Physics-guided feature-enhanced machine learning for asphalt-pavement rutting prediction and SHAP-guided structural screening

Accurate asphalt-pavement rutting prediction is central to performance-oriented design, but mechanistic-empirical models may not capture section-specific responses, while data-driven models offer limited physical interpretability. This study proposes a physics-guided feature-enhanced framework that incorporates the output of a low-fidelity mechanistic-empirical model as an engineered feature alongside layer-specific pavement variables. Long-term observations from 19 RIOHTrack sections spanning seven structural types were evaluated using pavement-section-based five-fold grouped cross-validation. PHY-XGBoost achieved the best performance (R² = 0.9391, RMSE = 4.6189 mm, MAE = 3.5680 mm, and MAPE = 6.9411%), reducing RMSE by 53.26% compared with the baseline XGBoost model. SHAP analysis identified the mechanistic-empirical feature, cumulative traffic loading, temperature, and third-layer dynamic modulus as the four most influential features; together, they accounted for 93.0% of total mean absolute SHAP importance. In an illustrative case, screening a bounded library of RIOHTrack-observed parameter templates reduced predicted rutting depth from 89.20 mm to 71.08 mm under a modulus-adjustment scheme and to 56.87 mm under a separate thickness-position scheme. The framework integrates section-level prediction, interpretable feature analysis, and transparent screening of reference pavement configurations.