Interpretable Prediction of Rock Strength: A Study Based on the COA-LightGBM-SHAP Framework

Uniaxial compressive strength (UCS) is a fundamental parameter for characterizing the mechanical behavior of rock masses which is essential for geotechnical stability analysis. However, the efficient and precise determination of UCS remains a significant challenge due to the complexity of rock materials. This study develops a robust predictive framework by integrating the Light Gradient Boosting Machine (LightGBM) with the Coati Optimization Algorithm (COA). Based on a comprehensive database of 273 rock samples, the COA is utilized to systematically optimize the hyperparameters of the LightGBM model. The results indicate that the proposed COA-LightGBM model exhibits exceptional predictive accuracy, achieving an RMSE of 20.48, an R² of 0.83, and an MAE of 15.24. Comparative analyses demonstrate that the COA-LightGBM model significantly outperforms conventional models such as K-Nearest Neighbors, Decision Trees, and Random Forests. Furthermore, the Shapley Additive Explanations (SHAP) method is introduced to elucidate the internal decision-making mechanism and bridge the gap between "black-box" models and physical interpretability. Feature analysis reveals that porosity (n) is the most critical factor governing rock strength. This research not only validates the superior optimization capability of COA but also provides a scientific and interpretable solution for the rapid assessment of rock mechanical parameters in engineering practice.