Ensemble learning for well-logging prediction of shale mineral composition and brittleness index: a case from the Qingshankou Formation, Songliao Basin

Accurate evaluation of rock mechanical properties in shale reservoirs is essential for optimizing hydraulic fracturing design and achieving efficient development. To address the challenges in conventional petrophysical evaluation caused by complex mineral composition and strong heterogeneity of shale in the Qingshankou Formation, Songliao Basin, a novel well logging evaluation framework based on ensemble learning algorithms is proposed. This framework utilizes conventional well logging data to quantitatively predict mineral composition and incorporates an improved brittleness index model with organic matter correction. Based on X-ray diffraction (XRD) and total organic carbon (TOC) analysis data from 174 core samples collected from seven cored wells in the Sanzhao Sag, a dataset comprising seven sensitive logging curves—including gamma ray (GR), interval transit time (AC), density (DEN), compensated neutron (CNL), and microspherically focused resistivity (MSFL)—was constructed. Through comparison of XGBoost, Random Forest (RF), and AdaBoost algorithms, XGBoost was selected as the core prediction tool. The results demonstrate that the XGBoost model exhibits robust generalization performance in addressing strongly nonlinear geological regression problems. Specifically, coefficients of determination (R 2 ) exceeded 0.6 for clay, felsic, and carbonate mineral predictions in the test set, while root mean square errors (RMSE) were controlled within 3.8%, indicating improved accuracy relative to the other evaluated models. Feature importance analysis reveals that CNL, which reflects the neutron moderation effect of interlayer water, serves as a key feature for predicting clay and felsic content. And AC and MSFL provide indicative significance for carbonate cement identification. Based on the prediction results, significant vertical heterogeneity in material composition of the Qingshankou Formation shale is revealed, and the Q 7 -Q 9 oil layer groups are identified as Class I high-brittleness sweet spots (BI 0.65, felsic mineral content 60%). Fracturing target layers of the Qingshankou Formation shale were optimized based on the evaluation results, and differentiated engineering parameter schemes were formulated. This method provides a scientific basis and effective technical approach for sweet spot optimization and engineering decision-making in continental shale oil.