WeatherBridge is a unified, source‑conditioned adapter that maps heterogeneous operational NWP initial states to ERA5 reference states without updating the pretrained forecasting model. It predicts structured residual corrections with a Swin‑UNet and injects source embeddings across multiple spatial scales, enabling the capture of shared atmospheric structures and source‑specific discrepancies.
The training objective combines a latitude‑weighted ERA5 reconstruction loss with One‑step Rollout Alignment (ORA). ORA back‑propagates the one‑step forecast error through a frozen FuXi model, optimizing the adapted states for both initial‑state fidelity and downstream forecast compatibility.
We evaluate WeatherBridge on daily GFS, HRES, and CMA initializations from 1 January to 22 November 2025, covering six variables and lead times up to 240 h. Compared with raw source initialization of FuXi, WeatherBridge reduces the mean RMSE over the 6‑240 h rollout for 17 of 18 source‑variable combinations, with a median reduction of 10.2 % and a maximum reduction of 46.3 %. Moreover, when trained with a frozen FuXi, the same adapter improves FengWu forecasts on GFS inputs without retraining either the adapter or FengWu.
These findings demonstrate that forecast‑aware input adaptation can enhance forecasts from heterogeneous operational NWP inputs and transfer across pretrained ERA5‑based forecasting models.
Review