NOAA AIGFS

From GraphCast to AIGFS

In December 2022 Google DeepMind described GraphCast, a neural network trained on ERA5, ECMWF's reanalysis of past weather, that produces a 10-day global forecast at 0.25° in under a minute on a single Google TPU v4 processor. Published in Science in November 2023, it beat ECMWF's high-resolution physics-based forecast on 90% of 1,380 verification targets. But it had learned from ECMWF's data and took ECMWF's data as its starting state, so the open question for other centres was whether it would work from their own analyses.

NOAA's Environmental Modeling Center (EMC) set out to answer it. Experimental GraphCastGFS runs, GraphCast started from NOAA's own analysis, were published from 5 February 2024. EMC then fine-tuned GraphCast on the analyses of NOAA's Global Data Assimilation System (GDAS): in its March 2025 report the fine-tuned model had larger errors than the GFS in the first two days but smaller ones beyond on the wind, the 500 hPa height and precipitation, about 10% less on the 500 hPa height at day 10, and better tropical cyclone tracks, but worse cyclone intensities. Real-time runs continued under Project EAGLE, a joint effort of the National Weather Service, NOAA's research laboratories, EMC and the Earth Prediction Innovation Center, joined by an experimental ensemble in April 2025.

AIGFS v1.0 became operational on 17 December 2025, with the 12 UTC run, together with two ensembles, AIGEFS and HGEFS. NOAA announced better skill than the GFS for many large-scale features and smaller tropical cyclone track errors at longer lead times, and warned that v1.0 degraded cyclone intensity forecasts. The experimental EAGLE forecasts it replaced stopped on 17 and 18 December, while the older experimental GraphCastGFS runs went on until 00 UTC on 5 May 2026, their last in NOAA's cloud archive.

AIGFS v1.1 went live on 27 July 2026, again with the 12 UTC run. Its network is scored differently during training: through a spherical harmonic loss, computed on the patterns of each field scale by scale rather than point by point, and on wind speed and direction instead of the east-west and north-south components; it was fine-tuned on four years of GDAS analyses, and trained on forecasts chained out to 72 hours. The National Weather Service says these changes improve tropical cyclone intensity and precipitation forecasts and reduce the blurring of forecasts at longer lead times.

A network that learned the atmosphere

The AIGFS solves no equation. Like the GraphCast it comes from, it has learned from past states of the atmosphere how one state turns into the next six hours later. Given the state now and the state six hours ago, it predicts the state six hours ahead; it then feeds its own forecast back in and repeats, 64 times to reach 16 days. This chaining is called autoregression, and every error made early is carried into the following steps.

The network has three parts. An encoder, a single layer of a graph neural network, gathers the 1,038,240 points of the 0.25° grid onto a hidden mesh built by dividing the faces of an icosahedron six times, down to 40,962 evenly spaced nodes. A processor of 16 graph layers passes messages along the edges of that mesh, short ones for local weather and long ones for distant influences. A decoder brings the result back onto the latitude-longitude grid as a change to the latest state. In all, the network has 36.7 million parameters.

What goes in and out is a compact picture of the atmosphere: wind, temperature, geopotential height, specific humidity and vertical velocity on 13 pressure levels from 1000 to 50 hPa, plus surface fields: the 10 m wind, the 2 m temperature, the sea-level pressure and precipitation. There is no gust, no wave and no cloud field.

GraphCast was trained by Google DeepMind on ERA5 from 1979 to 2017, and its 13-level version was then fine-tuned on ECMWF's operational analyses from 2016 to 2021. NCEP started from that version and fine-tuned it on its own GDAS analyses, so that it learns the particular features of the starting state it receives every day. GraphCast was scored on a mean squared error, and that choice explains its main flaw: when it is unsure, the error is smallest if it predicts the average of the possible outcomes, so its fields grow smoother with lead time. The changes of v1.1 are meant to reduce that blurring.

Four runs a day, at a fraction of the GFS's cost

The AIGFS does not read observations itself. Each run starts from NOAA's GFS-based analysis, produced by its Global Data Assimilation System (GDAS), which combines observations from satellites, aircraft, radiosondes and surface stations into the best estimate of the present state of the atmosphere. The AIGFS therefore depends entirely on the physics-based system and its observation network for its starting point.

The forecast itself is cheap: according to NOAA, a 16-day forecast takes about 40 minutes and uses 0.3% of the computing resources of the operational GFS. Runs start at 00, 06, 12 and 18 UTC and produce a state every 6 hours, from the analysis to 384 hours, 16 days ahead.

The output is published in GRIB2 on the 0.25° grid on NOMADS, NOAA's model distribution server, in two files per step: one on pressure levels and one at the surface, with the 10 m wind, the 2 m temperature, the sea-level pressure and precipitation. Since April 2026 the runs are also copied, a few hours later, to NOAA's open-data archive on Amazon Web Services, which also holds the experimental development versions. As information produced by the US government, it is in the public domain.

Its siblings AIGEFS and HGEFS, implemented on the same day, deal with uncertainty. AIGEFS runs 31 machine-learning forecasts out to 16 days; HGEFS combines them with the 31 members of the physics-based GEFS into a 62-member ensemble out to 10 days, which NOAA describes as the first hybrid physical and machine-learning ensemble run operationally by a weather centre. Saily WR uses the AIGFS, not the ensembles.

Reading it at sea

Where the AIGFS shines is the large-scale picture several days ahead: where the lows and highs will be, when a front will arrive, which way a tropical cyclone will go. NOAA reports better skill than the GFS for many large-scale features and smaller cyclone track errors at longer lead times, and NOAA's verification of v1.0 found a statistically significant gain over the GFS on the anomaly correlation of geopotential height. For an ocean passage, where the choice of a route is mostly the choice of a weather system to ride or avoid, that is the information that counts.

Its weakness is the strength of the wind. A network trained on an average error smooths: wind peaks are blunted, tight gradients are softened, and the effect grows with lead time. Version 1.0 underestimated the intensity of tropical cyclones; v1.1 was built to correct it and to reduce the blurring, without a published figure for how much. In a gale, near a cyclone or along a sharp front, read its wind speed with caution and its timing and position as the useful part.

At 0.25°, about 28 km, and with one state every 6 hours, it sees nothing of what happens near land: no sea breeze, no acceleration around a cape, no wind shadow behind an island, no squall line, and it gives no gusts. A front that passes between two steps is only known through what happens before and after it. For coastal sailing or a decision measured in hours, a fine-mesh regional model tells far more.

Its best use is as a second opinion next to the GFS. Both start from the same NOAA analysis, so the AIGFS is not independent of the GFS: when they part ways the difference comes from the models alone, which is a clean signal that the forecast is uncertain, but when they agree they may share the same error in the starting state. A model started from another centre's analysis, such as the IFS or the AIFS, gives the independent view; when all of them agree on a route, confidence is high.

AIGFS in Saily WR

Saily WR downloads each new AIGFS run from NOAA's NOMADS server, four times a day, on the 0.25° grid: the 10 m wind, as its east-west and north-south components, every 6 hours out to 16 days (384 hours). A run is complete about 4 h 20 after its cycle time. Because the model is global, every route gets an AIGFS version, wherever it starts and ends.

The isochrone route is sailed from your boat's polar in that wind, interpolated in space between grid points and in time between the 6-hourly steps. It is drawn next to the routes of the other models covering the same water, the GFS among them, so a difference between the AIGFS and GFS routes shows at a glance where the two forecasts part ways.

The Data page shows live which AIGFS run the app is on and how far ahead it reaches.