NASA and IBM unveil Surya AI for solar flare forecasts two-hour lead

NASA and IBM have unveiled Surya, a new artificial intelligence model that can predict solar flares and related geomagnetic storms up to two hours before they erupt, offering earlier and more precise warnings for satellite operators, power grid managers and other infrastructure managers. Built as a heliophysics foundation model and trained on nearly a decade of high resolution observations from NASA’s Solar Dynamics Observatory, Surya improves short term forecasting accuracy and produces visual maps indicating where on the Sun a flare is likely to occur.

Surya, named after the Sanskrit word for Sun, is a 366 million parameter model that uses spatiotemporal transformer techniques to learn general purpose representations of solar behavior. The model was trained principally on SDO data, a continuous record of solar images captured every 12 seconds since 2010, and supplemented with datasets from multiple research centers. Developers say this long, well calibrated record spanning an entire solar cycle is uniquely suited to teach AI systems to detect subtle, predictive patterns that shorter datasets would miss.

Researchers describe Surya as the first foundation AI trained specifically...