Editor's Note: This article was updated Aug. 20, 2025, to correct the number of years of training data used.

NASA’s newly released Surya Heliophysics Foundation Model (Surya) marks a pivotal moment in how scientists study and predict the Sun’s behavior. Trained on 9 years of high-resolution observations from the Solar Dynamics Observatory (SDO), Surya can forecast solar activity like flares, winds, and irradiance with unprecedented speed and accuracy.

This open-source model combines technical ingenuity, months of problem solving, and collaboration across multiple institutions and disciplines.

A Cross-Division Collaboration

“Surya is the next model in our 5+1 strategy and marks an exciting step forward in how we use AI for data-driven science. We are developing powerful AI models for each of NASA’s major science areas, plus one large language model that connects them all,” said Kevin Murphy, Chief Science Data Officer for NASA’s Science Mission Directorate at NASA Headquarters in Washington. “We’re embedding NASA’s scientific expertise directly into these models to turn our vast amount of data into faster discoveries and smarter decisions with real-world impact.”

The project began as a joint effort between NASA’s Office of the Chief Science Data Officer (OCSDO) and the Heliophysics Division, aligned with NASA's 5+1 strategy for advancing science with artificial intelligence. This strategy commits the agency to developing foundation models as flexible community serving tools, platforms that can transform research across domains.

"Domain scientists and computer engineers worked side-by-side with petabytes of SDO data, and the results are remarkable," said Madhulika “Lika” Guhathakurta, heliophysicist and SDO program scientist. "Surya wasn’t just built for heliophysics—it was built with heliophysics. It speaks the language of our discipline from day one and it’s already teaching us new ways to listen to our star."

Led By the IMPACT AI Team

The Interagency Implementation and Advanced Concepts Team (IMPACT) under NASA’s Office of Data Science and Informatics (ODSI) at NASA’s Marshall Space Flight Center in Huntsville, Alabama, led the development as part of the agency’s AI for Science initiative. 

According to Dr. Rahul Ramachandran, the AI for Science lead, Surya is the third foundation model his team has delivered. The group has developed a process to move from defining science requirements to releasing a working model in under a year, less than half the time of most traditional science projects.

"This success hinges on a highly collaborative team with complementary skills and a shared goal. This focus on valuable tools for science echoes a sentiment from atmospheric scientist Warren Wiscombe, who noted that while there is no shortage of ideas in science, there is a constant shortage of good tools. The history of science shows that great leaps forward are often enabled by the creation of new tools,” said Ramachandran. “I am immensely proud of our team's contributions, as we are acting as pathfinders, creating new ways to apply NASA data and AI to build broadly useful models. Releases like this are personally deeply rewarding because they represent not just an idea, but a tangible, collaborative tool that can advance our understanding of the Sun.”

Partnering for Innovation

Surya was co-developed with IBM Research, extending a partnership that began with the Prithvi family of foundation models in 2023.

“With this new model, we continue that journey, now turning our focus to the Sun,” said Juan Bernabe-Moreno, Director IBM Research Europe - Ireland in Dublin, Ireland. “It’s the same spirit of embracing bold challenges, breaking new ground, and pushing technology forward to advance science.”

Image

The Challenge of Modeling the Sun

Modeling the Sun is not like modeling the Earth. “Things happen at many different sizes and durations all at once,” said Andrés Muñoz-Jaramillo, project scientist for Surya at the Southwest Research Institute in Boulder, Colorado. “In the past, we had to break the system into smaller pieces. AI lets us capture more of it at once, but that creates new scientific and engineering challenges.”

The team designed Surya to combine spectral block layers with a long-short transformer backbone, enabling it to detect both broad, long-term solar patterns and fleeting, subtle changes like emerging sunspots. Overcoming memory constraints and blending frequency-aware with time-series modeling were crucial to enabling this breadth.

Surya’s development demanded creativity. “There was no playbook for building a foundation model for heliophysics,” said Johannes Schmude, Surya project co-lead at IBM. “We had to make quite a few choices about pretraining paradigms and model architectures while clearly being a newcomer to the domain and its intricacies. Fortunately we were able to work with a wonderful team of heliophysicists who helped us focus on the most pressing questions while putting our results into context. It's these interdisciplinary interactions that made this collaboration so rewarding and successful.”

“Working with SDO was exhilarating and, at times, frustrating,” said Sujit Roy, Surya’s project co-lead for NASA IMPACT. “The Sun’s dynamic nature and SDO’s extreme resolution demanded careful design under high memory constraints. We had to blend techniques from frequency-aware and time-series modeling so Surya could learn both fine spatial detail and long-term temporal dynamics. It carries the imprint of late nights, stubborn bugs, and weekend experiments.”

Surya uses ground-truth SDO data to forecast solar evolution. These images compare the ground-truth data (right) with model output (center) for solar flares which are the events behind most space weather. Preliminary results suggest that Surya has learned enough solar physics to predict the structure and evolution of a solar flare by looking at its beginning phase (left).

Image

Surya was able to reproduce the St. Patrick’s Day geomagnetic storm that occurred on March 17, 2015. Surya’s output (shown below) captures the coronal mass ejection that occurred around 2:10. The figure below shows model inputs and outputs for the 131 Angstrom band. The first two images are model input (2015-03-14 23:12 and 2015-03-15 0:12); the latter two are model outputs (2015-03-15 1:12 and 2015-03-15 2:12). The coronal mass ejection event occurred around 2:10. The output of Surya captures this in the brightening in the rightmost image.

Image

Science Use Cases for Validation

Surya’s versatility was tested in four core research applications:

• Active Region Emergence Forecasting: A 24-hour generative rollout to predict the intensity and spatial distribution of emerging solar regions, which are often precursors to major flares.
• Solar Flare Forecasting: A classification task that predicts the likelihood of strong flares (M- or X-class) in the upcoming 24 hours. Surya outperformed state-of-the-art models by up to 16 percent.
• Solar Wind Speed Prediction: A forecasting task using multi-channel SDO data to predict solar wind speeds up to 4 days in advance, a crucial task for satellite operations and grid resilience.
• EUV Spectra Prediction: Forecasting irradiance across 1,343 EUV channels to better understand the Sun’s energy output and its impact on Earth’s upper atmosphere.

“One of our main goals is to elevate the entire heliophysics community into an age of AI assisted scientific progress,’ Muñoz-Jaramillo said. “Anyone in our community can find an example of how to use Surya that they can adapt to their science.”

It Took a Village

The Surya project also brought together experts from NASA’s Goddard Space Flight Center in Greenbelt, Maryland, the Jet Propulsion Laboratory in Pasadena, California, the Southwest Research Institute in San Antonio, Texas, the SETI Institute Mountain View, California, multiple universities, and industry. The team included:

Sujit Roy, Johannes Schmude, Marcus Freitag, Julian Kuehnert, Theodore Van Kessel, Johannes Jakubik, Etienne Vos, Joao Lucas de Sousa Almeida, Campbell Watson, Juan Bernabe-Moreno, Vishal Gaur, Rohit Lal, Dinesha V. Hegde, Amy Lin, Talwinder Singh, Berkay Aydin, Andrés Muñoz-Jaramillo, Vishal Upendran, Daniel da Silva, Shah Bahaudding, Spiridon Kasapis, Kang Yang, Chetraj Pandey, Jinsu Hong, Nikolai Pogorelov, Ryan McGranaghan, Manil Maskey, and Rahul Ramachandran.

Training was supported by the National Science Foundation’s National AI Research Resource (NAIRR) in partnership with NVIDIA in Santa Clara, California.

An Open Invitation

"In line with our commitment to promoting open science, Hugging Face is pleased to make the Surya Heliophysics Foundation Model openly available to the community on our platform,” said Clem Delangue, chief executive officer of Hugging Face, in New York City, New York.

By sharing both the model and the data, the team hopes to speed the adoption of AI tools in heliophysics and inspire new approaches in other scientific domains. 
“Surya shows what’s possible when you combine cutting-edge AI with a mission-driven team,” Ramachandran said. “This is just the beginning.”