Astronomy and Astrophysics Seminars

Supervised convolutional neural networks for understanding solar flares

by Dr. Shamik Bhattacharjee (TIFR, Mumbai)

Monday, February 24, 2020 from to (Asia/Kolkata)
at D-406
Description
Solar flares are explosions in the solar atmosphere that release intense bursts of short-wavelength radiation and are capable of producing severe space-weather consequences. The exact processes that lead to flares are not fully known and therefore reliable forecasting of flares is challenging. In the first part of the talk, I will discuss how we trained convolutional neural networks (CNNs) to classify line-of-sight magnetograms of ARs into ARs producing at least one M- or X-class flare or as non-flaring. We find that flaring ARs remain in flare-productive states --- marked by recall > 60% with a peak of 80% --- days before and after flares. We use occlusion maps and statistical analysis to show that CNN pays attention to regions between the opposite polarities from ARs and the CNN output is dominantly decided by the total line-of-sight flux of ARs. In the second part of the talk, I will discuss how we trained a CNN to obtain vector-magnetic-field features of ARs, highly correlated with flaring activity, from line-of-sight magnetograms. We set up a regression problem for CNN to output AR features such as total unsigned flux, mean free energy, electric current helicity and flux near polarity inversion line for the input line-of-sight magnetogram. The trained CNN yields Pearson correlation ~95% for extensive AR features as well as features that depend on the mean free energy. The correlation is ~83% for flux near the polarity inversion line and ~60% for electric current helicity. Our results can be readily extended to line-of-sight magnetograms from earlier ground-and space-based observatories for calculating vector-magnetic-field features of ARs to advance understanding of flares.