Sarcouncil Journal of Engineering and Computer Sciences

Abstract: Precise estimation of asphalt pavement degradation was found to require the integration of material breakdown at the mesoscale with structural behaviour at the macroscale — a task that had been computationally impractical for conventional finite element approaches. In this study, MesoPINN-Pave was presented, a physics-informed neural network framework in which mesoscale cohesive fracture mechanics and viscoelastic constitutive laws were embedded into a deep learning architecture to enable real-time macroscale pavement performance forecasting. The framework was composed of three interconnected neural networks: (i) a mesoscale surrogate that was trained on extended finite element method (XFEM) simulations of asphalt–aggregate interface fracture, (ii) a macroscale network by which pavement stress–strain behaviour was predicted, and (iii) a coupling network through which damage state variables were transferred across scales. Physical consistency was ensured by thermodynamically constrained loss functions, which incorporated viscoelastic PDE residuals, cohesive zone evolution, and dissipation inequality constraints. When validated against Long-Term Pavement Performance (LTPP) data and field falling weight deflectometer measurements, a 4.8% mean absolute error in fatigue life prediction was achieved, together with a 1,200-fold acceleration relative to full multi-scale finite element simulation. Leave-one-climate-out testing confirmed that robust generalisation was maintained across freeze–thaw, hot-arid, and moderate climate zones (R² > 0.87). The proposed framework was thereby shown to offer a paradigm shift away from empirical pavement design and toward physics-consistent, interpretable, and computationally efficient deterioration modelling suitable for smart pavement management systems.