Interpretable tree-based machine learning models with SHAP-GAM analysis for predicting blasting vibration

Abstract

Peak particle velocity (PPV) is a critical metric for evaluating the environmental impact of blasting in open-pit mines, and its accurate prediction holds significant value for optimizing blast design, controlling environmental vibrations, and ensuring slope stability. Based on 192 sets of field data from the Mirador copper mine in Ecuador, this study proposes a PPV prediction method that integrates machine learning and interpretability analysis. The research employs five decision tree ensemble algorithms (CatBoost, Extra Trees, NGBoost, RF, and XGBoost) combined with random search and Optuna hyperparameter optimization techniques to construct predictive models. Innovatively, a joint interpretability analysis approach using SHapley Additive exPlanations (SHAP) and generalized additive models (GAM) is applied. The results demonstrate that the CatBoost model optimized by Optuna performs best on the test set, achieving a coefficient of determination (R²) of 0.963, root mean square error (RMSE) of 1.085, and mean absolute error (MAE) of 0.784, showcasing excellent generalization capability. SHAP analysis reveals that the maximum charge per delay (q) contributes the most to model predictions (57.0%), followed by the distance from the blast center (R) and hole depth (HD), while other features such as stemming length (SL), total charge quantity (Q), hole spacing (a), and row spacing (b) exhibit relatively lower contributions. Furthermore, the GAM model uncovers nonlinear patterns, including a directional reversal in the influence of q (499.03 kg) and threshold effects for R (189.05 m) and HD (18.05 m). The proposed SHAP-GAM joint analysis method provides a novel approach for optimizing blasting parameters, and the established high-precision interpretable model offers substantial engineering value for ensuring blasting safety and slope stability.