Abstract
The introduction of deep learning techniques has revolutionized the automated segmentation of digital rock images. These methods enable precise evaluations of critical properties such as porosity and fluid flow characteristics, thereby enhancing the efficiency of reservoir characterization. This study explores the application of state-of-the-art Convolutional Neural Network (CNN) architectures for analyzing rock micro-CT images, aiming to enhance reservoir characterization efficiency. Specifically, we implement various deep learning models, including Fully Convolutional Networks, Encoder-Decoder Models, Multi-Scale Networks, Dilated Convolution Models, and Attention-Based Models. The segmentation performance of these CNN architectures is benchmarked against the traditional Otsu thresholding method using a dataset of 5,000 2D slices of ten distinct sandstone types, each with a voxel resolution of 2.25 × 2.25 × 2.25 µm. Our evaluation utilizes pixel-wise accuracy metrics such as F1-score, binary-IOU, Recall, and Precision. To replicate the physics of pore-scale fluid movement, various numerical simulation methods such as the Lattice Boltzmann Method (LBM), Pore Network Modeling (PNM), and Computational Fluid Dynamics (CFD) are employed to predict the permeability and rock formation factor of a blind sample, using CNNs for image segmentation. Our findings reveal that advanced CNNs significantly outperform the Otsu method in both pixel-wise segmentation accuracy and fluid flow simulation performance. Among all CNNs, EfficientNetB0-Unet, VGG16-Unet, and Enet exhibit exceptional performance in segmenting complex pore structures, as evidenced by their high F1-scores and binary-IOU metrics as well as accurate predictions of porosity, permeability, and formation factor.