Identifying Ultracool Binary Systems using Machine Learning Methods

Abstract

Abstract Ultracool spectral binaries are unresolved pairs of low-mass stars and brown dwarfs revealed by peculiarities in their combined light spectra. Methods to identify these systems have relied on spectral indices, which have known selection biases. We report on a pilot study examining the application of machine learning methods to identify ultracool spectral binaries. Using a sample of single and binary templates constructed from low-resolution, near-infrared spectra, we trained a random forest model to identify binaries composed of M7–L7 primaries and T1–T8 secondaries. We find that uniform data preparation and balancing of the training sample are critical to building an effective model. Our model achieves precisions of ≥95%, confirms known spectral binaries, and identifies new spectral features sensitive to multiplicity, illustrating the utility of machine learning methods to identify these rare systems.