A calibrated-domain DEM–machine learning assessment of asphalt concrete stiffness and micromechanical response

Asphalt concrete is a heterogeneous composite whose stiffness and damage response depend on aggregate contacts, binder-mediated bonding, air voids, force chain evolution, and progressive bond failure. This study presents a calibrated-domain assessment of an AC-16C mixture using three two-dimensional Discrete Element Method contact/bond formulations: the linear parallel bond, soft-bond, and contact-bond models, together with a machine-learning stiffness surrogate. The 160 records constitute a hybrid laboratory–DEM/parametric dataset within a single AC-16C domain, rather than 160 physically distinct mixtures. The DEM outputs were used to interpret the contact forces, particle displacement, bond breakage, coordination number, energy dissipation, and fracture localization. The Linear Parallel Bond Model produced the stiffest response because it transferred both force and moment; the Soft-Bond Model represented gradual softening and energy dissipation and the Contact-Bond Model exhibited earlier localized bond failure. The machine learning models used uniaxial compressive strength, indirect tensile strength, dynamic modulus, and Poisson’s ratio to predict the experimental elastic modulus; DEM-derived descriptors were not included as predictors. Using a 112/24/24 training, validation, and holdout partition, Random Forest and XGBoost achieved holdout R 2 values of 0.940 and 0.928, respectively, whereas the ANN/MLP was less stable, with an R 2 of 0.847. Because the dynamic modulus dominated the feature importance, the high predictive accuracy primarily reflected the stiffness-variable collinearity within the calibrated domain. Therefore, the findings should be interpreted as internal calibrated-domain interpolation rather than external validation or evidence of a general microstructure–property relationship. Broader application requires physically independent mixtures, complete DEM reproducibility information, three-dimensional verification, raw data disclosure, and measured runtime comparisons.