AI weather forecasts retain skill beyond local heat records

If an AI weather system learns from past conditions, can its probability forecasts still be useful when a location sets a record? And does underprediction on selected extremes show that the system cannot handle weather beyond its experience?

The study compares four AI weather systems that issue multiple possible forecasts with ECMWF’s physics-based ensemble on heat, cold and wind records.

The authors compare four AI forecast ensembles with ECMWF’s IFS ensemble for land weather in 2022. They define local monthly records from ERA5 data for 1979–2021, extend thresholds beyond those records, and score both forecast probabilities and severity against the same verification reference.

At heat thresholds two local variability steps beyond the records, every AI system scores better than IFS-ENS at 2- and 10-day leads. Cold results are generally favorable but qualified, while wind results vary more.

Because the scores include cases without a record crossing, this advantage does not alone show better forecasts on realized extremes.

The main comparison covers one year of twice-daily forecasts on a 1.5° land grid, with few observed events at the deepest thresholds and dependent cases.

The systems differ in training history, initialization and design, so the study cannot attribute their score differences to any single feature or establish a causal benefit of probabilistic training.