Learning continuum strength models for meso-scale simulations of HMX from molecular dynamics using deep neural networks

The fidelity of meso-scale void-collapse computations for predicting detonation initiation in energetic materials (EMs) hinges critically on the constitutive models employed in simulations. Isotropic rate-dependent plasticity models are popular for meso-scale simulations of void-collapse in EMs under shocks. However, for these models to be physically correct, they must capture molecular-scale physics. In this work, we develop an isotropic rate-dependent Johnson-Cook (JC) plasticity model that is informed by molecular dynamics (MD) simulations. The overall goal is to capture the shear band and hot-spot dynamics observed in MD. To learn the model parameters, the power of deep learning (DL) is leveraged to create a predictive environment for an HMX sample with a void experiencing shock loading. This environment is trained by an ensemble of void collapse simulations; these use an isotropic rate-dependent JC model with varying JC constants. The rapid predictive nature of DL is then used iteratively to identify the JC constants such that meso-scale computations replicate the dynamics of MD void-collapse. This procedure will result in a strength model that incorporates molecular physics and is robust for performing mesoscale computations using isotropic plasticity models.