Abstract Understanding fluid flow in complex fractured porous media requires an accurate representation of the pore space, especially in the presence of both granular pores and fractures, which significantly differ in their geometries. The effect of such a complex fluid pathway is prominent in fractured sandstones and carbonates, which store a significant amount of energy resources. Digital core analysis allows three-dimensional imaging of rock cores at high resolutions to capture essential features such as granular pores, fractures, and minerals. These features are represented by a multi-modal grey-level histogram. Later, segmentation techniques are applied to these images to uniquely differentiate pores (fractures and granular pores) from minerals. However, these segmentation routines are purely based on gray-level intensities and hence, are incapable of automatically segregating the pore space into granular pores and fractures. This paper applies a computer vision technique called contour detection for structural analysis of the entire pore space. Since fractures are mainly considered planar features, we carry out a slice-by-slice analysis on 3D segmented images of fractured sandstones and carbonates. Contours of pores (granular pores and fractures) and their structural properties are read into a principal component analysis followed by a K-means clustering algorithm to segregate granular pores from fractures. The exploratory analysis showed that the optimum number of principal components required for segregation is 2 for both sandstones and carbonates. The test for generalization used the above information for K-means clustering in x, y, and z directions. The voxels identified as fractures and granular pores are then merged to give a 3D representation of fractures showing an accuracy greater than 89% compared to the ground truth labels. The permeability differences between reconstructed and labelled fractures for digitally fractured samples were less than 9% for all the directions. This analysis provides pathways to understand better competing fluid flow phenomena in pores and fractures and aids in the extraction of pore network models for complex fractured porous media.

中文翻译：

---

用于裂缝性多孔介质孔隙尺度结构分析的计算机视觉和无监督机器学习

摘要 要了解复杂裂缝性多孔介质中的流体流动，需要准确表示孔隙空间，尤其是在同时存在几何形状显着不同的颗粒状孔隙和裂缝的情况下。这种复杂的流体通路在裂缝性砂岩和碳酸盐岩中的作用尤为突出，它们储存了大量的能源。数字岩心分析允许以高分辨率对岩心进行三维成像，以捕捉颗粒孔隙、裂缝和矿物等基本特征。这些特征由多模态灰度直方图表示。后来，分割技术应用于这些图像，以独特地区分孔隙（裂缝和颗粒状孔隙）与矿物质。然而，这些分割程序纯粹基于灰度强度，因此，不能自动将孔隙空间分离成粒状孔隙和裂缝。本文应用称为轮廓检测的计算机视觉技术对整个孔隙空间进行结构分析。由于裂缝主要被认为是平面特征，我们对裂缝砂岩和碳酸盐岩的 3D 分割图像进行了逐片分析。孔隙（颗粒孔隙和裂缝）的轮廓及其结构特性被读入主成分分析，然后是 K-means 聚类算法，以从裂缝中分离颗粒孔隙。探索性分析表明，对于砂岩和碳酸盐岩，分离所需的最佳主成分数为 2。泛化测试将上述信息用于 x、y 和 z 方向的 K 均值聚类。然后将识别为裂缝和颗粒状孔隙的体素合并，以提供裂缝的 3D 表示，与地面实况标签相比，其精度超过 89%。数字裂缝样品的重建和标记裂缝之间的渗透率差异在所有方向上均小于 9%。该分析提供了更好地理解孔隙和裂缝中的竞争性流体流动现象的途径，并有助于提取复杂裂缝多孔介质的孔隙网络模型。