Reconstruction of 3D Nuclear Graphite Based on Generative Adversarial Neural Networks With Gradient Penalty

Abstract Nuclear graphite is an important structural material in reactor, and it is important to study the dehumidification method to improve the economy of reactor. In order to accurately obtain various physical property parameters of nuclear graphite from the pore scale, constructing a 3D digital graphite that can accurately characterizes the pore structure of micropores and nanopores is crucial. Due to the high experimental cost or a long reconstruction time of tradigital reconstruction methods, we use the generative adversarial neural networks with gradient penalty (WGAN-GP) to reconstruct 3D digital graphite. Taking binary image of graphite as training sample for model training, the generator model that can generate 3D graphite images is obtained, and then multiple graphite images are obtained. The generated images and original images were compared and analyzed from two-point statistics, pore distribution, porosity and specific surface area. The results show the reconstructed graphites are consistent with the original graphite, which proves the reliability of this method.