Overview

• NASA has made the model, code and SuryaBench datasets publicly available on Hugging Face, GitHub and IBM’s TerraTorch to enable community testing and applications.
• Surya was trained on high-resolution SDO/AIA data with images roughly ten times larger than typical AI inputs, using transformer-based, multi-architecture methods to handle the scale.
• A 12-member science team led by Southwest Research Institute’s Andrés Muñoz-Jaramillo curated and corrected the dataset and established performance metrics for validation.
• The model has been evaluated on tasks including predicting solar flares, estimating solar wind speeds, forecasting EUV spectra and identifying developing active regions.
• NASA and IBM describe the release as a step toward stronger space-weather services, with additional validation and the incorporation of magnetograms and irradiance data planned before operational use.