A research team led by the New Jersey Institute of Technology (NJIT) has released a new study, successfully developing a machine learning model called EarlyDetect that can identify precursors of active region formation from the acoustic activity and magnetic field changes of the Sun.

Active regions on the Sun are areas with strong magnetic fields, where sunspots typically form. The emergence of an active region may take only a few hours, but the full development process usually takes 1 to several days. Researchers hope to use earlier signals to provide supplementary information for space weather forecasting.

Using Transformer to Read SDO Data, Providing Early Warnings 9.24 Hours in Advance

EarlyDetect uses a Transformer architecture to process continuous solar observation data, analyzing data from the Helioseismic and Magnetic Imager (HMI) aboard NASA's Solar Dynamics Observatory (SDO). HMI records solar vibration observations every 45 seconds, and the research team used this to generate acoustic power maps, which were then combined with magnetic field measurements to predict continuous intensity changes. The best-performing version can identify signals indicating an active region is about to appear, on average, 9.24 hours in advance.

The researchers used the formation process of the solar active region AR11158 as a visualization case, observing the solar magnetic field, continuous intensity, and acoustic power simultaneously. They divided the target area into smaller blocks to continuously track changes. The results showed that acoustic power first decreased, followed by changes in continuous intensity and magnetic field, gradually making the active region visible. This suggests that when the Sun's internal magnetic field rises, it may leave faint traces in the propagation of sound waves first.

If the model is further validated, satellite communication companies and power grid operators may gain more time to adjust their operational strategies and assess the risk of solar storms. However, the research team emphasized that EarlyDetect is not yet in operation.