Investigating the Generalizability of Machine Learning Potentials Across Chemical Space

Institut

Professur für Multiscale Modeling of Fluid Materials

Typ

Bachelorarbeit / Semesterarbeit / Masterarbeit /

Inhalt

theoretisch /  

Beschreibung

Machine learning potentials (MLPs) have shown great promise in accurately describing interatomic

interactions while being computationally efficient. However, a critical challenge in their application is

ensuring generalizability—i.e., the ability to perform well across diverse chemical systems and

reproduce physically meaningful behaviour. Understanding how MLPs generalize across chemical

space is essential for their reliable use in molecular dynamics (MD) simulations and materials

discovery.

This thesis focuses on developing methods to investigate the generalizability of MLPs. The student

will analyze how model performance varies with chemical features (e.g., functional groups, bond

types, or atomic environments) and identify regions of chemical space where the model may fail.

Additionally, the student will design and test systems to evaluate whether the MLP can reproduce

known physical behaviour, or structures.

Voraussetzungen

We are looking for a highly motivated student with the following qualifications:

● Machine Learning Practice: Experience with training and evaluating machine learning

models, particularly for regression tasks.

● Programming Expertise: Strong proficiency in Python. Familiarity with machine learning

frameworks (e.g., PyTorch, JAX) and data analysis tools (e.g., pandas, matplotlib) is a plus.● Chemistry Knowledge: Basic understanding of chemical features, and molecular structures

is advantageous.

● Motivation and Curiosity: A strong interest in machine learning, computational chemistry,

and materials modeling.

Möglicher Beginn

sofort

Kontakt

Michał Sanocki
Tel.: 783880014
m.sanockitum.de

Ausschreibung