Identifying preflare spectral features using explainable artificial intelligence

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