Neural Network Potentials for Molecule- Surface Interactions

Abstract

A method of approximating the potential energy surface (PES) of molecules interacting with crystalline surfaces was used to find an analytic representation of the interactions between molecular hydrogen and a solid copper surface. The potential energy of the H2/Cu(111) system was sampled using density functional theory with the Perdew-Burke-Ernzerhof exchange-correlation functional. The machine learning framework TensorFlow was used to generate neural network representations of the energy surface from this data set. The final representation accurately reproduces the features of the underlying PES around potential entry channels for hydrogen dissociation, which agrees well with other studies. The model also accurately predicts the zero-point energy of the H2 molecule. In addition to examining the energy landscape around these entry channels the dissociative sticking probability at different translational kinetic energies for the system is estimated using classical trajectory methods. It is found that this produces sticking probabilities that disagree with comparable experimental results; they are in agreement with similar computational estimates, however, and reflect that the sticking probability increases with the kinetic energy in qualitative agreement with experiments.