Learning to cross barriers: Machine learning for rare event simulations

Rare transitions between long-lived states underlie many important processes in condensed matter, from nucleation to protein folding, yet simulating them remains challenging because of the vast separation of timescales involved. In this talk, I will discuss how machine learning can enhance rare event simulations. In particular, I will show how conditioned normalizing flows can improve path sampling by reducing correlations between sampled trajectories and enabling parallelization. A central theme is the committor function, the ideal reaction coordinate, which can be learned iteratively from transition path sampling data and used to guide the generation of new reactive trajectories. Symbolic regression then yields interpretable analytical expressions that identify the collective variables governing the transition. Finding these few essential variables is a form of physically informed dimensionality reduction. Applied to ice nucleation, this approach shows that nucleus size alone provides a non-Markovian description of the process, while incorporating structural descriptors such as crystalline order and tetrahedrality yields an improved reaction coordinate.