Machine Learning model data

ECMWF is now running a series of data-driven forecasts as part of its experimental suite. These machine-learning based models are very fast, and they produce a 10-day forecast with 6-hourly time steps in approximately one minute. The outputs are available in graphical form.

Currently, three of these models are available:

  • FourCastNetv2-small: the next iteration of the FourCastNet deep-learning system developed by Nvidia and collaborators. This model uses Spherical Fourier Neural Operators to capture spatial dependencies. This model is a reduced version that can fit into a single Nvidia A100 40GB for inference. It was trained using ERA5 to minimise mean-squared-error predictions and operates at 0.25°.
  • Graphcast: a deep learning-based system developed by Google Deepmind. It uses a graph neural network architecture with an encoder-processor-decoder structure with a multi-mesh representation. The model was trained on ERA5 at a 0.25° resolution and fine-tuned on the ECMWF HRES forecast to minimise mean-squared-error predictions. 

  • Pangu-Weather: a deep learning-based system developed by Huawei. It uses a 3D transformer architecture to capture spatial dependencies and is comprised of multiple models to make predictions across different temporal steps (e.g., 24 hours, 6 hours). It was trained using ERA5 to minimise mean-squared-error predictions for each model and operates at 0.25°.

More models will be added, and some will be discontinued when other versions are available. This web page will be updated accordingly.

You can also read our blog post about machine learning in weather forecasting, a newsletter on extreme event forecasting with machine learning and a preprint article on verification of machine learning models.

Available products

The output of these ML models are forecasts with 6-hourly time steps out to 10 days initialised from the ECMWF operational analysis. All forecasts are produced on a 0.25 x 0.25-degree grid.

FourCastNetv2-small:

Upper-air fields are z (geopotential), r (relative humidity), t (temperature), u (U component of wind), and v (V component of wind) on the following pressure levels: 1000hPa, 925hPa, 850hPa, 700hPa, 600hPa, 500hPa, 400hP, 300hPa, 250hPa, 200hPa, 150hPa, 100hPa and 50hPa

The single-level fields are msl (mean sea level pressure), sp (surface pressure), 10u (10 metre U wind component), 10v (10 metre U wind component), 100u (100 metre U wind component), 100v (100 metre V wind component, 2t (2 metre temperature) and tcwv (total column vertically-integrated water vapour).

Graphcast:

Upper-air fields are z (geopotential), q (specific humidity), t (temperature), u (U component of wind), v (V component of wind), w (vertical wind component) on the following pressure levels: 1000hPa, 925hPa, 850hPa, 700hPa, 600hPa, 500hPa, 400hP, 300hPa, 250hPa, 200hPa, 150hPa, 100hPa and 50hPa

The single-level fields are msl (mean sea level pressure), 10u (10 metre U wind component), 10v (10 metre U wind component), and 2t (2 metre temperature). 

Pangu-Weather:

Upper-air fields are z (geopotential), q (specific humidity), t (temperature), u (U component of wind), and v (V component of wind) on the following pressure levels: 1000hPa, 925hPa, 850hPa, 700hPa, 600hPa, 500hPa, 400hPa, 300hPa, 250hPa, 200hPa, 150hPa, 100hPa and 50hPa

The single-level fields are msl (mean sea level pressure), 10u (10 metre U wind component), 10v (10 metre U wind component) and 2t (2 metre temperature).

Graphical products

Graphical products are available as part of the charts catalogue and the ecCharts application. If the products are not found, please reload the page.

In both applications, search for “machine learning” to find the relevant products.

Running these models in your organisation

ECMWF has developed a tool to run these data-driven forecasts starting from the Centre’s data in MARS, the Climate Data Store or a GRIB file. The code is written in Python and can be install as:

pip install ai-models ai-models-panguweather ai-models-fourcastnetv2 ai-models-graphcast

FourCastNet and GraphCast are written in different frameworks (Pytorch and Jax) that may be incompatible currently. But you can install them in separate environments.

And then simply call:

ai-models panguweather

to run that model on the most recent operational analysis. This requires a MARS account to fetch the initial data. Although the code may run on a CPU, it will be very slow. It is strongly recommended to run it on a computer with a GPU.

For more information, see:

 

Period
Keywords
Machine learning
Range
Medium (15 days)
Provided by
ECMWF