Rock masses with certain shear strength are fundamental for ensuring the safety and stability of geotechnical engineering projects for geological disaster prevention. However, serrated jointed rock masses exhibit complex geometries and nonlinear mechanical properties, making accurate predictions of their shear strength challenging. To address this, an innovative machine learning-based prediction framework is proposed, integrating swarm intelligence optimization techniques with explainable data-driven methods to enhance prediction accuracy and reduce costs. This study utilizes experimental data of serrated jointed rock masses, covering key parameters such as internal friction angle, joint normal stress, ratio of normal stress to intact rock tensile strength, joint inclination, and shear strength. Based on this, various models were constructed, including Support Vector Regression (SVR), Backpropagation Neural Network (BPNN), Random Forest (RF), and Extreme Gradient Boosting (XGBoost). Furthermore, the models were optimized using Sparrow Search Algorithm (SSA), Chameleon Optimization Algorithm (CSA), Snake Optimization Algorithm (SO), and Kepler Optimization Algorithm (KOA). Results of statistical performance indicators showed that the KOA-XGBoost model performed best in predicting both the training and testing sets (R2 of 0.992, RMSE of 0.197 and 0.239), significantly outperforming other comparative models (R2 of 0.895 to 0.988, RMSE of 0.248 to 0.808). TreeSHAP analysis revealed that joint normal stress and joint inclination (with a cumulative importance score exceeding 0.6) were the most critical factors influencing shear strength. The findings provide an effective solution for ensuring the safety and stability of geotechnical projects involving serrated jointed rock masses.
Ma et al. (2026) studied this question.