What it is
Lattice gauge theory computes the physics of gauge theories, such as the one behind the strong nuclear force, by sampling field configurations on a grid, traditionally with Monte Carlo methods. The authors show that gauge-equivariant diffusion models can accurately model non-Abelian lattice gauge theory when sampled with the Metropolis-adjusted annealed Langevin algorithm (MAALA), using two-dimensional U(2) and SU(2) gauge theories as examples. The network is built from lattice gauge equivariant convolutional neural networks (L-CNNs), which respect the lattice's local and global symmetries, and each model is trained on a single ensemble generated with a traditional Monte Carlo method. Judged by Wilson loops of various sizes and by the topological susceptibility, the approach generalizes remarkably well to larger inverse couplings and lattice sizes with negligible loss of accuracy, while keeping moderately high acceptance rates (the share of proposed configurations the sampler's accept-or-reject step keeps).
Why it matters
Trained on a single ensemble, the model held its accuracy at larger inverse couplings (weaker coupling) and lattice sizes, the transfer a learned sampler needs. The network builds in the lattice's local and global symmetries and the sampler adds a Metropolis accept-or-reject correction; the abstract does not say which feature drives the generalization.
Every metric behind this entry is listed, with its source, under Sources and data below.
Filed underQuantum Chromodynamics and Particle Interactions, Particle physics theoretical and experimental studies, Theoretical and Computational Physics