ROOSTER, a machine learning tool to determine stellar surface rotation periods in Kepler data

The Kepler and, to a lesser extent, the CoRoT missions opened the era of large-scale photometric stellar surveys with space instruments. Corotating dark spots and bright faculae on the stellar surface lead to brightness variations and therefore those long-term photometric surveys provide ideal datasets to measure stellar surface rotation periods and build stellar rotation catalogs. Such catalogs can then be used to constrain gyrochronology models or to study the interplay between rotation and magnetic activity. Taking the best possible advantage of those large-scale surveys, the main challenge that we face today is finding efficient methods to analyse the large amount of data. It has been shown that a combination of methods (auto-correlation function, time-frequency analysis) applied on light curves with different high-pass filtering provide reliable rotation estimates. However, the results yielded by those methods require a significant amount of visual inspection.

In the work presented here, random forest learning abilities are exploited to automate the extraction of rotation periods and magnetic activity index in Kepler light curves and to reduce the number of required visual inspections in the dataset. We train three different classifiers: one to detect if rotation modulations are present in the light curve, one to flag classical pulsators or close binary candidates that can bias our rotation-period determination, and finally one classifier to provide the final rotation period. We test our machine learning pipeline, ROOSTER (Breton et al. 2021), on the Kepler K and M dwarf sample using the reference catalog of Santos et al (2019). We show that we are able to detect rotation modulations with an accuracy of 94.2% and to retrieve final rotation periods with an accuracy of 95.3%. This value is raised to 99.5% after visually inspecting 25.2% of the stars. Over the two main analysis steps, the pipeline yields a global accuracy of 92.1% before visual checks, 96.9% after. The method is then applied to analyse the F and G Kepler sample (Santos et al. 2021). This allowed us to derive the largest catalog of surface rotation periods for the Kepler targets with more than 55,000 entries. The work we performed used only time series from the Kepler mission, but the methodology presented here could be adapted to extract surface rotation periods for stars observed by other missions, like K2, TESS, or PLATO.