[paper] [website] [data & models] [live forecasts]
This repository contains the code and data for the paper Appa: Bending Weather Dynamics with Latent Diffusion Models for Global Data Assimilation by Gérôme Andry, Sacha Lewin, François Rozet, Omer Rochman, Victor Mangeleer, Matthias Pirlet, Elise Faulx, Marilaure Grégoire, and Gilles Louppe. The paper was published in 2025 in the Machine Learning and the Physical Sciences Workshop at NeurIPS.
Deep learning has advanced weather forecasting, but accurate prediction requires identifying the current state of the atmosphere from observational data. We introduce Appa, a score-based data assimilation model generating global atmospheric trajectories at 0.25° resolution and 1-hour intervals. Powered by a 565M-parameter latent diffusion model trained on ERA5, Appa can be conditioned on arbitrary observations to infer posterior distributions of plausible states without retraining. Our probabilistic framework handles reanalysis, filtering, and forecasting, within a single model, producing physically consistent reconstructions from various inputs. Results establish latent score-based data assimilation as a promising foundation for future global atmospheric modeling systems.
Most of our code should be well-documented. We split it in three parts:
appa: Main module containing the code for the data loading and processing, architectures, diffusion modules, losses, etc.experiments: Contains the scripts used to run each experiment in the paper.scripts: Contains data download and processing scripts, as well as plotting scripts.
The tests folder contains unit tests for the main parts of the code.
All the steps to run your own model or reproduce our results are detailed in the Wiki. Feel free to contact us through the issues or by mail if you have any further question.
Our work is based on the ERA5 reanalysis dataset, which is provided by Google through WeatherBench2. We provide scripts to download and process the data in the scripts/data folder. A latent dump of ERA5 encoded by our autoencoder is available on HuggingFace. This HuggingFace repository also contains our trained models and ERA5 statistics.
If you find this work useful in your research, please consider citing:
@article{andry2025appa,
title={Appa: Bending Weather Dynamics with Latent Diffusion Models for Global Data Assimilation},
author={Gérôme Andry and Sacha Lewin and François Rozet and Omer Rochman and Victor Mangeleer and Matthias Pirlet and Elise Faulx and Marilaure Grégoire and Gilles Louppe},
booktitle={Machine Learning and the Physical Sciences Workshop (NeurIPS)},
year={2025},
url={https://arxiv.org/abs/2504.18720},
}