Ultradeep carbonate reservoir lithofacies classification based on a deep convolutional neural network — A case study in the Tarim Basin, China

Abstract

Lithofacies identification is essential in reservoir evaluation, especially in ultradeep carbonate reservoirs. In general, coring samples are the best sources to identify carbonate lithofacies because they are taken directly from reservoirs. However, the core is expensive to obtain, and generally its availability is greatly limited. In recent years, deep learning has attracted enormous attention because of its robust nonlinear regression and classification ability. This study applies a deep-learning algorithm to identify the lithofacies using geophysical well-log data. Six types of well-log data, such as natural gamma ray, density (DEN), neutron porosity (CNL), acoustic (AC), and shallow and deep lateral resistivity well logs (RT/RXO), are smoothed by the average sliding method and converted to 2D data. Then, the 2D data are treated as inputs to predict the carbonate lithofacies through the convolutional neural network (CNN). The results indicate that the prediction accuracy rate is 90.2%. This indicates that the CNN can identify different carbonate lithofacies well.