Phase behavior of a machine-learning potential trained on stress–strain curves: The case of superionic water ice

Abstract

We analyze the transferability of a Deep Potential Machine Learning (DP-ML) model trained to reproduce stress–strain curves of high-temperature/high-pressure crystalline phases of water, determining the coexistence lines for the phase transitions between the insulating ice X and the superionic ice XVIII and that between ice XVIII and its melt. Using a set of various free-energy calculation techniques, we find the resulting coexistence lines to be in good agreement with previous data, indicating that the deformation-trained DP-ML model also transfers to thermodynamic properties. This suggests that the inclusion of deformed solid states in training sets may also be a beneficial general strategy in the development of ML interaction models for other condensed-matter systems. Furthermore, the DP-ML model should be useful to investigate other aspects of the considered phase transitions. One of these involves the possible characterization of the XVIII–liquid transition as weakly first-order, with its potentially associated continuous-like behavior. This is an interesting prospect since it might be the first example of such a transition in a three-dimensional structural solid–liquid transformation.