Prof. Dr. Nicolas Gauger, Chair for Scientific Computing (SciComp), TU Kaiserslautern

SciComp Seminar Series

Please contact Prof. Gauger, if you want to register for an online talk in our SciComp Seminar Series or just to register for the seminar.

A list of the already scheduled talks can be found –> here:

Guillermo Suárez, Chair for Scientific Computing (SciComp), University of Kaiserslautern-Landau (RPTU)

Title: Reinforcement Learning Discovers Efficient Strategies for Active Flow Control

Abstract:

We explore the use of reinforcement learning to develop effective strategies for active flow control in unsteady fluid dynamics. In a two-dimensional computational fluid dynamics simulation of flow past a circular cylinder at a Reynolds number of 100, a reinforcement learning agent learns to manipulate dual side jets to alter the vortex shedding dynamics. Without any prior knowledge of the flow physics, the agent discovers a control policy that suppresses vortex-induced oscillations and achieves a drag reduction of nearly 10%. This performance is attained with minimal actuation effort, using jet mass flow rates of less than 0.5% of the incoming flow.

Building on these results, ongoing work investigates the integration of model-based reinforcement learning, aiming to reduce training time and improve generalization.

How to join online

You can join online via Zoom, using the following link:
https://uni-kl-de.zoom-x.de/j/69269239534?pwd=Z9UOzMpkhMjrxVhll3d49sNHFe9Fd1.1

Prof. Dr. Alexander Heinlein, Numerical Analysis Group, TU Delft, The Netherlands

Title: Domain decomposition for neural networks

Abstract:

-> Download Slides  
Scientific machine learning (SciML) is a rapidly evolving research field that combines techniques from scientific computing and machine learning. This talk focuses on the application of domain decomposition methods to design neural network architectures and enhance neural network training. In particular, it first explores how domain decomposition techniques can be employed in neural network-based discretizations that can address forward and inverse problems involving partial differential equations, using physics-informed neural networks (PINNs) as well as neural operators. It further discusses domain decomposition-based neural networks and preconditioning strategies for randomized neural networks, where the resulting optimization problem becomes linear in both data-driven settings and PINNs involving linear differential operators. Finally, the talk explores the use of domain decomposition methods for traditional machine learning tasks, such as semantic image segmentation with convolutional neural networks (CNNs). Computational results show that domain decomposition methods can improve efficiency—both in terms of time and memory—as well as enhance accuracy and robustness.

How to join online

You can join online via Zoom, using the following link:
https://uni-kl-de.zoom-x.de/j/69269239534?pwd=Z9UOzMpkhMjrxVhll3d49sNHFe9Fd1.1

Yan Muller, Department of Computer Science, University of Kaiserslautern-Landau (RPTU)

Title: Partial Least Squares Regression

Abstract:

This thesis investigates the use of Partial Least Squares (PLS) regression for dimensionality
reduction in machine learning applications. A simple linear experiment first
demonstrates the basic functionality of PLS. Subsequently, Linear Regression (LR) and
Gaussian Process (GP) models are evaluated on a strictly nonlinear target function under
three scenarios: without dimensionality reduction, with PLS, and with Principal
Component Analysis (PCA). The experiments, conducted on datasets with low to moderate
dimensionality, show that dimensionality reduction through PLS and PCA does not
improve runtime or predictive performance in these settings. Gaussian Process models
without dimensionality reduction achieved the best results. While dimensionality reduction
can, in principle, approximate the predictive performance of a full-dimensional GP
model with fewer input variables, our findings emphasize that its success strongly depends
on the dataset properties and the number of components selected.

Julia Manger, Department of Computer Science, University of Kaiserslautern-Landau (RPTU)

Title: Bayesian Optimization with Inequality Constraints

Abstract:

Bayesian Optimization (BO) is an effective method for optimizing expensive-to-evaluate
black-box functions, often applied in engineering and machine learning. Many real-world
problems involve constraints that standard BO is not designed to manage.
This thesis explores the integration of inequality constraints in BO, extending it to con-
strained problems. In simulations of both one- and two-dimensional search spaces, con-
strained Bayesian Optimization (cBO) demonstrates an effective balance between explo-
ration and exploitation, simultaneously ensuring the feasibility of solutions.
cBO is also compared with the Quadratic Penalty (QPF) method and with a method
using projection to show cBO’s advantages compared to penalty- or projection-based
optimization.

Raju Ram, Bosch, Leonberg, Germany

Title: Hybrid Parallel Preconditioners in Domain Decomposition Methods: Achieving Robustness and Scalability in the Solution of Large-Scale Linear Systems

Abstract:

Krylov subspace solvers like GMRES and preconditioners such as ILU are widely used for solving large-scale linear systems in simulations. Scalable parallel implementations are required to exploit the increasing parallelism provided by modern hardware. We develop a hybrid parallel Crout ILU preconditioner within the linear solver library GaspiLS, leveraging the GASPI programming model. At the process level, domain decomposition methods such as additive Schwarz (AS) and restricted additive Schwarz (RAS) are employed, and the multilevel nested dissection method is applied at the thread level.
Our implementation achieves 80 percent parallel efficiency on 1280 cores (64 sockets) when solving the convection-diffusion equation. Compared to the PETSc linear solver library, it delivers comparable baseline performance and better scalability at higher socket counts. When solving large-scale real-world computational fluid dynamics problems, scalability is maintained up to 1920 cores (96 sockets), with the RAS preconditioner successfully controlling iteration growth, unlike the AS variant. Further tests on ill-conditioned matrices from the SuiteSparse collection demonstrate the robustness and generality of the preconditioner, successfully solving problems that the traditional ILU(0) preconditioner fails to solve.
Our approach offers a generic, algebraic, and scalable preconditioner that enables productivity for the domain expert in solving large-scale, sparse linear systems.

How to join online

You can join online via Zoom, using the following link:
https://uni-kl-de.zoom-x.de/j/69269239534?pwd=Z9UOzMpkhMjrxVhll3d49sNHFe9Fd1.1

Alexander Schilling, Chair for Scientific Computing (SciComp), University of Kaiserslautern-Landau (RPTU)

Title: Proton Therapy Range Verification with a Digital Tracking Calorimeter

Abstract:

Proton therapy is highly sensitive to range uncertainties due to the nature of the dose deposition of charged particles. To ensure treatment quality, range verification methods can be used to verify that the individual spots in a pencil beam scanning treatment fraction match the treatment plan. For this purpose, secondary particles emitted by nuclear interactions of the proton beam with the patient can be detected and compared to expected values from Monte Carlo simulations. To eliminate the need for an additional detector in the treatment room, we investigate the feasibility of using a digital tracking calorimeter designed for proton computed tomography for this purpose. Using uncertainty-aware machine learning, we define a quality measure for a treatment fraction, based on the amount of erroneously treated spots. With this method, it is possible to detect patient misalignments of 1 mm with statistical significance after around 1400 treated spots; sufficient for most treatment plans.

How to join online

You can join online via Zoom, using the following link:
https://uni-kl-de.zoom-x.de/j/69269239534?pwd=Z9UOzMpkhMjrxVhll3d49sNHFe9Fd1.1

Dr. Roman Pöschl, Laboratoire de Physique des 2 Infinis Ir`ene Joliot-Curie (IJCLab) / Universit´e Paris-Saclay, 91405 Orsay, France

Title: The Detector R&D (DRD) on Calorimetry – Introduction and overview

Abstract:

Calorimeters play a pivotal role in past, present and future experiments in particle physics. Final states of particle physics collision consist to a large fraction of jets. These jets are composed of electrons, photons and charged and neutral hadrons. A central requirement to meet scientific goals at future experiments is to keep the jet energy resolution at a level of 3-4% for jet energies between 45 GeV and around a TeV (or more). There are several proposal to meet this goal, by increasing the granularity of the calorimeters by dedicated precise measurements of hadrons and electromagnetic particles within a jet or by a combination of these features. This seminar will review the requirements to calorimeters in future experiments and the status and outlook on the current R&D to meet these requirements. The seminar will also sketch the potential to apply machine learning for calorimetry and how quantum sensing may dramatically change the design of future calorimeters. The R&D is organised within the worldwide Detector R&D (DRD) Collaboration on Calorimetry that is hosted by CERN. The seminar will include an overview on the structure and the mission of the DRD.

How to join online

You can join online via Zoom, using the following link:
https://uni-kl-de.zoom-x.de/j/69269239534?pwd=Z9UOzMpkhMjrxVhll3d49sNHFe9Fd1.1

Dr. Zebang Shen, Lecturer at the Department of Computer Science, Institute for Machine Learning, ETH Zürich, Switzerland

Title: tba

Abstract:

tba

How to join online

You can join online via Zoom, using the following link:
https://uni-kl-de.zoom-x.de/j/69269239534?pwd=Z9UOzMpkhMjrxVhll3d49sNHFe9Fd1.1