López-Martí et al. (2025) Can data-driven weather models accurately forecast atmospheric rivers?

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

Identification

• Journal: Environmental Research Letters
• Year: 2025
• Date: 2025-11-12
• Authors: Ferran López-Martí, Leonardo Olivetti, Sara M. Vallejo‐Bernal, Anna Rutgersson, Gabriele Messori
• DOI: 10.1088/1748-9326/ae1e8e

Research Groups

Not explicitly mentioned in the abstract, but implies involvement of institutions developing and operating the models and reanalysis data (e.g., ECMWF).

Short Summary

This study assesses the performance of leading data-driven weather models (GraphCast, Pangu-Weather) against a physics-based model (IFS-HRES) in forecasting integrated vapour transport (IVT) and atmospheric rivers (ARs). While data-driven models show comparable IVT skill, they struggle with higher IVT quantiles and geometrically stricter AR detection, suggesting physics-based models may retain advantages for complex derived features.

Objective

• To assess the performance of two leading operational data-driven weather models (GraphCast and Pangu-Weather) in forecasting integrated vapour transport (IVT) and atmospheric rivers (ARs) relative to a traditional physics-based numerical weather prediction system (ECMWF’s IFS-HRES), particularly concerning physically complex, derived variables and extreme events.

Study Configuration

• Spatial Scale: Global
• Temporal Scale: One year of global data; lead times ranging from 1 day to 10 days.

Methodology and Data

• Models used: GraphCast (data-driven), Pangu-Weather (data-driven), ECMWF’s IFS-HRES (physics-based reference).
• Data sources: ERA5 reanalysis (for evaluation); three distinct atmospheric river (AR) detection algorithms.

Main Results

• Data-driven models achieved root-mean-square errors for integrated vapour transport (IVT) comparable to or slightly better than IFS-HRES, particularly in the tropics and at shorter lead times.
• Data-driven models demonstrated a poorer representation of the higher quantiles of the IVT distribution.
• All models could forecast the main characteristics of a high-impact atmospheric river (AR) event up to five days in advance.
• AR characteristics and detection performance varied substantially across the three AR detection algorithms used.
• A geometrically stricter AR detection method highlighted a clearer advantage for IFS-HRES, especially in the midlatitudes and at shorter lead times.
• Overall, no single model systematically outperformed the others across all AR detection algorithms.

Contributions

• Provides a targeted evaluation of leading operational data-driven weather models (GraphCast and Pangu-Weather) for forecasting physically complex, derived variables like integrated vapour transport (IVT) and atmospheric rivers (ARs), which is crucial for understanding their capabilities beyond standard forecast metrics.
• Highlights specific strengths and weaknesses of data-driven models compared to physics-based models for extreme events and derived phenomena, particularly their struggle with higher IVT quantiles and geometrically stricter AR detection.
• Underscores the importance of developing and utilizing targeted evaluation frameworks for derived and extreme phenomena, especially under strict detection criteria, as data-driven models become more central in operational forecasting.

Funding

Not mentioned in the abstract.

Citation

@article{LópezMartí2025Can,
  author = {López-Martí, Ferran and Olivetti, Leonardo and Vallejo‐Bernal, Sara M. and Rutgersson, Anna and Messori, Gabriele},
  title = {Can data-driven weather models accurately forecast atmospheric rivers?},
  journal = {Environmental Research Letters},
  year = {2025},
  doi = {10.1088/1748-9326/ae1e8e},
  url = {https://doi.org/10.1088/1748-9326/ae1e8e}
}