Carbonate Lithology Identification with Generative Adversarial Networks

Abstract

Carbonate sedimentary rocks form the reservoir rocks of many oil and gas fields. The largest oil and gas fields in the world, such as the Ghawar field in Saudi Arabia and the Zakum field in Abu Dhabi, consist of carbonate reservoirs. Therefore, understanding the structure of carbonate sedimentary rocks is important to estimate the reservoir quality and distribution in the oil and gas field. However, carbonate sedimentary rocks have complex sedimentary structures that comprise various kinds of carbonate minerals. In addition, carbonate reservoirs often undergo diagenesis after deposition. Therefore, a detailed carbonate facies analysis requires great expertise. Additionally, traditional thin section analysis approaches such as the point counting method are extremely time intensive. In this context, machine learning, including deep learning, is attracting significant attention. In particular, image analysis using convolutional neural networks (CNNs) has seen dramatic development since the emergence of AlexNet in 2012. CNNs achieve superhuman image recognition capability by utilizing a deep layer structure that consists of a convolutional layer, activation function, etc. In the field of petroleum exploration and production, several studies on image analysis using CNNs have been performed by petroleum exploration and production companies and universities . Nanjo and Tanaka (in press) attempted carbonate lithology identification with pixel-wise segmentation in thin section images; the average accuracy of their category identification for each components [grain, cement, pore, and lime mud areas] was 83.9% and the automatic carbonate lithology identification based on the category identification was over 90%. They showed that machine learning is effective for carbonate lithology identification. However, the model is still not perfect with respect to both of category identification and automatic carbonate lithology identification. Generative Adversarial Networks (GAN) are unique and thought as useful tool to improve the model. GAN has already been studied in various fields (e.g., image generation and analysis). However, few studies have attempted to use GAN for carbonate lithology identification. In this study, the authors attempted to conduct carbonate lithology identification with a GAN and to review the potential of applying GAN for FMI imaging.