Using Ternary Adversarial Networks to Capture Geometric Information in the Reconstruction of Porous Media

Modern X-ray Computed Tomography (CT) technology allows us to show the three-dimensional microstructure of rocks. The rock itself is highly complex, and traditional pore-scale structure modeling methods have been unable to meet the modern demand for high-quality rock microscopic images. Generative adversarial networks (GANs) are now one of the most promising unsupervised learning techniques. Some researchers have begun to apply GANs to porous media reconstruction. The GAN constructs a mapping of the random probability distribution (mean distribution, Gaussian distribution) to the real data distribution. Using the GAN, three-dimensional rock microscopic images can be synthesized from the randomly distributed numbers. Their research results demonstrate the ability of GAN technology in the reconstruction of porous media. However, the generated adversarial network may converge to a local minimum, and it is difficult for the synthesized image to obtain the geometric information of the real image. Therefore, the geometric characteristics of the images generated by general research are often different from the real images. In this study, we built a ternary generation adversarial network structure including a generator and a two-discriminator. The new discriminator is used to discriminate the authenticity of the geometric parameters of the solved image. Research shows that this ternary adversarial neural network can effectively reduce the difference in geometric features between synthetic and real images.