Supervisors: Christoforos Rekatsinas, Panagiotis Krokidas, Vaia Prassia
Description:
Diffusion models (DMs) have demonstrated impressive performance in generating high-quality data by learning a probabilistic process that iteratively transforms noise into meaningful samples. In the realm of image generation, models like Latent Diffusion Models (LDMs) have achieved state-of-the-art results by operating in latent space, greatly reducing computational requirements while preserving detail. Inspired by this success, this project explores the application of Latent Diffusion to molecular dynamics (MD) simulations to improve the efficiency of biomolecular process simulations. DIFFMD, inspired by thermodynamic diffusion processes, provides a novel and computationally efficient method for simulating biomolecular dynamics without relying on traditional energy or force calculations. The goal of this thesis is to evaluate and enhance the performance of DIFFMD by applying latent diffusion techniques and comparing it to existing deep learning-based MD models. The work will involve integrating latent diffusion with a geometric Transformer and optimizing it for biomolecular applications. The project will also explore the potential for further scaling and improving the method to simulate a broader range of biomolecular processes.
Qualifications required: Python programming, Machine Learning and Deep Learning algorithms
References:
crek[at] iit [dot] demokritos [dot] gr