Detecting quasi-periodic pulsations in solar and stellar flares with a neural network

Abstract

Quasi-periodic pulsations (QPP) are often detected in solar and stellar flare lightcurves. These events may contain valuable information about the underlying fundamental plasma dynamics as they are not described by the standard flare model. The detection of QPP signals in flare lightcurves is hindered by their intrinsically non-stationary nature, contamination by noise, and the continuously increasing amount of flare observations. Hence, the creation of automated techniques for QPP detection is imperative. We implemented the Fully Convolution Network (FCN) architecture to classify the flare lightcurves whether they have exponentially decaying harmonic QPP or not. To train the FCN, 90,000 synthetic flare lightcurves with and without QPP were generated. After training, it showed an accuracy of 87.2% on the synthetic test data and did not experience overfitting. To test the FCN performance on real data, we used the subset of stellar flare lightcurves observed by Kepler, with strong evidence of decaying QPP identified hitherto with other methods. Then, the FCN was applied to find QPPs in a larger-scale Kepler flare catalogue comprised of 2274 events, resulting in a 7% QPP detection rate with a probability above 95%. The FCN, implemented in Python, is accessible through a browser application with a user-friendly graphical interface and detailed installation and usage guide. The obtained results demonstrate that the developed FCN performs well and successfully detects exponentially decaying harmonic QPP in real flare data, and can be used as a tool for preliminary sifting of the QPP events of this type in future large-scale observational surveys.