Insight on Multimodal Solar Flare Forecast With Explainable Deep Learning

Solar Flares are sudden and violent release of magnetic energy from the Sun’s atmosphere in the form of electromagnetic radiation bursts. Being able to forecast Solar Flares accurately is essential to mitigate the risks associated with Space Weather, but it has been proven to be particularly challenging. Recently, Deep Learning Methods (DLMs) captured the interest of many re- searchers in the field and new data-driven approaches to the problem have been developed. With Explainability Methods (EMs) we can reverse this data-driven approach to gather insight on the actual physics of the events. We apply EMs to DLMs trained on Solar Corona images which perform as well as models trained on Solar Photosphere magnetograms alone to forecast Solar Flares in a 24h forecasting window. In these preliminary results we confirm that DL models efficiently learn known physical precursors, such as the presence of sigmoidal coronal structures, which are observational signatures of highly twisted and sheared magnetic fields in the Solar Corona. From this proof of concept we also verified that DL models trained on Solar Corona images can be efficient to forecast Flares near the Solar Limb, where it is harder for models using magnetograms alone. Model combining magnetograms and images of the different layers of the Solar Atmosphere seem to provide the best performances.