Climate2Weather / README.md
zhangrenchao's picture
Add English model card
699b17a verified
|
Raw
History Blame Contribute Delete
2.7 kB
metadata
license: apache-2.0
language:
  - en
tags:
  - OneScience
  - Earth Science
  - Climate Downscaling
  - Diffusion
  - Probabilistic Forecasting
  - Coherence
frameworks: PyTorch

Climate2Weather

# Model Introduction Climate2Weather uses conditional score diffusion to transform coarse climate simulations into probabilistic high-resolution weather trajectories.

Paper: A Generative Framework for Probabilistic, Spatiotemporally Coherent Downscaling of Climate Simulation
https://doi.org/10.1038/s41612-025-01157-y

Model Description

The method was proposed by the University of Tübingen and Tübingen AI Center. It was trained with 2006–2013 COSMO-REA6 reanalysis and conditions on climate-model fields only during inference. Score-based data assimilation jointly downscales four variables from 8x8 to 128x128 and from six-hourly to hourly resolution.

Use Cases

Use Case Description
Probabilistic downscaling Generate high-resolution ensemble trajectories.
Coherent generation Jointly model variables and time.
Multivariate generation Downscale wind, temperature, and sea-level pressure jointly.
Climate impacts Generate fine-scale drivers for regional impact studies.
ModelScope/OneCode execution Validate data, training, inference, metrics, and visualization.
Multi-GPU training Start multi-process training through torchrun.

Usage Instructions

Use a GPU or DCU when available; CPU supports the default smoke configuration.

hf download OneScience-Group/Climate2Weather --local-dir ./Climate2Weather
cd Climate2Weather
python scripts/fake_data.py

For single-process training, use:

python scripts/train.py

For multi-process training, use:

torchrun --standalone --nproc_per_node=2 scripts/train.py

Run inference and evaluation with:

python scripts/inference.py
python scripts/result.py

Training minimizes denoising score matching. Inference produces a finite [8,3,4,128,128] ensemble with positive spread; evaluation reports RMSE, spread, PIT, and temporal differences.

Trained Weights

No weights are bundled under weight/. The authors provide trained diffusion-model weights and experiment code at https://github.com/schmidtjonathan/Climate2Weather; this compact implementation does not claim compatibility.

Citation and License

This repository is an independent engineering reproduction of the public Climate2Weather specifications.

The original paper is licensed under CC BY 4.0; the original paper, official code, model weights, and related data remain subject to their respective licenses and terms.