Speaker
Description
Gauge-field generation in lattice QCD is dominated by repeated solves of the Dirac equation. While multigrid preconditioners reduce solve costs significantly, their expensive setup phase limits overall efficiency. We introduce a gauge-equivariant neural network that accelerates this setup by using approximate low modes from previous configurations, exploiting the autocorrelation along the Markov chain. We tested the method on quenched ensembles at beta=6 with two lattice sizes (8^3 x 16 and 16^3 x 32). Comparing our setup method to a standard multigrid setup with similar computational cost, we find that our setup method leads to a substantially faster multigrid solve. Remarkably, networks trained on the smaller volume generalize to the larger lattice without the need for retraining. These results demonstrate a promising path toward more efficient large-scale lattice QCD simulations.