19.11.24 - Chris Chipot "Discovering Reaction Pathways and Committor Probabilities with Machine Learning"

Atomistic simulations encounter significant challenges in sampling transitions between metastable states in the free-energy landscapes that underlie slow molecular processes. Importance-sampling techniques offer a promising approach to expedite these dynamics by reducing free-energy barriers, yet they require low-dimensional reaction-coordinate (RC) models defined through collective variables (CVs). Traditionally, CV selection has relied heavily on human intuition, but machine-learning (ML) algorithms now offer more rigorous alternatives. Here, we introduce variational committor-based neural networks to identify robust RC models for transitions between two metastable states, and further extend this framework to jointly learn the reaction pathways linking these states. Our approach shows strong potential for discovering key descriptors of slow molecular dynamics, and can be incorporated into importance-sampling schemes through reweighting to approximate the kinetics of transitions.