Identifying preflare spectral features using explainable artificial intelligence

The prediction of solar flares is practical and scientific interest; however, many machine learning methods used for this task do not provide the physical explanations behind a model's performance. We made use two recently developed explainable artificial intelligence techniques called gradient-weighted class activation mapping (Grad-CAM) expected gradients (EG) to reveal decision-making process high-performance neural network that has been trained distinguish between MgII spectra derived from flaring nonflaring active regions, fact can be applied short timescale flare forecasting. generate visual (heatmaps) projected back onto spectra, allowing identification features are strongly associated with precursory activity. automated search interpretations on level individual wavelengths, multiple examples using IRIS spectral data, finding scores in general increase before onset. Large rasters cover significant portion region coincide small preflare brightenings both SDO/AIA images tend lead better forecasts. models triplet emission, flows, as well broad highly asymmetric all important prediction. Additionally, we find intensity only weakly correlated spectrum's score, meaning low still great importance task, $78$% time, position maximum attention along slit during phase predictive location flare's UV emission