Structure-property-performance linkage using machine learning based multiscale models for shocked heterogenous materials

Several engineering applications require design and development of new materials with controlled performance which makes it essential to model the Structure-Property-Performance linkages. However, it is challenging to quantify the material structure since it requires an understanding of the underlying physics governing the material behavior. This work helps to develop multiscale models using both conventional machine learning and deep learning techniques to efficiently and accurately link the salient features of the material microstructure to its effective properties and ultimately to the macroscale performance for shocked heterogenous materials.

The material microstructure quantification is performed on real images using both conventional techniques and deep convolutional neural networks. The material response to shock loading at the mesoscale 𝑂(𝜇𝑚) are numerically computed using a reactive parallel sharp interface hydrocode. The homogenized response or properties of interest are quantified using surrogate models and passed on to the macroscale 𝑂(𝑚). The macroscale performance is computed using the same framework and validated against previously performed experiments.