Physics vs. AI in weather prediction: evaluating GraphCast, AIFS, and FuXi against an observation-corrected WRF model for flash floods

Rehman, Haseeb ur; Teferle, Félicia Norma Rebecca

doi:10.5194/gmd-19-9831-2026

Articles | Volume 19, issue 19

https://doi.org/10.5194/gmd-19-9831-2026

© Author(s) 2026. This work is distributed under
the Creative Commons Attribution 4.0 License.

https://doi.org/10.5194/gmd-19-9831-2026

© Author(s) 2026. This work is distributed under
the Creative Commons Attribution 4.0 License.

Articles | Volume 19, issue 19

Model evaluation paper

|

13 Oct 2026

Model evaluation paper |

| 13 Oct 2026

Physics vs. AI in weather prediction: evaluating GraphCast, AIFS, and FuXi against an observation-corrected WRF model for flash floods

Haseeb ur Rehman and Félicia Norma Rebecca Teferle

Data sets

Observational rainfall data of the 2021 mid-July flood event in Belgium - Part 2. Radar product RADFLOOD21 E. Goudenhoofdt et al. https://doi.org/10.5281/zenodo.7740059

Model code and software

Python scripts for WRF vs. AI weather model evaluation over Luxembourg flood events H. u. Rehman https://doi.org/10.5281/zenodo.20794936

Video supplement

Evaluating RADAR vs Physics based NWP vs AI Models H. u. Rehman https://doi.org/10.5446/73607

Short summary

Artificial intelligence (AI) weather models are now competing with traditional physics-based models for flood forecasting. We tested three of them and a physics-based model against observations from three flood events in Luxembourg. The AI models detected light rain and predicted temperatures better, but the physics-based model was more skilful at the heavy rainfall that actually causes floods. One AI model kept its skill on an event after its training period.