ICON

From GME to ICON

Deutscher Wetterdienst (DWD), Germany's national weather service in Offenbach, has been running a global model on an icosahedral grid since December 1999. That model, GME, was the first operational weather model to cut the globe into cells built from an icosahedron, a solid of 20 triangular faces, rather than into latitude and longitude boxes that crowd together at the poles. Its cells were hexagons, its grid about 20 km and its equations hydrostatic, which means vertical accelerations were neglected: an acceptable shortcut at that scale, but one that closes the door to fine meshes.

ICON, for ICOsahedral Nonhydrostatic, was developed jointly by DWD and the Max Planck Institute for Meteorology in Hamburg as a single framework for weather and climate. Its designers wanted exact conservation of mass, the ability to refine the grid over a region, and equations valid down to the kilometre and below. ICON became DWD's operational global model on 20 January 2015, replacing GME, with a 13 km grid and 90 levels. A refined region over Europe, ICON-EU, was added in the summer of 2015 and has run routinely since 21 July 2015, taking over from the regional model COSMO-EU.

The system has grown step by step since. In January 2016 the analysis moved from 3D-Var to a hybrid ensemble-variational scheme; in January 2018 the 40-member global ensemble ICON-EPS entered service; on 10 February 2021 ICON-D2, a limited-area version at about 2 km over Germany and its neighbours, replaced COSMO-D2, so that DWD's whole forecasting chain now rests on ICON. In November 2022 the global model went from 90 to 120 levels and the European nest from 60 to 74, most of the new levels being placed in the stratosphere, while the ensemble was sharpened from about 40 to about 26 km.

Today ICON is developed by DWD, the Max Planck Institute for Meteorology, the German Climate Computing Center (DKRZ), MeteoSwiss and the Karlsruhe Institute of Technology. MeteoSwiss forecasts with it as well, and the Max Planck Institute uses it as the core of its climate model. Its source code has been published under an open source licence since 31 January 2024.

Triangles, a non-hydrostatic core and 120 levels

ICON's grid starts from an icosahedron projected onto the sphere. Each edge of its 20 triangles is first divided into n equal parts (the "root" division, Rn), then the resulting triangles are cut in four k times by joining the middles of their sides (bisections, Bk). The global grid, R3B07, holds 2,949,120 triangles with an average size of 13.2 km, measured as the side of a square of the same area. There is no pole where meshes pile up, and the resolution is almost the same everywhere on Earth.

The dynamical core, the part that solves the equations of motion, uses the fully compressible, non-hydrostatic equations: vertical accelerations are computed, not assumed to be zero, which is what lets the same code run at 13 km or at a few hundred metres. Its prognostic variables are the wind normal to each triangle edge, the vertical wind, the air density and the virtual potential temperature. Scalars sit at the centre of each triangle and normal winds at the middle of each edge, an arrangement called a C-grid. Time integration uses a two-time-level predictor-corrector scheme, explicit in the horizontal and implicit in the vertical for sound waves, and the scheme conserves mass exactly.

In the vertical, the global model has 120 levels from the surface to 75 km, deep into the mesosphere, which helps assimilate satellite data. The levels follow the terrain near the ground and flatten gradually with height, using a smooth coordinate called SLEVE, until they are perfectly flat above 16 km. Over the sea the lowest layer is about 20 m thick, so the model's first level sits close to the height at which the 10 m wind is reported. The European nest has 74 levels up to 22.8 km, the same as the lowest 74 of the global grid.

Processes smaller than a mesh are parameterised, that is represented by their average effect. Showers and thunderstorms are handled by a mass-flux convection scheme, a branch of the Tiedtke-Bechtold scheme of ECMWF's IFS, because at 13 km and even at 6.5 km a convective cell is far smaller than a triangle. Turbulence in the boundary layer is driven by a prognostic turbulent kinetic energy scheme inherited from COSMO; clouds and rain by a single-moment microphysics scheme; the drag of unresolved mountains by the Lott and Miller scheme; land surfaces by the TERRA soil model, lakes by FLake and sea ice by a simple ice model.

The ocean is not modelled. The sea surface temperature is analysed once a day and then held constant for the whole forecast, so ICON does not see a storm cooling the sea behind it or an upwelling developing during the run. Waves are not computed either: DWD runs its wave models separately, driven by ICON's winds.

How a run is computed

ICON's analysis, the best estimate of the current state of the atmosphere from which a forecast starts, is renewed every 3 hours, at 00, 03, 06 and so on to 21 UTC. Each analysis combines a 3-hour forecast from the previous one, called the first guess, with the observations of a 3-hour window centred on the analysis time: surface stations, ships, buoys, radiosondes, aircraft and satellites. The weight given to each depends on how wrong each is expected to be.

Since 20 January 2016 this is done by a hybrid ensemble-variational system, EnVar. A variational analysis finds the atmospheric state that best fits both the first guess and the observations at the full 13 km resolution; "hybrid" means that the expected errors of the first guess come partly from fixed statistics and partly from the spread of a 40-member ensemble of 3-hour forecasts, so that they follow the weather of the day. That ensemble is itself updated by a Localized Ensemble Transform Kalman Filter (LETKF), which corrects each member using nearby observations only, and it provides the starting states of the ICON-EPS ensemble.

To avoid jolting the model with spurious waves, the corrections are not added at once but dripped in over 3 hours centred on the analysis time, a method called incremental analysis update: the model actually starts 90 minutes before the nominal time. Separate analyses update the sea surface temperature and the soil moisture once a day and the snow cover every 3 hours.

Four forecasts a day start from these analyses, at 00, 06, 12 and 18 UTC. Those of 00 and 12 UTC run to 180 hours (7.5 days), those of 06 and 18 UTC to 120 hours, with hourly output up to 78 hours and 3-hourly beyond. The European nest runs inside the global model in a two-way coupling: the global model feeds the edges of the nest at every time step, and the nest feeds its finer solution back to the global grid. The nest also runs four shorter forecasts, at 03, 09, 15 and 21 UTC, to 51 hours.

Since an amendment to the DWD Act came into force on 25 July 2017, DWD publishes its model output free of charge on its open data server, opendata.dwd.de, in GRIB2 format, reusable under the CC BY 4.0 licence as long as DWD is credited. The global model is offered on its native triangles, ICON-EU and ICON-D2 on regular latitude-longitude grids of 0.0625° and 0.02°. According to DWD, more than 30 national weather services drive their own regional models with ICON's output.

At sea

ICON's main value for a sailor is that it is an independent opinion. It comes from a different centre than ECMWF's IFS or NOAA's GFS, with its own grid, its own physics and its own assimilation, so its errors are not those of the others. When ICON, IFS and GFS place a low or a front in the same spot, the forecast deserves more confidence; when they disagree, the disagreement itself says that the situation is uncertain and that a route depending on a narrow window is risky.

At 13 km the global model draws depressions, fronts, trade winds and the large wind shadows of islands well, but not the wind along a given coast: a cape, a strait or a bay a few kilometres wide is only a handful of triangles. Over Europe the 6.5 km nest does better with winds shaped by relief, such as the Mistral and Tramontane in the Gulf of Lion, the Bora in the Adriatic, the Meltemi between the Cyclades or the Levante in the Strait of Gibraltar, but a sea breeze that only reaches a few kilometres offshore remains beyond its reach. Close to the coast, a kilometre-scale model such as AROME is a better guide.

Convection is parameterised in the global model and in ICON-EU, so a squall line or a thunderstorm gust front appears there as a smoothed area of showers rather than as the wind shift a crew actually meets; ICON-D2 computes showers and thunderstorms itself, though their timing and place stay uncertain. The 10 m wind is a mean wind over a mesh and over time: gusts, which can be much stronger under a cloud or in a sudden acceleration, are not in it. A tropical cyclone's eyewall is narrower than a few 13 km triangles, so its maximum winds are usually underestimated; for a cyclone the official advisories of the warning centres come first.

Because the sea temperature is frozen at its value of the analysis day, the model can lag behind when sea temperature drives the wind, for instance in a cold-air outbreak over a warm current or over a sea cooled by a previous storm. On long passages, the 7.5-day range of the 00 and 12 UTC runs is shorter than that of IFS or GFS, and as with any deterministic forecast its last days should be read as a trend, not as a precise wind.

In Saily WR

Saily WR uses ICON three times, as three separate models: the global ICON, the European ICON-EU and ICON-D2 over Germany and its seas. Each computes its own route whenever its area covers both the start and the end, and that route is sailed in the model's 10 m wind, interpolated in space and in time, then shown next to the routes of the other models so that they can be compared.

For the global model, the app downloads DWD's open data on the native triangles and 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 app takes the four runs of the day, hourly up to 78 hours, then 3-hourly up to 180 hours for the 00 and 12 UTC runs or 120 hours for the 06 and 18 UTC runs; past a 06 or 18 UTC run, a route goes on in the 00 or 12 UTC run before.

For ICON-EU, the app uses DWD's 0.0625° grid (about 6.5 km) as published, over 29.5–70.5°N and 23.5°W–45°E. It takes the eight runs of the day: those of 00, 06, 12 and 18 UTC up to 120 hours, hourly to 78 hours then 3-hourly, and the intermediate ones up to 51 hours, past which a route goes on in the main run before.

For ICON-D2, the app uses DWD's 0.02° grid as published, over 43.18–58.08°N and 3.94°W–20.34°E, and takes its eight runs of the day, hourly up to 48 hours. That rectangle is wider than the model's domain: the points outside it carry no value in DWD's files, so the app draws the domain's real edge from them, and computes an ICON-D2 route only when the start and the end both lie within it. A server fetches each new run as soon as DWD publishes it, and the Data page shows live which run each of the three models is on, how far it reaches and the area it covers.