Can Artificial Intelligence Global Weather Forecasting Models Capture Extreme Events? A Case Study of the 2022 Pakistan Floods (Papers Track)
Rodrigo Almeida (Fraunhofer HHI); Noelia Otero Felipe (Fraunhofer HHI); Miguel-Ángel Fernández-Torres (UC3M); Jackie Ma (Fraunhofer HHI)
Abstract
Climate change is increasing extreme events. Despite advances in artificial intelligence-based global weather forecasting, most benchmarks remain deterministic, which limits uncertainty representation and extremes prediction skill assessment. This study examines how three state-of-the-art deterministic data-driven models, namely FourCastNet v2/SFNO, GraphCast, and FuXi, respond to input perturbations, evaluating the corresponding ensembles in forecasting extremes. 50-member ensembles are created using perturbation methods based on spherical Gaussian noise, hemispherical centered bred vectors, and huge ensembles. Focusing on August 2022, when devastating floods hit Pakistan, we compare our ensembles against deterministic ERA5-initialized forecasts and the ECMWF Integrated Forecasting System Ensemble for numerical weather prediction. While the huge ensembles method outperforms those based on Gaussian noise and hemispherical centered bred vectors in detecting the associated extreme precipitation event, all models still underperform the numerical weather model, suggesting promising research avenues.