Multi-Modal Foundation · nasa-ibm-ai4science
Surya 1.0
A 366M-parameter foundation model trained on solar observatory data for space weather forecasting.
What Surya 1.0 is
Surya 1.0 is an open-source foundation model for heliophysics developed collaboratively by NASA, IBM, and partner organisations. It is designed to model solar dynamics and anticipate space weather events, such as solar flares and solar wind shifts, which can disrupt power grids, communications networks, orbital satellites, and astronaut safety.
The model is pretrained on roughly 218 TB of harmonised, multi-instrument observations collected across nearly a full solar cycle (2010 to 2019). By capturing both local and global solar behaviours directly from multi-channel observations, it serves as a base network that researchers can fine-tune for varied heliophysics and operational forecasting tasks.
How it works
Surya is a 366-million-parameter transformer trained on native 4096×4096 resolution data from NASA's Solar Dynamics Observatory (SDO). Its architecture combines spectral gating—using frequency-domain filtering—with long–short range attention mechanisms to process spatial patterns across wide scales efficiently.
The model's input spans 13 harmonised observation channels, consisting of eight Atmospheric Imaging Assembly (AIA) channels and five Helioseismic and Magnetic Imager (HMI) products. Pretraining uses a one-hour-ahead forecasting objective paired with autoregressive rollout tuning extending up to 12 hours.
Key facts
- Model architecture
- Transformer with spectral gating and long–short range attention
- Parameter count
- 366M
- Training data
- Approximately 218 TB (9 years, 2010–2019) of NASA SDO AIA and HMI multi-channel data
- Native resolution
- 4096×4096
- Input channels
- 13 (8 AIA channels, 5 HMI products)
- Initial release
- 2025
What you can use it for
- Solar flare forecasting
- Fine-tuning the model for 24-hour binary classification of major solar events (M-class and X-class flares).
- Active region segmentation
- Identifying and segmenting active solar regions on the Sun's surface, reaching an intersection over union (IoU) of 0.768 and a Dice score of 0.853.
- Solar wind speed prediction
- Adapting the pretrained representations to predict downstream solar wind velocity against physics-based baselines.
- Extreme ultraviolet (EUV) spectral forecasting
- Forecasting solar emission spectra across extreme ultraviolet wavelengths.
Run it here
We list this model but cannot run it yet: no pipeline_tag — needs a human.
Everything on this page still applies. Running it needs a loader for this repository, not a bigger machine.
Strengths and limits
Good at
- Trained directly on full native-resolution (4096×4096) multi-instrument solar images without spatial downsampling.
- Pretrained across 13 harmonised channels spanning nine years of observations covering nearly an entire solar cycle.
- Employs spectral gating and hybrid long–short range attention to capture both localised and large-scale solar dynamics efficiently.
- Outperforms standard UNet architectures on active region segmentation and improves over existing solar flare forecasting benchmarks by 15% in preliminary tests.
Watch out for
- Pretrained specifically on NASA SDO instrument products (AIA and HMI), requiring inputs formatted according to its multi-channel normalisation pipelines.
- Autoregressive rollouts are limited to 12 hours during tuning.
Inputs and outputs
You provide
Sequence or text
A protein or nucleotide sequence, or free text — whichever this model was trained on.
You get back
Mean-pooled representation, one vector per input.
Licence
Check the licence before commercial use · apache-2.0
The apache-2.0 licence carries terms worth reading before you depend on it.
Questions
Can I use Surya 1.0 commercially?
Check the licence before commercial use. The apache-2.0 licence carries terms worth reading before you depend on it.
What instruments provide the training data for Surya 1.0?
Surya 1.0 is trained on observations from two instruments aboard NASA's Solar Dynamics Observatory (SDO): eight channels from the Atmospheric Imaging Assembly (AIA) and five products from the Helioseismic and Magnetic Imager (HMI).
What is the primary pretraining objective?
The pretraining pipeline uses a one-hour-ahead forecasting objective along with autoregressive rollout tuning up to 12 hours.
Can the model be applied zero-shot?
Yes, the model supports zero-shot rollout for solar dynamics, though fine-tuned checkpoints can also be developed for specialised tasks like flare classification and region segmentation.
How to cite it
Roy, S., Schmude, J., Lal, R., Gaur, V., Freitag, M., Kuehnert, J., van Kessel, T., Hegde, D. V., Muñoz-Jaramillo, A., Jakubik, J., Vos, E., Mandal, K., Asanjan, A. A., de Sousa Almeida, J. L., Lin, A., Singh, T., Yang, K., Pandey, C., Hong, J., Aydin, B., Kurth, T., McGranaghan, R., Kasapis, S., Upendran, V., Bahauddin, S., da Silva, D., Pogorelov, N. V., Watson, C., Maskey, M., Guhathakurta, M., Bernabe-Moreno, J., & Ramachandran, R. (2025). Surya: Foundation Model for Heliophysics. arXiv:XXXX.XXXXX [astro-ph.SR].
Written from Surya 1.0's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.