A neural network potential for shock-induced decomposition of CHNO energetic materials

Early NO₂ generation is identified as a common early-stage decomposition feature across diverse CHNO energetic materials under shock loading, whereas the molecular-level initiation events exhibit strong structural dependence. Herein, we construct a transferable deep neural network potential (NNP) trained on over 650 000 SCC-DFTB + MSST nonequilibrium configurations. This NNP achieves SCC-DFTB-consistent energies, forces, virials and shock dynamics with over 50-fold computational acceleration, and exhibits reliable predictive capability for chemically related unseen energetic compounds. Large-scale MSST simulations reveal two key findings: the investigated CHNO explosives exhibit early NO₂ accumulation, while subsequent initiation mechanisms and reaction pathways depend strongly on molecular structures. Moreover, finite-size effects strongly influence local thermal heterogeneity and product evolution. RDX and TNT exhibit size-dependent changes in OH-to-H₂O conversion together with distinct thermal responses, whereas DNTF shows a different trend characterized by increasing high-temperature volume fraction with system size. Increasing the system size substantially reduces finite-size fluctuations in product evolution, although the system size required to obtain statistically stable behavior remains material dependent. This work identifies common shock-initiation characteristics of CHNO explosives, clarifies size-dependent shock thermochemistry, and provides new opportunities to explore the chemical evolution of energetic materials under extreme shock conditions.