Evolution and Influence Analysis of Skid Resistance Performance in Steel Slag Ultra-Thin Wearing Course Based on XGBoost-SHAP Framework

Abstract

The integration of steel slag as a substitute for natural aggregate in ultra-thin wearing courses (UTWC) presents a sustainable approach to pavement maintenance. However, the evolution of skid resistance under multiple loading factors remains insufficiently quantified. This study introduces an interpretable machine learning framework that combines Extreme Gradient Boosting (XGBoost) with SHapley Additive exPlanations (SHAP) to predict and explain the skid resistance decay of steel slag asphalt mixtures subjected to MLS11 accelerated loading. A dataset comprising 168 observations from 24 specimens—featuring three steel slag replacement ratios across seven loading stages—was constructed, incorporating variables related to material composition, aggregate meso-morphology, pavement surface texture, and loading history. Four predictive models—multiple linear regression (MLR), support vector regression (SVR), random forest (RF), and XGBoost—were evaluated using a strict group-aware train-test split to prevent data leakage. Among the evaluated models, RF achieved the highest raw predictive accuracy, while XGBoost showed comparable performance with an R² of 0.979, RMSE of 1.202, and MAPE of 1.64%. Considering its near-optimal predictive accuracy, built-in regularization, and strong compatibility with the SHAP framework, XGBoost was selected as the principal model for subsequent interpretable analysis. Dependence plots revealed significant non-linear saturation effects, with BPN decay accelerating sharply during the first 20 × 10⁴ cycles before stabilizing. Steel slag content showed a threshold effect, beyond which additional anti-skid improvements diminished. Local SHAP waterfall decomposition illustrated that high Rz and low N values contributed to elevated BPN predictions, highlighting the “buffer mechanism” of steel slag against polishing-induced texture loss. This framework provides a data-driven tool for optimizing steel slag dosage and predicting residual skid resistance in pavement maintenance.