Deep Neural Networks Applied to Predictions of 4-Class Solar Flares and Eruptions

We developed a solar flare prediction model using deep neural networks named Deep Flare Net (DeFN), which is an operational model to predict flares occurring in the following 24 h after observation images. The solar flares are classified by GOES classes (X, M, C, etc.), and it is said that larger flares tend to trigger larger coronal mass ejections (CMEs). Here we developed a prediction model of 4-class of flares and eruptions, by extending DeFN model. The model can predict the occurrence probabilities of 4 classes of flares and eruptions occurring in the following 24 h. From 4$\times$10$^5$ observation images taken during 2010-2017 by Solar Dynamic Observatory, we automatically detected sunspots and calculated 79 features for each region, to which flare occurrence labels of X-, M-, C-class and CME were attached. We adopted the features used in Nishizuka et al. (2017, 2018), e.g., line-of-sight$/$vector magnetogram in the photosphere, coronal hot brightening at 131 Å (T$\ge$10$^7$ K) and the X-ray and 131 Å intensity data 1 and 2 h before an image. For operational evaluation, we divided the database into two by the chronological split. The DeFN model consists of deep multilayer neural networks, formed by adapting skip connections, batch normalizations and weighted cross entropy. To statistically predict 4-class of flares, the DeFN model was trained to optimize the skill score, i.e., the Gandin Murphy-Gerrity score (GMGS). As a result, we succeeded in predicting flares with GMGS=0.63. Furthermore, we applied the DeFN model to predict eruptions. In the case of CMEs, they occur not only in active regions with strong magnetic field but also in ones between strong and weak magnetic fields and in quiet regions. We found that these characteristics makes CME predictions more difficult than flare predictions. We also compared features effective for predicting solar flares and CMEs.