Improving Numerical Fluid Flow Simulation by Ring Artifact Removal in Micro-CT Images of Porous Media Using Attention Autoencoder–Decoders

Abstract

The emergence of micro-computed tomography has significantly enhanced our ability to examine the morphology of porous materials and the dynamics of fluid flow within pore spaces. However, image-based analyses can be compromised by various artifacts, particularly ring artifacts, which appear as concentric rings in the images. These artifacts can be misinterpreted as part of the pore space, artificially connecting pores and thus influencing numerical simulations. This study examines the influence of ring artifacts on pore network modeling (PNM), direct numerical simulation (DNS), and prominent numerical techniques, and presents a computing approach for their effective mitigation. For this purpose, a dataset was compiled from the literature that includes the images of Fontainebleau, Boise, and Belgian sandstones. Data augmentation was implemented by extracting real ring patterns from Fontainebleau samples and superimposing them onto clean images of the sandstones. Two U-Net autoencoder architectures (base and Attention U-Net) were trained for a regression task aimed at removing ring artifacts while reconstructing the underlying pore morphologies. The Attention U-Net outperformed the base model, achieving a mean squared error of 0.07 (calculated based on the grayscale values between 0 and 255). Visual evaluations confirmed the model’s effectiveness in artifact removal and pore morphology reconstruction. The model was further tested on unseen pore-scale data containing real ring artifacts, which indicated a high performance in removing the artifacts. DNS and PNM were performed on both original (with real rings) and improved 3D samples (2003 voxels) to assess the impact of artifact removal on transport properties. The results revealed that ring artifacts, identified as flow pathways, significantly influence the velocity profiles. While the presence of the artifact had a minimal effect on porosity (a 1.68% error) and the number of pores (1.45% error), it significantly increased the permeability by 34%.