Multi-dimensional neural network interatomic potentials for CO on NaCl(100)

The advent of machine learning (ML) models has unlocked new possibilities in the realm of interatomic potentials. An equivariant, graph neural network (NN), NequIP, recently developed by Batzner et al., [1,2] is utilized for a versatile system, CO on a NaCl(100) surface, to construct interatomic potentials and mediate efficient large-scale atomistic simulations with ab initio molecular dynamics accuracy. We report two neural network potentials, one trained on equilibrium configurations at finite temperatures (T = 30, 300 K), and the other additionally trained upon non-equilibrium trajectories of pre-excited CO adsorbates. We demonstrate first applications of the ML potentials for (i) adsorption energies and barriers for reactions, (ii) potential energy landscapes for submonolayer and monolayer coverages, (iii) vibrational spectra at finite temperatures as well as (iv) vibrational relaxation dynamics. Further possible applications are discussed.