Detailed guide for downloading forecast data

Overview

All forecasts submitted to the AI Weather Quest are openly available via the AI Weather Quest Data Portal. Participants submit quintile-based probabilistic forecasts on a 1.5 degree latitude/longitude grid for near-surface air temperature (tas), mean sea level pressure (mslp), and accumulated precipitation (pr), targeting lead times of days 19 to 25 (week 3) and 26 to 32 (week 4). This guide explains how to download, extract, and use individual forecast files or archives.

The following teams webpage contains information regarding individual teams and their associated models.

Data publication timetable

The first year of the AI Weather Quest is divided into four 13-week competitive periods. Once the last set of forecasts within a competition period have been evaluated, all submitted predictions from that phase are made openly available. Publication dates for each period are: 

Portal directory structure

Individual files

Forecasts are organised in two ways:

1. Forecast initialisation date (by_fc_date): by_fc_date/ -> by_team/ -> by_model/ -> forecasts
2. Participating team (by_team): by_team/ -> by_model/ -> by_fc_date/ -> forecasts

Each forecast variable and lead time is stored as a separate NetCDF (.nc) file.

Zip/Tar Archives

For simplify downloading, we also provide .tar.gz archives containing all forecasts either:

• submitted by a given team (by_team_zipfiles)
• issued for a given initialisation date (by_fc_date_zipfiles) 

Filename conventions

Each forecast file is in a NetCDF format with the following naming convention:

[variable]_[fc_start_date]_p[fc_window]_[teamname]_[modelname].nc 

where: 

• variable: The forecasted variable. Options are:
  • 'tas': Near-surface temperature (weekly-mean)
  • 'mslp': Mean sea level pressure (weekly_mean)
  • 'pr': Precipitation (weekly-accumulated)
• fc_start_date: The forecast initialisation date in format YYYYMMDD (e.g., ‘20250515’ for 15th May 2025).
• fc_window: The selected forecasting window. Valid options are:
  • '1': Weekly-mean forecasts for days 19 to 25 inclusive.
  • '2': Weekly-mean forecasts for days 26 to 32 inclusive.
• teamname : The associated team name with the submitted forecast.
• modelname: The associated model name with the submitted forecast.

Forecast download

Via the web browser

Click on individual files or zip/tar archives to download. Files will download automatically.

Via Linux/Unix terminal

To download forecasts in a Linux environment we recommend using wget or curl. When using either set of functions you will require the top-level URL of the portal (https://data.ecmwf.int/ai-weatherquest/) and the full directory path to the requested file.

Directory paths to individual forecasts are as follows: 

• Categorised by forecast initialisation date: 

     [ROOT]/by_fc_date/[fc_init_date]/[teamname]/[modelname]/[forecast_file]

• Categorised by team:

     [ROOT]/by_team/[teamname]/[modelname]/[fc_init_date]/[forecast_file]

Additionally, to download all forecasts submitted by a certain team or for a particular forecast initialisation date, use the following directory paths:

• Forecast initialisation date:

     [ROOT]/by_fc_date_zipfiles/[fc_init_date].zip

• Team:

     [ROOT]/by_team_zipfiles/[teamname].zip

The following are examples use the wget functionality:

Download near-surface temperature forecasts initialised on the 14th August 2025, for the first forecasting period, and submitted by team AIFS under the model name AIFSgaia:

     wget https://data.ecmwf.int/ai-weatherquest/by_fc_date/20250814/AIFS/AIFSgaia/tas_20250814_p1_AIFS_AIFSgaia.nc

Download all forecasts submitted by team AIFS throughout the competition:

     wget https://data.ecmwf.int/ai-weatherquest/by_team_zipfiles/AIFS.tar.gz

Extracting Archives

Finally, it is recommend that the following tar command is used to unzip zip files.

     tar -czf [zip_file] [new_directory]

     tar -czf AIFS.tar.gz AIFS_forecasts/

Forecast file content

All forecasts are saved in NetCDF files created using the AI_WQ_create_empty_dataarray function in the forecast_submission module of AI Weather Quest Python Package.

All data is stored in a fractional format.

Dimensions

The files include additional attributes to support file storage and forecast visualisation. Attributes important to forecast users include:

Examples of opening forecasts

Forecasts are easy to open in a Python environment. It is highly recommend to load in individual forecasts using array:

import xarray as xr
fc = xr.load_dataarray([filename])

It is also recommend that xarray is used to open multiple forecasts of the same variable and forecasting window from different models. As an example:

fcs = xr.load_dataarray('tas_20250814_p1_*.nc',combine='nested',concat_dim='modelname')

This output has an additionally dimension called modelname.

Licence information

This dataset is released under ECMWF's general licence and the Creative Commons Attribution 4.0 International (CC BY 4.0) licence. 

Contact / support information

For support with accessing AI WQ sub-seasonal forecast data, please use one of the following communication channels:

• Public forum.
• Email: aiweatherquest@ecmwf.int

Dataset citation / DOI

When using this dataset, please cite the dataset DOI (provided on the product webpage) and the following reference:

Loegel, O., Talib, J., Vitart, F., Hoffmann, J. and Chantry, M., 2025. The AI Weather Quest: an international competition for sub-seasonal forecasting with AI. Machine Learning: Earth, 1(1), p.010701. https://doi.org/10.1088/3049-4753/adf649 

Acknowledgement

If applicable, please acknowledge the AI Weather Quest competition as follows: “This dataset was produced as part of the AI Weather Quest competition organised by ECMWF (2025-2026)".