Machine Learning‑Based Prediction and Interpretability of Asphalt Mixture Stiffness Using Random Forest and SHAP Analysis

Accurate prediction of the indirect tensile stiffness modulus (ITSM) of asphalt mixtures is essential for optimizing pavement performance and durability. Traditional regression approaches often fail to capture nonlinear interactions among mixture variables. This study aimed to evaluate the predictive performance of ensemble machine learning models, specifically Random Forest and XGBoost, and to enhance model transparency using SHAP (SHapley Additive exPlanations) analysis. A dataset of asphalt mixture properties, including % Air Voids, % Asphalt Content (%AC), and Unit Weight, was analyzed. Random Forest models were tuned through hyperparameter optimization, and their performance was compared with XGBoost using five‑fold cross‑validation. SHAP summary and dependence plots were employed to interpret feature contributions in engineering units (N/mm²). Hyperparameter tuning improved Random Forest performance (median R² = 0.54 at 20 °C; 0.80 at 30 °C), but XGBoost achieved superior accuracy (median R² = 0.58 at 20 °C; 0.89 at 30 °C). SHAP analysis identified % Air Voids as the most influential feature, exerting a strong negative effect on ITSM, while %AC and Unit Weight contributed positively. Dependence plots confirmed that higher air voids consistently reduced stiffness, particularly in low‑density mixtures. XGBoost provided the most reliable ITSM predictions, and SHAP analysis enhanced interpretability by linking model outputs to mechanistic mixture behavior. These findings support the integration of explainable AI into asphalt design workflows, enabling transparent, data‑driven decisions that align with engineering principles.