AI weather models
Machine-learning models such as AIFS, GraphCast and Pangu: how they differ from physics-based models, and where they are weaker.
How they work
A machine-learning model is trained on a long record of past atmospheric states, typically a reanalysis covering several decades. It learns to predict the state a few hours ahead from the current one, then repeats that step to reach days ahead. There are no equations of motion inside: the model has learned, statistically, how weather patterns evolve. Once trained, a forecast takes minutes on a single graphics processor, where a physics-based run needs a supercomputer for hours.
Strengths
- Speed and cost. Cheap runs make large ensembles and frequent updates affordable.
- Skill on large patterns. In published comparisons, several AI models match or exceed leading physics-based models in the accuracy of pressure patterns and winds at medium range.
- Operational use. Some, such as ECMWF’s AIFS and its ensemble, now run routinely next to the centre’s physics-based model.
Weaknesses
- Smooth fields. Many AI models blur small-scale features, so they can underplay the sharpest winds, rain bands and gusts.
- Extremes and rare events. They learn from what happened before. Events outside the training record, or very intense ones such as the strongest tropical cyclones, are harder.
- Coarse grids and few outputs. Many run at about 0.25° (roughly 28 km) and provide fewer variables than a physics-based model, for example often no direct wind-gust or wave output.
- They still need a starting point. Most are started from an analysis produced by a physics-based assimilation system, so they inherit its errors.
AI models in this catalog
| Model | From | Notes |
|---|---|---|
| ECMWF AIFS | ECMWF | Runs operationally; there is also an ensemble version |
| NOAA AI-GFS | NOAA | AI model started from GFS analyses |
| GraphCast / GenCast | Google DeepMind | A deterministic model and a probabilistic one |
| Pangu-Weather | Huawei | Early model that showed AI could match the best |
| FourCastNet | NVIDIA | Fast model with public experimental runs |
| Aurora | Microsoft | A foundation model used for weather and for other Earth-system tasks |
| FuXi | Fudan University | Medium-range model, with a cascade of models for different lead times |
| FengWu | Shanghai AI Laboratory | Global medium-range model |
Most of these are listed without publication times because they are research or partner services; check the model page for what is known.