Stochastic reconstruction of porous media based on attention mechanisms and multi-stage generative adversarial network

The accurate reconstruction of porous media is difficult to accomplish in practical research due to the intricacy of the internal structure of porous media, which cannot be described only using some equations or languages. The emerging deep learning methods open up a new research field for the reconstruction of porous media. One of the classic deep learning methods is generative adversarial network (GAN), which has been applied to the reconstruction of porous media, but also requires sufficient training samples as well as a long simulation process. Some GAN's variants, such as the single-image GAN (SinGAN), are proposed to achieve desired results with only one training image (TI), but its serial structure still necessitates lengthy training time. Based on SinGAN, a stochastic reconstruction method of porous media combining attention mechanisms with multiple-stage GAN is proposed, focusing on important features of porous media with a single TI to achieve favorable simulation quality and using two discriminators to maintain enough diversity (one discriminator prefers real data, and the other favors fake data). Experiments prove that our method outperforms some numerical reconstruction methods and SinGAN in terms of practicality and efficiency.