Characterization of microfracture systems and nanopore networks in organic-rich shales is critical for understanding fluid transport, hydrocarbon storage capacity, and the feasibility of carbon dioxide (CO₂) sequestration. However, the inherent heterogeneity, complex morphology, and nano-scale features of shale microstructures present significant challenges for conventional image segmentation techniques. This study reviews and evaluates the performance of state-of-the-art deep learning architectures for automated high-precision segmentation of microfractures and pore systems in FIB-SEM images of organic-rich shales. To enhance the generalizability of each model, a hybrid training dataset comprising 5,000 real and 5,000 synthetically generated GAN-based FIB-SEM images is evaluated. Quantitative analysis reveals that Kite-Net (KiU-Net) outperforms both Swin UNET Transformers (Swin-UNETR) and Attention U-Net, achieving an overall segmentation accuracy of 94%, precision of 94%, and recall of 93%. Notably, KiU-Net excels in accurately delineating microfractures and complex pore geometries within kerogen-rich matrices. Based on KiU-Net's superior validation performance compared to the two other deep learning models, we employed it to segment 3D FIB-SEM image stacks, enabling volumetric reconstruction and analysis of pore connectivity. Results revealed marked morphological distinctions between organic and inorganic pores, with over 94% of pores existing as isolated, non-percolating clusters, a finding consistent with prior geological investigations. Cross-validation considering various shales and coals further validates the model's effectiveness across a range of lithofacies. Our study presents a scalable deep learning framework for analyzing nanoscale shale microstructures from images.
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号