Generative Gibbs sampling: Composing generative models with explicit physical context

Generative models are powerful when training data for the target distribution are available, but in scientific computing the target is often a joint system for which no such data exist. We propose Generative Gibbs Sampling, a framework that decomposes the joint system into subsystems, trains a generative model on each, and composes these pretrained models at inference time with the physical context they were not trained on. We prove this composition is exact at any noise level when the coupling between subsystems is quadratic. This exactness result identifies and removes the finite-noise bias of the split Gibbs sampler for linear inverse problems. In the quantum setting, the framework recovers the quantum ensemble of liquid water from a model trained on classical statistics, at an effective sample rate up to two orders of magnitude beyond that of path-integral molecular dynamics. With non-quadratic interactions, it forms the correct peptide dimer complexes from monomer models that have no knowledge of the complex.

Speaker bio

Weizhou Wang is a PhD student in theoretical chemistry at the University of Chicago, advised by Aaron Dinner (Chicago) and Jonathan Weare (NYU). He works on generative models for statistical physics, exploiting the mathematical structure of the target. He holds the Neubauer Family Distinguished Doctoral Fellowship and the Josef Fried Graduate Fellowship.

Series: MSR New England Generative Modeling & Sampling Seminar