Mechanistic-feature-enhanced data-driven prediction of structural responses in bi-modulus asphalt pavements

Objectives This study aims to predict bi-modulus asphalt pavement responses efficiently while accounting for temperature- and frequency-dependent material properties.Methods A temperature–frequency-dependent constitutive model based on tensile and compressive dynamic modulus master curves was implemented in a user material subroutine (UMAT). Finite element simulations incorporating speed-to-frequency conversion and ambient-to-layer temperature mapping generated a database covering temperature, vehicle speed, and axle load. Mechanistic features were progressively added to extreme gradient boosting (XGB) models evaluated using equal-budget randomized hyperparameter tuning and repeated nested cross-validation.Findings The model incorporating layer temperatures and loading frequency achieved a mean coefficient of determination of 0.9805 and a mean normalized root-mean-square error (NRMSE) of 0.0270. NRMSE was 21.6% lower than that of the baseline using only operating conditions; learning-curve generalization gaps also narrowed. Grouped shapley additive explanations (SHAP) analysis attributed 64.11% and 75.20% of feature importance to thermal state for peak surface deflection and peak tensile strain at the asphalt-layer bottom, respectively. Axle load ranked second; loading frequency contributed approximately 2%. Adding bi-modulus parameters and derived features yielded no further accuracy gains. Preliminary comparisons with limited field measurements showed good agreement in bottom tensile strain.Novelty The method improves data-driven prediction by representing temperature–frequency-dependent material behavior through complementary, physically interpretable features.