Data-driven multiscale modeling of crack evolution in compressed energetic composites

Understanding crack evolution from microscale initiation to macroscopic failure has long been a multiscale challenge, particularly in high-energy-density polymer-bonded explosives (PBXs) with pronounced heterogeneity. Here, we develop an advanced hybrid framework integrating discrete element method (DEM) simulations with machine learning (ML) to investigate crack nucleation, propagation, and coalescence in a prototypical HMX-based PBX under uniaxial compression. DEM simulations, calibrated against experimental SEM observations and stress-strain curves, yield high-fidelity spatiotemporal data on crack initiation and evolution, revealing that cracks predominantly nucleate at interfaces between HMX grains and the F2314 binder, driven by combined shear-tensile stresses. The 4.7 million entries of DEM simulation data were subsequently leveraged as input for our customized ML models to predict material failure, delivering satisfactory performance with an accuracy of 0.9776 for spatial prediction and an R² value of 0.9024 for temporal prediction on the validation set. The hybrid framework revealed that fine-grained samples exhibit dense, localized, and short interfacial microcracks that effectively dissipate external energy, whereas coarse-grained samples develop more extensive, interconnected cracks penetrating both interfaces and the F2314 binder. This study develops a multiscale modeling framework for PBXs, which provides mechanistic insight into the spatiotemporal crack evolution and can be generally applied to other anisotropic composite materials.