Deep-Learning-Based Vuggy Facies Identification from Borehole Images

Summary Identification of vuggy intervals and understanding their connectivity are critical for predicting carbonate reservoir performance. Although core samples and conventional well logs have been traditionally used to classify vuggy facies, this process is labor intensive and often suffers from data inadequacies. Recently, convolutional neural network (CNN) algorithms have approached human-level performance at multiimage classification and identification tasks. In this study, CNNs were trained to identify vuggy facies from a well in the Arbuckle Group in Kansas, USA. Borehole-resistivity images were preprocessed into half-foot intervals; this complete data set was culled by removing poor-quality images to generate a cleaned data set for comparison. Core descriptions along with conventional gamma ray, neutron/density porosity, photoelectric factor (PEF), and nuclear magnetic resonance (NMR) T2 data were used to label these data sets for supervised learning. Hyperparameters defining the CNN network size (numbers of convolutional layers/filters and the numbers of fully connected layers/neurons) and minimize overfitting (dropout rates, patience, and minimum delta) were optimized. The median losses and accuracies from five Monte Carlo realizations of each hyperparameter combination were the metrics defining CNN performance. After hyperparameter optimization, median accuracy for vuggy/nonvuggy facies classification was 0.847 for the cleaned data set (0.813 for the complete data set). This study demonstrated the effectiveness of using microresistivity image logs in a CNN to classify facies as either vuggy or nonvuggy, while highlighting the importance of data quality control. This effort lays the foundation for developing CNNs to segment images to estimate vuggy porosity.