Structure-property-performance linkages for heterogeneous energetic materials using deep-learning generated synthetic microstructures

This work investigates the shock sensitivity of synthetic microstructures generated using a deep neural network specifically Transfer Learning based approach. The structure-property linkages obtained from synthetic microstructures are compared with that of real microstructures obtained from SEM images of pressed HMX materials belonging to three classes (Class-3, Class-5 and FEM). We show that they closely mimic the global and local morphologies (quantified in terms of void sizes, shapes and orientations) of the real microstructures. To investigate energy localization (quantified in terms of hotspot ignition and growth rates), direct numerical simulations are performed on synthetic microstructures. We show that they perform realistically both qualitatively and quantitatively when compared to real microstructures. The ability to generate synthetic stochastic microstructures for ensemble simulations provides a route for energetic material designers to perform in silico experiments on synthetic microstructures and manipulate microstructural characteristics to achieve performance design outcomes.