From synthetic SEDs to stellar origins: A deep learning model for physical parameter retrieval in hot subdwarf stars

The formation mechanisms of spectrally diverse hot subdwarfs remain unclear. While existing mass distribution analyses suggest additional channels beyond helium white dwarf (He-WD) mergers contribute to He-rich subdwarf formation, these conclusions are constrained by the limited sample sizes of mass-measured He-rich objects. We developed a deep learning model that combines a convolutional neural network (CNN) with a squeeze-and-excitation (SE) block to calculate synthetic spectral energy distributions (SEDs) for 1,012 spectroscopically confirmed hot subdwarfs. By directly comparing synthetic SEDs and the observed flux density, we derived stellar parameters (mass, radius, and luminosity) for an unprecedented number of hot subdwarf stars, enabling more conclusive channel discrimination than prior studies. The mass distribution of sdB/sdOB stars confirmed the results from model predictions of binary population synthesis (BPS). A primary and secondary peak (i.e., around 0.56 and 0.4 $ ̊m M _⊙$) is obviously presented in the mass distribution of He-rich hot subdwarf stars. By comparing this with the results from the predictions of the recent BPS model, we propose that the merger of two He-WDs could produce most of the observed He-rich hot subdwarf stars, but the mass transfer during the stable Roche lobe overflow phase in binary evolution should be partially conserved.