A novel hybrid CNN-SVM method for lithology identification in shale reservoirs based on logging measurements

Shale oil and gas reservoirs are lithologically complex and heterogeneous and accurate identification of lithology is necessary for the correct identification of lithofacies. There is also a poor mapping relation between logging response and lithology, making reliable lithological identification difficult. We propose a hybrid CNN-SVM model for lithological identification of the shale reservoir in the southern Songliao Basin of Northeast China. As training data, seven conventional logging curves are used, including spontaneous potential (SP) and deep and shallow lateral resistivity (RD, RS) logs. The CNN automatically extracts feature information from well log data and lithology, while the SVM overcomes the problem of limited sample size. Using the receiver operator characteristic (ROC) curves and the area under the curve (AUC) values, we assess the effect of lithological classification. The accuracy of lithological identification for test well H2 is 91.95%, and the AUC of the hybrid model is 0.94, 0.98, and 0.99 in mudstone, shale, and sandstone, respectively. The hybrid model outperforms CNN and SVM in terms of the identification of three types of lithologies and is more stable in terms of AUC value and ROC curve shape. The lithological identification accuracy for test well H3 is 89.49%, which demonstrates that the method has much capacity for generalization and may be extensively utilized in the study area. Finally, from the perspective of model interpretability, SHapley Additive exPlanations (SHAP) is developed to increase transparency and further confirm the reliability of the hybrid model.