AI model forecasts solar active regions nine hours before they appear

Researchers at the New Jersey Institute of Technology (NJIT), with collaborators at Princeton University and NASA's Ames Research Center, have developed an artificial intelligence model that can forecast the emergence of solar active regions hours before they become visible on the Sun's surface. The model, named EarlyDetect, analyzes hourly acoustic power maps and magnetic field measurements from NASA's Solar Dynamics Observatory (SDO). In tests on previously unseen events, the best-performing version identified precursor signals an average of 9.24 hours in advance.

Active regions are magnetically intense areas where sunspots form and from which solar flares and coronal mass ejections can originate. As magnetic fields rise toward the visible surface, they leave faint signatures in the Sun's acoustic waves, detectable through helioseismology. EarlyDetect uses a Transformer architecture — the same type of neural network behind large language models — to learn patterns in these solar observations. The acoustic maps are derived from sound-wave data recorded every 45 seconds by the Helioseismic and Magnetic Imager aboard SDO.

The project was led by NJIT undergraduate researcher Jonas Tirona, an incoming senior computer science major, working with NJIT physics professor Alexander Kosovichev and data science professor Mengjia Xu. A key finding emerged when the team discovered that a filtering technique intended to isolate short-timescale patterns actually degraded forecasts by averaging away the very fluctuations that provided the earliest warning. Removing the filter improved performance across almost every test case.

The researchers have released the Solar Active Region Emergence Dataset (SolARED), a public collection of SDO observations compiled for this task, along with an interactive web portal for exploring the data. Xu described it as the first public dataset dedicated to solar active region emergence, intended to let both machine-learning and heliophysics communities develop and benchmark new prediction approaches.

EarlyDetect is not yet ready for operational forecasting. The model was trained on known emergence events and still produces occasional false alarms or late predictions. An emergence warning also does not guarantee that a flare or coronal mass ejection will follow, since many active regions never produce major eruptions. The team says further validation across many more solar events is needed before the system could support real-time space-weather alerts.