| Abstract Scope |
Machine-learned interatomic potentials (MLIPs) offer first-principles accuracy at classical molecular dynamics cost, creating opportunities for the design of energetic molecular solids. We report an E(3)-equivariant graph-neural-network potential for the high-energy-density crystal LLM-105, trained with the Allegro framework on >1,600,000 density-functional-theory configurations from the SPICE dataset, Materials Project trajectories, and stress-controlled deformations tailored to shock-responsive materials. The model attains root mean square energy and force errors of ~120 meV and ~130 meV/Å, and predicts equilibrium lattice parameters within 5% of experiment. Large-scale molecular-dynamics simulations (up to 600K and multi-nanosecond windows) reproduce thermal stability, anisotropic expansion, and elastic constants of LLM-105, while remaining two orders of magnitude faster than ab-initio MD and more accurate than state-of-the-art ReaxFF. These capabilities unlock mesoscale studies of hot-spot formation and shock response. This study contributes to the advancement of data-driven modeling of molecular crystals and highlights the applicability of equivariant neural networks to energetic materials research. |