(677c) A Molecular View of Methane Activation on Nickel Surfaces through Enhanced Sampling and Machine Learning Potentials

Computational modeling using first-principles significantly advances catalyst design. However, accurately capturing transient species and dynamic catalytic surfaces at relevant temperatures remains challenging. Machine learned interatomic potentials (MLIPs) balance computational efficiency and accuracy, enabling large-scale molecular dynamics simulations. However, constructing accurate MLIPs for modeling heterogeneous-catalytic reactions is challenging due to the dimensionality of the potential energy surfaces and the need for diverse training datasets. In this work, we address these challenges by integrating enhanced sampling with MLIP development to investigate methane activation on Ni(111) surface. We show that incorporating collective variables (CVs) that capture the behavior of methane molecule improves MLIP initialization, except for configurations requiring significant dynamic responses of the catalytic surface. These discrepancies arise because the CVs primarily capture the bond properties of the reactant molecule but lack sufficient description of the surface dynamics. By adding frames accounting for Ni-atom elevation and performing two additional MLIP refinements, we significantly enhance energy and force predictions (RMSEs: 1.0 meV atom^–1 and 0.02 eV Å^–1). The resulting free energy landscapes at 500-1100 K provide detailed insights into the thermodynamics and dynamics of methane activation. As methane dissociates on the catalytic surface, the process involves a dynamic interplay between CH₄ and the Ni catalyst, including both enthalpic and entropic contributions. The progression toward the transition state is characterized by increasing restraint in the ability of CH₄ to rotate or translate. After the transition state, there is a notable elevation of the Ni atom interacting with the cleaved C–H bond, leading to increased mobility of the adsorbed species—a feature more pronounced at higher temperatures. By integrating enhanced sampling with the development of MLIPs, we underscore the potential of MLIPs to replicate DFT-level accuracy while enabling efficient and dynamic modeling of catalytic systems for mechanistic investigations.