Exploring Parametric Contributions to Solar Flare Forecasting with Machine Learning

Gupta, Om (2021) Exploring Parametric Contributions to Solar Flare Forecasting with Machine Learning. Masters thesis, Indian Institute of Science Education and Research Kolkata.

Text (MS Dissertation of Om Gupta (16MS102)) 16MS102_Thesis_file.pdf - Submitted Version Restricted to Repository staff only Download (1MB)

Abstract

Solar ares are huge explosions that occur in the solar atmosphere and are associated with regions of high magnetic activity called Solar Active Regions. They release huge quantities of radiation and energetic particles, which can severely affect Earth's spaceweather depending on the magnitude and the location of occurrence of the are. Severe spaceweather events have the potential to result in large-scale economic losses. Solar are forecasting with Machine Learning is an emerging field that exhibits great promise of delivering highly accurate forecasts of solar ares. It can help humanity in minimizing the losses from severe spaceweather events by providing suffciently early warnings. We conduct a comparative study of four popular and effective machine learning algorithms by training them to differentiate between pre- are observations of active regions that have the capability to produce large magnitude solar ares and observations of those active regions that do not possess this capability. The classifiers are provided with parametric inputs derived from the magnetic field maps of Solar Active Regions. We utilize are data from Hinode XRT are catalog between May 2010 to April 2020 and use active region observations obtained from the Helioseismic and Magnetic Imager instrument onboard NASA's Solar Dynamics Observatory satellite. We find that the classifier called Support Vector Machine (SVM) performs the best among the four algorithms in this classi- fication task. The SVM is used to probe individual parametric contributions to the classification task. High correlations are found among features that scale with the active region area, which also seemingly have great classification contributions. We hypothesize that these correlations exist due to the area dependency. Eliminating these correlations, we uncover features whose high contributions were originally being obscured by extensive features. The best features are finally determined to be the sum of unsigned magnetic ux near the active region Polarity Inversion Lines and the active region area. vii

Actions (login required)

View Item