (t) = f (x(t)), ∀t ∈ [0, T ],

x(0) = x0 .

One way to employ a network is to model the function x : t 7→ x(t) directly as a neural network.

For the training process, the choice of the loss function is crucial. Some authors successfully applied
the residual of the ODE in a suitable norm, i.e.

||x0 (t) − f (x(t)) + (x(0) − x0 )|| (1)

as a loss function, but there exist other approaches that focus on existing numerical methods, e.g.
euler solvers.

Task
The project consists of the following tasks:
• Explore different methods to set up a neural network ODE (or PDE) solver, with focus on the
loss function.
• Implement the most promising solvers in a programming language of your choice (Python is
preferable).
• Compare the different solution with each other and with a traditional reference solution.

Contact
Prof. Dr. Martin Frank, Steffen Schotthöfer
Faculty of Mathematics
Steinbuch Centre for Computing
www.scc.kit.edu
[email protected]
 Steinbuch Centre for Computing
Scientific Computing & Mathematics

Bachelor Thesis and Master Thesis in

Computer Science or Mathematics
Development of an interface for various implicit methods

Background
We develop the NAStJA framework, that provides
an easy way to enable massively parallel
simulations for a wide range of multi-physics
applications based on stencil algorithms.
IIt includes a cellular Potts model for the
simulation of tissue growth with a geometric cell
resolution. Signals, e.g., nutrients or drugs,
follows a diffusion law. Due to metabolism, cells
act as a signal source or sink and can have
various diffusion coefficients.
Tasks
● Modeling of multi-signal diffusion.
● Handling high gradients at cell boundaries.
● Implementation of an interface for different solution methods.
● Performance analysis for the different solution methods.
Prerequisite
● Basic knowledge of diffusive processes.
● Basic knowledge of programming, preferably C++.
● Interest in numerical simulations and applied sciences.
About us
At the SCC, we do research as an interdisciplinary team with other international
research groups at the interface between mathematics, computer science, physics
and biology. We offer:
● Intensive supervision in german or english.
● Access to high-performance computing systems.
● Productive and dynamic atmosphere in a team of employees.
● Career perspectives as a young scientist.
● Excellent home working possibilities and online networking opportunities.
Contact
Dr. Marco Berghoff
NAStJA development team
SimLab NanoMicro
[email protected]
https://nastja.gitlab.io

KIT – The Research University in the Helmholtz Association

 Master’s Thesis
Flow simulations with uncertainty

Courses: Mathematics, Computer Science, Physics

Topic
Fluid dynamics is a frequently studied topic, which plays a key role in a vast number of engineering
applications. Since real world problems are usually prone to uncertainty, one is interested in including
these uncertainties in computer simulations. The goal is to compute the expected flow properties and
the corresponding variance. Currently, the research of several methods used to quantify uncertainties
has been limited to one- or two-dimensional spatial flow domains. This heavily limits the applicability
of the developed methods and makes them unusable for real world applications.
In this project, we are interested in extending a given two-dimensional code framework to three-
dimensional domains. Furthermore, we are interested in studying the effects the increased dimension
has on the performance of individual methods for uncertainty quantification. Below, you can find the
variance of the gas density around a NACA airfoil with uncertain angle of attack that has been
computed with the current framework. Questions that arise once the code has been extended to
three dimensions are: Which method is the fastest for three-dimensional problems? How can we
extend concepts developed for two-dimensional domains to three dimensions?
Task
The task of this thesis project is to (partially) answer these questions. The individual steps are:
• Literature study to understand how to efficiently determine neighboring cells in a three dimen-
sional mesh.
• Literature study on three dimensional finite volume methods.
• Extending the given code framework to three dimensions.
• Test different methods from Uncertainty Quantification for the extended code.

Contact
Prof. Dr. Martin Frank, Jonas Kusch
Faculty of Mathematics
Steinbuch Centre for Computing
[email protected],[email protected]
 Steinbuch Centre for Computing
Scientific Computing & Mathematics

Bachelor Thesis and Master Thesis in

Computer Science or Mathematics
Learned noise model for Denoising Autoencoder

Background
A Denoising Autoencoder (DAE) is a type of Neural Network, which is trained to
reconstruct the original from partially corrupted data and is commonly deployed in
many fields, from medical imaging to Astroparticle Physics. This technique requires
the type of noise to be known, which for this reason is often simply assumed to be
white noise.
This assumption does not necessarily hold true for real world applications and thus
the questions arises, whether it is possible to improve the performance of DAEs by
utilizing modern generative Methods like Variational Autoencoders (VAEs) or
Generative Adversarial Networks (GANs).

Tasks
● Design and implement data dependent noise models to be used as benchmarks.
● Implement and optimize DAEs for those noise models.
● Develop learned Noise models using an approach of your choice.
● Compare to traditional methods and analyze the efficiency of the new approach.

Prerequisite
● Basic knowledge of programming, preferably Python.
● Interest in Machine Learning, especially Deep Learning.

About us
At the SCC, we do research as an interdisciplinary team with other international
research groups at the interface between mathematics, computer science, physics
and biology. We offer:
● Intensive supervision in german or english.
● Access to high-performance computing systems.
● Productive and dynamic atmosphere in a team of employees.
● Career perspectives as a young scientist.
● Excellent home working possibilities and online networking opportunities.

Contact
Dr. Marco Berghoff
NAStJA development team
SimLab NanoMicro
[email protected]
https://nastja.gitlab.io

KIT – The Research University in the Helmholtz Association