Just another WordPress site - Ruhr-Universität Bochum

Abstract

We show the predictive power of the atomic cluster expansion (ACE) for modeling challenging systems such as water. We trained ACE on data from ab-initio molecular dynamics simulations (AIMD) of water at 300 K. ACE displays excellent agreement to the first-principles reference data in predicting radial distribution functions and covalent and hydrogen bonding characteristics. However, our investigation confirms that describing properties that require sampling a larger fraction of phase space, for example, self-diffusion or the free energies, requires longer AIMD simulation times at different pressures/densities and temperatures. We then show that ice structures provide a more efficient means of sampling the phase space of water: training ACE to diverse ice phases only describes liquid water in quantitative and qualitative agreement with first principles reference data. This reveals a powerful and efficient strategy for building transferable water interatomic potentials without running expensive AIMD.