<div align="center">

☀️ Surya: Foundation Model for Heliophysics ☀️

The first foundation model for heliophysics trained on full-resolution Solar Dynamics Observatory data

</div> <p align="center"> <img src="https://i.imgur.com/waxVImv.png" alt="Oryx Prithvi-EO-2.0"> </p>

📖 Overview

Surya (Sanskrit for "Sun") is a 366M-parameter foundation model for heliophysics, trained on full-resolution multi-instrument SDO observations (AIA & HMI). It learns general-purpose solar representations through spatiotemporal transformers, enabling state-of-the-art performance in solar flare forecasting, active region segmentation, solar wind prediction, and EUV spectra modeling.

Key Features

• Multi-instrument Learning: Trained on 13 channels from SDO's AIA (8 channels) and HMI (5 channels) instruments
• Full Resolution: Native 4096×4096 pixel resolution with 12-minute cadence
• Novel Architecture: Spatiotemporal transformer with spectral gating and long-short range attention
• Zero-shot Capabilities: Forecasts solar dynamics and flare events without additional training
• Versatile Fine-tuning: Parameter-efficient LoRA adaptation for diverse downstream tasks

What Makes Surya Special?

Unlike traditional task-specific models, Surya learns physics-aware representations that generalize across multiple solar phenomena:

• Solar Flare Forecasting
• Active Region Segmentation
• Solar Wind Prediction
• EUV Spectra Modeling

🚀 Quick Start

Prerequisites

• Python 3.11+
• CUDA-capable GPU (recommended)
• uv package manager (recommended)

🛠️ Installation

1. Clone the repository

git clone https://github.com/NASA-IMPACT/Surya.git
cd Surya

2. Install uv package manager (optional)

curl -LsSf https://astral.sh/uv/install.sh | sh
source ~/.bashrc

3. Set up the environment

uv sync
source .venv/bin/activate

4. Zero-Shot Inference

python easy_inference/run_easy_inference.py --config-path easy_inference/config_easy.yaml

Note

The above command is the fastest path to run Surya foundation-model inference on a date range without any additional setup. Data is downloaded from the public S3 bucket nasa-surya-bench. This prompts for start/end UTC datetime, downloads needed .nc files, and writes:

• easy_inference/outputs_.../prediction.nc
• easy_inference/outputs_.../metrics/*

Device selection behavior

• Please refer to the config_easy.yaml file for the default inference configuration.
• advanced.device: auto uses priority: cuda -> mps -> cpu
• Works across CUDA GPUs, Apple Silicon/macOS MPS, and plain CPU systems

---

🧪 Verify Installation

Run the end-to-end test to ensure everything is working:

python -m pytest -s -o log_cli=true tests/test_surya.py

This will:

• Download the pretrained model and test data
• Generate 2-hour ahead forecasts for 2014-01-07
• Create a validation visualization (surya_model_validation.png)
• Verify model inference

Expected output:

============================= test session starts ==============================
INFO     test_surya:test_surya.py:188 GPU detected. Running the test on device 0.
INFO     test_surya:test_surya.py:195 Surya FM: 366.19 M total parameters.
INFO     test_surya:test_surya.py:199 Loaded weights.
INFO     test_surya:test_surya.py:201 Starting inference run.
INFO     test_surya:test_surya.py:215 Completed validation run. Local loss 0.31665.
PASSED                                                    [100%]

🎯 Downstream Applications

To download the Surya model and a sample dataset for downstream tasks, please follow these steps:

# Step 1: Run pytest to download the model and verify dependencies
python -m pytest -s -o log_cli=true tests/test_surya.py  

# Step 2: Navigate to the downstream examples
cd downstream_examples/

# Step 3: Download the sample dataset
python download_data.py

1. Solar Flare Forecasting

Predict M-class and X-class solar flares up to 24 hours in advance.

cd downstream_examples/solar_flare_forcasting
bash download_data.sh
torchrun --nnodes=1 --nproc_per_node=1 --standalone finetune.py

2. Active Region Segmentation

Segment solar active regions and polarity inversion lines from magnetograms.

cd downstream_examples/ar_segmentation  
bash download_data.sh
torchrun --nnodes=1 --nproc_per_node=1 --standalone finetune.py

3. Solar Wind Forecasting

Predict solar wind speed at L1 point with 4-day lead time.

cd downstream_examples/solar_wind_forcasting
bash download_data.sh
torchrun --nnodes=1 --nproc_per_node=1 --standalone finetune.py

4. EUV Spectra Modeling

Model extreme ultraviolet irradiance across 1343 spectral bands (5-35 nm).

cd downstream_examples/euv_spectra_prediction
bash download_data.sh
torchrun --nnodes=1 --nproc_per_node=1 --standalone finetune.py

📥 Data and Model Access

Pretrained Models

The Surya foundation model and datasets are available on HuggingFace 🤗 :

• Model Repository: nasa-ibm-ai4science/Surya-1.0
• Dataset Repository: nasa-ibm-ai4science/core-sdo

SDO Data Download

For downstream applications, download the preprocessed SDO data:

cd downstream_examples
python download_data.py

This will:

1. Download data from HuggingFace repository
2. Extract and organize validation/test datasets
3. Generate CSV index files for each downstream task
4. Set up data in the expected directory structure

📊 Model Architecture

Surya employs a novel spatiotemporal transformer architecture optimized for solar dynamics:

Core Components

1. Spectral Gating Blocks (2 layers)

   • Frequency-domain filtering with learnable complex weights
   • Adaptive re-weighting of spectral components
   • Noise suppression and feature enhancement

2. Long-Short Attention Blocks (8 layers)

   • Local attention: Fine-scale dependencies within spatial windows
   • Global attention: Long-range correlations via dynamic projection
   • Multi-scale representation learning

3. Decoder Block

   • Lightweight projection back to physical domain
   • Maintains spatial structure and channel relationships

Training Strategy

• Phase 1: One-step ahead forecasting (160k steps, 128 GPUs)
• Phase 2: Autoregressive rollout tuning (2-5 hour horizons)
• Objective: Mean Squared Error with signum-log normalization
• Data: 2011-2019 SDO observations (~257TB processed)

Data Processing Pipeline

Our preprocessing ensures ML-ready, physics-consistent data:

• Temporal alignment: 12-minute cadence across all instruments
• Spatial registration: Uniform 0.6"/pixel grid, solar north alignment
• Calibration: Instrument degradation correction, exposure normalization
• Quality control: Automated flagging and filtering

🏆 Performance Benchmarks

| Task | Metric | Surya | Baseline | Improvement | |------|---------|-------|----------|-------------| | Solar Flare Forecasting | TSS | 0.436 | 0.358 (AlexNet) | 22% | | Active Region Segmentation | IoU | 0.768 | 0.688 (UNet) | 12% |
| Solar Wind Prediction | RMSE | 75.92 | 93.76 (ResNet50) | 19% | | EUV Spectra Modeling | MAPE | 1.48% | 1.68% (AlexNet) | 12% |

📄 Citation

If you use Surya in your research, please cite our paper:

@article{roy2025surya,
  title={Surya: Foundation Model for Heliophysics},
  author={Roy, Sujit and Schmude, Johannes and Lal, Rohit and Gaur, Vishal and Freitag, Marcus and Kuehnert, Julian and van Kessel, Theodore and Hegde, Dinesha V and Mu{\~n}oz-Jaramillo, Andr{\'e}s and Jakubik, Johannes and others},
  journal={arXiv preprint arXiv:2508.14112},
  year={2025}
}

📜 License

This project is licensed under the Apache License 2.0. See the LICENSE file for details.

🤝 Contributing

We welcome contributions to the Surya repository! Please see our contribution guidelines and feel free to:

• 🐛 Report bugs and issues
• 💡 Suggest new features or applications
• 🔧 Submit pull requests for improvements
• 📖 Improve documentation and examples