Pore Space Reconstruction of Shale Using Improved Variational Autoencoders

Abstract

Pore space reconstruction is of great significance to some fields such as the study of seepage mechanisms in porous media and reservoir engineering. Shale oil and shale gas, as unconventional petroleum resources with abundant reserves in the whole world, attract extensive attention and have a rapid increase in production. Shale is a type of complex porous medium with evident fluctuations in various mineral compositions, dense structure, and low hardness, leading to a big challenge for the characterization and acquisition of the internal shale structure. Numerical reconstruction technology can achieve the purpose of studying the engineering problems and physical problems through numerical calculation and image display methods, which also can be used to reconstruct a pore structure similar to the real pore spaces through numerical simulation and have the advantages of low cost and good reusability, casting light on the characterization of the internal structure of shale. The recent branch of deep learning, variational auto-encoders (VAEs), has good capabilities of extracting characteristics for reconstructing similar images with the training image (TI). The theory of Fisher information can help to balance the encoder and decoder of VAE in information control. Therefore, this paper proposes an improved VAE to reconstruct shale based on VAE and Fisher information, using a real 3D shale image as a TI, and saves the parameters of neural networks to describe the probability distribution. Compared to some traditional methods, although this proposed method is slower in the first reconstruction, it is much faster in the subsequent reconstructions due to the reuse of the parameters. The proposed method also has advantages in terms of reconstruction quality over the original VAE. The findings of this study can help for better understanding of the seepage mechanisms in shale and the exploration of the shale gas industry.