Here comes a foundation model for the Sun

Ninety-three million miles away, a blast of charged particles has started screaming its way toward Earth. This solar flare has the potential to disrupt our power grids, GPS, and even the internet. During the Sun’s most active phases, this scenario has the potential to play out hundreds of times each month.

NASA’s Solar Dynamic Observatory (SDO) satellite has been staring at the Sun for 15 years straight, aiming to learn more about how the Sun works. But researchers on Earth have barely unpacked all that its seen. When SDO launched into space, the AI boom was just starting, and tools for analyzing the satellite’s continuous stream of imagery were limited.  

Now, Surya, a first-of-its-kind foundation model for solar physics, is here to help. Using raw data captured by SDO, scientists at IBM, NASA, and eight other research centers,¹ have built an AI model of our host star that can be used to predict the kinds of violent solar outbursts that can endanger astronauts in space and throw off satellites, power grids, and communications here on Earth. 

Sanskrit for “Sun,” Surya is available starting today on Hugging Face, GitHub, and via IBM’s TerraTorch library for fine-tuning geospatial AI models. In addition to Surya, the team is also open-sourcing SuryaBench, a set of curated datasets and benchmarks aimed at simplifying the task of building and evaluating applications for not only space weather forecasting but to learn more about the Sun itself.

Predicting severe storms on Earth is difficult. It gets even trickier when the storms are brewing on a giant ball of plasma millions of miles away. When solar flares erupt through the Sun’s magnetic field, it takes 8 minutes just for that light to reach our eyes. The lag means that scientists need to get even further ahead of solar storms than the ones originating on Earth.

“We want to give Earth the longest lead time possible,” said Andrés Muñoz-Jaramillo, a solar physicist at the SouthWest Research Institute and a lead scientist on the project. “Our hope is that the model has learned all the critical processes behind our star’s evolution through time so that we can extract actionable insights.”

Into the heliosphere

It’s not just the distance separating our planet from the Sun that makes solar storm prediction so challenging. Scientists know considerably less about the Sun’s underlying physics than they do Earth’s. NASA sent SDO into space to gather facts but piecing them together has been slow work.

IBM and NASA’s earlier Prithvi family of AI models abstracted mountains of satellite data into a representation of Earth easily adapted for weather and climate prediction, among other tasks. The challenge for the Surya team was to similarly translate historical SDO data into a digital twin of the Sun that could be easily customized for different use-cases.

Surya is part of a broader effort at IBM to embrace generative and automated approaches that empower algorithms to be discovered, tested, and evolved at scale. Surya is the most recent example of how IBM sees AI not just as a tool, but as an originator and driver of scientific discovery.

“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, the IBM director in charge of the scientific collaboration with NASA. “With Surya we have created the first foundation model to look the Sun in the eye and forecast its moods.”

Kathy Reeves, a solar physicist at the Harvard–Smithsonian Center for Astrophysics, has been waiting for a model like this since helping to onboard one of SDO’s instruments. She has already discussed with her team how the new model might be used. “This is an excellent way to realize the potential of this data,” she said. “Pulling features and events out of petabytes of data is a laborious process and now we can automate it.”

SDO orbits alongside Earth to get a steady view of the Sun, snapping pictures every 12 seconds. But these aren’t images from any old camera. They capture the Sun at different wavelength bands to take the temperature of its layers, varying from a relatively cool 5,500°C at the Sun’s surface, to up to 2 million °C at the top of its atmosphere, the corona.

The SDO satellite also maps the Sun’s magnetic activity. Emerging sunspots are revealed in white light while other imaging tools clock the speed of bubbles at the Sun’s surface and track the tangling and twisting of the Sun’s magnetic lines.

To train Surya, researchers took a nine-year slice of this data, first harmonizing the different data types, then experimenting with AI architectures to process it. They settled on a long-short vision transformer with a spectral gating mechanism. Long-short attention was needed to digest SDO’s high-resolution 4096 x 4096-pixel images, which contained up to 10 times more detail than typical image data. The spectral gating cut memory costs by 5% and may have helped to filter noise from the data.

In their previous work training Prithvi on satellite data of Earth, researchers had the model reconstruct partially blacked out images, honing its ability to fill in missing values. With Surya, they tried something different: They gave the model sequential images and had it picture what SDO would see an hour in the future. They then checked the accuracy of its prediction against the actual observations.

By forcing the model to infer knowledge essential for skillful forecasting — things like the Sun’s geometry, magnetic structure, and its unusual rotation — they hoped to prepare it for a wide range of scientific tasks.

The Sun spins faster at its equator than its poles, and at first, researchers tried to encode this knowledge directly into the model. But letting it learn from the data proved far more effective. “It’s remarkable that the model performed better when we let it figure out the rotation on its own,” said Johannes Schmude, an AI researcher who was the technical project lead on the IBM side.

Surya proved capable at many forecasting tasks, including solar flare prediction. Scientists can currently tell an hour in advance from solar cues whether an active region is likely to set off a solar flare. In experiments, Surya provided a two-hour lead by using visual information. The model is thought to be the first to provide a warning of this kind. In early testing of the model, the team said they achieved a 16% improvement in solar flare classification accuracy, a marked improvement over existing methods.

A bridge to new discoveries

To make Surya more approachable for scientists without AI expertise, researchers aligned SuryaBench’s datasets and benchmarks for well-known space weather prediction tasks. They also included tasks to address longstanding mysteries like the magnetic structure of the corona and why solar winds intensify during the Sun’s quiet phase.

“Each of these applications is like a bridge to help the scientific community cross to the other side of the river,” said Muñoz-Jaramillo.

The applications target the Sun’s solar cycle, which peaks every decade or so when the Sun’s magnetic polarity flips. In the run up to the solar maximum, dark spots called active regions start to speckle the Sun’s surface. As they grow larger, solar flares sporadically erupt, and rising energy gathering in the Sun’s atmosphere can set off explosions, sending plasma and magnetic lines into space as solar wind and coronal mass ejections (CMEs).

When the Sun’s extended magnetic lines tangle with Earth’s own magnetic field, storms follow, disrupting GPS systems and long-distance communications, overloading energy transformers, and knocking satellites off orbit as the atmosphere heats and densifies. All this activity ultimately originates in the Sun’s active regions, making them ground zero for space weather prediction.

“These are the regions that produce all the magnetic particles we’re interested in,” said Spiros Kasapis, a postdoc at Princeton who studies active regions and contributed to SuryaBench. “This is essentially the very beginning of our relationship with the Sun.”

Since active regions fuel space weather generally, and solar flares, in particular, SuryaBench includes applications for predicting both. “If we can predict there’s a big active region forming, we can give NASA an early warning that hey, it will probably generate a lot of storms,” said Kasapis.

Other applications include detecting the buildup of extreme ultraviolet radiation and magnetism in the Sun’s atmosphere and predicting solar wind speeds. Extreme UV can cause Earth’s atmosphere to expand, sometimes forcing satellites to fall from their orbit and burst into flames. And accumulating magnetic lines in the Sun’s corona can accelerate solar wind to up to 400 kilometers per second, also wreaking havoc when they interact with Earth’s own magnitude lines.

“Knowing what the solar wind will be when it rams our magnetic shield, or interacts with it, brings us one step closer to understanding its effects on our power distribution systems,” said Muñoz-Jaramillo.

Surya and SuryaBench are set to offer an entirely new way of looking at the Sun. The team’s next release of curated data and benchmarks will address forecasting the effects of space weather on Earth. And now, with this new model, scientists have a tool to predict and respond to potentially dangerous solar flares faster than ever before. It’s a new world for how we look at the Sun.