OneScience-Group/FengWu-W2S

FengWu-W2S

Model Introduction

FengWu-W2S (FengWu Weather-to-Subseasonal) is a seamless global weather-to-subseasonal forecasting model extending FengWu.

Paper: FengWu-W2S: A deep learning model for seamless weather-to-subseasonal forecast of global atmosphere

https://arxiv.org/abs/2411.10191

Model Description

The model uses a six-hour time step for autoregressive forecasts of up to 42 days. Coupled atmospheric, ocean, and land branches, together with diversity perturbations, are used to improve extended-range forecast skill. This repository contains a compact implementation of the coupled interfaces for reproducible workflow checks.

Use Cases

Scenario	Description
Weather-to-subseasonal forecast research	Train a 78-channel coupled model with six-hourly global atmospheric, ocean, and land fields.
Local quick validation	Use small-grid synthetic HDF5 data to check training, fine-tuning, inference, and forecast visualization.
ModelScope / OneCode execution	Download the standalone model package, install dependencies, and run the scripts directly.
Multi-GPU training	Launch PyTorch DistributedDataParallel with torchrun.

Usage Guide

1. OneCode Usage

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

Click to Experience Intelligent One-Click AI4S Programming

2. Manual Installation and Usage

Hardware Requirements

• A GPU or DCU is recommended.
• CPU can be used for import and small-scale connectivity verification; full training and inference will be slow.
• DCU users must install DTK in advance. DTK 25.04.2 or above, or the OneScience recommended version matching your cluster, is recommended.

Download the Model Package

hf download OneScience-Group/FengWu-W2S --local-dir ./FengWu-W2S
cd FengWu-W2S

Install the Runtime Environment

DCU Environment

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

GPU Environment

# Please 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
# uv installation is supported
pip install onescience[earth-gpu] -i http://mirrors.onescience.ai:3141/pypi/simple/  --trusted-host mirrors.onescience.ai

Training Data Introduction

The OneScience community provides an ERA5 data slice for training. Download it and confirm that the data path in conf/config.yaml is correct:

hf download --repo-type dataset OneScience-Group/ERA5 --local-dir ./data

Real files must contain all 78 channels listed in conf/config.yaml, including atmospheric, ocean, and land variables, with six-hour time spacing. The configured group indices define the coupled branches.

Generate Synthetic Data

The default configuration describes a full 721x1440 grid. For a practical local smoke test, generate a small fixture explicitly:

python scripts/fake_data.py --height 32 --width 64 --timesteps 12

Synthetic HDF5 files are for workflow validation only and do not represent the ERA5 reanalysis or the paper's forecast quality.

Training

Single GPU:

python scripts/train.py

Multi-GPU:

torchrun --nproc_per_node=2 scripts/train.py

For a short smoke run, reduce the workload while keeping the generated data and model grid consistent:

python scripts/train.py --max-epoch 1 --batch-size 1 --num-workers 0 --rollout-steps 1

The default checkpoint is saved to data/checkpoints/model_bak.pth.

Fine-tuning

Resume from the base checkpoint with the lower fine-tuning learning rate:

python scripts/train.py --finetune

Training Weights

This repository provides a weight/ directory for FengWu-W2S checkpoints. The weight files will be uploaded soon and are expected to be available in the near future.

Inference

Inference reads data/checkpoints/model_bak.pth by default and writes forecasts grouped by initialization year to result/output/<year>/, together with result/output/index.json:

python scripts/inference.py

Use --steps and --limit to bound a local smoke test; --stochastic enables perturbation sampling.

Evaluation and Visualization

python scripts/result.py

The script computes per-channel RMSE, normalized RMSE, and anomaly ACC, and generates forecast-comparison, lead-time skill, channel-ranking, and training-loss figures.

Official OneScience Resources

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

• Paper: https://arxiv.org/abs/2411.10191
• This directory is an independent reproduction of the paper method and does not represent official code, weights, or training results released by the authors.