Image analysis and quantification on solid oxide fuel cell anode through inspired CNN based U-Net architecture

This research aims to develop a high precision automated segmentation framework for solid oxide fuel cell microstructures, replacing labour-intensive manual labeling to optimize electrode performance metrics. A novel three-dimensional convolution neural network with an inspired U-Net architecture was developed. Unlike modified conventional U-Nets, the proposed architecture eliminates all pooling and upsampling operations to preserve the full spatial resolution of voxel data. Multi-scale convolutional kernels are used to simultaneously capture fine structural details and broader contextual information while retaining skip connections to support stable training and effective feature reuse. The model was evaluated on the three distinct solid oxide fuel cell dataset and consistently outperformed existing approaches, achieving the lowest pixel misclassification and mean squared errors with image quantification accuracy exceeding 90%. High precision quantification was obtained for key metrics such as average diameter and volume fraction. However, eliminating the structural bottleneck increased sensitivity to local voxel noise, resulting in minor overestimation of triple phase boundaries in complex and high contrast regions. Overall, by preserving the native resolution and utilizing multi-scale feature fusion, this study achieves highly efficient and accurate quantification of triple phase boundary and area interface density. This work establishes a robust deep learning groundwork for the rapid automated characterization of complex energy materials in fuel cell applications.