An Active Learning Algorithm for Identifying Transition States on a Potential Energy Surface

Mapping reaction pathways on complex potential energy surfaces (PESs) and locating transition states (TSs) is often used for understanding chemical reaction mechanism(s). The nudged elastic band (NEB) method is widely used for this purpose, but it becomes computationally expensive for large systems due to the repeated evaluation of energies and forces. We present an active learning algorithm coupled with the nudged elastic band, AL-NEB, for efficient convergence to the TS. AL-NEB constructs a surrogate PES and actively selects training points in two phases: (a) Exploration-Exploitation and (b) Renunciation. Strategies have been introduced for making the algorithm efficient and stable. We show the efficacy of the algorithm on several 2D analytical potentials, HCN isomerization, keto-enol tautomerization, and high-dimensional heptamer island diffusion (up to 525 degrees of freedom). In all cases, AL-NEB locates the "exact" TS on the chosen model chemistry with an order-of-magnitude fewer force evaluations than the standard NEB, demonstrating its scalability and efficiency.