Sampling and Boltzmann Generators

Boltzmann Generators pair generative models with exact likelihoods and importance sampling to rapidly generate uncorrelated equilibrium samples from molecular systems at thermodynamic equilibrium. This is a hallmark challenge in statistical physics with direct applications to drug discovery and molecular dynamics.

Our group has pioneered several advances in this area: iterated denoising energy matching (iDEM) and progressive inference-time annealing (PITA) for sampling from Boltzmann densities without differentiating the target potential, sequential Boltzmann generators for scalable equilibrium sampling, and amortized sampling with transferable normalizing flows (Prose).

Key Directions

• Diffusion-Based Samplers: Iterated denoising energy matching and progressive inference-time annealing for Boltzmann densities
• Scalable Boltzmann Generators: Sequential and transferable normalizing flows for equilibrium sampling of molecular systems
• Efficient Likelihoods: Few-step accurate likelihoods for continuous flows (FALCON) and forward-only regression training (FORT)
• Autoregressive Approaches: Departing from flow-based paradigms to leverage autoregressive architectures from large language models (ArBG)
• Amortized Sampling: Transferable normalizing flows for amortized sampling across molecular systems (Prose)