Abstract
Physics-based nowcasting cascades for urban air mobility (UAM) must deliver 30 min wind forecasts on a 5 min decision cycle, yet their dominant error is not the physics but the statistics of their own advection operator: motion fields estimated from analysis differences inherit observation and assimilation noise, so forecasts are systematically over-advected, and the bias drifts with the flow regime faster than any retraining cycle. This paper therefore treats systematic advection bias as an online learning problem. We formalize delay-verified self-supervision, in which the analysis produced by a later cycle labels the forecast issued by an earlier one, and instantiate it as a self-supervised online advection-damping (SSOAD) learner that keeps one coefficient per lead through exponentially forgetting least squares with variance shrinkage and hard physical bounds, at a cost of three scalars per lead. The evaluation uses a controlled low-level-jet benchmark at 1 km resolution, comprising 30 independent scenarios, seven 5 min analysis cycles each, and 18 heterogeneous in-window observations per cycle and compares the corrected cascade with a physics-only backbone and four external reference systems under a common protocol. The corrected cascade reaches u- and v-component RMSEs of 1.97 and 1.51 m s−1, and the matched ablation isolates a further 6.2% (u) and 5.6% (v) error reduction attributable to the learner itself; it is the only evaluated system that detects clear-air wind shear at all, and it raises vertiport takeoff/landing index accuracy to 0.967. Scientifically, the results show that horizon-resolved, label-free correction of advection residuals recovers wind-forecast error that is unavailable to offline-trained models; practically, they show that the gain survives the hazard-to-risk mapping into the products that drive go/delay/divert scheduling. Future work will extend the method to hectometer-resolution real-scene validation with in situ station, wind-profiler, and Doppler lidar observations.