NASA and IBM just built an AI model for the Sun

Surya is an open-source heliophysics foundation model built by NASA and IBM, designed to process decades of raw solar observations and forecast space weather.

Its core job is practical early warning: helping satellite operators, power grid managers, and researchers see intense solar eruptions coming before they knock out communications or damage electronics.

What it can do

Surya's flagship capability is visual solar flare forecasting. Instead of just spitting out probability numbers, it can generate visual predictions of how a flare will look up to two hours into the future.

In initial benchmarks, Surya beat existing state-of-the-art predictive baselines by 16%. In the world of heliophysics, where predicting flares has historically been an uphill battle, that is a massive leap.

Beyond flares, Surya tracks active solar regions, estimates solar wind speeds, and models how the Sun's ultraviolet radiation heats and expands Earth's upper atmosphere—a crucial factor for low Earth orbit satellite drag.

The underlying setup can also ingest observation streams from other missions, including the joint NASA-ESA SOHO observatory and NASA's Parker Solar Probe.

How it works

Surya was trained on 9 continuous years of observational data from NASA’s Solar Dynamics Observatory (SDO).

Since launching in 2010, SDO has snapped images of the Sun across multiple wavelengths every 12 seconds, paired with magnetic field measurements. This stable dataset spans nearly an entire solar cycle, exposing the model to patterns that shorter records simply miss.

Architecturally, Surya is a true foundation model. It learns self-supervised representations directly from raw multi-wavelength solar imagery rather than human-labeled datasets.

Training infrastructure was supported by the U.S. National AI Research Resource (NAIRR) Pilot, tapping into computing power provided by NVIDIA and the National Science Foundation.

Where to find it

NASA and IBM made the entire effort openly accessible.

You can pull the model weights directly from Hugging Face, inspect and clone the code repository on GitHub, and freely access the underlying training data.

If you work in astrophysics, space systems, or scientific machine learning, you can readily grab the checkpoints and fine-tune them on your own downstream solar forecasting tasks.

Why it matters

We think Surya is one of the clearest proofs yet that foundation models belong in hard physical sciences, not just text and generic images. Instead of rebuilding ad-hoc statistical pipelines from scratch, the scientific community now has a shared, pre-trained base model for solar dynamics.

Anyone managing satellite constellations or researching atmospheric drag should check out this repo immediately. That said, bridging the gap from a 2-hour visual forecast to automated operational defense systems will still require heavy validation against rare, catastrophic solar events.