Real-time flare prediction based on distinctions between flaring and non-flaring active region spectra

With machine learning entering into the awareness of the heliophysics community, solar flare prediction has become a topic of increased interest. Although machine learning models have advanced with each successive publication, the input data has remained largely fixed on magnetic features. Despite this increased model complexity, results seem to indicate that photospheric magnetic field data alone may not be a wholly sufficient source of data for flare prediction. For the first time we have extended the study of flare prediction to spectral data. In this work, we use Deep Neural Networks to monitor the changes of several features derived from the strong resonant Mg II h&k lines observed by IRIS. The features in descending order of predictive capability are: The triplet emission at 2798.77 $text{AA}$, line core intensity, total continuum emission between the h&k line cores, the k/h ratio, line-width, followed by several other line features such as asymmetry and line center. Regions that are about to flare generate spectra which are distinguishable from non-flaring active region spectra. Our algorithm can correctly identify pre-flare spectra approximately 35 minutes before the start of the flare, with an AUC of 86 % and an accuracy, precision and recall of 80 %. The accuracy and AUC monotonically increases to 90 % and 97 % respectively as we move closer in time to the start of the flare. Our study indicates that spectral data alone can lead to good predictive models and should be considered as an additional source of information alongside photospheric magnetograms.