CatFlow: An Automated Workflow for Training Machine Learning Potentials to Compute Free Energies in Dynamic Catalysis

Abstract

Dynamic effects of catalysts play a crucial role in catalytic reactions, necessitating the incorporation of statistical sampling and understanding of the impact of dynamic structures in free energy calculations. However, the complexity of catalytic systems poses challenges in effectively exploring the vast configurational space effectively. In this work, we propose CatFlow, an automated workflow for training machine learning potentials (MLPs) to compute free energies of catalytic reactions. CatFlow combines constrained molecular dynamics (MD) simulation with concurrent training of MLPs and sequential calculation of free energies with well trained MLPs. By rapidly generating reliable MLPs, CatFlow facilitates rigorous free energy calculations, enabling the determination of the reaction profiles in an end-to-end manner. We showcased the capabilities of CatFlow by investigating the activation of O₂ catalyzed by Pt clusters and demonstrated the effects of phase transition on the activities of the catalytic reaction. CatFlow offers an efficient and automated solution for studying the catalytic elementary reaction processes. It reduces the need for human intervention and provides researchers with a powerful tool to investigate free energies of dynamic catalysis.