Machine Learning Approaches for Predicting the 10.7 cm Radio Flux from Solar Magnetogram Data

Using solar magetogram data, we explore potential of machine learning in space weather forecasting. In particular, unsupervised and supervised machine learning techniques are used to investigate the structure of magnetograms for 2006-2018, and their relation with the 10.7 cm solar radio flux. The similarity structure of the magnetograms is characterized with perception-based state of the art measures (the MSSIM index) and it was found that the data are contained in a space of intrinsically low dimension. The properties of these spaces were explored with methods preserving both local dissimilarity relationships, as well as conditional probability distributions within neighbourhoods. They reveal a clear relation with the intensity of the 10.7 cm flux. The flux was modeled using data driven supervised approaches in the form of model trees and convolutional neural networks. Models were found that allow prediction of the 10.7 cm radio flux with high accuracy. The results demonstrate significant potential which machine learning has in the space weather field.