UC Davis

The advent of machine learning potentials (MLPs) provides a unique opportunity to access simulation time scales and to directly compute physicochemical properties that are typically intractable using density functional theory (DFT). In this study, we use an active learning curriculum to train a generalizable MLP using the DeepMD-kit architecture. By using sufficiently long MLP-based molecular dynamics (MD) simulations, which provide DFT-level accuracy, we investigate the diffusion of key surface-bound adsorbates on a Ag(111) facet. Detailed analysis of the MLP/MD-calculated diffusivities sheds light on the potential shortcomings of using DFT-based nudged elastic band to estimate surface diffusion barriers. More generally, while this study is focused on a specific system, we anticipate that the underlying workflows and the resulting models can be extended to other adsorbates and other materials in the future.