This paper introduces HDRL-FP, a reaction-agnostic framework using high-throughput deep reinforcement learning with first principles to explore catalytic reaction mechanisms. HDRL-FP represents reactions using atomic positions mapped to first-principles potential energy landscapes, enabling thousands of simultaneous simulations on a single GPU for rapid convergence. Its effectiveness is demonstrated by studying hydrogen and nitrogen migration in Haber-Bosch ammonia synthesis on Fe(111), revealing a shared transition state for Langmuir-Hinshelwood and Eley-Rideal mechanisms for H migration to NH2, forming ammonia with a lower energy barrier than nudged elastic band calculations.

Tian Lan, Huan Wang, Qi An