Deep convolutional neural network for quantification of tortuosity factor of solid oxide fuel cell anode

Abstract

A deep convolutional neural network model (DCNN) is developed to quantify the tortuosity factor of porous electrodes of solid oxide fuel cells (SOFCs). The DCNN model is first trained using synthetic three-dimensional (3D) structure datasets generated by either a generative adversarial network (GAN) or a simple sphere-packing algorithm. Validation of the trained DCNN model is performed by applying it to the analysis of real Ni-YSZ anode structure datasets obtained by focused ion beam scanning electron microscopy (FIB-SEM). Sufficient estimation accuracy is achieved when the DCNN model is trained using the structures generated by the GAN, with the coefficient of determination of 0.9586, 0.9839, and 0.8674 for the Ni, YSZ, and pore phases, respectively. To further improve the model accuracy, the data adjustment is applied to the training datasets to equalize the frequency distribution of the tortuosity factors. As a result, the estimation accuracy is notably improved for all phases, particularly the pore phase achieving the coefficient of determination of 0.9230. In addition, since the DCNN model is designed to analyze 3D structures with arbitrary sizes in each dimension by adopting the global average pooling layer, its estimation accuracy for analyzing structures with different volumes is investigated. The results demonstrate that the developed DCNN model has sufficient estimation accuracy even for larger structures than those used in the training. The constructed DCNN model represents a significant step towards surrogate modeling for structure quantification of energy materials.