Enhanced 3D Pore Segmentation and Multi-Model Pore-Scale Simulation by Deep Learning

Abstract Rock permeability characterization is crucial to understanding fluid flow in subsurface geological formations. It contributes to accurately simulating such processes that can address challenges like sustainable hydrocarbon production and geological CO2 sequestration. Recent advancements in deep learning have facilitated efficient permeability prediction in digital rock. However, existing methods often struggle to predict core-scale properties due to limitations in accommodating larger sub-volumes. This study introduces novel approaches integrating deep learning and physics-constrained methods to enhance rock segmentation, permeability prediction and upscaling. We first propose a 3D Inception U-Net model for 3D pore segmentation, which leverages the capability of the Inception block to capture multi-scale features in porous media and thus enhances segmentation accuracy. Further, we develop two different upscaling methods for permeability prediction. The first method is direct upscaling using deep learning, which directly predicts permeability across multiple scales by training with a combination of various sizes of sub-volumes; the second method is physics-constrained upscaling using deep learning, which imposes additional physical constraints on permeability predictions. We evaluate our deep-learning-based segmentation and upscaling approaches on diverse datasets, including Bentheimer, Leopard, and Parker sandstones. Our 3D Inception U-Net model achieves 0.99 accuracy for 3D pore segmentation. In upscaling, the direct upscaling using deep learning achieves R2 scores of 0.94, 0.83, and 0.84 at sub-volume sizes of 1503, 3003, and 6003, respectively, which demonstrates its potential to generalize permeability prediction across multiple sub-volume scales. On the other hand, with the permeability prediction of the base sub-volumes (size 1503) through the Lattice-Boltzman Method (LBM), the physics-constrained upscaling using deep learning achieves R2 values of 0.98 after upscaling from 1503 to 3003 sub-volumes, and further increases R2 to 0.99 after upscaling from 3003 to 6003 sub-volumes. Furthermore, when using 3D CNN-predicted permeability of 1503 sub-volumes, the second upscaling method achieves R2 scores of 0.96 and 0.94 for these two upscaling stages, respectively, highlighting its stable accuracy across scales. This research highlights the potential of integrating advanced deep learning with physics-constrained approaches to advance rapid and precise permeability prediction in digital rock physics, offering a promising framework for future core-scale applications and research endeavors.