Finding flares in Kepler data using machine learning tools

TL;DR

This paper presents a machine learning-based method using RANSAC to automatically identify stellar flares in Kepler data, overcoming challenges of traditional analysis for long-term, multi-target observations.

Contribution

Introduces a novel automated flare detection code employing RANSAC and a voting system, improving robustness and reducing false positives in large photometric datasets.

Findings

Successfully detected flares consistent with manual analysis

Effective in handling outliers with RANSAC modeling

Applicable to long-term, multi-target stellar observations

Abstract

Archives of long photometric surveys, like the Kepler database, are a gold mine for studying flares. However, identifying them is a complex task; while in the case of single-target observations it can be easily done manually by visual inspection, this is nearly impossible for years-long time series for several thousand targets. Although there exist automated methods for this task, several problems are difficult (or impossible) to overcome with traditional fitting and analysis approaches. We introduce a code for identifying and analyzing flares based on machine learning methods, which are intrinsically adept at handling such data sets. We used the RANSAC (RANdom SAmple Consensus) algorithm to model light curves, as it yields robust fits even in case of several outliers, like flares. The light curve is divided into search windows, approximately in the order of the stellar rotation period. This search window is shifted over the data set, and a voting system is used to keep false positives to a minimum: only those flare candidate points are kept that were identified in several windows as a flare. The code was tested on the K2 observations of the TRAPPIST-1, and on the long cadence data of KIC 1722506. The detected flare events and flare energies are consistent with earlier results from manual inspections.