Super-Resolution Reconstruction of Porous Media Using Concurrent Generative Adversarial Networks and Residual Blocks.

Accurate porous media reconstruction has always been one of the significant research hotspots in the numerical simulation of reservoirs. The traditional methods such as multi-point statistics perform porous media reconstruction based on the statistical features of training images, but the process is possibly cumbersome and the result is less effective. Porous media reconstruction has been greatly developed and benefited by applying current flourishing deep learning to its simulation process thanks to the strong capability of extracting features by deep learning. As a typical branch of deep learning methods, generative adversarial network (GAN) can simulate a two-person zero-sum game through confrontation between a generator and a discriminator. However, in real experiments, constrained by the resolution of physical equipment and the size of samples, it is difficult to physically obtain a large-scale image of porous media with high-resolution (HR) since HR and large field of view are usually contradictory for physical equipment. In this paper, a method is proposed based on multistage concurrent GAN to learn the structural features of porous media from one low-resolution 3D image and then stochastically reconstruct larger-sized porous media images. Experimental comparison with some typical methods proves that this method can reconstruct HR images with favorable quality. Article Highlights: Our method is quite fast in multiple reconstructions by reusing model parameters. Our method realizes the super-resolution reconstruction of porous media from one low-resolution image. Our method outperforms SNESIM and some GAN's variants in the accuracy of reconstruction. [ABSTRACT FROM AUTHOR]