Skip to content



Folders and files

Last commit message
Last commit date

Latest commit


Repository files navigation

Geometric deep learning for medium-range weather prediction

Icíar Lloréns Jover, Michaël Defferrard, Gionata Ghiggi, Natalie Bolón Brun

The code in this repository provides a framework for a deep learning medium range weather prediction method based on graph spherical convolutions.

[June 2020]: The results obtained with this code are detailed in the Masters thesis report and slides.

[September 2020]: Results have been improved from the initial basis thanks to:

  • Introduction of residual connections in the architecture
  • Inclusion of further consecutive steps in the loss with different weighting schemes to reduce the loss at long term predictions
Model Z500 (6h) t850 (6h) Z500 (120h) t850 (120h)
Weyn et al 103.17  1.0380 611.33  2.957
Iciar June 2020 67.46 0.7172 861.7 3.432
Ours Sep 2020 61.58 0.7110 680.024 2.901
  • Results can be checked at plot_results.ipynb



For a local installation, follow the below instructions.

  1. Clone this repository.

    git clone
    cd weather_prediction
  2. Install the dependencies.

    conda env create -f environment.yml
  3. Create the data folders

    mkdir data/equiangular/5.625deg/ data/healpix/5.625deg/
  4. Download the WeatherBench data on the data/equiangular/5.625deg/ folder by following instructions on the WeatherBench repository.

  5. Interpolate the WeatherBench data onto the HEALPix grid. Modify the paremeters in scripts/config_data_interpolation.yml as desired.

    python -m scripts.data_iterpolation -c scripts/config_data_interpolation.yml


  • If deepsphere is not properly installed:

    conda activate weather_modelling
    pip install git+ 
  • If an incompatibility with YAML raises, the following command should solve the problem:

    conda activate weather_modelling
    pip install git+ --ignore-installed PyYAML
  • If it does not find the module SphereHealpix from pygsp, install the development branch using:

    conda activate weather_modelling
    pip install git+

Reproduce latest results

Train model:

The model listed as "Ours 2020" is trained using the module An example of how to use it can be found on the notebook Restarting_weights_per_epoch.ipynb.

The config file to be used is configs/config_residual_multiple_steps.json. You may want to modify the model name and/or the data paths if the data has been relocated.

Evaluate model:

You can generate the model predictions using the notebook generate_evaluate_predictions.ipynb. The parameters to be modified are:

  • model name (third cell)
  • epochs to be evaluated (you can define a range or a single one)

In order to evaluate the performance of the model, you only need to run up to "Generate plots for evaluation". This sencond part will generate the skill and climatology plots (you may be interested in generate them for a single epoch usually, not all of them)

Compare models:

In order to compare the performance of different models, or the same model at different epochs or simply a model against different baselines, you can use the notebook plot_results.ipynb. Depending on the purpose of the comparison, you may want to run a different section of the notebook. An explanation of each section and its use case can be found under the heading of the notebook.



Allows to train, test, generate predictions and evaluate them for a model trained with a loss function that includes 2 steps. All parameters, except GPU configuration, are defined in a config file such as the ones stored on the folder configs/ .

To use the mail notification at the end of the process, you need to provide a confMail.json file which must have the following structure:

  "password": "yourMailPassword",
  "sender": "yourMail"

Attention: If you are using gmail and have activated a two-step verification process, you need to get permission to the application and generate a new password. Details on how to generate the password can be found here


Allows to to train and test a model
with a loss function that includes multiple steps that can be defined by the user. It saves the model after every epoch but does not generate the predictions (to save time since it can be done in parallel using the notebook generate_evaluate_predictions.ipynb ). The parameters are defined inside the main function, although it can be adapted to use a config file as in

It is important to remark that the update function that takes care of the weight's update is defined on top of the file and should be adapted to the number of lead steps taken into account in the loss function.


Contains pytorch models used for both and Previous architectures used can be found in the folder modules/old_architectures/


Contains different functions to generate evaluation plots.


Contains code to train model with 2step-ahead prediction such as the one used for Iciar2020 results.


The main notebooks to explore are:

  1. Restarting_weights_per_epoch.ipynb Contains an example of how to use the functions that train the model that reported the best results mentioned earlier.

  2. generate_evaluate_predictions.ipynb Generate values on validation set using the weights of the desired saved model

  3. plot_results.ipynb Generate loss plots and comparison plots against different benchmark models

  4. healpix_resampling.ipynb Generate healpix data from equiangular data

  5. generate_observations.ipynb Generate ground-truth data for evaluation of the models


The content of this repository is released under the terms of the MIT license.

Cleanup check list

  • Benchmark_time_training.txt
  • LICENSE.txt
  • configs
    • config_bottleneck.json
    • config_original_less_channels.json
    • config_residual_level2.json
    • config_residual_level3.json
    • config_residual_level3_long_connections.json
    • config_residual_level4.json
    • config_residual_level4_nolong.json
    • config_residual_multiple_steps.json
    • config_train.json
  • data
    • final_models_rmse.pkl
    • weatherbench_training.npy
  • environment.yml
  • modules
    • old_architectures
  • notebooks
    • Iciar-report-notebooks
      • error_video.ipynb
      • test_dynamic_features.ipynb
      • test_static_features-nearest.ipynb
      • test_static_features.ipynb
      • test_temporal_dimension.ipynb
    • Restarting_weights_per_epoch.ipynb
    • full_pipeline.ipynb
    • generate_evaluate_predictions.ipynb
    • generate_observations.ipynb
    • healpix_resampling.ipynb
    • other_notebooks
      • benchmark_chunks.ipynb
      • benchmark_training_time.ipynb
      • effect_size_training_ds.ipynb
      • pytorch_weatherbench.ipynb
      • standardize_data.ipynb
    • plot_results.ipynb
    • running with variable weights multiple steps.ipynb
    • train_direct_prediction.ipynb
    • train_last_model_l1.ipynb
  • scripts
    • config_data_interpolation.yml
    • config_data_preprocessing.yml
  • weather.yml


No description, website, or topics provided.







No releases published


No packages published


  • Jupyter Notebook 99.8%
  • Python 0.2%