Conditional GAN for Generation of 3D SOFC Electrode Microstructure Dataset

Synthetic three-dimensional porous structures of solid oxide fuel cell anode are generated using the deep convolutional conditional generative adversarial neural network (DCCGAN). The developed network consists of a generator that produces an artificial structure dataset from random numbers, so-called latent variables, and a discriminator that judges whether the input structure dataset is real or fake. The generator and discriminator are alternately trained to improve their performance in an adversarial manner so that the generator can eventually create realistic structures indistinguishable from the real ones. The training process is performed using the real structure dataset of conventional Ni-YSZ anodes with different solid volume ratios, i.e., Ni: YSZ = 70:30, 50:50, and 30:70. In this study, additional training is performed for the generator to control the volume fraction of the generated structures. First, the values of the volume fractions are input to the generator alongside the latent variables. Then, the volume fraction loss is calculated from the difference between the volume fractions of the generated structure and the input values. Finally, the volume fraction loss is used to train the generator in addition to the ordinal training process. The trained network is validated by comparing the generated structures with the real structures in terms of microstructural parameters such as volume fraction, surface area density, and triple-phase boundary density. This comparison is conducted not only for the structures used for the training but also those that are not used for the training, i.e., a solid volume ratio of Ni: YSZ = 40:60 and 60:40. It is found from the results that the volume fractions of the generated structures are in good agreement with the specified input values to the generator. Also, other microstructural parameters quantified from the artificial microstructures are similar to those of the real structures. In addition, since the developed network does not contain fully-connected layers, and is hence categorized as a fully-convoluted network, structures with arbitrary sizes can be generated using the same generator by changing the dimension of the latent variables. The generated structures are confirmed to be free from digital artifacts such as noise, disconnectivity, and periodicity in structures.