The Artificial Intelligence Forecasting System (AIFS), recently released by the European Centre for Medium-Range Weather Forecasts (ECMWF), represents a major shift in global weather prediction by replacing traditional physically based approaches with machine-learning methods. This study evaluates the impact of using AIFS as initial and lateral boundary conditions for the Weather Research and Forecasting (WRF) model, in contrast to the well-established physically based GFS. The aim of this work is to analyze the sensitivity of these different modelling configurations during three high-impact storms that affected Spain in 2025 and the effects of replacing GFS for AIFS as lateral and boundary conditions for WRF over the accuracy of operational forecasts. The analysis focuses on maximum wind gusts, accumulated precipitation, and the generation of meteorological warnings. Results show that AIFS substantially underestimates wind gusts with mean bias values between −13 and −25 km/h, and its forecasts differ markedly from those of GFS. When coupled with WRF, however, both AIFS-WRF and GFS-WRF produce similar results, with a general tendency to overestimate gusts, with mean bias values between 4 and 15 km/h. In all cases, WRF adds value, improving the representation of wind-related variables compared with the raw global model outputs. For accumulated precipitation, both WRF configurations reproduce the main rainfall patterns associated with the storms. AIFS-WRF shows a stronger tendency to overestimate precipitation, with RMSE values of 64, 23, and 12 mm for the different high-impact storms considered, although it also achieves the highest correlations. Finally, the analysis of meteorological warnings indicates that AIFS alone generates almost no wind gusts alerts. Once coupled with WRF, both configurations generate warnings in the regions where the most severe conditions occurred. Overall, while the added value of mesoscale models such as WRF is well established and confirmed here, the AI-based AIFS does not show clear advantages in comparison with traditional global models for these high-impact events being analyzed.
Agudo et al. (Sat,) studied this question.