AI weather ensembles beat ECMWF's physics model on record heat and cold, preprint finds

A preprint by Bastiaan Van Velthoven, Cas Decancq, Diego G. Miralles, Stijn Luca and Thomas Mortier, posted to arXiv on Oct. 18, compared four probabilistic AI weather models (ArchesWeatherGen, U-Cast, WeatherNext2 and GenCast-mini) with ECMWF’s physics-based IFS-ENS ensemble on new monthly heat, cold and wind records set over land in 2022, against 1979–2021 records, at lead times of two to ten days. Verified against ERA5 reanalysis, the AI models outperformed IFS-ENS for heat and cold records at almost every lead time and record depth, and for heat the advantage grew with how far the record was broken, partly because IFS-ENS over-forecast deep records, the authors report. For wind the gains eroded for deeper records and longer leads, and all advantages narrowed with lead time.