Four services, one forecast
HARMONIE-AROME grew out of two European research consortia: HIRLAM, founded in 1985 by mostly northern European weather services, and ALADIN, started in 1991 by Météo-France and the services of central Europe. After a cooperation agreement signed in 2005, the HIRLAM countries adopted AROME, the French fine-mesh model, and adapted its physics to northern conditions. By 2017 the result, HARMONIE-AROME, made the operational short-range forecasts of Denmark, Estonia, Finland, Iceland, Ireland, Lithuania, the Netherlands, Norway, Spain and Sweden. It is maintained today within ACCORD, a consortium of 26 national weather services.
For years each of these services ran the model on its own computer, over its own area. In February 2018 four of them, Denmark, Iceland, Ireland and the Netherlands, met to do it together; they signed a memorandum of understanding that July and a binding agreement in April 2020, creating United Weather Centres West, UWC-West. Their common area took the four countries' initials: DINI. An early common suite started in September 2021, for testing and for weekly feedback from forecasters.
In November 2021 the four services announced a shared supercomputer built by Hewlett Packard Enterprise and installed in Iceland, where hydropower and geothermal energy run it and the cool climate helps chill it. It was to perform 4,000 trillion calculations per second and to be in service in early 2023; the timetable slipped. Two HPE Cray systems were delivered, one for operational forecasting and one for weather and climate research.
According to Met Éireann, the common models have run operationally since March 2024; KNMI began its own operational runs on 20 June 2024 after a year of tests, and on 23 September 2024 the four services announced that UWC-West had entered its operational phase. The new model covers East Greenland to the Adriatic in 2 km squares with 90 levels, where the national models it replaced used 2.5 km and 65 levels; at DMI, DINI took over from the NEA model. It runs cycle 43h2.2 of HARMONIE-AROME, and KNMI has published test data from cycle 46 since August 2026.
The science: AROME's physics on a 2 km grid
HARMONIE-AROME uses the same dynamical core as AROME-France, developed within ALADIN. It solves the fully compressible equations of the atmosphere without assuming that vertical motion is in balance, which is what lets it compute the strong updrafts of a thunderstorm or the flow over a steep coast. Most fields are carried as waves (a spectral method), and air is followed along its path with a semi-implicit, semi-Lagrangian scheme inherited from ECMWF's IFS, which allows long time steps without the computation becoming unstable.
At this resolution deep convection is not parameterised: showers and thunderstorms are computed by the model's own dynamics, with their updrafts, downdrafts and gust fronts. Only shallow convection, the small cumulus of a fair-weather day, is described by a mass-flux scheme called EDMFm. Clouds and rain come from ICE3, a microphysics scheme that tracks water vapour, cloud droplets, rain, cloud ice, snow and graupel, and full radiation is computed every 15 minutes in the reference configuration.
Mixing in the lowest layers, which sets how much of the wind aloft reaches the sea surface, is handled by HARATU, a turbulence scheme brought in from the Dutch climate model RACMO. The consortium reported that it nearly removed the daily cycle of 10 m wind bias over the Netherlands, lowered that bias in every domain running the model, and improved the wind over sea when checked against scatterometers. The surface is modelled by SURFEX, which splits each grid box into sea, lakes, towns and natural land; over the sea, fluxes of heat, moisture and momentum come from a scheme called ECUME.
A 2 km grid draws the coast in detail: a fjord, a sound between two islands or a headland a few kilometres wide is several cells across rather than a fraction of one. The relief is smoothed before use, so the model really represents features somewhat larger than its 2 km spacing, but it sees the sea breeze, the wind funnelled between islands, the shelter of a high coast and the outflow of a thunderstorm, which a 10 to 30 km model cannot. The published reference configuration, at 2.5 km with 65 levels, used a 75-second time step.
How a run is computed
Every hour, a new run starts on UWC-West's supercomputer in Iceland. Its observations are prepared by SAPP, ECMWF's observation-processing system, which the four services share: 24 extractions a day, each with a cut-off of 60 minutes. They include conventional observations from weather stations, ships, buoys and balloons, Mode-S reports from aircraft gathered by the European centre EMADDC, satellite radiances relayed by EUMETSAT, and extra data supplied by each partner.
The system's 2022 design gives it 3D-Var data assimilation: the short forecast from the previous run is corrected towards the new observations, weighed against what is known of the errors of both, to make the starting state. DINI is a limited-area model, so it needs to know what happens beyond its edges: its lateral boundaries come from ECMWF's global IFS, blended in through a relaxation zone around the domain. The large weather systems that enter the domain are therefore those of the IFS.
On DINI the consortium runs the model as a continuous ensemble. Every hour, one main run and five forecasts started from slightly altered conditions are computed; every three and six hours these are gathered into an ensemble, 30 members over six hours, to estimate the chance of strong wind or heavy rain. The perturbations act on the model's physics (a method called SPP), and the perturbed members take their boundaries from the IFS ensemble. Each forecast reaches 60 hours.
DMI publishes the deterministic forecast of every third hour, at 00, 03, 06, 09, 12, 15, 18 and 21 UTC, hourly from the start to 60 hours, as GRIB files on its public cloud storage in the HARMONIE_DINI_SF collection, under the CC BY 4.0 licence. A run is complete about 2 hours 30 to 2 hours 40 after its nominal time. The published grid is a Lambert conformal projection of 1906 × 1606 points, 2 km apart, true at 55.5°N and centred on the meridian 8°W; seen on a map it is a fan from about 37.7°N to 69.9°N and from 43°W to 40°E. Its 10 m wind is given along the grid's axes, not towards true east and north.
Reading it at sea
HARMONIE DINI is at its best close to land, in the first two days. It brings 2 km wind to coasts that are otherwise covered only by coarser models: the fjords from Bergen to the Lofoten at Svolvær, Shetland around Lerwick, the west of Ireland at Galway, the Faroes at Tórshavn, Iceland from Reykjavik to Akureyri, the Gulf of Finland between Helsinki and Tallinn, the Gulf of Bothnia up to Oulu and the Adriatic islands from Split to Dubrovnik. There it shows the accelerations around headlands, the wind funnelled down a sound and the calm in the lee of high ground.
Over Biscay, the Channel, the North Sea, the Gulf of Lion and the Ligurian Sea it is a second fine-mesh opinion next to AROME, and to ICON-D2 in the Channel and the North Sea. When they agree on a sea breeze or a wind shift, it deserves confidence; when they disagree, the difference is a measure of what no model knows yet. But it is not independent at the large scale: its boundaries come from the IFS, so a low that the IFS places wrongly enters DINI misplaced too. For an independent view of the weather systems, look at ICON or the GFS.
Showers and thunderstorms are computed, not guessed, so a squall line can appear with its sudden wind shift. A single run, however, gives one possible storm: its place and timing are uncertain, and a thunderstorm drawn over your route is a warning, not a schedule. The 10 m wind is a mean over each 2 km cell: gusts, stronger under a shower or in a sudden acceleration, are not in it. Verification against scatterometers and buoys has also shown the model's near-surface wind bias growing in strong winds; in a gale, compare its speeds with those of the other models.
Its 60-hour range makes it a model for coastal sailing, day passages and crossings such as Holyhead to Dublin or Kiel to Copenhagen, or for timing a front on the last stretch of a longer route. Near the edges of its domain, off Greenland or along its southern border in the Mediterranean, it can add little to the IFS wind it receives there. Beyond two and a half days, the global models take over.
HARMONIE DINI in Saily WR
Saily WR downloads all eight runs of the day from DMI's public cloud storage as each is published, hourly from the start to 60 hours. Each forecast step is a file of about 600 MB, of which the app reads only the 10 m wind: two messages of about 6 MB each, fetched by byte range.
The app regrids that wind bilinearly onto a regular grid of 0.03° in longitude by 0.02° in latitude, about 2 km, covering the whole rectangle around the model's fan, and turns it from the grid's axes to east and north. Points of the rectangle outside the fan carry no forecast and are written calm.
A route gets a HARMONIE DINI version when its start and its end both lie inside the model's domain, whose edge is taken one cell in. It is sailed from your boat's polar in that wind, interpolated in space and between the hourly steps, and drawn beside the routes of the other models covering the same water, where they overlap: AROME, ARPEGE-EU, ICON-EU, ICON-D2 and the global models. The sea current and land avoidance you choose apply to it as to the others.
Only the start and the end are checked against the domain. A route between two points inside it may still pass through a calm corner of the rectangle, for instance off Portugal or south of Sardinia, where it finds no wind and the HARMONIE DINI route is not to be trusted. The Data page shows live which HARMONIE DINI run the app is on, how far ahead it reaches and the area it covers.
Sources
- The HARMONIE–AROME Model Configuration in the ALADIN–HIRLAM NWP System, Bengtsson et al., Monthly Weather Review (2017)
- UWC-West operational cooperation: organization aspects and status, Whelan, Tijm et al., EUMETNET SRNWP (2022)
- Joining forces in weather forecasting and climate research, Met Éireann (2021)
- New Icelandic Met Office weather and climate supercomputer becomes operational, Icelandic Met Office (2024)
- Using ECMWF's observation processing (SAPP) in four countries in northwest Europe, Darcy, ECMWF Newsletter 182 (2025)
- Improving cloud forecasting in HARMONIE-AROME, special project request, Gleeson, Met Éireann and ECMWF (2024)
- Weather Model (HARMONIE) for Netherlands, Europe and Caribbean, KNMI (2026)
- HARMONIE-AROME Cy46 forecast Netherlands meteorological parameters (test phase data), KNMI Data Platform (2026)
- Weather Model (HARMONIE) for DINI and IG, DMI
- Migration guide to DINI and IG, DMI
- Danish Meteorological Institute (DMI) Open Data Forecasts, Registry of Open Data on AWS (2026)