Machine Learning for Atomistic Simulation I: Applications in Materials Science

Pt-functionalized graphene combines graphene's high surface area and excellent electrical properties with platinum's catalytic activity in hydrogen reactions, making it a promising material for hydrogen sensing and storage. However, efficient hydrogen storage is hindered by hydrogen's low volumetric density and the clustering tendency of transition metal atoms, which reduces storage capacity. Atomistic simulations are crucial for understanding the structural dynamics and reactivity of these catalytic systems. However, the poor scaling of ab initio methods limits modeling to small structures, while Pt/graphene systems require large-scale simulations to capture metal nucleation, clustering, and thin-film formation. Recent advances in machine learning interatomic potentials enable large-scale simulations with DFT-like accuracy at reduced computational cost. In this study, we employ equivariant neural network potentials to perform extensive molecular dynamics simulations, predicting equilibrium crystal structures and revealing growth dynamics of Pt on graphene. We identify different growth modes across varied Pt concentrations and demonstrate excellent agreement with experimental measurements via transmission electron microscopy. Furthermore, we model hydrogen adsorption and storage on the optimized Pt/graphene structures, providing an accurate large-scale investigation of Pt-functionalized graphene for hydrogen sensing and storage applications.