Multiscale damage evolution mechanisms and life-cycle performance prediction modeling of asphalt pavement for resilient infrastructure

Asphalt pavements deteriorate under heavy loading, thermal extremes and freeze–thaw cycling, yet damage physics, life-cycle prediction and resilience assessment are still handled as separate problems. This study joins them. We define damage at the micro, meso and macro scales through a single effective-bearing-area variable, and we link the scales with viscoelastic-damage constitutive relations, a cohesive interfacial criterion and representative volume element homogenisation. Scale-separation limits are stated explicitly: a convergence study fixes the representative volume element at four times the nominal maximum aggregate size, and the homogenisation integral is restricted to meso damage below 0.4. Characterisation combines openly accessible condition, climate and traffic records with a laboratory programme of 216 independent specimens covering dynamic modulus, four-point bending fatigue, wheel tracking and interleaved freeze–thaw plus ultraviolet conditioning, supported by image-based mesostructural reconstruction that we validate against expert annotation and against a learned segmentation baseline. A hybrid model embeds the damage evolution equation and a monotonicity requirement in the training objective of a gated recurrent network with attention, and an adaptive weight keeps the data and physics gradients comparable. On a test partition isolated by section, by period and by region, the model reached a coefficient of determination of 0.91 ± 0.01 for rut depth and 0.87 ± 0.02 for cracking ratio over five seeds; the physical constraint raised extrapolation trend consistency from 0.79 to 0.93 and cut monotonicity violations from 11.6 to 0.3%, while keeping the empirical coverage of the nominal 90% interval at 89.4%. Cross-scale transfer was nonlinear in temperature for all three gradations tested, amplifying below freezing and above about 45 °C, although the magnitude was gradation-dependent and we do not claim generality beyond the materials examined. Maintenance strategies separated clearly across resistance, recovery and adaptation. Data, code and configuration files are provided with the paper so that the workflow can be rerun without licensed components.