Ensuring the mechanical durability of solid oxide fuel cell electrodes requires identifying where local thermal stresses develop within the porous microstructure during fabrication. Conventionally, numerical simulations such as finite element (FE) analysis are employed for this purpose. However, these methods typically require significant computational resources. This study developed a simulation-informed image-to-image learning framework to predict hotspot regions directly from 2-dimensional cross-sectional phase maps. A Pix2Pix-based conditional generative adversarial network (cGAN) was trained to generate binary hotspot maps, using ground-truth labels derived from FE stress fields computed on 3-dimensional microstructures. To evaluate the impact of morphological realism, 2 synthetic datasets generated via kinetic Monte Carlo sintering simulations were compared: (i) microstructures derived from idealized spherical particle packings and (ii) those from GAN-based packings that better represent irregular pore morphologies. When applied to cathode microstructures reconstructed via focused ion beam scanning electron microscopy, the model trained with the GAN-based dataset achieved a lower mean absolute error and more accurately captured high-stress neck regions. This clearly demonstrates the importance of incorporating realistic pore geometries into the training dataset, particularly when predicting the mechanical response of porous materials. The proposed framework, which combines physics-based simulations with a Pix2Pix-based cGAN, enables rapid and reliable hotspot prediction for porous microstructures. • Predicting thermal stress concentrations in porous bodies using deep learning model. • Development of Pix2Pix model based on generative adversarial networks. • Novel numerical methodology for rapid generation of extensive training datasets. • Successful application to actual cathode microstructures obtained via FIB-SEM.
SHI et al. (Thu,) studied this question.