A Novel AI-Assisted Framework for Microstructural Design of Shocked Materials

Advanced manufacturing technologies have enabled the fabrication of complex metamaterial microstructures. However, their complex structure-property-performance (SPP) relationships require costly and laborious experiments, making it impractical to explore the large and complex design space. In this study, we propose a new design framework to discover optimal microstructure designs using deep neural networks. The proposed framework is demonstrated on a problem of optimizing the shock sensitivity of pressed cyclotetramethylene-tetranitramine (HMX). In this framework, we learn micromorphology descriptors using generative adversarial networks (GAN) to depict the latent space and accelerate the estimation of material properties using Bayesian physics-aware recurrent convolutional neural networks (B-PARC). In conjunction with B-PARC, the differential evolution approach is employed to practically navigate the design space. As intermediate results, the optimal microstructure candidates produced by B-PARC and differential evolution are validated against direct numerical simulations, then the B-PARC is recalibrated with updated data before proceeding to the next iteration. The new framework achieved a significant design improvement compared to the best performing structure in the dataset.