Steering Langevin dynamics towards transition states using collective-variable-free resampling

Exploring the potential energy surface to sample transition state regions is crucial to understand the atomic processes that govern chemical reactivity. Ideally, the exploration does not require any collective variables that are based on prior chemical domain knowledge. With this in mind, we adapt the stochastic saddle point dynamics algorithm (SSPD) and assess its applicability to increasingly complex systems. We motivate the adaptation of the original SSPD algorithm using a simple 2D toy potential and illustrate why it is necessary to converge to reactive transition state regions. We then demonstrate that the adapted SSPD can efficiently sample the high-dimensional isomerization reaction of a seven-atom Lennard-Jones cluster. Combining SSPD with an automatically differentiable machine-learned interatomic potential, we study decomposition reactions of isopropanol. Thereby, we demonstrate how to target distinct reaction pathways in systems with multiple competing reactions. Finally, we apply the same framework to CO dissociation on a flat Co surface and show how finite-temperature effects influence the transition pathway from the initial minimum to the transition state.