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

Accurate prediction of severe weather events has never been more critical, as the frequency and intensity of extremes continue to rise under a changing climate. Recent advances in machine learning have transformed the landscape of weather forecasting, demonstrating remarkable skill and computational efficiency in short- and medium-range forecasts, increasing predictability and reducing forecast jumpiness. Nevertheless, physically based global models remain essential to ensure the interpretability and robustness of forecasts, particularly when dealing with rare and high-impact extreme events, where physical processes must be accurately represented.
In this context, the AIFS has shown clear improvements in skill for synoptic conditions and surface variables compared to the ECMWF IFS model. But do these improvements extend to extreme weather? Or are higher horizontal resolution and physically resolved processes still more crucial for predicting severe events accurately?
This presentation explores these questions by comparing AIFS and the DestinE Global Extremes Digital Twin (Extremes-DT) in their ability to forecast extreme 24-h precipitation, 10-m wind, and 2-m temperature extreme events in the extratropics. The analysis focuses on the added value and limitations of each system, aiming to guide users on which approach performs best under different extreme scenarios, and how AI-based and physics-based forecasts can complement each other. The evaluation employs the “scorecards for extremes” framework to quantify the performance of Extremes-DT and the AIFS deterministic approach in forecasting severe events relative to the IFS 9-km operational forecasts (49r1 model cycle). Extremes are defined in a percentile-based manner using SYNOP station climatology, and also with fixed thresholds, enabling consistent identification of rare events across all models. Forecast skill is assessed across multiple lead times, seasons, and orography (flat and complex terrain), and selected case studies are included to illustrate model behaviour in specific extreme events.