This study introduces a hybrid machine learning (ML) framework integrating the Extreme Gradient Boosting (XGBoost) model with three evolutionary optimization algorithm such as Differential Evolution (DE), Particle Swarm Optimization (PSO), and Salp Swarm Algorithm (SSA) to predict cohesion (c) and angle of internal friction ( \(\:\varnothing\:\) ) from indirect and routinely available rock properties. A curated experimental database of 808 samples was subjected to statistical preprocessing, including outlier detection through interquartile range filtering, resulting in a refined dataset of 586 samples. Statistical diagnostics, including Z-tests and Sobol sensitivity analysis, confirmed the robustness of the input variables. Among the developed models, DE-XGBoost exhibited superior predictive capability, with coefficient of determination exceeding 0.84 for c and 0.73 for \(\:\varnothing\:\) , while maintaining low error values. Model robustness was verified through regression error characteristics curve analysis, residual distribution checks, and curve-fitting tests. Shapely additive explanations, Local Interpretable Model-Agnostic Explanations, and Partial Dependence Plot for Explanation interpretability analyses revealed tensile strength (TS) as the most influential factor for \(\:\varnothing\:\) , whereas both TS and UCS equally governed c. These findings demonstrate the potential of evolutionary optimized XGBoost models as reliable, interpretable, and cost-effective tools for estimating rock shear strength parameters, thereby supporting safer and more efficient geotechnical design.
