Deep learning enhanced estimation of local three-phase microstructural characteristics from ideal 2D or low-resolution 3D data

Abstract Three-dimensional (3D) microstructure reconstruction is essential for understanding processing-microstructure-property relationships in solid oxide cell (SOC) electrodes. While high-resolution 3D techniques like FIB-SEM and nano-CT offer precise local quantification, they are costly, destructive, and limited to small volumes. To help overcome these limitations, convolutional neural networks (CNNs) were investigated for estimating 3D microstructural characteristics from more accessible imaging data: phase-segmented high-resolution 2D SEM images and low-resolution 3D µ-CT scans. Two models were developed: a 2D3D CNN for 2D SEM and a LRHR CNN for µ-CT. Each was trained on an open-source library of synthetic SOC microstructures, evaluated across homogeneous and heterogeneous electrode distributions, and compared to traditional stereological methods for 2D SEM. Up to a resolution of 1.0 µm, the LRHR model performs better than the 2D3D model, and both are significantly better than stereology in estimating either local microstructural characteristics or local electrochemical properties computed from those characteristics. These accurate and precise deep-learning-enhanced estimates of local microstructural characteristics will enable the characterization of large numbers and large volumes of heterogeneous functional materials and devices, which is crucial to determining processing-microstructure-property relationships and will guide further design and optimization.