README.md

---
title: CRA5 Extreme ERA5 Compression
emoji: 🌍
colorFrom: green
colorTo: blue
sdk: gradio
sdk_version: 6.22.0
app_file: app.py
short_description: Decompress ERA5 weather data with the CRA5 VAEformer codec
python_version: "3.12"
startup_duration_timeout: 1h
pinned: false
license: cc-by-4.0
---

# 🌍 CRA5 — Extreme Compression of ERA5

This Space demonstrates **CRA5**, an efficient Variational Transformer (**VAEformer**)
for compressing the ECMWF **ERA5** reanalysis dataset.

A single hourly ERA5 snapshot contains **268 atmospheric variables** on a
**721 × 1440** global grid — roughly **1.1 GB** of `float32` data. CRA5 compresses
it into a **~2.4 MB** binary stream (a **~460× compression ratio**) while
preserving the physical fields well enough to train numerical weather-prediction
models.

The demo:

1. Fetches a tiny `.bin` stream for a chosen timestamp from the
   [CRA5 dataset](https://huggingface.co/datasets/taohan10200/CRA5-Dataset).
2. **Decompresses** it on the GPU with the VAEformer decoder.
3. Reconstructs the full 268-variable global field and visualises selected
   weather variables (geopotential, temperature, humidity, winds, …).

## Links

- **Model:** [taohan10200/CRA5-model](https://huggingface.co/taohan10200/CRA5-model)
- **Dataset:** [taohan10200/CRA5-Dataset](https://huggingface.co/datasets/taohan10200/CRA5-Dataset) (CC-BY-4.0)
- **Paper:** [CRA5: Extreme Compression of ERA5 (arXiv:2405.03376)](https://arxiv.org/abs/2405.03376)
- **Code:** [github.com/taohan10200/CRA5](https://github.com/taohan10200/CRA5)

The `cra5` package source (VAEformer + entropy models) is vendored from the
official repository under its BSD-3-Clause-Clear license; example weather data
is served from the CRA5 dataset (CC-BY-4.0), courtesy of Tao Han et al.

## Citation

```bibtex
@article{han2024cra5extremecompressionera5,
  title={CRA5: Extreme Compression of ERA5 for Portable Global Climate and Weather Research via an Efficient Variational Transformer},
  author={Tao Han and Zhenghao Chen and Song Guo and Wanghan Xu and Lei Bai},
  year={2024}, eprint={2405.03376}, archivePrefix={arXiv}, primaryClass={cs.LG}
}
```