Extracting ice phases from liquid water: why a machine-learning water model generalizes so well

TL;DR

This study shows that liquid water's local atomic environments resemble those of various ice phases, explaining why machine-learning water models trained on liquid water can accurately predict properties of ice phases.

Contribution

The paper demonstrates that local environments in liquid water encompass those of multiple ice phases, enabling transferability of machine-learning models across water's different phases.

Findings

Machine-learning water models predict ice properties accurately.

Liquid water explores environments similar to diverse ice phases.

Local atomic environments in water are highly transferable.

Abstract

We investigate the structural similarities between liquid water and 53 ices, including 20 knowncrystalline phases. We base such similarity comparison on the local environments that consist of atoms within a certain cutoff radius of a central atom. We reveal that liquid water explores the localenvironments of the diverse ice phases, by directly comparing the environments in these phases using general atomic descriptors, and also by demonstrating that a machine-learning potential trained on liquid water alone can predict the densities, the lattice energies, and vibrational properties of theices. The finding that the local environments characterising the different ice phases are found in water sheds light on water phase behaviors, and rationalizes the transferability of water models between different phases.