Mind the gap: Machine learning for rare event simulations

Speaker: Prof. Christoph Dellago, University of Vienna
Series: Lennard-Jones Centre

Rare transitions between long-lived states govern processes ranging from crystal nucleation to protein folding. Simulating these events remains challenging because of the vast separation between molecular timescales and transition waiting times. In this talk, I will discuss how machine learning can help overcome this challenge. Central to this approach is the committor function, the ideal reaction coordinate, which is learned from transition path sampling data and used to guide the generation of new reactive trajectories. Symbolic regression distills the learned committor into interpretable analytical expressions, identifying the collective variables that govern the transition. This combination turns the search for a reaction coordinate into a problem of physically informed dimensionality reduction. Applied to ice nucleation, the approach reveals that dynamics described by nucleus size alone are non-Markovian. Including structural descriptors such as crystalline order and tetrahedrality yields an improved reaction coordinate.