Thermomechanical properties of energetic materials (EMs) exhibit strong dependence on the material microstructures; hence, establishing the structure-property-performance (SPP) linkages is critical for enabling the ‘materials-by-design’ of EMs. However, the process of establishing SPP linkages is currently limited by time-consuming and expensive physical and numerical experiments, making it impossible to explore the design space effectively. In this work, we present a novel deep learning-based framework to dramatically reduce the time and efforts to explore the design space and enhance the chance to engineer material microstructures with desired properties. The proposed novel physic-aware recurrent convolutional neural network (PARC) is trained to rapidly estimate the hotspot ignition and growth of EMs during their shock-to-detonation transition (SDT). PARC is designed to predict the time evolution of temperature and pressure fields by modeling and solving the governing differential equations using convolutional neural networks (CNN), making it “physics-aware” and interpretable. PARC is validated by comparing the prediction results with direct numerical simulation (DNS) calculations. The validation results show that, compared to DNS, PARC can predict the thermomechanical behavior of EM during SDT with high accuracy (within 5% error), despite a significantly reduced computation time (up to 3000 times). In addition, a design optimization scenario was demonstrated where the optimal designs of EM microstructures are derived based on the high-throughput, high-accuracy prediction of PARC. Finally, we show that the interpretable architecture of PARC provides additional lenses for the study of SPP linkages by shedding light on identifying the types of microstructures that lead to high energy concentration.
2022-04-06 08:00:00
