A Generative Deep Learning Framework for Microstructure-Driven Reliability Analysis

Microstructure informatics is increasingly employed in materials engineering to predict the mechanical behavior of multiphase and polycrystalline composites. Data-driven approaches enable the identification of microstructural features governing the homogenized properties of such microstructured materials. However, their application to reliability assessment in the presence of stochastic effects at the microscale remains limited, particularly when data availability is scarce. This work presents a strategy to address problems involving stress and strain field gradients, where standard homogenization fails. The focus is on detecting micromechanical features, such as defects, that may affect structural reliability. A multiscale-inspired approach is adopted: at the macroscale, convolutional neural networks (CNNs) extract key patterns from micrographs; at the microscale, high-resolution images of selected sub-domains statistically represent the inclusion/pattern geometry. To augment data and improve prediction, a generative adversarial network (GAN) is used to create realistic microstructural samples and simulate the response to the external loading, so to estimate the stress state and its spatial distribution induced by the microstructure. The capability of the proposed approach is demonstrated on polycrystalline silicon showing a columnar microstructure and on a periodic composite, to finally assess the accuracy of the predicted overall elastic properties.