Prediction of interlocking parameters for a physics-enhanced hybrid modeling of asphalt mixture’s dynamic modulus

Accurate characterization of dynamic, storage, and loss moduli is critical for understanding the viscoelastic behavior and mechanical performance of asphalt mixtures. However, conventional empirical, micromechanical, and data-driven models either neglect viscous contributions, fail to capture aggregate skeleton interlocking, or operate as black boxes with limited extrapolation and interpretability. To address these challenges, this study introduces a physics-enhanced machine learning framework to identify and predict the interlocking parameters p and q , which characterize aggregate skeleton interactions and load-transfer behavior. These parameters are difficult to measure directly and have not been systematically modeled previously. The p and q values were obtained through micromechanical inverse analysis combining dynamic modulus master curves with viscoelastic characterization, and were then used as physically grounded targets for supervised learning. Ensemble learning models, incorporating independent and joint prediction strategies, were trained to map mixture design variables to p and q , with GroupKFold cross-validation used to evaluate group-wise generalization while preventing data leakage. Once predicted, p and q can be directly used as forward inputs in micromechanical formulations to compute dynamic, storage, and loss moduli for new mixtures, eliminating the need for mixture-specific inverse analysis. Results indicate that XGBoost achieves the highest prediction accuracy, while joint prediction provides a complementary strategy for simultaneously predicting p and q . Feature interpretation further reveals distinct governing mechanisms for each parameter and confirms their dependence on mixture composition and binder rheology. This framework provides an interpretable route to quantify previously inaccessible interlocking parameters and supports performance-oriented asphalt mixture design.