Identification of Ap candidates from LAMOST DR12 based on deep learning

ABSTRACT Ap stars are a subclass of early-type chemically peculiar stars characterized by abnormal surface elemental abundances and strong, global magnetic fields. These features make them valuable tracers for studying stellar magnetic field origins and evolution. In recent years, the identification of Ap stars has become a research hotspot. We present a novel method using low-resolution Large Sky Area Multi-Object Fiber Spectroscopic Telescope (LAMOST) Data Release 12 (DR12) spectra. We developed ECACNet, a deep neural network incorporating the ECA (efficient channel attention) mechanism to enhance sensitivity to key spectral features of magnetic Ap stars. On a balanced data set covering seven stellar classes, ECACNet achieved 96.80 per cent precision and 94.05 per cent recall. Evaluated on seven independent validation sets, it maintained robust performance (96.13 per cent precision and 92.03 per cent recall), significantly outperforming conventional methods. Applying the model to LAMOST DR12 B- to F-type spectra identified 8130 high-confidence spectra. Visual inspection confirmed 4826 unique Ap candidates via the characteristic 5200 Å flux depression and metallic lines. Cross-matching with eight catalogues revealed 4032 known and 794 new candidates. mkclassrefined the new candidates’ spectral classifications. Statistical analyses of atmospheric parameters ($T_{\mathrm{eff}}$, $\log g$, and [Fe/H]) for the new candidates showed distributions consistent with known Ap stars, further validating our method’s reliability.