Learning reaction mechanisms from trajectory sampling

The problems of sampling rare molecular events and of understanding the underlying reaction mechanism are intertwined: any progress in solving one problem should help solve the other. Building on this connection, we combine enhanced sampling with machine learning techniques to solve both problems at once. We use transition path sampling to learn the commitment probability (or, in short, committor) of reaching the product state instead of returning to the reactant state. In turn, we use the learned commitor to boost the efficiency of transition path sampling by a judicious choice of the points from which to initiate new trajectories. The learned committor model is validated on the fly by comparing its predictions for the trajectory endpoints to the observed outcome of actual simulation trajectories. Using these outcomes as input, the committor model is updated if needed. In this presentation, we will present new developments of the aimmd methodology of AI-based molecular mechanism discovery and its applications to autonomous rare-event sampling, state identification, and mechanism learning.