Ensemble learning for well logging evaluation of the hybrid shale brittleness index: A case from the Gaoyou sag, Subei basin

Accurate evaluation of the brittleness index (BI) is crucial for optimizing hydraulic fracturing during the extraction of hydrocarbon fluids in hybrid shale reservoirs, yet conventional petrophysical methods face limitations in scalability and generalizability. This study presents an integrated evaluation framework utilizing four ensemble learning algorithms—Random Forest (RF), AdaBoost, XGBoost, and CatBoost—to predict BI from well logging data in the second member of Funing Formation (E 1 f 2 ) shale from the Gaoyou Sag, Subei Basin, eastern China. A dataset comprising 1295 well logging data points and mineralogical compositions from 174 core samples was used to train and validate the models. These models were optimized by Particle Swarm Optimization (PSO) to resolve nonlinear interdependencies between logging responses and mechanical brittleness. Comparative analysis demonstrates that RF achieves the highest prediction accuracy (R 2 = 0.84 on the test set), outperforming CatBoost (R 2 = 0.79), AdaBoost (R 2 = 0.71), and XGBoost (R 2 = 0.65). The superior performance of RF is attributed to its robustness against overfitting and its ability to effectively capture complex nonlinear relationships in logging responses. SHapley Additive exPlanations (SHAP) analysis identifies acoustic (AC) and resistivity (Rt) logs as the most influential predictors, reinforcing their strong physical correlations with mineralogical brittleness. This study represents the application of ensemble learning for BI evaluation in the Funing Formation shale, providing a cost-effective alternative to laboratory-based methods and demonstrating the viability of data-driven approaches for fracturability assessment. The proposed framework offers significant potential for extension to other unconventional reservoirs, contributing to enhanced hydraulic fracturing design and improved reservoir development strategies for unconventional hydrocarbon fluid development. • Superior brittleness index prediction in lacustrine hybrid shale is revealed by the Random Forest. • Vertical brittleness zonation framework is proposed to enable targeted hydraulic fracturing strategies. • Ensemble learning is established for brittleness evaluation in lacustrine shale systems, offering a cost-effective and scalable alternative to laboratory-based methods.