Artificial intelligence can be used to detect coronal holes in the Sun’s upper atmosphere, an international research team has shown. Robert Jarolim at the University of Graz in Austria, Tatiana Podladchikova at Skoltech in Russia and colleagues have demonstrated a strong agreement between the holes identified by their convolutional neural network, and those picked up manually by astronomers. The system could lead to more reliable forecasting of disruptive space weather and an improved understanding of the Sun’s complex evolution.
When observed at extreme ultraviolet (EUV) wavelengths, holes can appear in the Sun’s corona – its upper atmosphere. These holes are cooler and less dense than surrounding material in the corona and comprise many smaller-scale magnetic funnels. These funnels are rooted deeper in the Sun, in the star’s light-radiating photosphere, and they extend far into interplanetary space. Along these magnetic field lines, solar plasma is rapidly accelerated away from the Sun, producing high-speed solar wind that can trigger powerful geomagnetic storms as they interact with Earth’s magnetosphere.
Currently, the shapes, sizes, and locations of coronal holes must be identified manually in EUV images gathered by space-based telescopes. This is a challenging process, both due to strong variations in the corona’s brightness throughout the Sun’s 11-year activity cycle, and because coronal holes can be difficult to differentiate from other dark features such as solar filaments.
Real-time maps
To address these shortcomings, Jarolim’s team developed a new artificial neural network called CHRONNOS, which can be trained to recognize the boundaries of coronal holes within images taken at several different EUV wavelengths. In addition, it can pick out the structures from real-time maps of the Sun’s magnetic field. By comparing these images, the artificial intelligence algorithm can then identify the boundaries of coronal holes by their intensities, shapes, and magnetic field properties.
Smartphones could create distributed space weather observatory
After training the neural network, the team used it to study 1700 EUV and magnetic field images of the Sun, taken by NASA’s Solar Dynamics Observatory between November 2010 and December 2016. Out of the 261 coronal holes identified manually by astronomers in the images, CHRONNOS picked up 256 – a 98.1% success rate. In addition, while the neural network performed best when the information from all EUV and magnetic images was combined, coronal holes could also be identified from magnetic field maps alone – which are far more difficult for humans to analyse.
The results showed that CHRONNOS could provide reliable detections of coronal holes, regardless of the level of solar activity. Over shorter timescales of days and weeks, the model even exceeded human performance in its consistency and reliability. Through future improvements, Jarolim’s team hope that their model could soon provide organizations that manage electrical and telecoms infrastructure with better warnings of damaging geomagnetic storms. CHRONNOS could also help astronomers to learn more about the long-term evolution of the Sun’s complex and ever-dynamic magnetic field.
The research is described in a paper to be published in Astronomy and Astrophysics and the video below is an animated representation of coronal holes detected by the system in solar observations taken over a period of almost 11 years.