Agent-driven multi-scale simulation for predicting the catalytic activity of complexes

This paper presents an agent-driven multi-scale simulation framework for efficiently and accurately predicting the catalytic activity of complexes.This framework constructs the reaction path search as a Markov decision process, adopts hierarchical reinforcement learning agents to actively explore the potential energy surface, and combines the equivariant graph neural network potential function to ensure quantum accuracy.Experiments on the open catalyst project (OC20) dataset show that the average absolute error of this framework in adsorption energy prediction is significantly reduced to 0.291 eV, the force prediction error is 0.072 eV/Å, and it can converge to a stable configuration in an average of only 18.3 steps.It is superior to the existing mainstream methods in both accuracy and efficiency.This research provides a new paradigm of intelligent computing for catalyst design and promotes the development of multi-scale simulation towards autonomous decision making and efficient exploration.