What a model cannot see: sub-grid physics
Squalls, turbulence, small clouds and sea-breeze fronts are smaller than the grid. Why they are missed or smeared, and how to allow for it.
The model only knows the box
A model keeps one value per grid box for each layer of the atmosphere. Inside the box nothing is resolved. Everything the weather does on a smaller scale has to be estimated from the box average, using a rule written by the model’s developers. That rule is a parametrisation. It describes what the small-scale processes do on average, not where each one happens. The box sizes and the five-times rule are explained in What km and NM mean on a model.
| Process | What the model does | What you may meet at sea |
|---|---|---|
| Showers and thunderstorms (convection) | A coarse model estimates how much convection a box would produce and spreads the effect over the whole box | A squall line that is weaker, wider or later than reality, or missing; rain in the model where there is none, and the reverse |
| Turbulence in the air above the sea | A rule gives the mixing of fast air from above down to the surface, based on how stable the air is | Gusts and lulls that follow the model’s estimate on average, but not minute by minute |
| Clouds and fog | Cloud forms when the box average reaches a set humidity, with extra rules for partly cloudy boxes | Fog banks, low cloud and sea fret that arrive early, late or not at all |
| Friction of the sea surface | Drag is taken from the wind speed and, in some models, from the waves | Wind over steep young seas, or near a coast, differs from the open-sea value |
| Sea breeze fronts and outflows | Only partly resolved; a coarse grid smooths the sharp edge into a gentle change | A wind shift and a rise in wind of several knots within minutes |
The details differ from model to model, which is one reason their forecasts differ.
Why local features are missed or smeared
- Size. A single thunderstorm cell is a few kilometres across. A model with 13 km boxes cannot hold it. It can only say that the box is favourable for storms and give a smoothed amount of rain.
- Timing and place. Whether a storm starts at 14:00 or 16:00, or over the headland or the bay, depends on small differences that the model cannot know. Even a very fine model puts individual cells in slightly wrong places.
- Sharp edges. A sea-breeze front or a squall gust front is a change over a few hundred metres. A model draws edges over at least a few grid spacings, so the change looks gradual and the peak is lower.
- Gusts are estimated. The model does not measure gusts. It calculates them from the mean wind, the stability of the air and the turbulence rule. A downdraft from a storm may not be in a coarse model at all, so the forecast gust can be far too low.
- Averages hide extremes. A box that is 20 kn on average can contain a patch of 35 kn under a cell and calm beside it.
How finer models help
At about 4 km and finer, models can let large thunderstorms form directly in the equations instead of through a rule. These are called convection-permitting models. They show the shape and organisation of squall lines, the way storms start along a sea breeze, and the cold pools that spread out from them, much better than coarse models do. Many national weather services run them for their coasts. They are also better at sharp coastal contrasts and at the sea breeze (see Wind: mean, gusts and coasts).
- They are not exact. They draw realistic storms, but where and when each one forms is still uncertain. A good fine-scale run is a picture of what could happen, not a promise of the position.
- They cover small areas for a short time. Most reach one to three days, so for later days you return to the coarser models.
- They need good input. They take their edges and starting state from a larger model. If that is wrong, the detail is wrong too.
- Ensembles help with the rest. A set of fine runs shows how likely it is that a storm affects a given area: Ensembles and probabilities.
For the wider picture of how a model is built, see How a model is computed. To compare several models for the same hours, see Using several models together.