Abstract：This study proposes an artificial intelligence-based constitutive modeling framework that integrates discrete element modeling (DEM) with machine learning algorithms to efficiently predict the mechanical response of energetic crystals during the compaction process. The primary objective is to overcome the limitations of traditional experimental methods and pure numerical simulations in capturing complex mesoscopic mechanisms and achieving rapid performance forecasting. The research methodology commenced with the development and meticulous calibration of a high-fidelity DEM model capable of accurately replicating the realistic morphology and mechanical response of crystalline particles. This validated model was subsequently employed to systematically generate a comprehensive numerical experimental dataset, comprising 75 distinct samples covering a wide range of median particle sizes (D??), uniformity coefficients (C?), and porosity levels. This process yielded an extensive and reliable database containing 4,500 stress-strain data points, providing a robust foundation for subsequent data-driven analysis. Leveraging this dataset, a comparative investigation was conducted to evaluate the predictive performance of two machine learning algorithms under small-sample conditions: a conventional Artificial Neural Network (ANN) and a Physics-Informed Neural Network (PINN). The study demonstrates that the PINN, which effectively incorporates physical prior knowledge by embedding constitutive constraints—specifically strain-hardening monotonicity and a stress-zero constraint at zero strain—as residual terms into the loss function, significantly outperforms the data-driven ANN. The PINN achieved a remarkable coefficient of determination (R2) of 0.99 on the test set, substantially surpassing the ANN's performance (R2 = 0.92). This result underscores the superior predictive accuracy, enhanced generalization capability, and improved physical consistency of the physics-informed approach. Furthermore, parameter sensitivity analysis confirmed that the established model reliably captures the underlying physical relationships: an increase in median particle size or uniformity coefficient, or a decrease in porosity, leads to a corresponding increase in compressive stress. Crucially, even for parameter combinations not present in the training dataset, the model's predictions remain strictly consistent with established physical trends, demonstrating its powerful extrapolation capabilities. This research provides a reliable data-driven modeling framework for predicting mechanical properties and optimizing processing parameters for energetic crystal materials.
2025, Vol. 46 
