Finding White Dwarfs' Hidden Companions using an Unsupervised Machine Learning Technique

Abstract: White dwarfs (WD) with main-sequence (MS) companions are crucial probes of stellar evolution. However, due to the significant difference in their luminosities, the WD is often outshined by the MS star. The aim of this work is to find hidden companions in Gaia's sample of WD candidates. Our methodology involves applying an unsupervised machine learning algorithm for dimensionality reduction and clustering, known as Self-Organizing Map (SOM), to Gaia BP/RP (XP) spectra. This strategy allows us to naturally separate WDMS binaries from single WDs from the detection of subtle red flux excesses in the XP spectra that are indicative of low-mass MS companions. We validate our approach using confirmed WDMS binaries from the SDSS and LAMOST surveys, achieving a precision of $\sim 90\%$. We demonstrated that the luminosity of the faint companions in the missed systems is $\sim 50$ times lower than that of their WD primaries. Applying our SOM to 90,667 sources, we identify 993 WDMS candidates, 506 of which have not been previously reported in the literature. If confirmed, our sample will increase the known WDMS binaries by $20\%$. Additionally, we use the Virtual Observatory Spectral Energy Distribution Analyzer (VOSA) tool to refine and parameterize a ``golden sample'' of 136 WDMS binaries through multi-wavelength photometry and a two-body Spectral Energy Distribution fitting. These high-confidence WDMS binaries are composed by low-mass WDs ($\sim 0.42 M_{\odot}$), with cool MS companions ($\sim 2800$ K). Finally, 13 systems exhibit periodic variability consistent with eclipsing binaries, making them prime targets for further follow-up observations.