Application of machine learning to study the effect of damage on sensitivity of energetic materials at the meso-scale

Damage in energetic material (EM) microstructures can impact performance in—or even cause failure of—devices critical to national security and safety. Therefore, it is important to be able to model microstructural damage and to study the effect of loading on the response of the material at various levels of damage. Here, we present a framework called HEDS (Heterogeneous Energetic Damage Simulator), a tool to generate varied levels of damage in microstructure images of one type of plastic bonded explosive (PBX). The workflow in HEDS starts with preprocessing and importing scanned cross-sectional images of the PBXs. HEDS uses deep learning to identify areas of damage in existing images of PBX, remove them from the image, and allows the user to adjust the extent (volume fraction) of damage in the same microstructure. This architecture consists of two separate U-Nets to perform the tasks of semantic segmentation (classification and extraction of damage from PBX images) and image inpainting (generating images of pristine microstructures with no damage). HEDS stores a library of extracted damage patterns that it draws from the segmentation, which can be enriched by the user through various affine transformations of the damage pattern. When damage needs to be added back into the microstructure, the patterns of damage are drawn from this “damage library”. By progressively adding back damage into the undamaged (in-painted) microstructure comparison of the sensitivity to shock loading of PBX with varying degrees of damage is performed. Finally, we present results of shock simulations from damaged microstructures generated by HEDS.