Uncertainty Propagation from Imperfect MLIP Simulators to Rare-Event Kinetics

Machine-learned interatomic potentials (MLIPs) require reliable uncertainty quantification, but existing approaches have largely focused on pointwise predictions, leaving uncertainty in kinetic observables largely unexplored. We introduce a framework to propagate parametric MLIP uncertainty to rare-event transition probabilities by combining Adaptive Multilevel Splitting (AMS) with Girsanov path-space reweighting. Reactive trajectories are generated once under a nominal potential and then reweighted to estimate transition probabilities for alternative parameter realizations without additional rare-event simulations. For MLIPs linear in their uncertain parameters, we derive an efficient and numerically stable first-order approximation. We validate the approach on toy models and apply it to the conformational transition of butane using MACE foundation models and POPS uncertainty quantification. In both accurate and misspecified surrogate regimes, the propagated uncertainty encompasses the reference prediction while reducing the computational cost by approximately three orders of magnitude compared with independently rerunning AMS for each parameter realization.