Before a sunspot darkens the visible surface of the Sun, the magnetic activity that creates it is already forming covertly. These early changes are very weak, but researchers say AI can now detect them hours before a new solar active region appears.
According to ITNA and citing SciTechDaily, a model called "EarlyDetect" has detected signs of forming solar active regions on average nearly nine hours before they emerge.
It is said that "EarlyDetect" looks for precursor patterns both in the Sun's acoustic activity and in its magnetic field - signals previously difficult for scientists to identify reliably.
Jonas Tirona, an undergraduate researcher at NGIT and the corresponding author of the study, says the most valuable outcome of this research is showing that machine learning can be used to predict when solar active regions will form in advance.
He says this early warning could help satellite communications companies or power grids prepare and prevent potential damage from solar storms.
Active regions are areas of intense magnetic activity where sunspots form. The formation of these regions occurs over several hours, while their full development may take from one to several days.
Before these regions reach the solar surface, the growing magnetic fields subtly alter the acoustic waves passing through the Sun.
Scientists can examine these oscillations using helioseismology, a method that reveals processes not directly observable by studying waves inside the Sun.
"EarlyDetect" identifies these signals by analyzing hourly acoustic power maps alongside solar magnetic field data obtained from NASA's SDO observatory. These acoustic maps come from measurements of sound waves recorded every forty-five seconds by the helioseismic and magnetic imager called "HMI", mounted on the "SDO" observatory.
Alexander Kosovichov, one of the lead researchers on the project, says the main difficulty is that an active region begins forming below the Sun's visible surface, where its magnetic structure is not directly observable.
He says, instead, researchers look for very small changes in the magnetic field and in the pattern of acoustic waves that continuously move within the Sun - a task he likens to detecting a slight change in the rhythm of a very loud orchestra.
"EarlyDetect" is built on a transformer architecture; the same broad type of AI technology used in large language models like "ChatGPT". The difference is that this model, rather than learning statistical patterns of language, seeks patterns in solar data that precede changes in solar activity.
When Tirona joined the project last year, the research group was using a data conditioning method to make important solar patterns more discernible to the model. But experiments showed the opposite effect: conditioning the data actually reduced prediction quality.
Kosovichov said this surprised them most because they had expected conditioning to help identify useful short-term patterns, but instead it removed those very weak oscillations that provided the earliest warnings.
Tirona also said this process was somewhat like noise removal that usually filters out loud sounds to reveal the overall trend better, but in this case it was harmful in nearly all instances, and the signals the conditioning removed were crucial for predicting the timing of active region formation.
Tirona concludes that he hopes the project will raise awareness of the role of machine learning in heliophysics, and it would be very interesting if one day a model like this could help forecast space weather events; although, he says they have not reached that point yet, it is an exciting step.