Deep potential-driven structure exploration of ice polymorphs.

Abstract

Ice, a ubiquitous substance in nature, exhibits diverse forms under varying temperature and pressure conditions. However, our understanding of ice polymorphs remains incomplete. The directional nature of hydrogen bonding and the complexity of the networks they form pose significant challenges to computational studies of ice structures. In this work, we present an extensive exploration of ice polymorphs under pressure conditions ranging from 1 bar to 10 GPa. We employ an advanced crystal-structure-prediction scheme that integrates an evolutionary algorithm, an active-learning deep neural network potential, and molecular dynamics simulations with ab initio accuracy. Among the 131,481 predicted structures, we successfully identify all experimentally known ice phases within the target pressure range, including the particularly challenging ice IV and V. These phases feature highly intricate H-bond networks, which have hindered previous efforts to fully explore ice structures. Additionally, we identify 34 new ice polymorphs that are potential candidates for experimental discovery. Notably, we predict the existence of a new stable ice phase, ice L, within the temperature range of 253-291 K and pressure range of 0.38-0.57 GPa, exhibiting a unique topology unseen in any known crystals. Our findings highlight the potential for experimental discovery of new ice phases. Furthermore, our approach can be applied to other complex systems, particularly those with network structures.