ECMWF AIFS

From physics to learning

For seventy years, a weather forecast meant solving the equations of fluid physics on a supercomputer. In 2022 three preprints from outside the meteorological services shook that idea: FourCastNet from NVIDIA in February, Pangu-Weather from Huawei in November and GraphCast from Google DeepMind in December. All three were neural networks trained on ERA5, ECMWF's own reanalysis of past weather, and all three claimed to rival the IFS, ECMWF's physics-based model, on many scores for a tiny fraction of its computing cost.

ECMWF answered by running these outside models in real time from its own analyses, and by building its own: the Artificial Intelligence Forecasting System, AIFS. A first version ran in experimental mode from October 2023 at about 1°; in February 2024 it moved to about 0.25° and to the transformer-based design it still uses. It is built on Anemoi, an open-source framework that ECMWF and several European national weather services launched in 2024 to train such models on their own data.

AIFS Single v1, the deterministic version (one forecast per run), became operational on 25 February 2025, running alongside the IFS rather than replacing it. ECMWF reported gains of up to 20% over physics-based models on some measures, tropical cyclone tracks among them, for about 1,000 times less energy per forecast. Version 1.1 followed on 27 August 2025, mainly to remove spurious spots of rain that came from an unphysical link the network had learned between soil moisture and precipitation. An ensemble version, AIFS ENS, became operational on 1 July 2025.

AIFS Single v2 went live on 12 May 2026, together with cycle 50r1 of the IFS. Its architecture is unchanged, but it was fine-tuned on two more years of recent analyses, gained a 10 hPa level in the stratosphere and, for the first time at ECMWF, forecasts waves and snow cover from a network instead of a physical model. On the same day ECMWF stopped running the outside experimental models, its own AIFS having taken their place.

A network that learned the atmosphere

The AIFS solves no equation. It has learned, from decades of 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, step after step, out to 15 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 graph neural network with attention (a mechanism that lets each point weigh what its neighbours tell it), gathers the 542,080 points of the N320 grid, about 31 km apart, onto a coarser hidden grid of about 40,000 points. A processor, a transformer whose attention works on sliding windows of that hidden grid, carries the atmosphere forward six hours; the published description gives it 16 layers. A decoder, another graph network, brings the result back onto the 31 km grid.

What goes in and out is a compact picture of the atmosphere: geopotential, temperature, wind and humidity on 14 pressure levels from 1000 to 10 hPa, plus surface fields such as the 10 m wind, 2 m temperature, sea-level pressure, total column water and the temperature and moisture of two soil layers. Some outputs are only predicted, never fed back: precipitation, the 100 m wind, cloud cover, solar radiation and, since v2, eleven wave variables such as significant wave height, period and direction.

Training happens in two phases. The network is first pre-trained on ERA5 from 1979 to 2022, then fine-tuned on ECMWF's operational analyses (2018 to 2024 for v2), several steps chained together so that it learns to keep its own drift in check and to match today's sharper analyses. One training took about a week on 64 graphics processors. The network is scored on a weighted 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.

Four runs a day, in minutes

The AIFS does not read observations: none were used to train it, and none enter a run. Each run starts from the IFS control analysis, the best estimate of the present state of the atmosphere that ECMWF builds with 4D-Var, a method that adjusts a physical model's trajectory to fit millions of observations from satellites, aircraft, ships, buoys and weather stations over a time window. That analysis, made on a grid of about 9 km, is regridded to the AIFS's 31 km grid. The AIFS therefore depends entirely on the physics-based system and its observation network for its starting point.

The forecast itself runs on graphics processors in minutes: the 2024 description of the system reports about two and a half minutes for 10 days on a single processor, where the IFS needs a large share of a supercomputer. That speed is the root of the 1,000-fold energy saving ECMWF quotes. Runs start at 00, 06, 12 and 18 UTC and produce a state every 6 hours out to 15 days.

The output, computed on the N320 grid, is published in GRIB2 on a regular 0.25° grid on ECMWF's open data server as soon as each forecast is produced, without the delay that once applied. It is released under the Creative Commons CC BY 4.0 licence, which allows redistribution and commercial use as long as ECMWF is credited. The trained weights of the network are published too, under the same licence, so anyone with a graphics processor can run the AIFS.

Its sibling AIFS ENS runs 51 forecasts at once, 50 perturbed members and a control, on the same 31 km grid and to the same 15 days. It is trained with the continuous ranked probability score, a score that rewards a set of forecasts whose spread matches the real uncertainty, so its members stay sharp instead of collapsing into one smooth average. Saily WR uses AIFS Single, not the ensemble.

Reading it at sea

Where the AIFS 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. ECMWF's evaluation found upper-air scores about 10% better than the IFS through most of the troposphere and tropical cyclone track errors substantially lower. 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. Because it learned to minimise an average error, it smooths: wind peaks are blunted, tight gradients are softened, and the effect grows with lead time. ECMWF notes that the AIFS produces less intense tropical cyclones than the IFS on average. In a gale, near a cyclone or along a sharp front, read its wind speed as probably too low and its timing and position as the useful part.

At about 31 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. 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 IFS. Both start from exactly the same analysis, so when they part ways the difference comes from the models alone, which is a clean signal that the forecast is uncertain. When the AIFS, the IFS and a model from another centre such as the GFS agree on a route, confidence is high; when the AIFS alone disagrees, look at whether it simply smooths a feature the others keep sharp.

AIFS in Saily WR

Saily WR downloads each new run of AIFS Single from ECMWF's open data as soon as it is published, 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 15 days. Because the model is global, every route gets an AIFS 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 IFS among them, so a difference between the AIFS and IFS routes shows at a glance where the two forecasts part ways. The sea current and land avoidance you choose apply to it exactly as to the others.

Expect a smooth route from a smooth wind: the AIFS route rarely zigzags after a local gust, and in strong-wind situations it may be optimistic about how comfortable the passage will be. The Data page shows live which AIFS run the app is on and how far ahead it reaches.