Researchers compared two weather‑prediction systems: the AI‑driven AIFS‑CRPS and the physics‑based Met Office MOGREPS‑UK. Each system’s raw forecasts were first calibrated with Ensemble Model Output Statistics (EMOS). After calibration, the two ensembles were combined using a simple constant weighting: 80 % from the better‑performing model and 20 % from the other, as implemented in the IMPROVER post‑processing system. The blended forecast was evaluated against a 41‑threshold exceedance probability scheme.

Across every lead time, the blended model showed lower Continuous Ranked Probability Scores (CRPS) and root‑mean‑squared errors (RMSE) than either AIFS‑CRPS or MOGREPS‑UK alone. The improvement held for both probabilistic and deterministic metrics. A visual comparison of mean CRPS and RMSE over all days and stations confirms that the blend consistently outperforms the individual models (see Figure 1).