IBM and NASA have open-sourced Surya, the first heliophysics AI foundation model designed to read high-resolution solar imagery and forecast space weather events that can disrupt satellites, GPS, telecom, aviation, and power grids. Announced on August 20, 2025, Surya is available on Hugging Face for researchers and builders worldwide.
What’s new
- First-of-its-kind foundation model for the Sun, trained on nine years of NASA Solar Dynamics Observatory (SDO) data and released openly to accelerate scientific discovery.
- Early results show a ~16% improvement in solar flare classification accuracy vs. prior methods, plus visual forecasts up to two hours ahead—a novel capability for flare localization.
- IBM and NASA also released what they describe as the largest curated heliophysics dataset to advance data-driven space-weather research.
Under the hood
Surya is a ~366M-parameter transformer trained at SDO’s native 4096×4096 resolution across 13 channels (AIA + HMI). The team used forecasting objectives (e.g., predicting the next hour) and rollout tuning to help the model learn dynamics needed for downstream tasks like flare prediction, active-region segmentation, EUV spectral forecasting, and solar-wind speed estimation. The model is released under Apache-2.0 on Hugging Face, with weights, configs, and preprocessing pipelines for reproducibility.
Why it matters
- Operational impact: Better lead time and localization for solar flares can inform satellite maneuvers, grid hardening, and flight routing—reducing risk to critical infrastructure.
- Open science multiplier: Open weights + open datasets lower the barrier for space-weather startups, government labs, and universities to fine-tune models for regional or mission-specific needs.
- Template for science models: Surya extends IBM–NASA’s foundation-model playbook (e.g., Prithvi for geospatial & weather) from Earth observation to heliophysics, pointing to a future of domain-specific scientific FMs.
Get hands-on
- Press Release (overview & results): IBM Newsroom
- Model Card (tech details, weights, license): nasa-ibm-ai4science/Surya-1.0 on Hugging Face. Hugging Face
- Deep dive (training approach & SuryaBench): IBM Research
What to watch next
Expect fine-tuned variants for specific tasks (e.g., flare nowcasting and active-region tracking), community benchmarks via SuryaBench, and integrations into operational toolchains used by agencies and satellite operators. If the early numbers hold up in broader testing, Surya could become a go-to open baseline for space-weather forecasting.
Related reads on DYDC
- Agentic AI: A Beginner’s Guide — how agent frameworks turn models into doers.
- Salesforce + AI: The Future of Intelligent CRM — patterns for putting AI to work in the enterprise.
- Getting Started with Agentforce — practical steps to ship AI-assisted workflows.
Discover more from DYDC
Subscribe to get the latest posts sent to your email.
