Understanding high pressure molecular hydrogen with a hierarchical machine-learned potential

The hydrogen phase diagram has several unusual features which are well reproduced by density functional calculations. Unfortunately, these calculations do not provide good physical insights into why those features occur. Here, we present a fast interatomic potential, which reproduces the molecular hydrogen phases: orientationally disordered Phase I; broken-symmetry Phase II and reentrant melt curve. The H-2 vibrational frequency drops at high pressure because of increased coupling between neighbouring molecules, not bond weakening. Liquid H-2 is denser than coexisting close-packed solid at high pressure because the favored molecular orientation switches from quadrupole-energy-minimizing to steric-repulsion-minimizing. The latter allows molecules to get closer together, without the atoms getting closer, but cannot be achieved within in a close-packed layer due to frustration. A similar effect causes negative thermal expansion. At high pressure, rotation is hindered in Phase I, such that it cannot be regarded as a molecular rotor phase. Hydrogen has multiple molecular phases which are challenging to explore computationally. The authors develop a machine-learning approach, learning from reference ab initio molecular dynamics simulations, to derive a transferable hierarchical force model that provides insight into high pressure phases and the melting line of H-2.