Separate Classification Prediction Model for Lithofacies Identification of Paleogene Yingxiongling Shale, Qaidam Basin

Abundant shale oil resources have been discovered in the upper member of the Paleogene Lower Ganchaigou Formation of the Yingxiongling area from the Qaidam Basin, China. The lithofacies of Yingxiongling shale oil exhibit strong heterogeneity vertically. Accurate lithofacies identification is the key to characterizing the potential of unconventional oil and gas resources. Traditional lithofacies identification is limited by factors such as the duration of experiments and the subjectivity of the scholars. Only a limited amount of coring section data is available for analysis, while a sea of logging data remains underutilized. Therefore, utilizing machine learning algorithms to effectively leverage logging data for constructing the accurate lithofacies identification model has become a crucial area in both academia and industry. In this paper, 15 basic logging curves were used, and algorithms of random forest (RF), support vector machine (SVM), and extreme gradient boosting (XGBoost) were selected through Python programming to establish machine learning classification models, identifying the lithofacies types of Yingxiongling shale and analyzing the results. The lithofacies classification scheme of Yingxiongling shale is based on “rock structure + mineral composition”, developing 8 lithofacies types: thin-bedded/laminated dolomitic limestone, thin-bedded/laminated limy dolostone, thin-bedded sandstone, laminated shale, and thin-bedded/laminated mixed rock. Due to the differing sensitivities of various logging data in identifying rock structures and mineral compositions, the corresponding algorithms and parameters vary accordingly. Hence, an innovative stepwise prediction model integrating “sedimentary structures and mineral composition” is proposed. The model first identified the rock structure through the genetic algorithm-RF and 15 logging curves, yielding thin-bedded/laminated structures. Then, SVM and 9 logging curves were used to identify mineral composition, yielding limy dolostone, dolomitic limestone, sandstone, shale, and mixed rock. The lithofacies were obtained by integrating the predicted results from the two models. The maximum accuracy of identifying rock structure and mineral composition can reach 87.3% and 78.7%, respectively, and the maximum prediction accuracy of the separate prediction model reached 73.2%, which is 22% higher than that of the direct prediction model. The relationship between the well logging curves and the predicted results is discussed, and the reasons for errors will be explained. These understandings can further help provide new ideas and methods for the identification of shale lithofacies types and can provide scientific guidance and technical support for the exploration and development of the Qaidam Basin.