Generative adversarial network-based super-resolution of subsurface rock images: Visual, petrophysical, and flow simulation assessment

Abstract

High-resolution (HR) micro-CT imaging of porous reservoir rocks plays a critical role in digital rock physics and flow simulations, yet it is limited by the resolution–field-of-view trade-off. To address this challenge, we propose a Pore-Preserving Denoised Generative Adversarial Network (PPD-GAN), trained on denoised, HR two-dimensional computed tomography (CT) slices to recover fine-scale pore structures from low-resolution (LR) images. The PPD-GAN model is systematically compared against seven convolutional neural network (CNN)-based SR methods and bicubic interpolation using both image-based metrics—including Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM) and Learned Perceptual Image Patch Similarity (LPIPS)—and a comprehensive set of petro-physical metrics, including porosity, connectivity, pore size distribution, skeleton topology, and grain boundary preservation. Results show that although CNN models yield high pixel-wise fidelity, the PPD-GAN model achieves superior perceptual quality and reconstructs structural features that are consistently closer to ground truth across all physical metrics. Furthermore, pore-scale transport simulations on three-dimensional (3D) core images confirm that PPD-GAN outputs accurately replicate ground-truth (GT) properties such as permeability, diffusivity, and relative permeability—substantially outperforming bicubic interpolation. These findings demonstrate that the proposed PPD-GAN model not only enhances visual resolution but also preserves physically meaningful characteristics, enabling reliable downstream simulation and analysis.