Consistent Sampling and Simulation - Molecular Dynamics with Energy-Based Diffusion Models

Learn about energy-based diffusion models for molecular dynamics simulation in this 44-minute conference talk that addresses critical inconsistencies between diffusion sampling and energy-based interpretations in biomolecular systems. Explore how classical diffusion models, while effective at recovering training distributions, often fail to maintain consistency with the underlying energy landscape, particularly at small diffusion timesteps where models struggle to satisfy the Fokker-Planck equation. Discover a novel approach that introduces Fokker-Planck-derived regularization terms to enforce consistency between sampling and simulation processes. Examine practical applications through demonstrations on fast-folding proteins and the development of a state-of-the-art transferable Boltzmann emulator for dipeptides that achieves improved consistency and efficient sampling. Gain insights into the theoretical foundations connecting diffusion models to molecular dynamics, understand the sources of observed inconsistencies in current approaches, and learn about solutions that enable more reliable energy-based interpretations of learned scores for biomolecular simulation and drug discovery applications.