Automated lithology classification from drill core images using convolutional neural networks

In hydrocarbon reservoir evaluation, lithology is a key characteristic for determination of storage capacity and rock properties. Lithology is usually predicted from well log data or directly identified through manual inspection by geologists during core logging. Automatic prediction of lithology has been a focus of many studies inspired by rapid developments in machine learning. Studies mostly focus on predicting lithofacies from well logs using neural networks, and more recently from borehole image logs using convolutional neural network (CNN). Although core is more representative of reservoir lithology than well log data, employing machine learning for automatic lithology identification from core is limited. A CNN-based approach is developed to classify core images into three lithologies: sandstone, limestone, and shale. The model is based on ResNeXt-50 architecture, which had better performance in lithology prediction than ResNet-18 and Inception-v3 architectures. Our method processes core tray images and outputs lithology logs in a fully automated procedure and predicts the lithologies of new (unseen) core with an accuracy of 93.12%.