Carbon Star Identification in LAMOST DR12 Based on a Multiscale Global and Local Spectral Feature Fusion Model

Abstract Carbon stars are important kinematic tracers for investigating the structure and evolutionary history of the Milky Way. Therefore, the efficient identification of carbon stars from large-scale spectroscopic surveys is of great scientific importance. In this work, we propose a carbon star identification model, CarbonNet, which fuses multiscale global and local spectral features. The proposed model achieves a precision of 99.80% and a recall of 98.80% on the carbon star test set. When evaluated on an independent validation set, it maintains robust performance, with a precision of 99.02% and a recall of 93.15%. Subsequently, the trained model was applied to LAMOST DR12, and through visual inspection, we identified a total of 5369 carbon star candidates. These candidates are further classified into 2137 Ba-type, 1510 CH-type, 384 CN-type, and 1195 CR-type carbon stars, while an additional 143 stars are assigned to an “Unknown” category due to their relatively low signal-to-noise ratios. Based on this sample, we investigate the Galactic distributions of different carbon star subtypes. Ba stars and CN- and CR-type carbon stars are mainly concentrated at low Galactic latitudes (∣ b ∣ ≤ 30°), whereas CH-type stars preferentially occur at higher latitudes, indicating distinct formation channels and stellar populations.