Applying Machine Learning \to Understanding the Physical Processes of Solar Flare Onset

Abstract:

We apply deep learning algorithms (Long Short Term Memory -LSTM) to train binary strong/weak classification models using active region parameters provided in HMI/Space-Weather HMI-Active Region Patch (SHARP) data files. We identify 35 active regions which show a sudden transition of the prediction score (above 0.7) indicating the likelihood of an M/X class flare several hours before the event. We examine the HMI vector magnetogram data for these events to determine if there are common circumstances and physical processes responsible for the magnetic energy and magnetic shear that informs the predictions. We further examine extrapolated nonlinear force-free coronal fields and Atmospheric Imaging Assembly (AIA) images to determine the structure of the coronal field and the associated electric currents and free magnetic energy increase prior to the flaring events. In many cases, we find development of strong horizontal fields running nearly parallel to the polarity inversion line to be a precursor to the flaring events.