Speaker
Description
In the last few years, generative machine learning methods have been explored as possible alternatives to classic Markov chain Monte Carlo (MCMC) methods in lattice field theory. A key advantage being that generative models such as normalizing flows and diffusion models have by design no autocorrelation between generated samples. In this vein we have recently introduced gauge-equivariant diffusion models for lattice gauge theories. Using lattice gauge equivariant convolutional neural networks (L-CNNs) and the Metropolis-adjusted annealed Langevin scheme we are able to accurately sample 2D U(2) and SU(2) field configurations, and most recently also 4D SU(3) configurations. Additionally, our models extrapolate remarkably well to larger lattices and larger inverse couplings than originally trained with. We confirm this by comparing observables to analytically known values and to hybrid Monte Carlo results.