Do AI weather models miss extremes?
Abstract
AI weather models are often reported to underestimate extremes, but most evidence concerns deterministic regression models verified against reanalysis. We evaluate twelve physical and AI forecast models against ECMWF IFS using ten months of European station observations. The evaluation covers 10 m wind, 2 m temperature, solar radiation, and precipitation within regimes defined from a fixed ERA5 1991-2020 climatology. We find no uniform AI-specific deficit in the tails. Several AI models remain more accurate than IFS under extreme conditions, while others deteriorate markedly; comparable variation occurs among physical models. Every model nevertheless exhibits a common conditional-error pattern, overpredicting low observations and underpredicting high observations. Attenuation of extreme values therefore does not imply a uniform loss of relative skill: tail performance depends on the model, variable, and evaluation setting rather than on whether the forecast is produced by AI or physical numerical modelling.
Related
- Probabilistic AI models outperform numerical weather prediction models on record breaking heat extremes
- Can We Trust Evaluations of AI Weather Forecasts for Extremes? How Evaluation Protocols Shape Model Rankings
- Can AI Forecast the Extremes? Evaluating Probabilistic Weather Models on Tropical Cyclones
- Air Quality in Cities: What Do Sparse Sensor Networks Miss?
- Distinguishing leaf scorching from senescence under climate extremes
BibTeX
@misc{gabler2026weather,
title = {{Do AI weather models miss extremes?}},
author = {Marvin Vincent Gabler and Roberto Molinaro and Niall Siegenheim and Henry Martin and Mark Frey and Niels Poulsen and Philipp Seitz and Olivier Lam},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2608.09972},
eprint = {2608.09972},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2608.09972},
note = {CC BY-NC-SA 4.0},
}
RIS
TY - UNPB
TI - Do AI weather models miss extremes?
AU - Marvin Vincent Gabler
AU - Roberto Molinaro
AU - Niall Siegenheim
AU - Henry Martin
AU - Mark Frey
AU - Niels Poulsen
AU - Philipp Seitz
AU - Olivier Lam
PY - 2026
DA - 2026/10/06
PB - arXiv
DO - 10.48550/arxiv.2608.09972
UR - https://arxiv.org/abs/2608.09972
AB - AI weather models are often reported to underestimate extremes, but most evidence concerns deterministic regression models verified against reanalysis. We evaluate twelve physical and AI forecast models against ECMWF IFS using ten months of European station observations. The evaluation covers 10 m wind, 2 m temperature, solar radiation, and precipitation within regimes defined from a fixed ERA5 1991-2020 climatology. We find no uniform AI-specific deficit in the tails. Several AI models remain more accurate than IFS under extreme conditions, while others deteriorate markedly; comparable variation occurs among physical models. Every model nevertheless exhibits a common conditional-error pattern, overpredicting low observations and underpredicting high observations. Attenuation of extreme values therefore does not imply a uniform loss of relative skill: tail performance depends on the model, variable, and evaluation setting rather than on whether the forecast is produced by AI or physical numerical modelling.
ER -