Stable solid molecular hydrogen above 900K from a machine-learned potential trained with diffusion Quantum Monte Carlo

Predicting the phase diagram of molecular hydrogen remains an important challenge for computational condensed matter physics. Accurate quantum Monte Carlo (QMC) methods are restricted to small system sizes where finite-size effects make it difficult to study the solid-solid and melting phase transitions. Recent advances allowed us to train a machine learned (ML) interatomic potential with many QMC calculations. We have produced a substantial publicly accessible QMC database for training interatomic potentials and used it to train a deep-neural network ML potential using QMC forces. We used this new potential in large-scale path integral molecular dynamics simulations to study molecular hydrogen. We find a phase diagram with HCP and C2/c-24 phases and two new structures with Fmmm-4 molecular centers. The Fmmm-4 structures show a molecular orientational order transition from an ordered low-temperature structure to an isotropic high-temperature phase which melts to a molecular liquid with a maximum melting temperature of 1450K at 150 GPa. This finding will likely lead to new experimental studies of the melting curve for molecular hydrogen.