Geometrical Characterization of Multiscale Pore Structures in Reservoir Rocks Using Machine Learning on Images

The geometrical and fractal features of pore structures are crucial for understanding reservoir rock heterogeneity and its impact on fluid transport properties. However, characterizing pore geometrical and fractal features remains challenging with conventional methods, as they provide only a broad overview, lacking critical microstructural insights. This paper proposes an effective method to describe the geometrical and fractal features of the multiscale pore structure of reservoir rock samples based on scanning electron microscope images. The workflow includes four steps: (1) image-based rock typing to classify the target heterogeneous rock sample with multiscale pore structures into different rock types while each rock type denotes a homogeneous porous medium; (2) image segmentation to distinguish pores and solids; (3) pore partitioning to isolate individual pores; and (4) calculating and comparing geometrical and fractal features across rock types. The innovations of this study lie in the first and fourth steps, while segmentation and partitioning use the watershed algorithm and skeleton extension erosion grain partitioning (SEEGP) method. Image-based rock typing uses the U-net model, with K-means clustering, support vector machine, random forest (RF), Gaussian mixture model as reference methods. The U-net model achieved the highest accuracy (97%), followed by RF (88.8%). The watershed algorithm segmented images into pore and solid phases, and the SEEGP method isolated individual pores successfully. The experimental results demonstrated that geometrical and fractal features distinguish rock types clearly.