OneScience-Group/Pangu-ICON-DKE

Pangu-ICON-DKE

Model Introduction

Pangu-ICON-DKE implements the non-training evaluation protocol from a study of ensemble perturbation growth in Pangu-Weather and ICON, diagnosing hourly evolution through difference kinetic energy (DKE) maps, global means, and total-wavenumber spectra.

Paper: Can Artificial Intelligence-Based Weather Prediction Models Simulate the Butterfly Effect?
https://doi.org/10.1029/2023GL105747

Model Description

The study reproduced by Pangu-ICON-DKE was proposed by research teams from the German Aerospace Center (DLR) and Ludwig Maximilian University of Munich (LMU). The paper analyzes 73-hour ensemble forecasts from five experiments with five members each, using fields at 300 hPa on a 721×1440 global grid and T719 spectra. The task requires no model training and compares initial-condition perturbation growth, butterfly effects, and intrinsic predictability in Pangu-Weather and ICON.

Use Cases

Use Case	Description
Butterfly-effect evaluation	Compare initial-condition perturbation growth in Pangu-Weather and ICON.
DKE and spectral diagnosis	Compute global DKE maps, hourly growth, and T719 spectra.
ModelScope/OneCode execution	Validate synthetic-data generation, inference, evaluation, and visualization in ModelScope or OneCode environments.
Hugging Face/OneCode execution	Validate the complete non-training evaluation workflow in Hugging Face or OneCode environments.
Multi-GPU protocol checks	Partition five experiments without duplication through torchrun.

Usage Instructions

1.OneCode

Experience intelligent, one-click AI4S programming through the OneCode online environment:

Try intelligent, one-click AI4S programming

2. Download and Installation

hf download OneScience-Group/Pangu-ICON-DKE --local-dir ./Pangu-ICON-DKE
cd Pangu-ICON-DKE

Environment Dependencies

Hardware Requirements

• A GPU or DCU is recommended.
• A CPU can be used for connectivity and protocol checks with the default synthetic-data configuration.
• DCU users must install DTK first. DTK 25.04.2 or later, or the OneScience-recommended version matching the cluster, is recommended.

DCU Environment

# Activate DTK and Conda first
conda create -n onescience311 python=3.11 -y
conda activate onescience311
pip install onescience[earth-dcu] -i http://mirrors.onescience.ai:3141/pypi/simple/ --trusted-host mirrors.onescience.ai

GPU Environment

# Activate Conda first
conda create -n onescience311 python=3.11 -y libstdcxx-ng=12 libgcc-ng=12 gcc_linux-64=12 gxx_linux-64=12
conda activate onescience311
pip install onescience[earth-gpu] -i http://mirrors.onescience.ai:3141/pypi/simple/ --trusted-host mirrors.onescience.ai

Training Data

This repository uses parameterized synthetic data to validate the engineering workflow while preserving five experiments, five members per experiment, 73 hourly times, one 300 hPa level, a 721×1440 global grid, and the T719 spectral protocol. Data are generated lazily in chunks, with analytic parameters replacing real forecasts instead of materializing complete raw fields. Synthetic data validate the non-training evaluation, inference, and visualization workflows only and do not represent official Pangu, ICON, or ECMWF data or paper results.

python scripts/fake_data.py --force

Training

For a single-device protocol check, use:

python scripts/train.py

For an eight-GPU protocol check, use:

torchrun --nproc_per_node=8 --nnodes=1 --rdzv_id=1000 --rdzv_backend=c10d --max_restarts=0 --master_addr="localhost" --master_port=29500 scripts/train.py

The paper has no model-training stage. train.py only checks the data protocol, experiment partitioning, and parameter-free DKE forward diagnostic; it creates no optimizer, executes no backward pass, and produces no trained parameters. Protocol-check artifacts are saved to:

weight/pangu_icon_dke.pt
result/training/metrics.json

Trained Weights

This repository does not include official weights under weight/. Because this is a non-training evaluation study, the protocol-check checkpoint stores configuration and protocol state only; official Pangu weights and ICON software are not redistributed.

Inference

python scripts/inference.py

Inference generates global DKE, hourly spatial fields, and T719 spectra for the five experiments, saving the artifact to:

result/inference_results.npz

Evaluation and Visualization

python scripts/result.py

Evaluation processes each of all 73 time steps and reports DKE, stepwise growth, scaling compensation, and inter-experiment spatial correlation; the task has no classes, so it produces no per-class metrics. It generates a DKE time-series plot, a 72-hour global map, and a T719-spectrum plot. Synthetic-data results validate the engineering workflow only and do not represent formal paper performance; artifacts are saved to:

result/evaluation/metrics.json
result/evaluation/dke_diagnostics.png

Official OneScience Information

Platform	OneScience Main Repository	Skills Repository
Gitee	https://gitee.com/onescience-ai/onescience	https://gitee.com/onescience-ai/oneskills
GitHub	https://github.com/onescience-ai/OneScience	https://github.com/onescience-ai/oneskills

Citation and License

This repository is an independent engineering reproduction of the public Pangu-ICON-DKE specifications.

Use of this repository's code, official model weights, and data remains subject to the licenses and terms of their respective projects.