Rehman et al. (2026) Physics vs. AI in weather prediction: evaluating GraphCast, AIFS, and FuXi against an observation-corrected WRF model for flash floods

Identification

• Journal: Geoscientific model development
• Year: 2026
• Date: 2026-10-13
• Authors: Haseeb ur Rehman, Felix Norman Teferle
• DOI: 10.5194/gmd-19-9831-2026

Research Groups

• Department of Engineering, Faculty of Science, Technology and Medicine, University of Luxembourg
• ECMWF (European Centre for Medium-Range Weather Forecasts)
• Google DeepMind
• Fudan University and the Shanghai AI Lab

Short Summary

This study compares the performance of a physics-based numerical weather prediction (NWP) model, WRF, with three artificial intelligence (AI)-based weather forecasting models: GraphCast, AIFS, and FuXi. The evaluation focuses on their ability to predict flash floods in Luxembourg and the Greater Region over three high-impact flood events.

Objective

• Evaluate the performance of AI-based weather forecasting models against a physics-based NWP model for predicting flash floods.
• Investigate the strengths and limitations of both approaches in predicting extreme weather events.

Study Configuration

• Spatial Scale: Regional scale, covering Luxembourg and the Greater Region (12 km horizontal grid spacing).
• Temporal Scale: 6-hourly cycling interval, with forecasts initialized every 6 hours.

Methodology and Data

• Models used:
  • WRF (Weather Research and Forecasting model)
  • GraphCast (multi-scale graph neural network)
  • AIFS (Artificial Intelligence Forecasting System)
  • FuXi (cascaded transformer-based forecasting system)
• Data sources:
  • ERA5 reanalysis
  • Conventional observations (SYNOP, METAR, TEMP, TAMDAR, and GNSS ZTD)
  • Weather radar data (RADFLOOD21 product)

Main Results

• The assimilated WRF configuration shows improved categorical precipitation skill at the 1 mm threshold.
• AIFS achieves the highest detection rate (POD) at the 1 mm threshold.
• GraphCast has the highest net categorical skill (CSI) at the 1 mm threshold.
• FuXi produces no exceedances at all at the 20 mm threshold.

Contributions

• This study provides a replicable model evaluation methodology for benchmarking emerging data-driven NWP systems against observation-corrected regional models.
• The results highlight the strengths and limitations of both physics-based and AI-based approaches in predicting flash floods.

Funding

• Not specified.

Citation

@article{Rehman2026Physics,
  author = {Rehman, Haseeb ur and Teferle, Felix Norman},
  title = {Physics vs. AI in weather prediction: evaluating GraphCast, AIFS, and FuXi against an observation-corrected WRF model for flash floods},
  journal = {Geoscientific model development},
  year = {2026},
  doi = {10.5194/gmd-19-9831-2026},
  url = {https://doi.org/10.5194/gmd-19-9831-2026}
}