State-aware supervised learning for post-disaster road repair sequencing: A case study of the Dujiangyan road network

Road repair order determines the structural recovery available before a disrupted network is fully restored. This study develops and evaluates a state-aware supervised learning method for post-disaster road repair sequencing. A processed subset of the Dujiangyan road network is evaluated under simulated closures and homogeneous serial repair assumptions. Three greedy rules generate complete initial sequences. A supervised query gate combines road relationships, current-sequence information and the network state at the proposed adjustment to prioritize candidates for exact evaluation. Acceptance requires an improvement in the preference score and a guard against a decrease in normalized recovery area. Across 27 fixed confirmation scenarios, the accepted adjustments improve the initial sequences. However, the mean preference-score increment over a strong non-learning baseline remains statistically uncertain after correction for multiple comparisons. A higher preference score can also accompany weaker early recovery. Re-evaluation of saved sequences identifies how an illustrative road endpoint-sharing preference changes sequence selection and the associated structural recovery. The method connects candidate screening with verified acceptance, while the assessment distinguishes accumulated recovery from recovery at intermediate stages. These outputs support the review of structural repair priorities before operational scheduling.