FFLUX is a multipolar machine-learned force field that uses Gaussian process regression models trained on data from quantum chemical topology calculations. It offers an efficient way of predicting both lattice and free energies of polymorphs, allowing their stability to be assessed at finite temperatures. Here the Ih, II and XV phases of ice are studied, building on previous work on formamide crystals and liquid water. A Gaussian process regression model of the water monomer was trained, achieving sub-kJ mol-1 accuracy. The model was then employed in simulations with a Lennard-Jones potential to represent intermolecular repulsion and dispersion. Lattice constants of the FFLUX-optimized crystal structures were comparable with those calculated by PBE+D3, with FFLUX calculations estimated to be 103-105 times faster. Lattice dynamics calculations were performed on each phase, with ices Ih and XV found to be dynamically stable through phonon dispersion curves. However, ice II was incorrectly identified as unstable due to the non-bonded potential used, with a new phase (labelled here as II' and to our knowledge not found experimentally) identified as more stable. This new phase was also found to be dynamically stable using density functional theory but, unlike in FFLUX calculations, II remained the more stable phase. Finally, Gibbs free energies were accessed through the quasi-harmonic approximation for the first time using FFLUX, allowing thermodynamic stability to be assessed at different temperatures and pressures through the construction of a phase diagram.
Keywords: ice structures; machine learning; polymorphism; quantum chemical topology; quasi-harmonic approximation.
open access.