LangYa Ocean Large Model

This repository contains the code and mdoels used for "LangYa: Revolutionizing Cross-Spatiotemporal Ocean Forecasting" [preprint paper]

The term Langya is taken from the Ci Hai (Chinese Dictionary), and refers to treasures with fine textures, crystal-clear and translucent like jade. Historically, Langya Terrace, located to the south of the Guzhenkou Campus of the IOCAS, had served as an important center for observing celestial bodies, including the sun, moon, and stars, and is a key site related to the twenty-four solar terms. Our ocean large model is named "LangYa," which aligns perfectly with the mission it carries in the field of modern oceanography.

Getting Started

1. Environment Setup

conda env create -f environment.yml

2. Model Testing

Our model training & testing is implemented using distributed computing with Slurm workload manager, utilizing 16 GPUs across 4 nodes (4 GPUs per node). You can customize the testing configuration by modifying the parameters in submit_test.sh according to your hardware setup.

To start the testing process, simply run:

sbatch submit_test.sh

Note: Before testing, please download the model weights, and place them in the weights folder.

weights/
    ├── langya_v1.tar

Version Notes

Release Version	Model Weights	Training Data	OSV
Initial release v1.0	pan.baidu, google drive	ERA5, GLORY12	Temperature, Salinity, Velocity-U, Velocity-V

Updates

• [2025/03/31] LangYa Test Script and Dataset has been released.
• [2024/12/28] The launch event of LangYa v1.0 was successfully held at the Guzhenkou Campus of the Institute of Oceanology, Chinese Academy of Sciences (IOCAS), in Qingdao, China. (CCTV News, CAS News, China daily)

References

For training and testing LangYa v1.0, we downloaded the GLORYS12 and ERA5. For comparison with other methods, we downloaded the IV-TT Class 4 framework. All these data are publicly available for research purposes.

If you find this work useful, cite it using:

@article{yang2024lang,
      title={LangYa: Revolutionizing Cross-Spatiotemporal Ocean Forecasting}, 
      author={Nan Yang and Chong Wang and Meihua Zhao and Zimeng Zhao and Huiling Zheng and Bin Zhang and Jianing Wang and Xiaofeng Li},
      year={2024},
      eprint={2412.18097},
      archivePrefix={arXiv},
      primaryClass={physics.ao-ph},
}