Transferable Learning of Reaction Pathways from Geometric Priors

Identifying minimum-energy paths (MEPs) is a crucial application of molecular simulations to understand chemical reaction mechanisms but is computationally demanding. We introduce MEPIN, a scalable machine-learning method that predicts MEPs from reactant and product configurations, without relying on transition-state geometries or preoptimized reaction paths during training. The MEPIN task is defined as predicting the deviation between ground-truth MEPs and purely geometric interpolations along the reaction coordinates. The model predicts a continuous reaction path using a symmetry-broken equivariant neural network architecture that generates a flexible number of intermediate structures. MEPIN is trained on an energy-based objective, and we report efficiency gains of also using geometric priors from geodesic interpolation as initial interpolations or as pretraining objectives. The approach generalizes across diverse chemical reactions and achieves accurate alignment with reference intrinsic reaction coordinates, as demonstrated in various small molecule reactions and [3 + 2] cycloadditions.