Interpretable Machine Learning for Predicting Splitting Strength of Asphalt Concrete: Insights from SHAP Analysis

This paper proposes an interpretable machine learning approach for predicting the splitting strength of asphalt concrete and supporting data-driven mixture design. A database consisting of 296 samples was constructed, and 14 input variables related to asphalt properties, aggregate gradation, and fiber characteristics were selected for modeling. Eight machine learning models, namely TabPFN, ANN, SVR, RF, XGBoost, LightGBM, FLAML, and FT-Transformer, were developed and compared. The results show that all eight models achieved satisfactory predictive capability, whereas TabPFN overall achieved the best performance in the Monte Carlo cross-validation, with the lowest average RMSE of 0.34 ± 0.10, the lowest average MAE of 0.21 ± 0.02, a relatively low average MAD of 0.14 ± 0.03, the highest average R² of 0.85 ± 0.08, and the highest composite score of 0.81. SHAP analysis further indicated that splitting strength prediction was mainly governed by a limited number of dominant variables, among which Ag9.5, AC, Du, FT, and Ag4.75 were the most effective parameters. In addition, favorable parameter ranges for improving splitting strength were quantified, such as Ag9.5 < 66.8%, AC < 5.4 wt.%, Du > 134.7 cm and Ag4.75 < 45.0%. Finally, a graphic user interface platform integrating prediction and SHapley Additive exPlanations-based interpretation was developed to improve the accessibility and practical applicability of the proposed framework.