ARPEGE

A global model with a zoom on France

In 1987 Philippe Courtier and Mats Hamrud started writing, at ECMWF, the code of a new global model shared by ECMWF and Météo-France; the joint project's kick-off meeting was held in October 1988. ECMWF calls the result the IFS, Météo-France calls it ARPEGE, for Action de Recherche Petite Echelle Grande Echelle, a "small-scale large-scale research action". The main motive was variational data assimilation, which needs a model written together with its tangent-linear and adjoint versions, and Météo-France wanted two more things: a semi-Lagrangian spectral code, and a global grid that could be stretched over France.

ARPEGE replaced Météo-France's previous global spectral model, EMERAUDE, in September 1992, at the same T79 resolution and without stretching. In October 1993 the stretched version went operational, with a stretching factor of 3.5, and also replaced PERIDOT, the limited-area model that covered France: from then on one global model did both jobs. It became semi-Lagrangian in October 1995 and moved to 4D-Var in June 2000, after ECMWF in 1997; these were the only two operational 4D-Var systems before the end of the 20^th century.

Resolution climbed with each supercomputer: TL358 with the stretching reduced to 2.4 in 2003, then by 2015 TL1198 stretched by 2.2 on 105 levels, 7.5 km over France and 36 km in the South-West Pacific, and since July 2019 TL1798, about 5 km over France. Cycle 46t1 entered service in June 2022, and in November 2024 Météo-France deployed a new version of its whole forecasting suite. The code keeps evolving with ECMWF and the ACCORD consortium of European weather services, and the common IFS/ARPEGE code reached a "cycle 50" in 2025.

ARPEGE does several jobs at Météo-France: short-range forecasting over France and Europe to about four days, the global picture for the overseas territories and the open ocean, the boundary conditions of the fine-mesh AROME model over France and, run as an ensemble, PEARP. In 2025 and 2026 Météo-France is also testing a hybrid with artificial intelligence, in which ARPEGE's large scales are nudged towards those of ECMWF's AIFS while ARPEGE keeps its own fine scales.

A stretched spectral grid

ARPEGE solves the hydrostatic primitive equations with the same spectral method as the IFS: fields are carried both as waves over the sphere and as values on a reduced Gaussian grid, where the physics is computed. Time stepping is semi-Lagrangian and semi-implicit with two time levels: the model follows air parcels back along their path, which allows long steps (360 s in 2017). In the vertical it uses a hybrid coordinate that follows the ground near the surface and becomes pressure higher up, discretised by finite elements on 105 levels, from 10 m above the ground to 0.1 hPa.

Its signature is the stretched grid. A transformation due to Schmidt moves the pole of the computing grid to central France, near 46.5°N 2.6°E, and draws the grid points towards it: with a nominal truncation T and a stretching factor C, the model has the local resolution of a truncation T×C over France and of T/C at the antipode. With TL1798 and C = 2.2, that gives about 5 km over France and about 25 km near New Zealand, varying smoothly in between. The Bay of Biscay, the Channel and the Western Mediterranean are seen close to the finest mesh, the Caribbean or the Indian Ocean far more coarsely.

ARPEGE has its own physics, built for short-range forecasting, distinct from ECMWF's. As described in 2017, turbulence uses a scheme that carries the turbulent kinetic energy as a variable, shallow convection a mass-flux scheme, deep convection a mass-flux scheme derived from Bougeault's (1985), clouds and precipitation a prognostic microphysics, and infrared radiation the RRTM scheme. The surface is handled by SURFEX, which splits each cell into soil and vegetation, sea and sea ice, lakes and towns; these schemes have evolved since.

At 5 km, ARPEGE sees over France and its coasts the main effects of relief: the Mistral and the Tramontane channelled by the Rhone valley and the gap between the Pyrenees and the Massif Central, the deflection of the wind by the large headlands, a well-formed sea breeze in outline. It does not resolve convection, which a scheme represents on average, so a squall line or the gust front of a thunderstorm is placed and timed only approximately. Far from France, the coarser mesh treats coastal effects like any global model.

From observations to a run

ARPEGE assimilates a wide range of observations: land stations, ships and buoys, radiosondes and dropsondes, aircraft reports, winds tracked from cloud motion by geostationary satellites, winds over the sea from scatterometers, microwave and infrared sounders such as IASI and CrIS, and GPS signals. New instruments enter as they arrive: Metop-C and NOAA-20 sounders in October 2019, and from June 2020 the wind profiles of the Aeolus lidar satellite during its mission.

The analysis is a 4D-Var on a 6-hour cycle: the gap between the observations and the previous forecast is computed at full resolution, but the corrections are computed on lower, unstretched grids in two successive loops (TL224 then TL499 as of 2021), on the assumption that the small scales follow the corrected large scales. An ensemble of 50 lower-resolution 4D-Var assimilations supplies the day-to-day error statistics. In 2017 the sea surface temperature was analysed from ship and buoy reports and from the NCEP and OSTIA analyses.

Four runs a day start from 00, 06, 12 and 18 UTC and reach between 102 and 114 hours (as of 2026). They run in Toulouse on Météo-France's two supercomputers, Belenos and Taranis, Atos BullSequana XH2000 machines in service since 2021; the contract with Atos, now Eviden, runs until 2028 and their successor is being procured. In its 2025 benchmark for that next machine, Météo-France asks for a 102-hour ARPEGE forecast in 30 minutes.

After the run, the fields of the stretched grid are interpolated onto regular latitude-longitude grids: a global 0.25° grid and a 0.1° grid over Europe and the North Atlantic. Météo-France publishes them as open data under the Licence Ouverte, the French government's open licence, which allows reuse with attribution.

Reading ARPEGE at sea

ARPEGE is first a model for France and its neighbours. On the Atlantic coast, in the Channel and in the Western Mediterranean, its 5 km mesh sees the strengthening of the Mistral and the Tramontane, the passage of fronts and the turning of the wind around the main headlands in more detail than a global model with a uniform grid. For a passage of two to four days in these waters it is a sound reference, and AROME, which takes its boundaries from it, adds the detail of the coast for the first two days.

Its range is short, about four days, against 15 for ECMWF or 16 for GFS, so it says nothing about the end of an ocean crossing. Far from France, in the Caribbean, the Indian Ocean or the Pacific, its mesh is no finer than that of the other global models, and tropical cyclones are handled at that resolution with a convection scheme: there, prefer the regional models and the official advisories.

Like every hydrostatic model with parameterised convection, ARPEGE does not resolve squalls or gusts under cumulonimbus: an unstable sky can bring much more than the 10 m wind it shows. Compare its route with AROME close to the coast and with ECMWF, ICON-EU and GFS offshore; when ARPEGE and ECMWF agree on the timing of a front, confidence is higher.

In Saily WR

Saily WR uses ARPEGE twice, through its two open products: the global product on its 0.25° grid, which covers the whole planet, and the Europe product on its 0.1° grid, which covers 20 to 72°N and 32°W to 42°E. From each, the app takes the 10 m wind, as its east-west and north-south components, from the four daily runs, downloaded as soon as they are published.

The global product comes hourly from H+1 to H+48, then every 3 hours to H+102; the Europe product hourly from H0 to H+102. The 12 UTC run goes on to H+114 on both. A route is computed for every model whose area covers both the start and the finish, so inside the Europe domain you get two ARPEGE routes side by side, from the same model, differing by their grid and their time steps.

The boat sails, with its polar, in this wind interpolated in space and time, with a SMOC or RTOFS current if you choose one, kept off land by the OpenStreetMap coastline and, if you ask, by a minimum depth. The Data page shows live which run each ARPEGE product is on, its horizon and its coverage.