Forecasting extremes: what AIFS and the physics-based km-scale global IFS model can (and can’t) do

As AI-based weather models are rapidly closing the gap with physics-based systems in average forecast skill, their ability to predict extreme events (where accurate process representation may be critical) remains an open question. AIFS (ECMWF's global artificial intelligence model) has shown clear improvements in average forecast skill for synoptic conditions and surface variables compared to ECMWF's physics-based IFS. But do these improvements extend to the prediction of extreme events? Or do high horizontal resolution and the explicit representation of physical processes remain more important factors for accurately predicting severe events? This presentation addresses these questions by comparing AIFS and the experimental IFS at 4.4 km horizontal resolution in their ability to predict extremes of 24-hour accumulated precipitation, 10 m wind speed, and 2 m temperature over the extratropical regions of the Northern Hemisphere compared to the current operational IFS 9 km resolution model. AIFS-ENS (AIFS ensemble) is also compared against the operational IFS-ENS to evaluate their performance in predicting extreme events in a probabilistic framework.The analysis focuses on the added value and limitations of each system, with the aim of guiding users on which approach offers better performance under different extreme scenarios, and how AI-based and physics-based forecasts can complement each other. 

The evaluation uses the new "scorecards for extremes" framework to quantify the skill of the high-resolution IFS and the AIFS in predicting severe events, benchmarked against operational IFS forecasts at 9 km (model cycle 49r1). It allows to define extremes using both percentile-based thresholds derived from the SYNOP station climatology or fixed absolute thresholds, enabling the identification of events that are climatologically rare as well as those with high societal impact. Forecast skill is assessed across multiple lead times, seasons, and orographic complexities (flat and mountainous terrain). Selected case studies illustrate model behaviour during specific extreme events.