Phase Segmentation of Ni/YSZ Anode for Solid Oxide Fuel Cells by Deep Learning

Abstract

A deep learning network is implemented for the phase segmentation of Ni/YSZ anode of solid oxide fuel cells (SOFCs), as manual segmentation of focused ion beam-scanning electron microscopy (FIB-SEM) images is time-consuming. Segmentation is performed on two samples with different volume ratios of Ni to YSZ. The mean intersection over union (mIoU) reaches 0.9363, indicating good agreement between the predicted images and the manually segmented images. Furthermore, the impact of image augmentation on segmentation accuracy is investigated, and it is found that augmentation is necessary when the number of images in the training set is small or the training and testing sets come from different samples. Furthermore, mixing a sixth of manually segmented images from a new sample into the training set can significantly improve the accuracy of automated segmentation for the new sample. This strategy reduces the number of images requiring manual segmentation and improves processing efficiency.