Microstructural heterogeneity affects the macroscale behavior of materials. Therefore, optimizing macroscale material performance requires designing material at the micro-scale. However, conventional numerical approaches face significant challenges in the practical application of multiscale material design, optimization, and uncertainty quantification. To overcome this, we developed an integrated deep learning (DL) and numerical optimization framework for multiscale mechanics modeling and design of material microstructures with enhanced performance. Several numerical examples show applications of this approach in microstructure-resolved multiscale design, uncertainty quantification, and optimization for composites and metals.
Of particular interest is the applicability of this approach to design the material microstructures to maximize their spall strength in a laser-driven spall experiments on thin metal foils. The cost of estimating spall strength experimentally by performing very high strain rate (∼106 s-1) experiments is extremely high. Furthermore, it is computationally very expensive to estimate spall strength by performing microstructure-resolved crystal plasticity finite element (FE) simulations. Therefore, we used a 3D U-Net deep learning model as a surrogate model to efficiently predict the spall strength for any poly-crystalline microstructure and integrated this in a Bayesian optimization framework to design microstructures with enhanced spall strength.
