Preprocessing of MRI Data With Neural Networks

In medical imaging, magnetic resonance imaging (MRI) images are used to obtain detailed information about the human brain. To make these images from different brains comparable, preprocessing is required. However, previous preprocessing tools are often unreliable, highly computationally intensive, and not very user-friendly. This project aims to use machine learning to publish a simplified, fast, and user-friendly preprocessing tool. In addition, the project will develop a Python package that simplifies the application of Deep Learning to MRI data.

Contact: Lukas Fisch, M. Sc.,  l_fisc17@uni-muenster.de

Data-driven Identification of Macroscopic Evolution Equations for Molecular Dynamics Systems

Replacing the full dynamics of microscopic systems by macroscopic evolution equations describing the essence of the microscopic system is an integral part of modeling in science and engineering. Within the framework of the macroscopic description collective behavior and self-organization can be studied. In this project methods for extracting partial differential equations (PDEs) solely from experimental data. These methods are then applied to the results of molecular dynamic Simulations of Systems where a macroscopic description is unavailable or to compare existing theories to the results of the data analysis process.

Contact: Oliver Mai, M. Sc.,  o_mai001@uni-muenster.de

Data-driven Identification of Dynamical Systems

Under construction.

Data-based Learning of Force Fields

In the field of theoretical chemistry and physics, molecular dynamics simulations are part of a everyday tool for studying system properties at the atomic level, such as molecular structures or molecular structures or transport processes, for example. For this purpose, Newton's equations of motion for atoms are solved based on underlying analytical potentials. are solved. For all required potentials, in the whole also called force field, often all parameters are optimized individually and on the basis of various experimental and quantum chemical data. optimized. This requires a high understanding of the problem and a lot of chemical intuition.
is required.
This project aims to optimize all parameters in an automated way and simultaneously based on a single (quantum chemical) data set. Alternatively, the potential analytic form can be optimized by more complex functions, such as neural networks, Gaussian processes or more complicated polynomial functions.
These can then be systematically improved using machine learning.

Contact: Mirko Fischer, M. Sc., m_fisc38@uni-muenster.de

Machine Learning of Off-Shell Effects in Top Quark Production at the LHC

The properties of the top quark are of great importance for understanding many aspects of the universe. Therefore, an exact determination of the fundamental properties of the top quark is mandatory. To this end, methods already exist that allow accurate calculations. The most sophisticated of these calculations involve enhancements such as radiative corrections or off-shell effects, which makes their evaluation extremely computationally expensive. Modern machine learning techniques such as neural networks could help make these critical calculations more efficient and ultimately feasible on a large scale. The goal of the research project is therefore to explore the application of these techniques to greatly reduce the computational cost of these calculations.

Contact: Mathias Kuschick, M. Sc., mathias.kuschick@uni-muenster.de

Application of Emergent Self-Organizing Maps (ESOM) for the Development of New Molecular Compounds

The development of pharmaceuticals is accompanied by a great need for new chemical compounds. As the experimental collection of chemical and biochemical data is extremely cost- and time-intensive, applications of unsupervised machine learning are of great interest. They enable to recognize connections between the data points even without prior labeling. These are of importance in drug development since, based on the “principle of similar properties”, molecules with similar structures are predicted to have similar properties.

The focus of this research project lies in the development and application of an implementation of the ESOM algorithm that meets the special requirements of cheminformatics and exploits the possibilities of modern computer architectures. This will then be used to visualize and gain knowledge from molecule libraries with several hundred million compounds, in doing so, generative approaches based on the learned maps will be pursued later.

Contakt: Johannes Kaminski, M.Sc. :  j.kaminski@uni-muenster.de

Potentials, Challenges and Conflicting Goals in the Relationship Between AI and Sustainability

Developments in the field of machine learning and artificial intelligence cause high relevance for the question of possible benefits and influence of these technologies on sustainable development. This project considers both the application of AI technologies to improve sustainability efforts and the sustainability challenges that arise from the development and use of AI systems.

Therefore, a distinction is made between two perspectives: AI for sustainability and the sustainability of AI, whereby the effects of AI systems are analyzed in ecological, social and economic terms. In addition, the project examines the role of the regulatory framework in the development of sustainable AI technologies and practices, as well as the role that sustainability considerations play in regulatory processes, such as the European AI Act. Through a comprehensive analysis of the opportunities and challenges at the shared interface of AI and sustainability, the research provides insights into how AI can be used for sustainable development and alleviate its negative impacts at the same time.

Contact: Benedikt Lennartz, M.A., benedikt.lennartz@uni-muenster.de

Event-based Detection of Insects

Monitoring insect presence and behavior is not possible using conventional camera traps due to the size, speed and lack of infrared signatures of insects. Event-based cameras combine several properties that make them a very suitable sensor for the trigger mechanism of an insect camera trap. These include a very high temporal resolution, a high contrast range and reduced production of redundant data while maintaining low power consumption. However, the resulting data cannot be processed using classic computer vision methods. The aim of this project is to use machine learning to detect and track insects in event data.