Improved Visualization of Reactive Transport Dynamics in Fractured Shales using Computed Tomography and Deep Learning Super Resolution

Abstract:

The complex dynamic interaction of hydraulic fracturing fluid with fracture surfaces and existing reservoir fluids governs hydrocarbon production and performance. Acidic aqueous fracturing fluids as well as nonaqueous fluids, such as CO₂, may cause dissolution or precipitation upon contact with minerals. Such reactions have significant impact on porosity and permeability. While static geochemical models and numerical simulations exist for such studies, experimental observation of spatial and temporal changes in porosity and geochemistry of representative samples is necessary to validate model predictions. Computed tomography (CT) imaging conventionally is used to image flow during nonreactive (core-flooding) experiments. The resolution of CT images, however, is limited by the scanning device. There is a trade-off of spatial resolution and how rapidly an image is acquired. The field of view during dynamic experiments is typically insufficient to characterize accurately shale fabric features on the order of micrometers, or less. Micro CT scanners offer superior resolution but pose other challenges such as longer scan times and limited space for core-flooding experimental set-up. This work employs a multiscale image data set and shows that higher resolution 2D CT rock images can be inferred from lower resolution input using deep learning-based image super resolution. We train both feedforward convolution neural network and generative adversarial network models to predict micron-scale resolution CT images from input CT images with voxel resolution of about 200 microns. Alongside the deep learning image super-resolution approach, we also present a comprehensive image processing workflow for creating aligned and normalized image volumes that are usable for training deep-learning models. Due to the homogeneity and high noise level of CT rock images, we first test the models on synthetic low-resolution µCT data as benchmark. Results for benchmark data shows qualitative and quantitative improvements in image resolution in terms of the peak signal to noise ratio (PSNR) and structure similarity index (SSIM). We also perform an ablation study on components of the neural network architecture and compare the effectiveness of different training loss functions on model performance.