AIGFS, NOAA's machine-learned forecast

On a long ocean crossing, GFS is often the forecast a sailor downloads first, and the one that reaches furthest, sixteen days. But when it shows the trade wind failing in the second week, is that the weather or the model? A forecast from another centre mixes both answers: it starts from a different picture of the weather as well as computing it differently.

AIGFS, NOAA's machine-learned model, answers from the inside. Built on GraphCast, a network that learnt from decades of past weather how the atmosphere moves on, and refined by NOAA on its own analyses, it starts every run from the same picture of the weather as GFS. In service since December 2025, it now gives a route on every passage, beside GFS's.

It covers the whole world, four runs a day, a forecast every six hours up to sixteen days ahead, as far as GFS itself. NOAA reports it better than GFS on many of the large features several days out, where the lows and highs will lie, and on the track of tropical cyclones at longer range.

From the Canaries to Antigua, three weeks at sea, the first decision is how far south to dive for the trade wind. If AIGFS and GFS agree on where it fills in, the choice is clear; if they part, the two models alone disagree, since both start from the same weather, reason enough to keep a more cautious line. And where a model from another centre, ECMWF or AIFS, sides with one of them, that one gains weight.

At a quarter of a degree and with a step every six hours, it sees nothing of the coast and blurs a front that passes between two steps. Like every model learnt on an average error, it smooths: wind peaks are blunted, the more so the further ahead, and its first version underestimated the strength of tropical cyclones. And since it starts from GFS's picture, when the two agree they may share the same starting error: only a model from another centre tells you that.