Diffusion Accelerants: Towards Augmenting Molecular Dynamics with Learned Measure Transport

The computational cost of molecular dynamics simulation has motivated the development of surrogates based on generative models, yet the physical fidelity and generalization ability of such surrogates remain persistent challenges. Here, we introduce a hybrid algorithm synergizing numerical simulation with a learned non-Markovian biasing force, inspired by diffusion models, for accelerating the time evolution of simulated systems. Our algorithm admits two flavors---one based on successive intervals of accelerated evolution, and the other based on continuous acceleration of two replicates. Preliminary results on toy systems (double well and Müller-Brown) support the potential efficacy of both variants of our method.