NASA's Surya AI predicts solar flares

NASA and IBM released Surya in August 2025 as an open-source artificial intelligence model built to forecast solar activity, including solar flares, solar wind and changes in extreme ultraviolet radiation. NASA described it as a heliophysics foundation model trained on full-resolution observations from the Solar Dynamics Observatory, or SDO. The agency said early results showed Surya improved solar flare forecasting against existing benchmarks by 16%. The code, model materials and benchmark datasets were published for outside researchers to inspect and use. ### What exactly is Surya? Surya is a 366 million-parameter transformer model trained on multi-instrument data from NASA’s Solar Dynamics Observatory. NASA, IBM and partner researchers said the system learned from 13 SDO data channels — eight from the Atmospheric Imaging Assembly and five from the Helioseismic and Magnetic Imager — at native 4096-by-4096 resolution and 12-minute cadence. NASA’s technical materials describe it as a general-purpose model for heliophysics rather than a single-task forecasting tool. (science.nasa.gov) Nine years of SDO observations, totaling about 218 terabytes, were used to pretrain the model, according to NASA and the project’s Hugging Face documentation. NASA said that scale was intended to help the system learn broad solar patterns that can later be adapted to different downstream tasks. ### What problem is it supposed to solve? NASA says solar flares and other space-weather events can disrupt satellites, communications systems, power grids and astronaut operations. (github.com) The agency said flare forecasting has been a long-standing challenge in heliophysics because many models are narrow, task-specific and limited by labeled training data. Surya was built to generate forecasts from raw solar observations and to extend beyond flare prediction into related solar phenomena. (science.data.nasa.gov) NASA’s science and data teams said the model can be used for solar flare forecasting, active-region segmentation, solar wind prediction and extreme ultraviolet spectra modeling. Princeton’s Space Physics group, which highlighted the release, said the open publication was meant to let outside scientists test the model on other phenomena such as coronal mass ejections and solar energetic particles. ### How does it make flare forecasts? (science.nasa.gov) NASA said Surya predicts future SDO images, which researchers then use to infer the Sun’s near-term state, including flare risk. The agency’s materials say the model can generate visual predictions of solar flares up to two hours into the future. GitHub documentation for the benchmark suite says the flare-forecasting task uses rolling windows and labels events by maximum flare class and cumulative flare intensity over a defined period. (github.com) IBM said the model uses spatiotemporal transformers to learn solar dynamics directly from raw observations. NASA’s project pages say that approach is intended to reduce dependence on hand-labeled datasets and make the system more adaptable across forecasting tasks. ### What has NASA made public? NASA published project explainers, technical-report entries and code repositories tied to the release. (science.nasa.gov) The Surya model code is available through GitHub, while model documentation and weights were posted on Hugging Face. NASA also linked SuryaBench, a benchmark dataset collection for machine-learning work on active regions, flare forecasting, solar wind and related tasks. The NASA Technical Reports Server includes a paper and related records describing the model architecture, training setup and evaluation tasks. (science.data.nasa.gov) NASA said the release was part of a broader open-science push aimed at researchers and operational forecasters working on space weather. ### What happens next for researchers and operators? NASA said Surya is available now for zero-shot inference and for fine-tuning on downstream heliophysics tasks. (github.com) The GitHub repository includes installation instructions and an inference workflow, while SuryaBench provides standardized evaluation data for follow-on testing. NASA and partner institutions said researchers can use those materials to assess whether the model improves operational forecasting for flares, solar wind and other space-weather events. (ntrs.nasa.gov)