8 Minutes
A new AI to forecast solar storms
IBM and NASA have unveiled Surya, a first-of-its-kind foundation artificial intelligence (AI) model designed to predict violent solar flares and other space-weather phenomena with unprecedented speed and accuracy. Surya consumes raw, high-cadence imagery from NASA’s Solar Dynamics Observatory (SDO) and produces short-term forecasts of solar dynamics, including flare likelihood, solar wind behavior, and extreme-ultraviolet (EUV) spectral changes. The system is open-source and publicly available—accompanied by SuryaBench, a curated suite of datasets and benchmarks to accelerate heliophysics research and operational space-weather prediction.
Why accurate solar flare prediction matters
Space weather—driven by solar flares and coronal mass ejections (CMEs)—poses growing risks as human activity expands in orbit and society becomes more dependent on satellite-enabled infrastructure. Large flares and CMEs can damage satellites, disrupt radio and GPS signals used by aviation and maritime navigation, induce geomagnetically induced currents that threaten power grids, and increase radiation exposure for astronauts and high-altitude flight crews. Improving short-term forecasts of solar activity provides operators and mission planners the time needed to power down vulnerable systems, reorient spacecraft, protect astronauts, and mitigate downstream effects on communications and electrical networks.
Forecasting terrestrial weather is already challenging; predicting solar storms adds additional complexity. Light and electromagnetic signals from flare events travel at light speed, but there is an observation lag: photons from a flare take roughly eight minutes to reach Earth, limiting real-time awareness. Predictive models therefore must infer evolving conditions on the Sun and anticipate eruptions before their effects propagate through interplanetary space.
Scientific background: instruments, physics, and data
The Solar Dynamics Observatory (SDO) has been continuously observing the Sun for more than a decade, capturing full-disk images every 12 seconds across multiple wavelength bands. Two primary SDO instruments feed Surya:
Atmospheric Imaging Assembly (AIA)
AIA records the Sun’s upper atmosphere (the corona) in several EUV and UV channels. Each channel maps emission at different temperatures and heights in the solar atmosphere—together they reveal plasma structures, heating events, and eruptive behavior across roughly 1.3 solar diameters in view.
Helioseismic and Magnetic Imager (HMI)
HMI measures photospheric magnetic fields and helioseismic oscillations at the Sun’s visible surface. Magnetic field evolution—emergence, cancellation, and twisting of field lines—is a primary driver of flares and CMEs, so HMI’s vector magnetic data are essential for predictive modeling.
Surya integrates eight AIA channels and five HMI products to form a multi-layered representation of solar activity. The model is trained to recognize patterns in plasma emission, magnetic topology, and surface dynamics that presage energetic events.
Surya model: architecture, training, and capabilities
Surya is an open-source foundation model with roughly 360 million parameters, engineered to learn a compact, physically meaningful representation of solar behavior. Researchers selected a nine-year segment of harmonized SDO data to train the system: images and magnetic products were preprocessed and aligned so the model could learn cross-channel correlations and multi-timescale dynamics.
The training task used sequential image input and required the model to predict SDO observations one hour into the future. During development, teams experimented with architectures and data harmonization strategies; one notable finding was that Surya learned solar idiosyncrasies—such as differential rotation (faster spin at the equator than the poles)—more effectively from the data itself than by hardcoded rules.
Technical capabilities:
- Short-term forecasting: Surya predicts solar imagery, magnetic evolution, and derived quantities like EUV spectra and solar-wind-relevant features.
- Flare prediction: In tests the model could flag active regions likely to produce a flare within one hour and generated useful predictions out to two hours under certain visual-driven conditions.
- Operational potential: Faster-than-human automated feature extraction from petabytes of SDO imagery enables near-real-time alerts and downstream operational workflows.
Surya’s creators reported a roughly 16% improvement in flare-prediction performance compared with existing methods in their experiments. Results and model details were posted to arXiv on Aug. 18; the paper is currently a preprint and under peer review.
"We’ve been on this journey of pushing the limits of technology with NASA since 2023, delivering pioneering foundational AI models to gain an unprecedented understanding of our planet Earth," said Juan Bernabé-Moreno, director of IBM Research Europe for the U.K. and Ireland. "With Surya we have created the first foundation model to look the sun in the eye and forecast its moods."
"This is an excellent way to realize the potential of this data," said Kathy Reeves, a solar physicist at the Harvard–Smithsonian Center for Astrophysics, who was not involved in the study. "Pulling features and events out of petabytes of data is a laborious process and now we can automate it."
Testing, results, and operational implications
In benchmarking tasks, Surya produced hour-ahead imagery and probabilistic flare forecasts that matched or outperformed state-of-the-art techniques. The model’s ability to synthesize multi-channel inputs (AIA EUV bands and HMI magnetograms) enables it to detect subtle precursors—magnetic shear, emerging flux, and rapid coronal heating—that often precede energetic eruptions.
Operational benefits include:
- More lead time for satellite operators and mission control to switch systems to safe modes.
- Improved radiation risk forecasts for crewed missions and high-altitude aviation.
- Earlier warnings for electric utilities and GNSS-dependent services to prepare for possible geomagnetic effects.
The team has released Surya as open-source code on GitHub and hosted a copy on Hugging Face, alongside SuryaBench: a curated, documented collection of datasets and evaluation benchmarks meant to help other researchers reproduce results and extend the work.
Expert Insight
Dr. Elena Morales, a fictional senior heliophysicist and mission analyst (Expert Insight), comments: "Surya represents a step change in how we translate continuous observational streams into actionable forecasts. By learning directly from multi-wavelength imagery and magnetic products, the model can identify precursors that are difficult to extract with manual feature engineering. The open-source release and SuryaBench will be critical for building community trust, validating performance across solar cycles, and integrating predictions into operational pipelines that protect spacecraft and infrastructure."
Note: The quote above is a constructed expert perspective intended to illustrate how a practicing heliophysicist might assess the system’s capabilities and implications.
Related technologies and future prospects
Surya joins a growing family of foundation models for Earth and space science. IBM’s "Prithvi" models, for example, focus on terrestrial climate and weather tasks—mapping deforestation, modeling floods, and forecasting extreme heat—by ingesting terabytes to petabytes of satellite data. Integrating solar and terrestrial AI systems could enable end-to-end assessments of how solar storms propagate to the near-Earth environment and affect infrastructure.
Future directions include:
- Extending forecasts beyond short-term horizons by coupling Surya with heliospheric propagation models to predict CME arrival times and geomagnetic storm intensity.
- Continual learning to adapt to new solar conditions and instrument updates.
- Broader community validation across different phases of the solar cycle to quantify real-world operational value.
Conclusion
Surya is a milestone in space-weather forecasting: an open-source, 360-million-parameter AI trained on multi-channel SDO data that can predict short-term solar activity with improved accuracy. By automating feature extraction from petabytes of solar imagery, Surya provides a promising tool for protecting satellites, astronauts, and Earth-based infrastructure from harmful space-weather effects. The public release of the model and SuryaBench invites the scientific community to reproduce, validate, and extend these methods—an essential step for transitioning AI-driven solar forecasting from research to operational use. Continued evaluation, peer review, and integration with heliospheric and geomagnetic models will determine how much additional lead time Surya can reliably provide in real-world scenarios.
Comments