Transfer Learning of Graph Neural Networks for Implicit Solvent Modeling

Institut

Professur für Multiscale Modeling of Fluid Materials (TUM-ED)

Typ

Semesterarbeit / Masterarbeit /

Inhalt

 

Beschreibung

Background and Motivation
Molecular dynamics (MD) simulations are essential for understanding the dynamical properties of proteins and other molecular systems. While highly accurate, traditional simulations are computationally expensive. Machine learning potentials (MLPs), particularly those based on advanced Graph Neural Networks (GNNs), provide a promising alternative by balancing physical accuracy with computational efficiency.

A major bottleneck is explicit solvent simulation: tracking every water molecule is very costly. Implicit solvent models avoid this by representing the solvent as a continuous medium, which keeps key physics while reducing computation.

In this thesis, you will combine GNNs with parameter-efficient transfer learning (e.g., LoRA) to adapt pre-trained MLPs to implicit solvent settings. The goal is a method that is accurate, scalable, and efficient for implicit solvent simulation.

Objectives

• Understand modern GNN architectures for molecular potentials.
• Apply LoRA/adapters to transfer pre-trained models.
• Build implicit solvent representations in the model.
• Evaluate accuracy, transfer performance, and runtime efficiency.

Voraussetzungen

Candidate Profile
We are looking for technically strong Master's students with a keen interest in AI/ML methods and their scientific applications. Preferred skills and interests include:

• Familiarity with probability theory and statistics.
• Interest in dynamical systems, physics, chemistry, or applied mathematics.
• Strong time management and a proactive mindset.

Möglicher Beginn

sofort

Kontakt

Linying Zhang
Tel.: 015257290423
linying.zhangtum.de

Ausschreibung