Tropical cyclones: probabilistic AI weather models are better at tracking trajectories than the CEPMMT ensemble

Compared with IFS-ENS, GenCast, FourCastNet 3 and FGN are more accurate at locating cyclones; FGN performs better for 90 per cent of the storms in 2025. Intensity and dispersion remain their weak points.

AI-based weather forecasting models already beat, on average, the best numerical systems for a fraction of the computational cost. One disputed question remains: do they hold up against extreme events? A study filed on arXiv compares three probabilistic models, GenCast, FourCastNet 3 and FGN, to IFS-ENS, the European Central Operational Ensemble Forecast (ECMWF), on tropical cyclones. The authors designed a battery of probabilistic indicators covering the track, intensity biases, uncertainty and the wind-pressure relationship that reflects the storm's structure.

On the trajectory, AI models clearly prevail: FGN obtains better probabilistic location scores than the physical model for 90% of the storms assessed in 2025. However, they still underestimate the intensity and dispersion of their ensembles remains insufficient. The analysis also reveals failures specific to each model and a persistent influence of pre-training data. For the authors, these results nevertheless show that AI models capture physical relationships that remain predictive well beyond ordinary weather conditions.