WeatherBridge and the limits of global weather interpolation at held-out query times

Abstract Many forecast archives provide six-hourly fields, but applications need hourly states. We ask whether a learned interpolator can recover intermediate trajecto-ries from two endpoints at query hours omitted from training. We compare linear interpolation and six architectures at six-and twelve-hour horizons, and test transfer to IFS HRES forecasts. A matched six-hour WeatherBridge control receives supervision at every interior hour. WeatherBridge combines query-conditioned bidirectional transport with multiscale and spectral corrections. It has the lowest aggregate error at held-out query hours at both horizons, although no model wins every field. The matched control improves held-out query hours but barely changes trained ones, measuring the cost of withholding query times. Models trained on ERA5 lose this advantage on forecast trajectories; HRES adaptation recovers about 7% over linear interpolation in a retrospective audit. Learned interpolation captures nonlinear motion and small-scale structure between sparse anchors, while temporal coverage and anchor-domain mismatch bound its benefit.