Generative AI for Molecular Simulations

Abstract

In statistical mechanics, computing the average behavior of microscopic states is crucial, for example, in estimating observables for equilibrium distributions in molecular systems. The challenge lies in the difficulty of sampling, as the density is known but hard to sample from. Typically, sampling of molecular conformations is performed using molecular dynamics, which faces challenges in obtaining iid samples due to the problem of rare events.

Various enhanced sampling methods have been proposed to tackle this issue. Machine learning, specifically continuous-time generative models, offers a new perspective for tackling this problem. In our thesis, we propose two generative models using the recent Stochastic Interpolants framework. The first learns to transform between equilibrium distributions with different temperatures, which can be further applied with the current replica exchange method. The second model learns transition probability densities across time scales, which can be used as a surrogate model to accelerate MD simulations.

We highlight the ability of Stochastic Interpolants to design efficient sampling methods for many-body systems in different ways, making it a powerful tool for advancing molecular simulation. Our results are two-fold. First, we present our Stochastic Interpolant ITO model and show how it reduces the VAMP-2 score gaps when benchmarked against the original ITO architecture. Next, we showcase our Thermodynamic Interpolant model, that to some extent manages to perform temperature transformations in a setting where it has to generalize beyond the training data. Our advancements show potential and could benefit various fields such as drug discovery, material science, catalysis, and green chemistry.