Deep learning assisted denoising of fast polychromatic X-ray micro-CT imaging of multiphase flow in porous media

Understanding the flow of fluids in the subsurface and their interaction with different solid surfaces is crucial for addressing challenging geological applications such as CO 2 sequestration, enhanced oil recovery, and environmental remediation of polluted aquifers. Synchrotron-based 3D X-ray micro-computed tomography (micro-CT) has enabled the visualization of dynamic pore-filling events in multiphase flow experiments at sub-second time resolutions. However, the limited accessibility of synchrotron facilities has driven the use of low-flux polychromatic micro-CT systems, which often produce relatively noisy images during fast scans. To overcome this limitation, we propose a deep learning workflow using a cycleGAN network trained on unpaired datasets as no direct pixel-wise correspondence exists between the noisy domain and the high-quality domain. This approach transforms noisy fast polychromatic micro-CT scans into high-quality images, enabling detailed analysis of multiphase flow dynamics. The effectiveness of the denoising process was verified using blind image quality evaluators and Minkowski functionals for the non-wetting phases. The results indicate that the cycleGAN network achieves an average 1 to 6 percentage error difference for 3D morphological analysis parameters and outperforms other filtering methods such as non-local means and the adaptive Weiner filter, demonstrating its potential as a reliable technique for restoring noisy fast scans from polychromatic micro-CT systems. • CycleGAN assisted denoising of fast polychromatic micro-CT images. • Beneficial for fast imaging of multiphase flow in porous media as cycleGAN requires only unpaired datasets for its training. • Results of ‘no reference image’ quality evaluators, and the morphological analysis indicate average percentage error difference of 1 to 6% for the denoised images. • Outperform traditional noise removal techniques like non-local means filter and adaptive Wiener filter by accurately capturing 2D Minkowski functionals with -3% average error. • Open-source code that enable reproducibility of the workflow. A micro computed tomography (micro-CT) is a well-known technique for investigating porous media. The study of flow of fluids and its interaction with each other in porous media in real time is important for addressing several geological applications such as gas storage, and oil recovery. Such studies in the past have often been carried out at synchrotron facilities which are high flux x-ray sources, but these facilities are scarce around the world. However, when such fast scans are done at commonly accessible low flux micro-CT facilities, noisy images are produced. We propose the use of a deep learning network called cycle consistent generative adversarial network (cycleGAN) for transforming noisy, fast micro-CT scans into high-quality images enabling analysis of multiphase flow dynamics.