Modelpublic

OneScience-Group/TropiCycloneNet

sourceHugging Faceapache-2.0updated 22d agoView on Hugging Face

0likes23downloads

Model Card

TropiCycloneNet

Model Introduction

TropiCycloneNet is a multimodal deep-learning method for joint global tropical-cyclone track and intensity forecasting. It fuses intrinsic cyclone attributes, local meteorological fields, and environmental information while generating multiple possible developments.

Paper: Benchmark dataset and deep learning method for global tropical cyclone forecasting https://doi.org/10.1038/s41467-025-61087-4

Model Description

The model was proposed by researchers from Zhejiang University of Technology, Shandong University, Tianjin University of Technology, and the Zhejiang Key Laboratory of Visual Information Intelligent Processing. It was trained with nearly 70 years of TCND best-track records, ERA5 fields, and environmental features spanning six ocean basins. Through joint temporal, spatial, and environmental encoding with multiple generators, the model supports 24-hour global tropical-cyclone track, central-pressure, and maximum-wind forecasting.

Use Cases

Use Case	Description
Track forecasting	Predict cyclone longitude and latitude at 6–24-hour leads.
Intensity forecasting	Predict central pressure and maximum sustained wind.
Multimodal modeling	Fuse Data1d, Data3d, and environmental inputs.
Multiple tendencies	Produce candidate paths from multiple generators.
ModelScope/OneCode execution	Validate structured data, training, inference, probabilistic precipitation metrics, and visualization in ModelScope or OneCode.
Multi-GPU training	Start multi-process training through `torchrun`.

Usage Instructions

Download

bash

hf download OneScience-Group/TropiCycloneNet --local-dir ./TropiCycloneNet
cd TropiCycloneNet

Environment Dependencies

Hardware Requirements

—A GPU or DCU is recommended.
—A CPU can run the default small-sample connectivity configuration.
—DCU users should install DTK 25.04.2 or a compatible OneScience-recommended version first.

DCU Environment

bash

# 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

bash

# 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

Synthetic Data

The paper uses eight six-hourly states containing four cyclone attributes, an 81x81 500 hPa geopotential-height field, and environmental variables to predict four future states. Synthetic data preserve the modality shapes, temporal order, six-basin protocol, and six generators while reducing the cyclone count, hidden width, and epochs. Results verify the engineering workflow only and do not represent paper performance.

bash

python scripts/fake_data.py

Training

Single-process training:

bash

python scripts/train.py

Distributed training:

bash

torchrun --nproc_per_node=2 --nnodes=1 --master_addr="localhost" --master_port=29500 scripts/train.py

The default synthetic run completes multimodal forward propagation, backpropagation, and parameter updates, and both single-process and two-process DDP training have been verified. Training produces one recoverable checkpoint containing the model state and configuration and records the training loss. Training results are saved to:

text

result/checkpoints/tropicyclonenet.pt
result/training/metrics.json

Weights

The paper's code is available at https://github.com/xiaochengfuhuo/TropiCycloneNet. This reproduction does not provide a link to an official pretrained checkpoint because a directly loadable checkpoint with an explicit license was not confirmed.

Inference

bash

python scripts/inference.py

Inference restores the checkpoint and produces six candidate track and intensity forecasts at 6, 12, 18, and 24 hours. The default output has passed shape and finite-value checks, and generator probabilities are retained. Inference results are saved to:

text

result/output/predictions.npz

Evaluation

bash

python scripts/result.py

Evaluation reports great-circle track MAE, pressure MAE, and wind-speed MAE and creates a lead-time and track comparison figure. All metrics are finite, and the PNG has passed format and non-empty-pixel checks; synthetic results do not represent paper performance. Evaluation results are saved to:

text

result/evaluation/metrics.json
result/evaluation/comparison.png

Official OneScience Information

Platform	OneScience	OneSkills
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 TropiCycloneNet specifications, with code licensed under the Apache License 2.0.

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