Pasquini et al. (2026) Assessing the ability of a stretched‐grid deep‐learning weather prediction model to capture physical balances

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: Quarterly Journal of the Royal Meteorological Society
• Year: 2026
• Date: 2026-10-15
• Authors: Francesco Pasquini, Michiel Baatsen, Bastien François, Natalie Theeuwes, Maurice J. Schmeits
• DOI: 10.1002/qj.70315

Research Groups

• Norwegian Meteorological Institute (Met Norway)
• University of [Unknown]

Short Summary

This study compares the performance of a deep learning weather prediction model, Bris, with a traditional numerical weather prediction model, MEPS, in forecasting an extreme extratropical cyclone event. The results show that Bris struggles to capture key mesoscale features and disrupts atmospheric balances due to fine-scale noise.

Objective

• Investigate the physical consistency of deep learning weather prediction models during extreme events

Study Configuration

• Spatial Scale: Regional (Europe)
• Temporal Scale: 6 days (forecast period)

Methodology and Data

• Models used: Bris (deep learning), MEPS (numerical weather prediction)
• Data sources: Not specified

Main Results

• Bris struggles to capture key mesoscale features of the event, including incorrect and unrealistic spatial gradients.
• The model disrupts several atmospheric balances significantly due to fine-scale noise in its output fields.

Contributions

• This study highlights the limitations of deep learning weather prediction models during extreme events and emphasizes the need for improved physical consistency in their predictions.

Funding

• Not specified

Citation

@article{Pasquini2026Assessing,
  author = {Pasquini, Francesco and Baatsen, Michiel and François, Bastien and Theeuwes, Natalie and Schmeits, Maurice J.},
  title = {Assessing the ability of a stretched‐grid deep‐learning weather prediction model to capture physical balances},
  journal = {Quarterly Journal of the Royal Meteorological Society},
  year = {2026},
  doi = {10.1002/qj.70315},
  url = {https://doi.org/10.1002/qj.70315}
}