Neural network potentials (NNPs) provide an efficient and accurate approach for predicting the mechanical and thermophysical properties of high-energy-density materials (HEDMs) under extreme temperature and pressure conditions, where experimental characterization is often limited and first-principles simulations become computationally demanding. In this work, an NNP for HEDMs is developed within a new active-learning framework that automatically eliminates structurally redundant configurations. The resulting NNP predicts the mechanical and thermophysical properties of β-cyclotetramethylene-tetranitramine (β-HMX) with density functional theory (DFT)-level accuracy while achieving orders-of-magnitude improvements in computational efficiency. The predicted bulk modulus of β-HMX at zero pressure agrees with experimental measurements within 13%, representing a substantial improvement over the classical molecular dynamics. Likewise, the volumetric thermal expansion coefficient α, constant-pressure heat capacity Cp, and constant-volume heat capacity Cv are all reproduced with deviations below 1%, demonstrating quantitative agreement with experiments. Beyond reproducing known benchmarks, the NNP enables efficient mapping of the coupled pressure–temperature dependence of thermophysical properties. Specifically, α decreases with pressure in two distinct regimes: a rapid drop at 0–6 GPa, which correlates with densification and the suppression of low-frequency phonon modes, followed by a more gradual decline at higher pressures where phonon stiffening becomes more uniform. Overall, this work addresses an important gap in the characterization of temperature- and pressure-dependent mechanical and thermophysical properties of HEDMs. The proposed training strategy provides a transferable framework for modeling the thermomechanical responses under extreme conditions.
Wei et al. (Tue,) studied this question.