Applying Machine Learning Methods Based on Panoptic Segmentation, Context Registration, and Octree Indexing for Multiscale Pore Structure and Connectivity of the Organic-rich shales in Bohai Bay Basin, East China

    The characterization of the microscopic pore systems in organic-rich shales is crucial for comprehending the occurrence mechanisms and flow behaviors of shale oil. Although various image processing techniques have advanced the study of shale pore systems recently, challenges such as unclear boundaries of pore structures, interwoven connectivity, high similarity, and complex topological structures remain unresolved. In this study, a comprehensive investigation of the multiscale pore structure in organic-rich shale is presented, through the examination of 20 lacustrine shale samples from the Paleogene Kongdian Formation. These samples were analyzed using a variety of techniques, including N2 adsorption, mercury intrusion capillary pressure (MICP), nano X-ray CT, and Focused Ion Beam-Scanning Electron Microscopy (FIB-SEM). Furthermore, We innovatively propose a machine learning-based multi-objective panoramic segmentation modeling method. This approach allows for precise segmentation and rapid panoramic modeling of various instances across different semantic categories and the gaps between these instances. It enables the creation of a more comprehensive multi-scale porous network model, which will be more conducive to future simulations of multi-physical processes such as fluid dynamics permeation models.
    We combine the dilated convolutions suitable for semantic segmentation with the feature pyramid structure favorable for instance segmentation to achieve precise panoramic segmentation. This approach accurately segments and represents various components in SEM images, including interparticle pores, intraparticle pores, organic pores, microfractures, feldspar, quartz, dolomite, calcite, clay minerals, and organic matter.In the three-dimensional reconstruction of FIB images, we innovatively employ registration based on contextual relationship sequences to accurately expand the reconstruction scope of pore pathways. Simultaneously, the use of an octree data structure index in constructing pore network structures enhances efficiency and speed. 
   The results show that the overall pore sizes range from 5 nm to more than 50 μm, consisting of abundant nanopores and a small quantity of micropores, and the dominant pores are in the range of 5 nm -200 nm. Through three-dimensional characterization of different types of pore networks, the transport behavior of shale oil within nanoslits was simulated, and it is proposed that fluid migration path is mainly controlled by the content of minerals, whether laminae are developed, and organic matter content. This study offers a promising solution for optimizing the automatic processing of microscopic images for pores, the combination of methods can provide pore structure characterization from sub-nanoscale to macroscale, spanning four orders of magnitude, which is crucial for improving the understanding of reservoir mechanisms and the hydrocarbon potential of lacustrine shale.