Hot subdwarfs are essential for understanding the structure and evolution of low-mass stars, binary systems, astroseismology, and atmospheric diffusion processes. In recent years, deep learning has driven significant progress in hot subdwarf searches. However, most approaches tend to focus on modelling with spectral data, which are inherently more costly and scarce compared to photometric data. To maximise the reliable candidates, we used Sloan Digital Sky Survey (SDSS) photometric images to construct a two-stage hot subdwarf search model called SwinBayesNet, which combines the Swin Transformer and Bayesian neural networks. This model not only provides classification results but also estimates uncertainty. As negative examples for the model, we selected five classes of stars prone to confusion with hot subdwarfs, including O-type stars, B-type stars, A-type stars, white dwarfs (WDs), and blue horizontal branch stars (BHBs). On the test set, the two-stage model achieved F1 scores of 0.90 and 0.89 in the two-class and three-class classification stages, respectively. Subsequently, with the help of $Gaia$ DR3, a large-scale candidate search was conducted in SDSS DR17. We found 6804 hot-subdwarf candidates, including 601 new discoveries. Based on this, we applied a model threshold of 0.95 and Bayesian uncertainty estimation for further screening, refining the candidates to 3413 high-confidence objects, which include 331 new discoveries.
Read full abstract