Estimation of Mechanical Properties of the Bakken Shales Through Convolutional Neural Networks

Effective mechanical properties of shale rocks can be determined by knowing the mechanical properties and distribution pattern of each comprising constituent. However, building the relationship between them is complicated and requires mathematical manipulations. In this study, by taking advantage of machine learning (ML) that is capable of delineating hidden patterns with the least sophistication, a new approach to estimate Young's modulus of shales by integrating deep learning convolutional neural networks (CNNs) into 2D elemental intensity distribution maps is presented. The generated SEM-EDX maps contain spatial distribution and intensity information of nine major elements abundant in a shale, Al, Ca, C, Fe, K, Mg, Na, S, and Si. The ground truth data are Young's modulus based on laboratory microindentation tests from ten samples. A total amount of 800 images were used for training and testing, and the trained CNNs were then used to predict Young's modulus of shale samples by feeding the elemental images. The predicted Young's modulus exhibited an acceptable relative error of 6.5% and in a much faster time and less effort compared to the laboratory tests. Ultimately, we believe that this novel method has great potential for field applications due to simplified requirements for sample preparation and laboratory apparatus.