Draft

Types of weather models

Global, regional, ensemble, wave, ocean, AI and nowcasting models: what each is for and how they fit together.

The main types at a glance

TypeWhat it is forGrid spacing (km)Forecast length (days)ExamplesIn catalog
GlobalCovers the whole Earth. The big picture and the longest range.7 to 254 to 16ECMWF IFSGFSICON15
Regional / high-resOne area in more detail; takes its edges from a global model. Shorter range.1 to 131 to 4ICON-D2AROMEHRRR52
EnsemblesMany slightly different runs that show the range of possible outcomes.1 to 332 to 16ECMWF ENSGEFSICON-EPS21
WavesSea state, driven by wind from an atmospheric model.1 to 257 to 16GFS-WaveECMWF WAMMFWAM18
Ocean, currents & iceCurrents, sea level, temperature and ice.1 to 258RTOFSCopernicus Global OceanCopernicus Baltic21
AI / machine learningMachine-learning models trained on decades of past weather; very fast to run.11 to 2815 to 16ECMWF AIFSGraphCast / GenCastPangu-Weather10

Resolution and length ranges are taken from the models in this catalog that state them, so they change as the catalog grows.

Global and regional models

A global model has to cover the whole planet, so its grid is relatively coarse and it can run out to ten days or more. A regional model (also called a limited-area model) covers one part of the world with a much finer grid. It cannot see beyond its own edges, so at every step it takes its boundary conditions from a global model. That is why a regional model is only as good as the global model feeding it, and why it is usually run for a shorter time.

Global modelWhole Earth, roughly 9 to 25 km grid, up to 10 to 16 daysRegional model (nest)One area, roughly 2 to 7 km grid, 1 to 5 days.Takes its edges from the global modelLocal, convection-permittingA small area, about 1 to 2.5 km grid, 1 to 3 days
Nesting. The finest model sits inside a coarser one, which sits inside the global model. Sizes are illustrative.

Below roughly 4 km grid spacing a model can start to simulate individual thunderstorms instead of approximating them. These are called convection-permitting models, and they are most valuable for squalls, sea breezes and winds around coasts. See How a model is computed.

Ensembles

An ensemble is not a different physics. It is the same model run many times, each time from a slightly different starting point and sometimes with slightly different physics. The spread between the runs measures how uncertain the forecast is. Because an ensemble needs many runs, it usually has a coarser grid than the main run of the same centre. Ensembles and probabilities shows how to read one.

Wave and ocean models

Waves are calculated by a separate wave model that is fed the wind from an atmospheric model and works out how much wave energy is generated, travels and decays. Ocean models do the same for currents, temperature, sea level and ice. A wind error therefore becomes a wave error: Waves explains the consequences.

AI models

Machine-learning models learn from decades of past weather instead of solving the physical equations. They run in minutes and are competitive on many measures, but they work differently and have different weaknesses: see AI weather models.

Nowcasting and blends

  • Nowcasting extends the latest radar and satellite images forward for the next hour or so. It is not a full model, but it is the best tool for the next squall.
  • Blends and post-processing combine several models and correct their known errors against observations. Many apps and national services do this on top of the raw models: see Why different apps show different forecasts.
  • Specialist models exist for tropical cyclones, storm surge, sea ice and other tasks: see Tropical cyclone models.