Long before dark sunspots appear on the surface of the Sun, a new active region — where powerful solar flares may occur — begins to show barely noticeable signs of its formation. Researchers now say that a new artificial intelligence model can detect these early signals and predict the appearance of solar active regions on average almost nine hours in advance.
New EarlyDetect AI Model
In a study published in the journal Journal of Geophysical Research: Machine Learning and Computation, a team of scientists led by the New Jersey Institute of Technology (NJIT) reports that an artificial intelligence model called EarlyDetect can identify warning signals of emerging active regions in the Sun’s acoustic activity and magnetic field. Until now, scientists had been unable to detect such signals. This was reported by Phys.org.
NJIT student researcher Jonas Tirona, the corresponding author of the study, developed the approach together with NJIT computer scientists and solar physicists, as well as colleagues from Princeton University and NASA’s Ames Research Center, using observational data from NASA’s Solar Dynamics Observatory (SDO).
Detecting Hidden Changes on the Sun
Active regions — areas of high magnetic intensity where sunspots form — begin to appear within several hours, while their full formation may take from one to several days.
As magnetic fields rise toward the Sun’s surface, they leave faint traces in acoustic waves that scientists can detect using helioseismology — the science that studies oscillations inside the Sun.
To detect these traces, the EarlyDetect model developed by the team analyzes hourly maps of acoustic power and magnetic-field measurements obtained from NASA’s Solar Dynamics Observatory. The acoustic maps are based on observations of sound waves recorded every 45 seconds by the Helioseismic and Magnetic Imager (HMI) aboard NASA’s SDO satellite.
“The main challenge is that an active region begins forming below the visible surface of the Sun, where we cannot directly observe the magnetic structure,” said Alexander Kosovichev, Distinguished Professor of Physics at NJIT and co-leader of the project. “Instead, we look for very subtle changes in the magnetic field and in the structure of acoustic waves that continuously travel through the Sun. It is more like detecting a barely noticeable change in rhythm in a very noisy orchestra.”
The Filter Turned Out to Be Harmful
The team’s model uses a Transformer architecture — the same type of artificial intelligence technology that underlies large language models such as ChatGPT. While those systems learn patterns in text, EarlyDetect learns patterns in solar observations to predict future changes in solar activity.
After joining the project last year, Tirona and the team discovered that a filtering method they had been using to help the AI model identify important patterns in the solar data was actually making its predictions worse.
“That surprised us the most,” Kosovichev said. “Initially, we expected it to help isolate useful short-term patterns. Instead, the filter smoothed out exactly the subtle fluctuations that provided the earliest warning. Scientists explain that it is something like a noise-reduction function. Usually it removes loud noise so that the overall trend is easier to see. But the signals the filter removed turned out to be genuinely important in helping the model predict when an active region would appear.”
Model Warning Time and Data on the Emergence of Active Regions on the Sun
After training on observational data from NASA’s SDO/HMI satellite, the researchers tested the EarlyDetect model on active regions that the model had never seen before.
The most effective version of the model detected warning signals on average 9.24 hours before the active regions became visible, outperforming both a standard Transformer model and the previous benchmark approach.
To help other researchers build on this work, the team also released the Solar Active Region Emergence Dataset (SolARED) — a publicly available collection of observations of emerging solar active regions gathered using the SDO satellite — as well as the Solar Active Region (SAR) portal, an interactive web platform for exploring the data. This is the first publicly available dataset on the emergence of active regions on the Sun. It serves as a shared resource for both the machine-learning community and heliophysicists, making it possible to develop and test new forecasting methods.
Need for Further Testing of the AI Model
Although the system looks promising, Tirona noted that EarlyDetect is not yet ready for real-time forecasting. The model was trained on known active-region emergence events and still occasionally produces false alarms or delayed predictions.
A warning about the emergence of an active region is also not a prediction that a flare or coronal mass ejection will occur — many active regions never produce major eruptions. The team notes that the approach still needs to be tested on a much larger number of solar events.
“I hope this project helps broaden awareness of how machine learning can contribute to the development of heliophysics,” Tirona said. “It would be really cool if a model like this could one day help forecast space-weather events. We are not there yet, but this is an exciting step.”
