A systematic framework for constructing reactive machine learning interatomic potentials

Machine learning interatomic potentials (MLIPs) have emerged as an essential tool for extending the reach of atomistic simulations, offering near quantum level accuracy at a fraction of the computational cost. However, constructing reactive and transferable MLIPs that describe rare events, transition states, and high energy intermediates remains a significant challenge. While the active learning (AL) protocol has proven effective for iterative MLIP refinement, its integration with enhanced sampling techniques has typically been implemented in an ad hoc and problem specific manner, which limits its reproducibility and generalizability. Here, we present a systematic methodology for automated construction of reactive MLIPs using SPARC, an open-source workflow package that integrates AL with enhanced sampling through the PLUMED library. SPARC combines (i) query-by-committee uncertainty estimations to identify configurations where the ensemble of ML models disagrees, with (ii) enhanced sampling methods to access reactive regions of the potential energy surface that are otherwise inaccessible in unbiased molecular dynamics. We demonstrate the robustness of the workflow across three chemically distinct reactive systems, showing that stable and accurate reactive MLIPs can be constructed from minimal initial training data. In all cases, the workflow discovers high-energy configurations, including transition states.