Searching for Short M-Dwarf Flares by Machine Learning Method

We propose a machine learning method to identify M-dwarf flares in astronomical observation data. A flare is a sudden increase of luminosity of a star's surface, and is thought to be the result of magnetic reconnection. Observations of the stellar flares play a crucial role in understanding stellar magnetic activity. In particular, analyzing flare time evolution (light curve) is essential. We use the data from Tomo-e Gozen camera, mounted on the Kiso Schmit telescope, with a cadence of approximately one second, which is shorter than the cadence of other telescopes such as NASA's Kepler space telescope and the Transiting Exoplanet Survey Satellite. The dataset is ideal for identifying fast flares. We develop a one-dimensional convolutional neural network (CNN) to detect fast and faint flares in optical light curves. We train the model on a limited number of real flares identified by human experts, augmented with a large number of artificially generated flares to detect sub-minute flare candidates within the light curves captured by Tomo-e Gozen camera, and subsequently fit these candidates to make the selections. Our novel CNN model has successfully identified potential flares characterized by a rise time in the range of 4 s less than or similar to t(rise) less than or similar to 88 s, and energy levels spanning 10(30) erg less than or similar to E-flare less than or similar to 10(33) erg. Notably these potential flares exhibit shorter duration and lower energy compared to those detected by human experts, who typically identify flares with a rise time of 5 s less than or similar to t(rise) less than or similar to 100 s and energy of 10(31) erg less than or similar to E-flare less than or similar to 10(34) erg.