Surya
NASA IMPACTSurya is a NASA-IBM heliophysics foundation model trained on aligned observations from the Solar Dynamics Observatory's AIA and HMI instruments. Its spatiotemporal transformer works at the instruments' native image resolution and supports solar forecasting, segmentation, solar-wind prediction, spectral modeling, and parameter-efficient downstream adaptation.
Openness
3 high confidence- weights
- open(Apache-2.0 ungated checkpoint)
- data
- open(core SDO dataset released on Hugging Face with download tooling)
- code
- partial(inference and four downstream fine-tuning paths
- license
- Apache-2.0(OSI)
Surya publishes the checkpoint and dataset, but the repository's training surface begins at downstream adaptation rather than reproducing foundation-model pretraining.
- https://huggingface.co/api/models/nasa-ibm-ai4science/Surya-1.0 recorded 2026-09-10
license: apache-2.0; gated: false; private: false; surya.366m.v1.pt present.
- https://raw.githubusercontent.com/NASA-IMPACT/Surya/main/README.md recorded 2026-09-10
Links the Apache-2.0 model and core-SDO dataset, provides data download scripts, inference, and four downstream fine-tuning examples.
Adoption
1 high confidenceAdoption is measured as trailing-30-day Hugging Face downloads of the declared checkpoint.
- https://huggingface.co/api/models/nasa-ibm-ai4science/Surya-1.0 recorded 2026-09-10
downloads: 407; gated: false; private: false.
Capability
4 high confidenceThree of the four capability attributes are established. Generation is not credited because the pretraining objective is one-hour-ahead forecasting with autoregressive rollout -- a deterministic future state rather than a generated scientific output -- which leaves TerraMind alone at the top of the category.
- https://raw.githubusercontent.com/NASA-IMPACT/Surya/main/README.md recorded 2026-09-10
Documents LoRA fine-tuning, 13 channels from AIA and HMI, two-hour forecasts, and local/global attention described as multi-scale representation learning.
Verified 2026-09-10