Reactive Active Learning: An Efficient Approach for Training Machine Learning Interatomic Potentials for Reacting Systems.

Discovering chemical reaction pathways using quantum mechanics is impractical for many systems of practical interest because of unfavorable scaling and computational cost. While machine learning interatomic potentials (MLIPs) trained on quantum mechanical data offer a promising alternative, they face challenges for reactive systems due to the need for extensive sampling of the potential energy surface in regions that are far from equilibrium geometries. Unfortunately, traditional MLIP training protocols are not designed for comprehensive reaction exploration. We present a reactive active learning (RAL) framework that is designed to efficiently train MLIPs to achieve near-quantum mechanical accuracy for reactive systems for situations where one may not have prior knowledge of the possible transition states, reaction pathways, or even the potential products. Our method combines automated reaction exploration, uncertainty-driven active learning, and transition state sampling to build accurate potentials. We demonstrate RAL's effectiveness across three different systems: uncatalyzed ammonia synthesis (gas-phase), methanimine hydrolysis (solution phase), and methane activation on titanium carbide surfaces (heterogeneous). The resulting MLIPs accurately predict reaction barriers and transition states. For catalysis, we show that RAL-trained MLIPs identify Ti