Coupled prediction of viscoelastic properties of asphalt mixtures using a multi-task physics-constrained neural network

Accurate characterization of viscoelastic properties is fundamental for assessing long-term pavement performance. However, conventional models typically treat dynamic modulus and phase angle as independent prediction tasks, neglecting their intrinsic physical correlations. This study introduces a novel multi-task physics-constrained neural network (MT-PCNN) for the coupled prediction of dynamic modulus and phase angle. The model incorporated the Kramers-Kronig relationship as a constraint in the loss function, ensuring that the predictions comply with constitutive equations. Binder properties, aggregate gradation, volumetric characteristics, and angular frequency were used as input features. The overall performance of the MT-PCNN model was evaluated against several baseline models in prediction accuracy, extrapolation ability, and interpretability. The results show that the proposed MT-PCNN model exhibits greater physical consistency and robustness than purely data-driven models. Moreover, it achieves superior extrapolation performance with lower prediction errors in out-of-distribution regions. Feature importance analysis reveal that the dynamic modulus is primarily influenced by loading frequency, shear modulus, and air void content, whereas the phase angle is largely governed by binder properties. This study provides a transferable methodology for operationalizing constitutive laws, promoting robust prediction in broader engineering informatics applications.