Baño-Medina et al. (2025) A regional high resolution AI weather model for the prediction of atmospheric rivers and extreme precipitation

Identification

• Journal: npj Climate and Atmospheric Science
• Year: 2025
• Date: 2025-12-12
• Authors: Jorge Baño-Medina, Agniv Sengupta, Daniel Steinhoff, Patrick Mulrooney, Thomas Nipen, Mario Santa-Cruz, Yanbo Nie, Luca Delle Monache
• DOI: 10.1038/s41612-025-01265-9

Research Groups

• Center for Western Weather and Water Extremes, Scripps Institution of Oceanography, University of California San Diego, La Jolla, CA, USA
• Norwegian Meteorological Institute, Oslo, Norway
• European Center for Medium-Range Weather Forecasts, Reading, UK

Short Summary

This study develops and evaluates a stretched-grid AI-driven weather model with high resolution over the Western United States for predicting atmospheric rivers and extreme precipitation. The model significantly reduces precipitation errors, performs competitively with regional numerical weather prediction (NWP) models, and effectively captures extreme events, outperforming coarser global models.

Objective

• To develop and comprehensively evaluate a regional, high-resolution (kilometer-scale) AI-driven stretched-grid weather model for forecasting atmospheric rivers and extreme precipitation events in the Western United States.
• To compare its performance against state-of-the-art global and regional numerical weather prediction (NWP) systems, as well as a lower-resolution AI model, assessing its accuracy, spatial detail, and computational efficiency.

Study Configuration

• Spatial Scale: Stretched-grid AI model with 6 km horizontal grid increments over the Western United States and North Pacific/North America, and approximately 31 km globally. Validation domain is located in the Western United States. Observational reference (PRISM) is at 4 km spatial resolution. West-WRF operates at 9 km horizontal grid spacing. IFS operates at 0.25 degree resolution.
• Temporal Scale: Forecasts generated at 6-hour intervals, with daily precipitation computed using a -12 to +12 UTC window. Forecast lead times extend up to 7 days. Training periods include ERA5 data from 1979–2020 and CW3E's 6-km reanalysis from 2012–2020. Validation was performed over three winter seasons (November 1 through March 31) for 2020–2021, 2021–2022, and 2022–2023.

Methodology and Data

• Models used:
  • AI-driven stretched-grid weather model (graph-transformer architecture).
  • Coarser version of the AI stretched-grid model (AI 31 km).
  • Integrated Forecasting System (IFS) from ECMWF (0.25 degree resolution).
  • Weather Research and Forecast (WRF) limited-area model (West-WRF, 9 km version).
• Data sources:
  • Training:
    • ERA5 reanalysis (European Center for Medium-Range Weather Forecasts Reanalysis version 5): Global, approximately 31 km horizontal resolution, N320 reduced Gaussian grid.
    • CW3E's 6-km high-resolution regional reanalysis: Covers the North Pacific and a significant portion of North America.
  • Observational Reference: Parameter-elevation Regressions on Independent Slopes Model (PRISM): Daily precipitation data at 4 km spatial resolution over the continental United States.
  • Atmospheric River Catalog: CW3E AR catalog (https://cw3e.ucsd.edu/Projects/ARCatalog/catalog.html).

Main Results

• The regional AI model (AI 6 km) significantly reduces 24-hour accumulated precipitation errors compared to global models.
• AI 6 km performs competitively with the regional NWP model (West-WRF) and substantially outperforms the global dynamical model (IFS) and the global AI model (AI 31 km) in precipitation forecasting.
• The AI 6 km model effectively captures extreme precipitation events, particularly those associated with atmospheric rivers, which global coarser models tend to underestimate.
• It exhibits minimal to no underestimation of extreme precipitation percentiles (up to the 99th percentile) and shows a power spectral density more consistent with observations than global models.
• Using neighborhood-based verification metrics (Fraction Skill Score, FSS), AI 6 km is competitive with West-WRF and outperforms IFS and AI 31 km, especially at longer lead times (e.g., 6-day forecasts).
• A case study of an atmospheric river event demonstrated AI 6 km's ability to reproduce fine-scale precipitation features and spatial distribution with high accuracy over the San Francisco Bay Area and the Sierra Nevada.
• The AI 6 km model generates a 7-day forecast in 6–7 minutes on a single H100 GPU, demonstrating significantly lower computational cost compared to traditional NWP methods.

Contributions

• Presents the first comprehensive evaluation of a regional, high-resolution (kilometer-scale) AI-driven stretched-grid weather model for precipitation forecasting, with a specific focus on atmospheric rivers and extreme precipitation in the Western United States.
• Demonstrates that regionally focused AI models can surpass coarser global AI counterparts and rival state-of-the-art limited-area dynamical models in performance, while operating at a fraction of the computational cost.
• Highlights the capability of AI models to accurately capture fine-scale atmospheric features and extreme precipitation events, addressing a known limitation of global AI models (overly smoothed fields).
• Introduces a novel stretched-grid AI model based on a graph-transformer architecture, trained using a combination of high-resolution regional and lower-resolution global datasets.

Funding

• Office of Naval Research (ONR) (Award Number N000142412731)
• California Department of Water Resources Atmospheric River Program Phase V (Grant 4600015671)
• US Army Corps of Engineers (USACE) Forecast Informed Reservoir Operations Phase 3 (USACE W912HZ-24-2-0001)
• National Artificial Intelligence Research Resource Pilot (NAIRR Award 240367)
• DeltaAI advanced computing and data resource (National Science Foundation (NSF-OAC 2320345))

Citation

@article{BañoMedina2025regional,
  author = {Baño-Medina, Jorge and Sengupta, Agniv and Steinhoff, Daniel and Mulrooney, Patrick and Nipen, Thomas and Santa-Cruz, Mario and Nie, Yanbo and Monache, Luca Delle},
  title = {A regional high resolution AI weather model for the prediction of atmospheric rivers and extreme precipitation},
  journal = {npj Climate and Atmospheric Science},
  year = {2025},
  doi = {10.1038/s41612-025-01265-9},
  url = {https://doi.org/10.1038/s41612-025-01265-9}
}