Enabling high throughput deep reinforcement learning with first principles to investigate catalytic reaction mechanisms

Abstract

Exploring catalytic reaction mechanisms is crucial for understanding chemical processes, optimizing reaction conditions, and developing more effective catalysts. We present a reaction-agnostic framework based on high-throughput deep reinforcement learning with first principles (HDRL-FP) that offers excellent generalizability for investigating catalytic reactions. HDRL-FP introduces a generalizable reinforcement learning representation of catalytic reactions constructed solely from atomic positions, which are subsequently mapped to first-principles-derived potential energy landscapes. By leveraging thousands of simultaneous simulations on a single GPU, HDRL-FP enables rapid convergence to the optimal reaction path at a low cost. Its effectiveness is demonstrated through the studies of hydrogen and nitrogen migration in Haber-Bosch ammonia synthesis on the Fe(111) surface. Our findings reveal that the Langmuir-Hinshelwood mechanism shares the same transition state as the Eley-Rideal mechanism for H migration to NH2, forming ammonia. Furthermore, the reaction path identified herein exhibits a lower energy barrier compared to that through nudged elastic band calculation.