Early recognition of Microlensing Events from Archival Photometry with Machine Learning Methods

TL;DR

This paper develops a machine learning-based method to distinguish microlensing events from other variable stars using archival multi-wavelength photometry, achieving 90% accuracy in identifying true microlensing events.

Contribution

It introduces a supervised machine learning approach trained on multi-wavelength data to classify microlensing candidates solely from archival photometry, improving early detection accuracy.

Findings

Correctly identified one-third of known microlensing events at 90% probability threshold.

Classified 38 microlensing candidates out of 368 Gaia alerts at 90% threshold.

Demonstrated universal applicability for current and future transient alert surveys.

Abstract

Gravitational microlensing method is a powerful method to detect isolated black holes in the Milky Way. During a microlensing event brightness of the source increases and this feature is used by many photometric surveys to alert on potential events. A typical microlensing event shows a characteristic light curve, however, some outbursting variable stars may show similar light curves to microlensing events especially when the cadence of observations is not dense enough. Our aim is to device a method for distinguishing candidates for microlensing events from any other types of alerts using solely their archival photometric multi-wavelength data. The most common contaminants in the microlensing event search are Classical Be-type stars, Young Stellar Objects and Asymptotic Giant Branch stars such as Miras. We build a training set using thousands of examples for the main classes of alerting stars combining optical to mid-infrared magnitudes from Gaia, 2MASS and AllWISE catalogues. We used supervised machine learning techniques to build models for classification of alerts. We verified our method on 120 microlensing events reported by Gaia Science Alerts which were studied spectroscopically and photometrically. With the use of only archival information at 90% probability threshold we correctly identified one-third of the microlensing events. We also run our classifier on positions of 368 Gaia alerts which were flagged as potential candidates for microlensing events. At the 90% probability threshold we classified 38 microlensing events and 29 other types of variables. The machine learning supported method we developed can be universally used for current and future alerting surveys in order to quickly assess the classification of galactic transients and help decide on further follow-up observations.