Vector‐Based and Machine Learning Approaches for Pore Network Parameters Analysis

ABSTRACT Accurate characterization of pore structures in carbonate rocks is critical for evaluating fluid flow and storage capacity in subsurface reservoirs, a key concern in geophysical exploration and reservoir engineering. This study proposes a hybrid digital rock physics workflow that integrates deep learning–based segmentation, vectorial geometric analysis and clustering techniques to investigate pore‐scale features using x‐ray micro‐computed tomography at resolutions of 22 and 42 m. A convolutional neural network (CNN) enhances the segmentation ofcomplex pore geometries, addressing the limitations of conventional thresholding methods. To estimate the representative elementary volume, two‐dimensional porosity () distributions were integrated into three‐dimensional space using Riemannian methods. Pore connectivity () was quantified via the coordination number (), derived from a vector‐based analysis of local tangents and orthogonals, enabling precise identification of throats and pore networks. CNN models were trained on two carbonate samples (IL033 and IL636), achieving training accuracies of 0.9850 and 0.9914 and validation accuracies of 0.9854 and 0.9918, respectively. Total porosity () estimates from the CNN and classical segmentation approaches were compared to experimental data, with the deep learning approach showing superior performance, especially in capturing isolated or poorly connected pores at higher resolutions. This integrated methodology offers a powerful framework for quantifying microstructural heterogeneity and its influence on pore connectivity and geometry, contributing to more realistic geophysical modelling and reservoir simulation.