Learning Molecular Dynamics with Generative Models: From Equilibrium to Nonequilibrium Systems

Licentiate thesis, 2026

Statistical mechanics provides a broad theoretical framework for modeling molecular and biological systems, but in practice, the underlying dynamical equations are often not analytically tractable. Consequently, molecular simulations have been the workhorse of statistical mechanics for the past seventy years. However, with the advent of machine learning and generative models, these data-driven methods are starting to complement simulations by providing fast surrogates in scenarios where direct simulation is prohibitively expensive.

In this thesis, we discuss how generative models can enhance traditional simulation methods in both equilibrium and nonequilibrium settings.

In equilibrium sampling with continuous normalizing flow-based Boltzmann generators, likelihood evaluations scale unfavorably with system size. We show how this issue can be alleviated, demonstrating speedups of up to 100 times on Lennard-Jones systems.

Nonequilibrium settings encompass a wider range of systems. We briefly discuss some of the generative modeling methods appropriate in this setting and present an extension of implicit transfer operator models to nonautonomous domains. By combining flow map matching with a physically grounded short-time inductive bias, we accurately model both long- and short-time behavior of nonautonomous systems.

This work concludes with a discussion of the broader role of generative machine learning methods in computational statistical mechanics, pointing out new applications and possible future research directions.