Systematic and unbiased pathway exploration by artificial force application to a generic neural network potential

Abstract

Computational pathway exploration can unravel complex catalytic mechanisms and even predict unexplored catalytic reactions when performed in a fully systematic and unbiased manner. However, such a comprehensive exploration typically requires years of computation or thousands of CPU cores even for small systems. Herein, a generic neural network potential (NNP) trained on a large structure–energy database and force- and kinetics-based pathway exploration algorithm were combined without any tuning. An interface code was developed to combine the NNP in the Matlantis platform with the search algorithm in the GRRM software. The combined approach automatically generated comprehensive pathway ensembles containing over 10,000 local minimum structures for methane oxidation on the Pd(111) and Pd(100) surfaces with reasonable computational effort. The kinetically most plausible mechanism derived from the ensemble was qualitatively consistent with that obtained by density functional theory. These results highlight the considerable predictive power of the proposed approach at low computational cost.