18–20 May 2026
James Clerk Maxwell Building
Europe/London timezone

Fighting critical slowing down with Neural Enhanced Out-of-Equilibrium Sampling.

18 May 2026, 15:30
30m
Higgs Centre Seminar Room (4305) (James Clerk Maxwell Building)

Higgs Centre Seminar Room (4305)

James Clerk Maxwell Building

School of Physics and Astronomy University of Edinburgh James Clerk Maxwell Building Peter Guthrie Tait Road Edinburgh EH9 3FD
Abstract for talk

Speaker

Elia Cellini (University of Edinburgh)

Description

Critical slowing down is the well-known phenomenon whereby the integrated autocorrelation time of Markov Chain Monte Carlo simulations, and therefore the computational cost, grows rapidly as one approaches the continuum limit. This effect is particularly severe for topological observables in non-Abelian gauge theories.

In this talk, I will present Stochastic Normalizing Flows (SNFs), hybrid algorithms in which non-equilibrium Markov Chain Monte Carlo algorithms are enhanced with normalizing flows, a class of deep generative models. I will show that SNFs significantly mitigate topological freezing in $3+1$ dimensional SU(3) gauge theory, while exhibiting favorable scaling with the lattice volume.

I will then introduce a novel neural enhanced out-of-equilibrium approach for sampling lattice field theory based on Langevin dynamics. I will discuss the conceptual advantages of this framework and present preliminary results for lattice non-linear $\sigma$ models.

Author

Elia Cellini (University of Edinburgh)

Presentation materials