AI Model Detects Solar Active Regions 9 Hours Early
The EarlyDetect model from the New Jersey Institute of Technology spots active regions on the Sun an average of 9.24 hours before they become visible, using NASA Solar Dynamics Observatory data.
A team of researchers at the New Jersey Institute of Technology unveiled an artificial intelligence model on August 14 that can tell an active region is forming on the Sun an average of 9.24 hours before it becomes visible. The paper appeared in the Journal of Geophysical Research: Machine Learning and Computation, under DOI 10.1029/2025JH001207, and the model is called EarlyDetect. Its input data comes from NASA’s Solar Dynamics Observatory.
Active regions are magnetically charged zones where sunspots form and where solar flares and coronal mass ejections originate. Their formation begins beneath the Sun’s visible surface, where the magnetic structure cannot be observed directly, so the event is normally registered only once sunspots reach the surface. The emergence itself takes a few hours, while a region’s full development can take anywhere from one to several days.
As the magnetic field rises toward the surface, it leaves a faint trace in the sound waves that constantly travel through the Sun. Those vibrations are tracked through helioseismology. Alexander Kosovichev, a physics professor at the institute and a coauthor of the paper, compared the task to picking out a small change in rhythm inside a very noisy orchestra.
The practical value of such a forecast was described by the paper’s lead author, Jonas Tirona, an undergraduate computer science student in his final year, in an institute statement:
Such an early warning could allow satellite communications companies, or companies that operate power grids, to prepare and potentially mitigate the damage from solar storms.
Satellite link and power grid operators, then, are the users the authors are counting on. Reaching that application, however, still takes more work.
How the Model Looks Beneath the Sun’s Surface
EarlyDetect runs on a transformer architecture, the same type that powers large language models, except that instead of learning patterns in text, it learns patterns in data about the Sun. The data comes from an instrument at the observatory that measures helioseismology and magnetic field strength. Acoustic readings are logged every 45 seconds, and from them the model builds maps of acoustic power that it compares against magnetic field measurements and surface brightness.
The model was trained on known cases of active regions forming, then tested on regions it had never seen before. Its best version picked up early signals an average of 9.24 hours before a region became visible, outperforming both a standard transformer model and an earlier benchmark approach.
One finding caught the team off guard. Filtering that was supposed to isolate short-term patterns actually made the predictions worse, so results improved once the filters were removed.
“It’s kind of like noise cancellation, which usually removes loud sounds so you can see the general trend better,” Tirona said. What the filter had been stripping out were precisely the faint fluctuations that carry the earliest signal.
Alongside the paper, the team released a database of active region emergences on the Sun, compiled from observations by the same observatory, along with an interactive web portal for browsing the data.
“This is the first public database on the emergence of active regions on the Sun,” said Mengjia Xu, an assistant professor of data science at the institute and the project’s principal investigator. Collaborators from Princeton University and NASA’s Ames Research Center also worked on the project.
What Doesn’t This Forecast Prove?
The model is not yet ready for live operation. It occasionally produces false alarms and forecasts that arrive too late, and it was trained on already known events, which is not the same as working on a raw, real time data stream. The authors say the approach still needs to be tested on a much larger number of solar events.
A second limitation lies in the task’s own definition. A forecast that an active region is emerging is not a forecast of a flare or a coronal mass ejection. Many active regions never produce a major eruption, so a nine hour heads up means knowing where to look, not knowing what will happen.
