Speaker
Description
We propose a multilevel generative sampling framework for lattice gauge theories designed to overcome topological freezing by explicitly sampling all relevant topological sectors. The target distribution is decomposed into coarse and fine scales using an RG-inspired blocking strategy. At each level, the generative sampler produces plaquette degrees of freedom and reconstructs consistent gauge-link configurations through gauge fixing. A mixture-model architecture, together with deterministic sector-changing transformations, enables efficient transitions between distinct winding-number sectors, avoiding the severe critical slowing down typically encountered in HMC. In numerical tests for the two-dimensional U(1) gauge theory, the method achieves good effective sample sizes covering all relevant topological sectors, and accurately reproduces the topological susceptibility.