FourCastNetGFS: Experimental Global Weather Forecasts on a 0.25-Degree Grid

NOAA's FourCastNet Global Forecast System (FourCastNetGFS) is an experimental machine learning model producing medium-range global weather forecasts. It runs on a 0.25-degree latitude-longitude grid and 13 pressure levels, generating 6-hourly forecasts up to 10 days four times daily. The system is built upon Nvidia's pre-trained FourCastNet model, using Adaptive Fourier Neural Operators trained on ECMWF ERA5 reanalysis data.

Use Cases

• Benchmarking machine learning weather models against traditional numerical forecasts based on atmospheric fields.
• Analyzing forecast skill for major atmospheric variables like temperature, wind, and geopotential height.
• Studying the application of Fourier Neural Operators and vision transformer architectures for global prediction.
• Investigating medium-range (10-day) forecast uncertainty and error growth in an experimental system.

Strengths

• Provides forecasts on a high-resolution 0.25-degree global grid (approximately 28 km).
• Includes major atmospheric and surface fields such as temperature, wind components, geopotential height, and relative humidity.
• Produces forecasts four times daily (00Z, 06Z, 12Z, 18Z cycles) for consistent temporal coverage.
• Leverages a state-of-the-art, pre-trained machine learning model (FourCastNet) architecture.

Limitations

• Dataset is explicitly labeled as experimental, indicating it is for research and not operational use.
• Row count, size, and last updated date are unknown, which may limit suitability assessment.
• Column-level documentation is absent; field semantics and data structure must be inferred after download.

Provenance

Source

National Centers for Environmental Prediction (NCEP), NOAA.

Collection Method

Machine learning forecasts generated by the FourCastNetGFS system, initialized with NCEP GDAS analysis data.

Time Range

Forecasts are produced up to 10 days ahead from each initialization cycle.

Freshness

Update frequency is four times daily, but the last update date is unknown.

Geography

Global coverage on a 0.25-degree latitude-longitude grid.