Steering Langevin Dynamics toward Transition States Using Collective-Variable-Free Resampling

Exploring the potential energy surface to sample transition-state regions is essential to understanding the atomic processes governing chemical reactivity. Ideally, the dividing surface between the educt and product states can be sampled without requiring predefined collective variables. Here, we adapt the stochastic saddle point dynamics (SSPD) algorithm by constraining the accessible configuration space according to the number of negative Hessian eigenvalues and evaluate its performance across increasingly complex systems. We motivate the adaptation using a simple two-dimensional model potential and demonstrate how the algorithm can efficiently sample the isomerization reaction of a Lennard-Jones cluster and the decomposition reactions of isopropyl alcohol. Combining the SSPD with automatically differentiable machine-learned interatomic potentials, we apply the approach to CO dissociation on a Co(001) surface both with and without explicit water solvation. The results highlight the role of SSPD as a framework for sampling transition-state regions in complex systems at finite temperatures and demonstrate its versatility in situations where it is not known a priori whether the reaction is governed by energetic or entropic contributions.