Predicting geomechanical properties of heterogeneous shale using ensemble learning methods

Accurate characterizing shale mechanical properties is crucial in oil and gas exploration and development. However, acquiring rock mechanical data remains challenging. This study investigates machine learning algorithms for predicting shale geomechanical properties using readily available data. A comprehensive dataset was collected, including confining pressure (CP), sampling orientation, Young's modulus ($E$) and Poisson's ratio ($\nu$) from triaxial compression tests, as well as core analysis and conventional logging data. Three ensemble learning models were constructed following two strategies to predict $E$ and $\nu$, with inputs from core analysis and logging parameters. The results indicate that the Random Forest, eXtreme Gradient Boosting and Light Gradient Boosting Machine (LightGBM) outperformed neural networks and other classical models. The LightGBM model exhibited the highest accuracy, with determination coefficient (R²) and mean relative error (MRE) being 0.82–0.85 and 6.70 %–8.36 % on test dataset. Density and orientation were the primary factors influencing shale mechanical properties, with relative importance being 0.285–0.301 and 0.178–0.230, respectively, while the CP, mineralogical composition and porosity are secondary controlling factors. Based on different core or logging parameter combinations, model performance was categorized into four levels: “optimal,” “suboptimal”, “poor” and “very poor”, ensuring adaptability to varying data conditions for mechanical property prediction. The LightGBM model was successfully applied in predicting Wufeng-Longmaxi shale mechanical properties, outperforming empirical formulas and demonstrating the advantages of ensemble learning. This study provides a practical tool for the rapid estimation of shale mechanical parameters, facilitating oil and gas development while improving economic efficiency.