GraphCast Global Forecast System (GFS GraphCast)

Status: GribStream currently serves graphcast as a historical experimental archive. The latest indexed GraphCastGFS run in GribStream is 2026-05-05 00:00 UTC. NOAA's EAGLE registry describes AIGFS and AIGEFS as the operational replacements for the EAGLE SOLO/Ensemble lineage, so use AIGFS or AIGEFS for current NOAA AI global guidance.

Description

GraphCast Global Forecast System is an experimental global forecast from NOAA/NCEP built on Google DeepMind’s GraphCast machine-learning model. When the feed was active, it ran four cycles per day (00, 06, 12, 18 UTC) and produced 6-hourly forecasts of atmosphere and surface fields. In GribStream, it is best treated as a historical comparison dataset, not as the current NOAA AI global forecast feed.

The core model is a message-passing graph neural network with an encoder–processor–decoder design. It was pre-trained on ECMWF’s ERA5 reanalysis and in NOAA’s implementation uses two GDAS analysis states, the current time and the one 6 hours earlier, as inputs to step the forecast forward in 6-hour increments. The indexed configuration focuses on a 13-level pressure grid and provides temperature, wind components, geopotential height, specific humidity, vertical velocity, and standard surface variables.

GraphCastGFS is related to GFS through its GDAS initial conditions and is a peer to other AI-based global models such as AIFS. It is not a replacement for numerical weather prediction and is best used alongside physics-based models for validation, backtesting, and risk assessment.

Detail

• Archive begins: 2024-04-25
• Product page: GraphCast GFS documentation
• Resolution: 0.25° global grid
• Vertical levels: 13 pressure levels (the earlier 37-level stream was retired)
• Cycles: historical feed with 00, 06, 12, 18 UTC cycles while active
• Forecast cadence and length: 6-hour steps to 384 hours
• Inputs: GDAS analyses at t0 and t0-6 h; model pre-trained on ERA5
• Status: Historical experimental archive in GribStream; latest indexed run 2026-05-05 00:00 UTC
• Primary uses: Medium-range scenario replay, AI-vs-physics checks, research, benchmarking, and backtesting

Parameters