An Innovative Application Of Generative Adversarial Networks For Physically Accurate Rock Images With An Unprecedented Field Of View

High-resolution X-ray microcomputed tomography (micro-CT) data are used for the accurate determination of rock petrophysical properties. High-resolution data, however, result in a small field of view, and thus, the representativeness of a simulation domain can be brought into question when dealing with geophysical applications. This paper applies a cycle-in-cycle generative adversarial network (CinCGAN) to improve the resolution of 3-D micro-CT data and create a super-resolution image using unpaired training images. Effective porosity, Euler characteristic, pore size distribution, and absolute permeability are measured on super-resolution and high-resolution ground-truth images to evaluate the physical accuracy of the proposed CinCGAN. The results demonstrate that CinCGAN provides physically accurate images with an order of magnitude larger field of view when compared to typical micro-CT methods. This unlocks new pathways for the geophysical characterization of subsurface rocks with broad implications for flow modeling in highly heterogeneous rocks or fundamental studies on nonlocal forces that extend beyond domain sizes typically used for pore-scale simulation.Plain Language Summary Digital rock analysis aims to characterize the physical structure of porous rocks with the aid of X-ray computed microtomography for the simulation of effective properties. The accuracy of these simulations to actual physical rock properties is highly dependent on how well the rock pores are resolved in addition to the particular setup and solvers used. We expect to obtain accurate results from high-resolution images with a large enough field of view to fully characterize microstructural heterogeneities given an appropriate numerical scheme. However, high-resolution data are obtained at the cost of a smaller field of view due to technological limitations. To circumvent this issue, super-resolution methods based on deep learning algorithms can be applied to improve image resolution by learning the mappings between low- and high-resolution data. The training data, however, often require the pairing of low-/high-resolution images, which is obtained by a time-consuming process of image registration. We apply a novel deep learning algorithm to achieve 3-D super-resolution images using unpaired training data. We demonstrate that super-resolution data can provide physically accurate rock images compared to high-resolution ground-truth images. The presented method provides new opportunities to study core-scale properties, such as heterogeneity and flow dynamics, at a resolution high enough to resolve the microscopic rock structure.Key PointsThe connectivity and petrophysical properties of super-resolution data are comparable to that provided by high-resolution ground truth data The presented approach offers a unique opportunity to provide physically accurate digital rock images with a large field of view The methodology narrows the gap between sample sizes used for digital rock studies and those used for petrophysical experiments in laboratory