SCRIPT

A General Deep Learning Framework for Satellite Data Assimilation

Overview

Data assimilation is an indispensable component of modern Numerical Weather Prediction (NWP) systems, playing a crucial role in generating analysis fields that significantly affect forecast performance. However, building an efficient DA system faces major challenges, particularly in establishing complex relationships between the background field and massive multi-source observational data within limited operational time windows. Here, we present FuXi-DA — a general deep learning–based data assimilation framework designed for assimilating satellite observations. By assimilating data from the Advanced Geosynchronous Radiation Imager (AGRI) aboard Fengyun-4B, FuXi-DA consistently reduces analysis errors and significantly improves forecast performance. Moreover, FuXi-DA has been validated against atmospheric physical theories, demonstrating its consistency and reliability.

Installation

We recommend installing SCRIPT in a virtual environment. For Conda, use the following command to create and activate a dedicated environment:

conda create -n script python=3.8 && source activate script

Next, install the required dependencies:

torch, xarray, dask, netCDF4, zarr

Tutorial

Run SCRIPT with a single command To run SCRIPT with one command, use:

cd fuxi-da/code_inference
bash run_inference.sh

The assimilation results are saved in fuxi-da/output.

reference

@article{xu2025fuxi,
  title={Fuxi-da: A generalized deep learning data assimilation framework for assimilating satellite observations},
  author={Xu, Xiaoze and Sun, Xiuyu and Han, Wei and Zhong, Xiaohui and Chen, Lei and Gao, Zhiqiu and Li, Hao},
  journal={npj Climate and Atmospheric Science},
  volume={8},
  number={1},
  pages={156},
  year={2025},
  publisher={Nature Publishing Group UK London}
}