(169t) Transferable Water Potentials Using Equivariant Machine Learning Interatomic Potentials

The ability of machine learning models to learn the underlying physics is critical for the transferability of models to different thermodynamic states(e.g., temperatures, phases, interfaces, etc.). Recent work^[1,2] suggests that MLIPs trained on solely liquid-phase water (utilizing the MB-Pol potential as a ground truth) cannot accurately describe other phases such as the vapor-liquid equilibrium or gas-phase. An analysis of the many-body decomposition analysis of gas-phase water clusters shows that MLIPs rely greatly on error cancelation rather than reproducing fundamental interactions, limiting transferability. In this work, we demonstrate it is possible to train equivariant Allegro MLIPs on only 3,200 liquid-phase water structures and for the resulting MLIP to reproduce liquid-phase water properties (e.g., density within 0.003 g/cm3 between 230 and 365 K), vapor-liquid equilibrium density up to 550 K, and the relative energy and the vibrational density of states of ice phases. The many-body decomposition analysis of gas-phase water clusters up to six-body interactions shows that fundamental interactions are well described when equivariance is enabled (l=2), but results similar to the previous works are observed when we remove equivariance (l=0). The equivariant MLIP architecture appears to be capable of encoding complex descriptors which allow for improved transferability to arbitrary phases of water while remaining stable for nanoseconds to allow for full equilibration.