Speaker: Alex Tong, AITHYRA

Title: Scalable Equilibrium Sampling via Generative Modelling

Abstract: Traditional equilibrium sampling methods are slow, sequential, and struggle to scale or cross high energy barriers. This talk reframes this challenge as a generative modeling task, focusing on models with fast, tractable likelihoods, such as normalising flows. This likelihood-centric approach provides an unprecedented level of control. It allows us to directly optimise the model against the target distribution, compute exact observables via unbiased importance re-weighting, and actively steer the generation process towards specific regions of interest, such as rare events. We will demonstrate how these methods enables the massively parallel and scalable generation of independent, physically valid configurations, opening a new path for tackling intractable sampling problems in science.

This meeting will take place remotely on Zoom.