44th COSPAR Scientific Assembly. Held 16-24 July

2022.

Blumenau M., Vorobev D., Fridman M.

We propose a method for recognizing several types of magneto-plasma structures at the Sun, employing deep machine learning. Various neural networks (namely, fully connected, recurrent and convolutional networks) have increasingly become popular to many fields of physics involving data analysis and pattern recognition [Krizhevsky et al., 2012, NeurIPS]. Meanwhile, in solar physics, traditional algorithms relying on the pixel intensity, the optical flow, and other similar techniques are mainly used with few exceptions [e.g. Illarionov et al., 2018, MNRAS; Mackovjak et al., 2021, MNRAS]. Our task is to simultaneously detect several types of magneto-plasma structures observed in the solar corona (in particular, prominences without rotation, prominences with rotation - magnetic tornadoes, coronal holes, and active regions). For magnetic tornadoes, evolution over time is important, so it is reasonable to use recurrent layers for building a model. Coronal holes can be detected with convolutional networks. For simultaneous detection of several objects of different types, we propose to combine specialized neural networks into one pipeline. Sequences of SDO/AIA images obtained at different wavelengths and available from open-access databases have become our model inputs. The events previously discussed in the literature are used as training objects. The method makes it possible to detect coronal holes, prominences, and tornadoes automatically. Examples of the results obtained will be shown and discussed.

Keywords: Neural networks, solar physics, solar magnetic tornadoes