Applied Computing and Geosciences (Dec 2024)
Enhancing prediction of fluid-saturated fracture characteristics using deep learning super resolution
Abstract
Utilization of subsurface resources is essential to achieve energy sustainability including large-scale CO2 sequestration, H2 storage, geothermal energy extraction, and hydrocarbon recovery. In-situ visualization of fluid flow in geological media is essential to understand complex, coupled, physical and chemical processes underlying fluid injection, storage, extraction. X-ray Computed Tomography (CT) in the laboratory has proven beneficial to visualize changes in the flow field with rapid temporal resolution (10’s s) and moderate spatial resolution (100’s μm). There is a trade-off between spatial and temporal resolution that limits accurate characterization of dynamics in rock features that are below spatial resolution of CT. While past literature has offered solutions to improve resolution of CT rock images, including deep learning-based algorithms, our study uniquely focuses on improving dynamic, partially and fully fluid-saturated geological images. Fluid-saturated CT images offer additional information, through augmented signals provided by the presence of fluid. Among challenges, CT images of geological media inherently possess limited information due to their single-channel gray-scale source. Additionally, fluid flows through partially saturated media frustrate existing super resolution techniques because unsaturated CT images are an inaccurate proxy for saturated dynamic rock images. The novelty of this work is the expansion of a generative adversarial network (GAN) for applications involving super resolution of partially saturated low resolution CT images using end-member, unsaturated high resolution μCT images. To this end, we acquired multiscale low- and high-resolution CT rock images in unsaturated and saturated states. Among GAN and convolutional neural networks, GAN’s produce realistic high-resolution reconstructions of saturated geological media when trained using high-resolution, unsaturated images and lower resolution images in various saturation states. The model has direct usefulness for interpretation of real-time images.
