Six-hourly archives lack the hours apps need: WeatherBridge fills them with bidirectional transport, cutting held-out-hour error 48.6% below linear interpolation

Reconstructing the interior hours omitted from training between two six-hour anchors keeps an aggregate error advantage for learned interpolation, rather than only reproducing trained hours. Linear interpolation treats each grid value as the chord between its endpoint values, so a moving pressure system, moisture filament or near-surface wind feature can be attenuated; weather time interpolation had received less controlled study than forecasting. On the 2020 ERA5 benchmark at held-out hours τ∈{2,4} of the six-hour task, WeatherBridge reaches normalised RMSE 0.07431, 9.0% below WeatherDCAE-14M's 0.082 and 48.6% below linear interpolation's 0.144; at twelve hours the held-out value is 0.13277, 5.8% and 52.3% below the same references.

Restoring hours 2 and 4 to training lowers error at those hours by about 8.8% while barely changing trained hours, isolating the cost of withholding query times. This control shares seed, update budget, batch size, precision and loss weights with the main model and changes only the training query set from {1,3,5} to {1,2,3,4,5} hours. Direct supervision lowers held-out-hour RMSE by 8.77% in 2020 and 8.78% in 2021, while trained hours 1, 3 and 5 change by less than 0.4%; the sparse-query checkpoint still transfers useful skill to both omitted hours.

The model moves resolved weather structure toward the requested hour, then uses additive branches for amplitude and shape changes that displacement cannot express. The motion head emits first- and second-order displacement coefficients shared across weather channels, a blend gate weights the two warped anchors per channel and grid point, and linear interpolation stays as an explicit fallback. On the same checkpoint, disabling warped-anchor transport raises full-year 2020 aggregate RMSE by 60.54%, disabling acceleration raises it by 2.16%, and disabling the hydrostatic head changes the aggregate by 0.04% while worsening all 25 geopotential/MSLP field-hour cells it modifies.

On IFS HRES forecast trajectories, ERA5-trained corrections transfer poorly, and only full-weight adaptation restores a gain over linear interpolation. Without adaptation, WeatherBridge is 1.23% better than linear interpolation on the fixed 2022 retrospective set, and WeatherDCAE-14M is 0.91% better. On the 24-initialisation 2022 retrospective audit, the full-range coarse-grid checkpoint lowers the mean of 120 field-hour RMSE values from 0.2902 to 0.2691, or 7.28%; that branch postdates inspection of 2022, so the figure is retrospective.

The evaluated scope is deterministic interpolation of 24 forecast fields on a nominal 0.5° grid, with six- and twelve-hour anchor gaps, covering the lower troposphere from 1000 to 700 hPa plus T2m, U10, V10 and MSLP, and ending at the later anchor. The authors state that the present checkpoint supports retrospective gap filling on a finer temporal grid, while warnings, aviation routing and other safety-critical decisions lie outside its evaluated scope.

The authors state that the evidence covers integer query hours, with continuous-time evaluation and a matched twelve-hour all-query control still open; at twelve hours WeatherBridge preserves less high-frequency energy than WeatherDCAE-14M. The 7.28% HRES figure is retrospective because the full-range branch was designed after 2022 had been inspected, and it combines temporal reconstruction with correction of systematic forecast error; the public WeatherBench-2 HRES archive ends in 2022, so confirmation requires a later operational archive or an independently sourced forecast trajectory.