Speaker
Description
Within the framework of deep learning, generative models are receiving increasing attention due to their ability to generate independent samples starting from a set of training examples. Their application to statistical mechanics is particularly promising, given the difficulty of generating decorrelated samples from complex physical distributions. This has led to a growing line of research in which generative models, and in particular normalizing flows, are integrated directly into atomistic simulation workflows, with encouraging results in condensed matter systems.
In this talk, I will present two approaches in which normalizing flows can be fruitfully applied to condensed matter physics. In the first case [1], we investigate equilibrium simulations of liquid systems by exploring physically informed choices of source distributions that more closely resemble the target distribution. This enables more efficient sampling of equilibrium configurations and facilitates exploration of thermodynamic variables, as well as transformations between different representations of the same physical system.
In the second case [2], we propose the use of conditional normalizing flows to enhance nested sampling in condensed matter systems. By replacing rejection-based Monte Carlo steps, which often constitute the primary computational bottleneck, this approach significantly improves sampling efficiency while preserving accuracy in the estimation of thermodynamic properties.
[2] AC, S. Falkner, P. L. Geissler, and C. Dellago, The Journal of Chemical Physics 162, 184102 (2025).
[3] AC, N. Unglert, S. Falkner, L.B. Pártay, G.K.H. Madsen, C. Dellago, manuscript in preparation (2026);