DMLAB-GSU/Surya-Flare

Surya Flare Forecasting Models

This repository contains the model checkpoints for the Flare Forecasting downstream application, built upon the Surya Solar Foundation Model.

For the official codebase and training scripts, please visit the main project repository: surya_arseg_flare on Bitbucket.

For more details on the foundation model, please refer to the Surya paper.

Project Motivation

Flare Forecasting: Flare intensities and classes are gathered from NOAA/GOES data. A 24-hour prediction window is used to identify the maximum intensity flare within that period. For example, given inputs at (t-1 hour, t) and a prediction window of (t, t+24 hours), the inputs are labeled with the maximum flare intensity occurring within the 24-hour window. These labels are then binarized using a specific threshold for binary classification. The objective is to determine if the model can accurately forecast the occurrence of flares above this threshold within the next 24 hours.

Data Preparation

Surya Input Data (4K SDO Images)

For the foundation model input data, please check the Surya Core SDO dataset.

Flare Forecasting Data

The flare forecasting experiments are based on the Surya Bench dataset. You can download the flare dataset from Hugging Face. The training/validation/testing requires hourly 13-channel SDO images from 2010 to 2024, which requires ~70TB of disk space.

To download the SDO images (.nc files) from S3:

aws s3 sync s3://nasa-surya-bench/2010/ test_nc/2010/ --no-sign-request --exclude "*" --include "*_0000.nc"
# Repeat for 2011-2024

Checkpoints and Usage

This repository hosts the fine-tuned baseline checkpoints and Surya foundation model fine-tuned checkpoints for Flare Forecasting (e.g., C-class, M-class, X-class).

These weights can be directly used with the Surya codebase to run inference and fine-tuning scripts:

# Example command to run a fine-tuning or inference script for Flare Forecasting
python scripts/flare/finetune.py experiment=flare/exp_surya

Support & Troubleshooting

If you encounter any issues or have general questions, please reach out:

• Jinsu Hong: jhong36@gsu.edu
• Kang Yang: kyang30@gsu.edu
• Berkay Aydin: baydin2@gsu.edu

Acknowledgements

This material is based upon work supported by NASA under Award No. 80NSSC24PC377. Computational resources were provided by NSF ACCESS, GSU ARCTIC, and NASA Advanced Supercomputing (NAS). The authors acknowledge the NASA Solar Dynamics Observatory (SDO) team and NOAA/GOES for providing the data used in this work.