An algorithmic view of solar flares

Published online 22 May 2019

Machine learning offers insight into the dynamics of solar flares

Sedeer el-Showk

NASA's Solar Dynamics Observatory captured this image of a solar flare – as seen in the bright flash on the right side – on Sept. 10, 2017.
NASA's Solar Dynamics Observatory captured this image of a solar flare – as seen in the bright flash on the right side – on Sept. 10, 2017.
A machine learning approach has refined understanding of the dynamics behind solar flares, improving the prospect of forecasting major flares in time to avoid damage.

The intense X-ray radiation released by solar flares can affect GPS and telecommunication equipment, as well as pose a hazard to astronauts and high altitude flight crews. Energetic, charged particles and coronal mass ejections associated with flares can also damage satellites and power grids. A major solar flare in March 1989 interrupted communication and control of several satellites, disrupted radio transmissions, and tripped the circuit breakers in Hydro-Québec’s power grid, leaving millions without electricity.

Researchers in India, the United Arab Emirates, and the United States examined the Sun’s active regions: magnetic features that sometimes produce solar flares. They used machine learning to characterise magnetic field features in active regions before and after major flares to try to identify patterns that could improve understanding and prediction of flares.

The analysis included active regions between 2010 and 2016, some of which served as training data and the remainder to test the algorithm’s performance. Once trained, the algorithm correctly identified 75% of the active regions that would flare, and 89% of the regions that would not. When fed a time series of observations, the algorithm could identify 60% of flaring regions days before and after the event, and the rate increased to 91% in the 24 hours before the flare.

The team analyzed which features the algorithm used to determine whether an active region would flare found that the main predictors were those depending on the size of the active region, such as total magnetic flux.

The work “highlights the flare precursors that are useful for the design of robust forecasting systems that can provide at least a few hours of warning time,” says Dattaraj Dhuri of India’s Tata Institute of Fundamental Research, and the study’s lead author. “It is important to develop protocols to take necessary action when we know that a solar storm is imminent,” he adds.

This research follows several recent machine learning studies of solar flares. Eric Jonas, a University of California, Berkeley scientist who was involved in some of the earlier studies, praised the team’s care in separating the training and testing data. “The immediate impact will be more on scientific questions than flare prediction, but that's just because flare prediction is an incredibly hard problem,” he says.


Dattaraj, D. B. et al. Machine learning reveals systematic accumulation of electric current in lead-up to solar flares. PNAS (2019).