An Efficient Framework for Automatic Carbon Star Detection Using Machine Learning Techniques on LAMOST DR9 and Gaia DR3

Abstract Carbon stars are characterized by carbon-rich atmospheres. They play a key role in stellar evolution, enrichment of the interstellar medium, and distance estimation. Therefore, assembling large samples of these stars is essential. Machine learning has become one of the most effective tools for identifying stars of specific types in large-scale surveys; however, its application to carbon stars remains challenging due to the small number of confirmed objects available for developing and validating such methods. In this work, we present a methodology for constructing a carbon star detection model, a set of algorithms capable of automatically identifying patterns in data. Our adopted model is both computationally efficient and accurate. The approach begins with the selection of features based on the C 2 and CN spectral indices measured from LAMOST DR9 spectra combined with absolute magnitudes and colors from Gaia DR3. The model is trained and evaluated using the subset of sources with data available from both surveys. This procedure enables reliable identification of carbon stars, even considering the large disparity in numbers between carbon and non-carbon stars. Our model reached over 97% performance in identifying known carbon stars, demonstrating strong consistency. When applied to the full LAMOST data set, it identified 6372 carbon-star candidates, of which 5665 were not previously reported. The model’s efficiency and reliability in recognizing carbon stars within large data sets were further supported by visual inspection of candidate spectra in the Orion region, where more than 94% of the model-identified stars exhibit clear spectral features consistent with carbon-rich atmospheres.