arXiv · 2611.23399
Generative diffusion models for lattice field theory
Abstract
Generative diffusion models offer a new route to ensemble generation in lattice field theory. The backward, or denoising, process underlying these models is a stochastic process with a learned, time-dependent drift, which leads to a clear connection with stochastic quantisation. After reviewing this correspondence, I turn to theories with a sign problem. There the complex Boltzmann weight obstructs direct sampling, but complex Langevin dynamics can sometimes evade the sign problem by generating configurations on a complexified manifold, distributed according to a real and semi-positive distribution. This distribution is not known in closed form. I discuss how this distribution can be learned from complex Langevin data using score-based and, in particular, energy-based diffusion models. The latter give direct access to the unnormalised distribution itself, which can subsequently be used as the weight in a standard Markov chain Monte Carlo simulation. Results are shown for a simple quartic model with a complex action.
Explore related subjects
Keep this discovery
Explore connections, maps & timelines
Gert Aarts, Diaa E. Habibi, Lingxiao Wang, Kai Zhou. 2026-11-17. Generative diffusion models for lattice field theory. https://arxiv.science/abs/2611.23399
Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite arXiv Science by Indiciae: BibTeX · RIS.