An Unsupervised Machine Learning Approach to Identify Spectral Energy Distribution Outliers: Application to the S-PLUS DR4 data

Identification of specific stellar populations using photometry for spectroscopic follow-up is a first step to confirm and better understand their nature. In this context, we present an unsupervised machine learning approach to identify candidates for spectroscopic follow-up using data from the Southern Photometric Local Universe Survey (S-PLUS). First, using an anomaly detection technique based on an autoencoder model, we select a large sample of objects ($\sim 19,000$) whose Spectral Energy Distribution (SED) is not well reconstructed by the model after training it on a well-behaved star sample. Then, we apply the t-distributed Stochastic Neighbor Embedding (t-SNE) algorithm to the 66 color measurements from S-PLUS, complemented by information from the SIMBAD database, to identify stellar populations. Our analysis reveals 69 carbon-rich star candidates that, based on their spatial and kinematic characteristics, may belong to the CH or Carbon-Enhanced Metal-Poor (CEMP) categories. Among these chemically peculiar candidates, we identify four as likely carbon dwarf stars. We show that it is feasible to identify three primary white dwarf (WD) populations: WDs with hydrogen-dominated atmospheres (DA), WDs with neutral helium-dominated atmospheres (DB), and the WDs main sequence binaries (WD + MS). Furthermore, by using eROSITA X-ray data, we also highlight the identification of candidates for very active low-mass stars. Finally, we identified a large number of binary systems using the autoencoder model, but did not observe a clear association between the overdensities in the t-SNE map and their orbital properties.