An Automated Machine-Learning Procedure For Robust Classification of SEM Images of Cross-Laminated Sandstones For Digital Rock Analysis

Depositional structures such as laminae in sandstones reflect local changes in grain size, shape, orientation and composition. Laminae can occur as a set of parallel or intersecting structures, depending on the depositional processes. Further aided by diagenetic modifications, pore structures and pore surface properties may vary to a large degree, in their topology and geometry as well as physicochemical nature, within each lamina and across a set of laminae. The combination of spatial arrangements and local properties means that laminations can greatly influence the flow of multi-phase fluids. As an example, it is well-known that sandstone laminae can trap a significant amount of hydrocarbon in reservoirs. Most of the previous studies, however, treat each lamina as a uniform continuum without taking into consideration the true grain-pore characteristics associated with them (see [1] and references therein), whereas fewer others did but for tabular lamina only [2].
To gain a fuller understanding of the aforementioned combinational effect on multi-phase fluid flow and to obtain more appropriate estimates of effective properties for cross-laminated reservoir rock samples, we are taking digital rock analysis approach by reconstructing the pore structures of representative samples and then numerically simulating fluid flow through the pore systems. Because a representative sample deems to be much larger than a usual core plug, the reconstruction calls for multi-scale imaging techniques (e.g. using industrial CT scanner, microCT, Scanning Electron Microscopy (SEM)) to capture the spatial arrangements of the laminae and associated diverse pore structures, and deterministic and stochastic integration techniques to fuse obtained images. In integration, the lamina structures, which are captured in low-resolution 3D images, need to be calibrated against those identified in high-resolution 2/3D images, in order to reconstruct fine-scale grains and pores in those coarser structures. However, it is non-trivial to identify those structures in a high-resolution image. In this report, we develop an automated machine-learning procedure for image classification to perform this task. We illustrate this procedure using an SEM image of a cross-laminated tight sandstone sample. This work is an attempt to extend multi-scale data integration for digital rock analysis beyond what has been proposed for core plugs [3] to larger and structurally more complex samples.