Prediction of microstructural-dependent mechanical properties, progressive damage, and stress distribution from X-ray computed tomography scans using a deep learning workflow

Creating computationally efficient models that link processing methods, material structures, and properties is essential for the development of new materials. Translating microstructural details to macro-level mechanical properties often proves to be an arduous challenge. This paper introduces a novel deep learning-based framework to predict 3D material stress fields, mechanical behavior, and progressive damage in ceramic materials informed by the microstructural features of the material. We construct a dataset of synthetic representative volume elements utilizing X-ray computed tomography scans and employ an automated finite element (FE) modeling approach to generate datasets of alumina ceramics with varying inclusion morphologies. The deep learning model, a U-Net based convolutional neural network (CNN), is trained to understand the structure-property linkages and mechanical responses directly from FE-generated data without transforming them into image format. The CNN's architecture is optimized for capturing both local and global contextual information from the microstructural data, enabling accurate prediction of stress fields and damage evolution. Inclusions within the material are shown to play a crucial role in the initiation and propagation of damage. The CNN model demonstrated robust performance in predicting the stress field, stress-strain curve, and progressive damage curve, with training and test data both showing high and consistent similarity between predictions and the ground truth. Overall, this research offers a generalized approach that can be adapted for different materials and structures toward creating efficient and accurate digital replicas for optimizing material performance in real-world applications.