Reproduce “Diffusion Models are Molecular Dynamics Simulators” on OpenResearch

About this paper

Justin Diamond, Markus Lill

Researchers demonstrate that denoising-diffusion models, when augmented with a harmonic adapter, are mathematically equivalent to Euler-Maruyama integrators for overdamped Langevin dynamics, creating a data-driven molecular dynamics simulator. This approach generates physically consistent, temporally correlated trajectories with substantial computational speed-ups compared to traditional methods.

Key insights

• A single reverse denoising step, when augmented with a specific harmonic coupling, is mathematically identical to an Euler-Maruyama update for overdamped Langevin dynamics.
• The total simulation error can be rigorously decomposed into a model error (accuracy of learned forces) and a discretization error (number of reverse steps), providing clear pathways for improvement.
• The quadratic coupling inherently enables full time-parallel execution of trajectory segments, a fundamental advantage over sequential traditional MD.

Summary by alphaXiv.