The Météo-France models, what they do well and what they do not
AROME sees the storm over your parish but stops at two days. ARPEGE reaches four days but averages your valley away. What each one is good at, where it goes wrong, and which to read for the question you actually have.
Météo-France does not run a model, it runs several, and they are not for the same thing. Knowing which one is answering your question avoids two common mistakes: believing a seven-day forecast made by a model built for two, and ignoring a six-hour forecast made by a model that sees things no other model can.
Two models, two jobs
ARPEGE is a global model. It computes the atmosphere of the whole planet, several times a day, out to four days. It is the model that tells you a low is coming in off the Atlantic.
AROME is a fine-mesh model over France and its surroundings. It computes far less ground, far more finely, and far less far ahead. It is the model that tells you whether the storm crosses your parish or the next one.
ARPEGE has a peculiarity that explains much of both its strength and its weakness: its grid is stretched. Instead of covering the globe at a constant mesh, it is tightened over France — around 5 km — and dilates gradually to about 24 km at the antipode, over the South Pacific.
This is a deliberate engineering choice: for a given computing budget, Météo-France concentrates the resolution where its users are. The price is that ARPEGE is a global model whose quality depends on where on the globe you look. Over France and western Europe it is excellent. Over Australia it is markedly worse than its billing as a "global model" would suggest.
What a mesh sees, and what it does not
This is the point that matters most, and the easiest to overlook.
A model does not know the terrain. It knows an average of the terrain over each cell. A model on a 25 km mesh does not know there is a valley where you are: it holds a single elevation for a 625 km² square in which the valley, the ridge and the lake have been melted together.
This is not a question of display precision, it is a question of physics. A model that cannot see the ridge cannot produce the wind that accelerates over it, nor the rain that triggers on the way up it, nor the pool of cold air that collects in the valley beneath it. Those phenomena do not exist in its world.
That is why AROME exists. At 1.3 km it sees the coastline, it sees the valley, it sees the town. And crucially, at that mesh it no longer has to parametrise convection: it can represent a storm's updraughts directly instead of estimating them with a formula. The term is "convection-permitting", and it is a step change rather than a gradual improvement.
In practice AROME is markedly better than any global model for:
- summer storms, both their triggering and their position;
- sea and land breezes;
- fog, which is a very shallow boundary-layer phenomenon;
- terrain effects: föhn, wind channelled along valleys, strengthening over ridges;
- gusts, which depend heavily on the ground underneath.
What AROME does not do
It stops. AROME runs out to 51 hours — a little over two days. Beyond that there is nothing to read, and that is deliberate: on a limited domain, information arriving from outside eventually dominates, so that past a few days you would no longer be reading AROME but the model feeding it its boundary conditions.
It also has a flaw its own advocates cheerfully admit: the double penalty. A model that forecasts a well-formed storm, at the right time, 20 km from the right place, scores worse than a model that smeared light rain across the whole region — because it is wrong twice, where the storm did not happen and where it did. Put another way: AROME can be more useful than ARPEGE while scoring worse. Do not judge a fine-mesh model on a mean error.
Finally, like any model, AROME needs a few hours for its starting state and its own physics to come to terms with each other. The very first hours of a run are the least reliable, which is somewhat ironic. That is why AROME-PI exists — a nowcasting version, restarted every fifteen minutes and reaching only six hours.
When each one is the one to read
The update cadences matter as much as the reaches. AROME is restarted every three hours, eight times a day: at midday, the forecast you are reading was computed from observations less than three hours old. ARPEGE runs four times a day, at 00, 06, 12 and 18 UTC, reaching 102 hours from the 00 and 12 runs and 72 hours from the 06 and 18 ones.
A simple rule for reading:
| Your question | The model |
|---|---|
| Is it going to rain in two hours? | AROME-PI |
| Can I spray tomorrow morning? | AROME |
| Does the site hold up until Thursday? | ARPEGE |
| And next week? | Neither — a long-range global model |
That last line deserves to be said plainly. No Météo-France model reaches beyond four and a half days. For next week you have to look elsewhere — ECMWF's IFS, the American GFS — and Météo-France makes no secret of it: at global medium range, the IFS is the world reference, including for Météo-France.
The ensembles
Both models have an ensemble version, which runs several forecasts from slightly different starting states to measure the uncertainty — the subject of another article.
PEARP is the ARPEGE ensemble: 35 members, twice a day, out to 108 hours. This is the one to read to find out whether Thursday's low is settled or still undecided.
AROME-EPS is the AROME ensemble: 16 members, fine mesh, out to 51 hours. Its question is a different one: not whether the storm happens but where. A fine-mesh ensemble is the only honest tool for a storm forecast, because the position of a convective cell is inherently uncertain by tens of kilometres, and a single line on a map hides that.
What is freely available
Since 2024 most of the Météo-France catalogue has been open data, with no commercial licence to negotiate. That is not a detail: it is what made a tool like this one possible.
The published grids are not exactly the computing grids. AROME runs at 1.3 km and is distributed on regular grids at 0.025° (about 2.5 km) and 0.01° (about 1.1 km). ARPEGE is distributed on a 0.1° European grid and a 0.25° global one. So you do not quite receive what the model computed, but an interpolation of its field — which is of no consequence for most uses, and worth knowing when you compare two sources and they do not give quite the same number.
In short
ARPEGE and AROME are not two grades of the same product, they are two tools for two questions. The first tells you what is arriving over France; the second tells you what that amounts to where you are.
The mistake to avoid is reading one while believing you are reading the other: asking a global model whether the storm will cross your field, or asking AROME what next Saturday holds. In both cases the answer looks like a forecast and is not one.