Abstract
Digital Rock Physics (DRP) serves as a powerful computational tool for analyzing the petrophysical properties of rock. Obtaining the properties of a fine-grained sample, such as shale is very challenging due to its highly variable and complex nature. Capturing the micro-features of this structure requires advanced microscopy techniques such as SEM (scanning electron microscopy) and FIB-SEM (Focused ion beam- scanning electron microscopy). However, performing advanced microscopy techniques to capture the heterogeneity of the sample is quite difficult; the slow speeds of data collection and analysis are two critical problems that limit more extensive use of this technology. In this study, an alternative approach is proposed to quantify the physical properties of the rock sample. This study aims to accelerate the process of SEM image analysis and reduce the computational cost by using machine learning. A deep learning-based method, Convolutional Neural Networks (CNN), is utilized to predict the properties from the 2D grayscale SEM images of Marcellus shale. The image data set is segmented by applying watershed segmentation to extract the pore network of the sample. Porosity and average pore size are the two properties computed for this study. The SEM images are down sampled to low-resolution images which are fed as an input, and the computed properties are used for training and validation of the CNN network. A detailed description of the image segmentation process, CNN architecture and the predicted results are discussed in this work.
