DWD AICON

ICON's apprentice

When neural networks began to rival physics-based forecasts in 2022, most of them learned from the same textbook: ERA5, ECMWF's reanalysis of past weather, used with a point about every 31 km and a picture every 6 hours. Deutscher Wetterdienst, Germany's national weather service, chose to train its own network on its own data. Its name, AI in front of ICON, says as much: ICON is the physics-based model whose output it learned from.

That data is ICON-DREAM, a reanalysis in which DWD re-ran the weather of the whole globe from 2010 onwards with its ICON forecasting system as it stood in March 2024: the past atmosphere rebuilt every 3 hours on ICON's 13 km triangles, from the observations of the time. AICON was trained on fifteen years of it. What decides the quality of an AI method is the data it is trained on, as DWD's president, Sarah Jones, put it at the launch. The network itself is built on Anemoi, the open-source framework that ECMWF and several European weather services launched in 2024, to which DWD added the pieces needed to work on ICON's grids.

AICON-Global started running round the clock on 3 September 2025, a trial period of six months in which DWD's forecasters compared it almost daily with ICON and with other physical and machine-learned models. On 2 March 2026 the Federal Transport Minister, Patrick Schnieder, put it into service at DWD's headquarters in Offenbach, making DWD one of the first national weather services with a fully operational machine-learned model. AICON complements ICON, which remains the core of DWD's forecasting, and its forecasts are now part of the basis of DWD's weather warnings.

Two changes followed. From 6 May 2026 AICON starts from the initialised version of ICON's analysis, the one it was trained on, instead of the raw one, a correction whose only significant effect was on surface pressure in the first hours. On 30 June 2026 DWD added AICON-EU, a regional version on ICON-EU's grid of about 6.5 km over Europe. In August 2026 twelve DWD scientists published the full description of the global model, its training and a year of verification.

A network on ICON's triangles

AICON solves no equation. Given the state of the atmosphere now and 3 hours earlier, it predicts the state 3 hours ahead; it then feeds its own forecast back in and repeats, sixty times for a 180-hour run. Every error made early is carried into the following steps, as in every model that works this way.

What sets it apart is its graph, drawn directly from ICON's grid. Its 2,949,120 triangles, about 13 km across, are the points the network reads and writes. Each is linked to the three corners above it on a coarser hidden mesh of 368,642 points, where the forecast is actually made; that mesh joins the edges of every level of refinement of ICON's grid, from the coarsest to the finest, so that information can travel across an ocean in one step as well as to the next cell. The network is a graph transformer: one attention layer to gather the triangles onto the hidden mesh, 16 to carry the atmosphere forward, one to bring it back, each with 16 attention heads.

It works on ICON's own vertical levels rather than on pressure levels: 13 of the terrain-following levels that hug the ground and the sea and flatten out with height, from near the surface to about 50 hPa. On them it predicts wind, temperature, humidity and pressure; at the surface, the 10 m wind, 2 m temperature and humidity, surface pressure, the temperature of the surface (that of the sea, over water), snow and two soil layers. Precipitation is the only output never fed back in.

Training went from coarse to fine: first on ICON-DREAM thinned to 52 km, then 26 km, then the full 13 km, the network keeping what it had learned at each stage. The stages listed in the paper add up to about 2,900 hours of graphics-processor time on a single node of eight NVIDIA A100s. Errors are scored as a weighted mean squared error, levels near the surface weighing five times as much as the top, which shows where DWD's priorities lie.

Its central choice is to train on a single 3-hour step only. Most machine-learned models are trained on forecasts of a day or more, chained step after step, which improves their headline scores but teaches them to blur whatever they cannot predict. DWD tried it: a 52 km test version trained on 24-hour sequences beat ICON on many headline scores, but lost small-scale detail, so DWD gave it up. AICON still smooths, since a mean squared error rewards hedging between possible outcomes, but DWD finds its fields less blurry than those of other deterministic models trained on ERA5.

Eight runs a day, four in the open

AICON reads no observations of its own. Each run starts from ICON's global analysis, the best estimate of the present atmosphere that DWD builds every 3 hours by blending a short ICON forecast with the latest observations from satellites, aircraft, ships, buoys and weather stations, using a hybrid ensemble-variational method. That analysis is cut down to AICON's 13 levels. AICON and ICON global therefore start each run from the same analysis.

It runs on the same schedule as ICON: every 3 hours, to 180 hours for the runs of 00 and 12 UTC, 120 hours for those of 06 and 18 UTC and 48 hours for the runs in between, with a forecast every 3 hours. At the launch the minister stressed this rhythm, a forecast every 3 hours being twice as frequent as in other global models. The model and its whole software chain run inside a container, a first in DWD's forecasting system.

The four main runs, at 00, 06, 12 and 18 UTC, are published in GRIB2 on DWD's open data server, opendata.dwd.de, about 3 hours after their start time, on the same native triangles as ICON global. They carry the 10 m wind, 2 m temperature and humidity, surface and sea-level pressure, precipitation, and wind, temperature, humidity and pressure on the 13 levels. Like all of DWD's open data, they are released under the Creative Commons CC BY 4.0 licence, which allows redistribution and commercial use as long as DWD is credited.

AICON-EU, the regional version, runs every 3 hours from AICON-Global, to 120 hours for its main runs, on ICON-EU's grid of about 6.5 km over Europe. It is not on the open data server, and Saily WR uses the global model only.

Reading it at sea

AICON was built for the weather near the surface in the first days, and that is where it does best. Over a year of verification against weather stations, it beat ICON on most surface variables during the first days of the forecast, more clearly in summer than in winter. DWD states that for near-surface variables over the first three days it is more accurate than every other machine-learned model it tested.

It is also steadier than ICON from one run to the next: for wind speed, humidity, pressure and rain, its successive forecasts for the same moment jump around significantly less. For a skipper watching the forecast over several days before a departure, a model that does not flip-flop is easier to plan with, and with a picture every 3 hours rather than every 6, a front that passes between two steps is pinned down more closely in time.

Its weakness is strong wind. DWD's forecasters found that it tends to underpredict strong-wind situations, and in a category 5 typhoon, Sinlaku in April 2026, AICON's storm was weaker than ICON's, which was itself not as deep as observed: its sustained winds were substantially too low. It followed the track well while the typhoon was weak, less well once it intensified rapidly. In a gale, near a tropical cyclone or along a sharp front, read its wind speed as probably too low and its timing and position as the useful part.

Further ahead it loses ground: beyond about two days it forecasts surface pressure worse than ICON, and the height of its upper-air pressure surfaces is its weakest score, so the position of the large weather systems late in a run may deserve less trust than its surface wind in the first days. Beyond three days it can also produce spurious rain stuck to a few coastal grid points. And at 13 km it resolves no sea breeze, no acceleration around a cape and no squall: for coastal sailing a fine-mesh regional model tells far more.

Its best use is next to ICON. Both start from exactly the same analysis, so when they part ways the difference comes from the models alone, a clean sign that the forecast is uncertain. When AICON, ICON and a model from another centre agree on a route, confidence is high; when AICON shows less wind than the others in a blow, it is most likely smoothing a peak the others keep sharp.

AICON in Saily WR

Saily WR downloads each new run of AICON from DWD's open data as soon as it is complete, about 3 hours after its start time, four times a day: the 10 m wind, as its east-west and north-south components, every 3 hours, to 180 hours for the runs of 00 and 12 UTC and 120 hours for those of 06 and 18 UTC. Because the model is global, every route gets an AICON version, wherever it starts and ends.

The files come on ICON's native triangles, the same 2,949,120 cells of about 13 km as ICON global. As it does for ICON global, the app places them itself on a regular 0.25° grid, taking for each point the value of the nearest triangle: nothing is smoothed, but with a point every 0.25°, about 28 km in latitude, some of the detail of the 13 km grid is lost.

The isochrone route is sailed from your boat's polar in that wind, interpolated in space between grid points and in time between the 3-hourly steps, and stops at the end of the run, or, past a 06 or 18 UTC run, at the end of the 00 or 12 UTC run before, which it goes on in. It is drawn next to the routes of the other models covering the same water, ICON among them, so a difference between the AICON and ICON 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.

In strong-wind situations expect the AICON route to be optimistic about how comfortable the passage will be. The Data page shows live which AICON run the app is on and how far ahead it reaches.