AI and the World's Hard Problems · Astronomy & space · arXiv · 2026-10-21
J-PLUS: Identifying Outlier Spectral Energy Distributions in the Fourth Data Release Using Unsupervised Machine Learning
Selecting candidates from rare or underrepresented stellar populations, as well as novel objects, is essential for spectroscopic follow-up, so the authors applied an autoencoder-based anomaly detection technique to twelve-band photometric data from J-PLUS to identify spectral energy distributions that deviate from the dominant pattern. They then applied t-SNE to the colors of the outlier population and classified the outliers by combining information from the SIMBAD database and other astronomical surveys, characterizing well-defined islands in the t-SNE map corresponding to specific populations. The study identified a large sample of white dwarfs and low-mass stars, including a small subset of polluted white dwarfs, and reports potential candidates for spectroscopic follow-up: 19 carbon-rich stars, 11 cataclysmic variables, and 170 white-dwarf-main-sequence binaries, all previously unreported in the literature.