Predicting adsorption energies and the physical properties of H, N, and O adsorbed on transition metal surfaces: A machine learning study

Abstract

We present a machine learning study utilizing the Hierarchically Interacting Particle Neural Network to predict the adsorption properties of atomic H, N, and O on various single metal and bimetallic single crystal FCC surfaces. Our trained HIP-NN models directly predict, from atomic species and coordinates of the clean surface, and initial adsorption site, adsorption energies with a MAE of 0.08 eV to 0.16 eV, adsorption heights with a MAE of 0.08 Å to 0.09 Å, and a buckling of the first layer of the substrate with a MAE of 0.07 Å to 0.12 Å, illustrating that a neural network can calculate the physical properties of a surface/adsorbate system along with adsorption energy. Our MAEs for adsorption energy also compare favorably with MAEs presented by other works predicting the adsorption energies of H, N, and O. We also demonstrate that while HIP-NN can predict on these systems without explicit periodic boundary conditions, the inclusion of explicit periodic boundary conditions drastically improves the results. Our results also reveal that the buckling of the first layer of the substrate can be utilized to prune datasets, removing unstable surfaces.