AI reconstruction of European weather from the Euro-Atlantic regimes

Description

Code release for the paper "AI reconstruction of European weather from the Euro-Atlantic regimes"

Camilletti, A., Tomasi, E., Franch, G., (2025). AI reconstruction of European weather from the Euro-Atlantic regimes. arXiv preprint 2506.13758.

Preprint: https://arxiv.org/abs/2506.13758

Data & Models: https://zenodo.org/records/16751720

📁 Folders

The project follows the folder structure of lightning-hydra-template:

• configs: configs files (.yaml) for configuring the hyperparameters
• data: containing the data used for training and testing the model. Initially is empty. Data can be downloaded from Zenodo: https://zenodo.org/records/16751720
• logs: training output, including the pretrained models. It will be created automatically during training.
• notebooks: jupyter notebooks for analyzing the data and testing the model
• scripts: python scripts for analyzing the data, testing the model and producing the plots
• src: PyTorch implementation of the Dataloader and the Model

For a detailed description of the folder structure refer to lightning-hydra-template.

📄 Requirements:

Experiments were run in a Python 3.12 environment with the following packages:

• python = "^3.12"
• numpy = "^2.0.1"
• torch = "^2.4.0"
• lightning = "^2.3.3"
• xarray = "^2024.6.0"
• cfgrib = "^0.9.14.0"
• ipykernel = "^6.29.5"
• netcdf4 = "^1.7.1.post1"
• matplotlib = "^3.9.1"
• einops = "^0.8.0"
• torchinfo = "^1.8.0"
• hydra-core = "^1.3.2"
• rootutils = "^1.0.7"
• rich = "^13.7.1"
• hydra-colorlog = "^1.2.0"
• tensorboard = "^2.17.0"
• xeofs = "^3.0.2"
• xclim = "^0.57.0"
• xsdba = "^0.5.0"
• cartopy = "^0.25.0"
• hydra-joblib-launcher = "^1.2.0"

📦 Installation

If python3.12 is not already installed in your system, you can install it by running the following command:

# install python3.12 on ubuntu
bash install_python_ubuntu.sh

I suggest to create a new environment:

# create environment with poetry
bash create_environment.sh

# activate the environment
source .venv/bin/activate 

⬇️ Download data

To download the data from zenodo, run:

cd data

# download the dataset
python download_data.py

or download the data manually from Zenodo (https://zenodo.org/records/16751720)

⚙️ Pre-processing of the data

The raw data in the Zenodo dataset must be preprocessed to obtain the anomalies and the indices used to train and validate the model.

Compute anomalies

To compute the anomalies and bias-correct the SEAS5 forecast, simply run:

cd scripts

# compute the anomalies and save them in the data folder
bash compute_anomalies.sh

The bias correction will take several minutes (~1h 30min).

Compute indices

To compute the seven WR, four WR and NAO indices from the daily geopotential height anomalies, run:

cd scripts

# compute the indices and save them in the data folder
bash compute_z500_indices.sh

🤖 Train the model

To train models for both temperature and precipitation reconstruction with the same hyperparameters used in the paper:

# train all models used in the paper
bash scripts/schedule.sh

To change the hyperparameter you can create a new configs/experiment/new-experiment.yaml overriding the hyperparamenters you want to modify. Then you can run

# run a custom experiment
python train.py experiment=new-experiment

📊 Reproduce the plots

To reproduce the plots shown in the paper, run:

cd scripts

# create the plots and tables
bash produce_results.sh

This will takes several hours (~3h).

📖 Citing

If you use this code or data, please cite:

Camilletti, A., Franch, G., Tomasi, E., & Cristoforetti, M. (2025). AI reconstruction of European weather from the Euro-Atlantic regimes. ArXiv. https://doi.org/10.48550/arXiv.2506.13758