Advanced hybrid models for predicting dynamic compressive strength of rocks: Exploring various input scenarios and introducing novel empirical approaches

• A scenario-based framework was developed accommodating both destructive and non-destructive testing parameters, enabling flexible model implementation under varying data availability constraints. • With R 2 = 0.962 and RMSE = 12.38 MPa, the BBOA-XGBoost model performed better than Random Forest. • Among six empirical formulations, Model 1 demonstrated optimal statistical performance (R² = 0.90) with significant parameters (p < 0.01) across all input variables. • SHAP analysis revealed scenario-dependent feature importance: strain rate dominated in full feature sets while static compressive strength became primary in reduced configurations. • Machine learning models substantially outperformed empirical approaches, demonstrating enhanced predictive accuracy and computational efficiency for dynamic rock strength assessment. This research presents comprehensive hybrid modeling methodologies for predicting the dynamic compressive strength of rocks, addressing the challenges associated with expensive and time-consuming Split-Hopkinson Pressure Bar testing through integrated empirical and machine learning approaches. The study employs a scenario-based framework with four distinct configurations: Scenario 1 (length-to-diameter ratio, density, P-wave velocity, strain rate, static uniaxial compressive strength), Scenario 2 (grain size, density, P-wave velocity, strain rate, static uniaxial compressive strength), Scenario 3 (length-to-diameter ratio, P-wave velocity, strain rate), and Scenario 4 (strain rate, static uniaxial compressive strength). Six empirical models were developed across these scenarios, with Model 1 achieving superior performance (R² = 0.90) and demonstrating statistical significance for all parameters (p < 0.01). The machine learning phase utilized XGBoost and LightGBM algorithms optimized through three metaheuristic approaches: Brown Bear Optimization Algorithm (BBOA), Particle Swarm Optimization (PSO), and Starfish Optimization Algorithm (SFOA). BBOA-XGBoost consistently demonstrated optimal performance across scenarios, with Scenario 2 yielding the highest test accuracy (R² = 0.962, RMSE = 12.38 MPa, MAE = 9.28 MPa), substantially outperforming previous Random Forest implementations. Critical Difference analysis confirmed BBOA-XGBoost as the superior configuration across multiple evaluation scenarios. SHAP analysis revealed scenario-dependent feature importance patterns: strain rate dominated in comprehensive feature sets (Scenarios 1 and 3), while static compressive strength became most significant in reduced feature configurations (Scenarios 2 and 4). The proposed hybrid models demonstrate substantial performance improvements over existing methodologies, offering practitioners computationally efficient alternatives to laboratory testing while maintaining engineering-acceptable accuracy levels for dynamic rock strength assessment applications.