ECMWF IFS

From Reading in 1975 to cycle 50r1

ECMWF was born of a European convention signed in 1973, which came into force on 1 November 1975: rather than each country building a supercomputer and a global model on its own, the member states pooled them in one centre devoted to forecasts several days ahead, the "medium range". The centre settled in Reading, in England, made its first real-time forecasts in June 1979 and its first operational medium-range forecast on 1 August 1979, on its first Cray supercomputer. It is today an intergovernmental organisation of 23 member states, France among them, and 12 co-operating states.

The 1979 model was a grid-point model of the global atmosphere with 15 vertical levels. In April 1983 it gave way to a spectral model, which describes the fields as sums of waves over the sphere, and on 24 November 1992 ECMWF issued its first ensemble forecasts, the same year as the American NCEP. In March 1994 the Integrated Forecasting System, the IFS, took over: a code begun in 1987 jointly with Météo-France, which runs its own version of it under the name ARPEGE. "Integrated" means that observation processing, data assimilation, the model and post-processing share a single code.

On 25 November 1997 ECMWF put 4D-Var into operation, a way of building the starting state that the sections below explain; Météo-France followed in 2000, and they were the only two operational 4D-Var systems of the 20^th century. Resolution kept climbing: in March 2016 cycle 41r2 took the grid from about 16 to 9 km with a new "cubic octahedral" grid, and in June 2023 cycle 48r1 brought the ensemble to the same 9 km. On 18 October 2022 production moved from Reading to a new data centre in Bologna, Italy.

The current cycle, 50r1, went live on 12 May 2026 from the 06 UTC run. It closes a long transition: the former high-resolution forecast, HRES, and the unperturbed "control" member of the ensemble had been made identical in cycle 49r1, and ECMWF now runs only one of them, called the ENS control. It is this run that is still distributed as the single high-resolution forecast, and that many apps, this one included, still call HRES. Since 2025 ECMWF also runs a machine-learning model, the AIFS, which has its own page.

A spectral model on a 9 km octahedral grid

The IFS solves the primitive equations of a hydrostatic atmosphere: it assumes that vertical accelerations are small next to gravity, an approximation ECMWF still keeps in operations because it is stable and cheap, whereas limited-area models such as AROME drop it to resolve convection. Wind, temperature, pressure and moisture are carried both as spectral coefficients, waves over the sphere in which horizontal derivatives are exact, and as values on a grid, where the physics is computed; spectral transforms switch between the two at every time step.

Since cycle 41r2 (2016) the grid is "cubic octahedral", written TCo1279: the spectral truncation keeps waves up to number 1279, and the grid, called O1280, has 2,560 latitude circles from pole to pole, with fewer points along each circle towards the poles so that spacing stays close to 9 km everywhere. "Cubic" means the shortest wave the model carries spans four grid points instead of two, which is more accurate and needs less artificial smoothing. In the vertical, 137 levels follow the ground near the surface and become pressure surfaces higher up; the lowest four lie at about 10, 31, 54 and 79 m, and the top at 0.01 hPa, in the mesosphere.

Time stepping is semi-Lagrangian and semi-implicit: instead of computing what flows through fixed grid boxes, the model follows each air parcel back along its path over one step, which allows long steps without the computation becoming unstable. What the grid cannot resolve is parameterised, that is represented by its average effect: convection by a mass-flux scheme descended from Tiedtke's (1989), boundary-layer turbulence, clouds and precipitation with their own prognostic variables, radiation by the ecRad scheme, and the drag of sub-grid mountains. The land surface is the HTESSEL model, with a lake model.

The atmosphere is not computed alone: since 2018 the same coupled system runs at all ranges, with the NEMO ocean model and its sea-ice model SI3 (NEMO4-SI3, fully coupled, since 50r1), and the ecWAM wave model, which hands the roughness of the sea back to the wind. For a sailor the main limit is scale: the shortest wave on a 9 km grid is about 36 km long, and a feature needs to be larger still to be well represented, so a sea breeze, the funnelling between two islands or the acceleration off a cape are only sketched, and a squall is not resolved at all, only its average effect.

From 60 million observations to a 15-day forecast

Every day ECMWF receives around 800 million observations and, after quality control, keeps about 60 million for the IFS, most of them from around 90 satellite instruments. Satellites measure radiances that depend on the temperature and moisture of the air, winds tracked from cloud motion and winds over the sea read from the roughness of its surface; to these are added radiosondes, aircraft, ships, buoys and land stations.

These observations are blended with the previous forecast by 4D-Var, four-dimensional variational assimilation: over a window of 12 hours, the model is run forwards, and its adjoint backwards, to find the starting state whose forecast best matches every observation at the time it was made, while obeying the model's physics. A smaller ensemble of data assimilations, the EDA, runs alongside to estimate from day to day how uncertain the first guess is. Cycle 50r1 cut the observation time slots from 30 to 15 minutes and lets more 2 m temperature observations into the boundary layer.

The forecast itself runs on four Atos BullSequana XH2000 clusters, more than a million processor cores in all, in ECMWF's Bologna data centre, which took over operational production on 18 October 2022. Runs start from 00, 06, 12 and 18 UTC: those of 00 and 12 UTC go out to 15 days (360 hours), those of 06 and 18 UTC to 6 days (144 hours). The same run is the control of the ensemble, ENS, whose 50 other members start from slightly perturbed states and use perturbed physics to sample the uncertainty.

ECMWF's Council decided in December 2020 to move to open data. The free open subset reached 0.25°, about 28 km, in March 2024; on 1 October 2025 the whole real-time catalogue came under the Creative Commons CC BY 4.0 licence with no information cost, although high-volume delivery is still charged. Since 20 October 2026 the open-data portal also publishes the same 9 km forecast on a 0.1° grid, about two hours after the 0.25° files. Both are interpolations of the native octahedral grid onto regular latitude-longitude grids, not separate runs.

Reading it at sea

The IFS is the usual reference for the medium range: the position and depth of the depressions, the extent of an anticyclone, the timing of a front several days ahead. For an ocean passage or a long offshore leg, it is the model whose broad pattern is worth comparing with first beyond three or four days. Its ensemble companion tells a confident forecast from a fragile one: when the 50 members spread apart, the deterministic run deserves less trust.

Its weaknesses are those of a 9 km global model, which the 0.1° grid of the open data, about 11 km, hardly coarsens. Coastal effects are smoothed: sea breezes, katabatic winds down valleys, the acceleration of the wind round headlands and between islands. ECMWF itself lists known issues that matter to sailors: extreme gusts are under-represented when vigorous convection is involved, strong gusts can be underestimated just offshore, downwind of the land, and winds over mountains are often too light.

In the tropics, the IFS is widely used to forecast the tracks of tropical cyclones, but at 9 km it underestimates their maximum winds compared with the warning centres, and errors on the central pressure of small systems can exceed 50 hPa in extreme cases: near a cyclone, follow the official advisories rather than the raw field. Spurious squall lines and convective rings can appear, a problem ECMWF says persists in 50r1.

Next to the others: along the coasts of France and Europe, the finer AROME, ARPEGE Europe, ICON-EU and ICON-D2 see what the IFS smooths, and deserve more weight over the first two days. Beyond, comparing the ECMWF route with those of GFS, ICON and the AIFS gives a sense of the spread: where the routes agree, the forecast is probably sound; where ECMWF parts from the others, the decision is open.

In Saily WR

Saily WR takes ECMWF's open data on its finest grid, 0.1°: the 10 m wind, as its east-west and north-south components, from each of the four daily runs at 00, 06, 12 and 18 UTC, downloaded as soon as they are published. Steps are every 3 hours to 144 hours, then every 6 hours to 360 hours, 15 days; at the source, the 06 and 18 UTC runs stop at 144 hours, and past them a route goes on in the 00 or 12 UTC run before, which the server keeps.

The model is global, so an ECMWF route is computed whatever the start and finish. The boat sails, with its polar, in this wind interpolated in space between grid points and in time between steps; with 3- then 6-hour steps, a quick wind shift between two steps comes out smoothed. The ECMWF route is drawn beside those of the other models, with a SMOC or RTOFS current if you choose one, and kept off land by the OpenStreetMap coastline and, if you ask, by a minimum depth. The Data page shows live which ECMWF run the app is on, its horizon and its coverage.