A Machine Learning Approach to Identify Solar Stokes Profiles in Flaring and Non-Flaring Active Regions

Abstract:

Solar flares are explosive events on the surface of the Sun that release electromagnetic radiation, which can disrupt the earth’s atmosphere and cause havoc in our communication system. Models for flare forecasting use properties of active region (AR) magnetic fields as predictors of flare occurrence. The magnetic field properties are obtained using inversion models that decode the information contained in Stokes Profiles (SP) as the radiation passes through the solar atmosphere. The inversion techniques ignore the rich information contained in the SP since they tend to use line fitting methods and derive average magnetic field properties. The line parameters can give better information on the magnetic field complexity of the AR atmosphere. We apply a modified K-means clustering method to Hinode spectropolarimetric data to identify and classify the Stokes V profiles in flaring and non-flaring ARs. The modified K-means method leads to a stable result, in which random initializations converge to comparable clustering. The profiles which characterize the centroids of the clusters are used to calculate three-line parameters: the amplitude asymmetry, the area asymmetry (associated with the degree of non-linear polarization), and the percentage of atypical profiles inside and outside the polarity inversion lines (PIL). We find that the amplitude asymmetry is higher in non-flaring vs. flaring regions; the area asymmetry is greater in flaring ARs vs. non-flaring ARs, and inside the PIL vs. outside. Our results indicate that flaring ARs, harbor a higher percentage of atypical profiles compared to non-flaring ARs & outside the PIL. These results are compatible with those found using the individual pixel profiles in the calculations. They indicate that the three parameters can be used to distinguish flaring from non-flaring ARs.