A Catalog of High-confidence Very Metal-poor Stars from DESI Based on a Deep Learning Method

Abstract Very metal-poor (VMP; [Fe/H] < −2) stars are critical tracers for understanding early star formation and Galactic chemical evolution. However, identifying these rare objects from the massive datasets generated by the Dark Energy Spectroscopic Instrument (DESI) presents significant challenges in efficiency and precision due to the scarcity of high-fidelity labels and low signal-to-noise ratios in the metal-poor regime. To address this, we propose a novel dual-model deep learning framework that integrates a 1D-ResNet binary classifier with a specialized parameter regression model. Leveraging a transfer learning strategy with high-quality labels from APOGEE and the Large Sky Area Multi-object Fiber Spectroscopic Telescope, we optimized the framework for DESI spectra. The classification model achieves an accuracy of 97.87%, while the regression model predicts stellar metallicity with an rms error of 0.093 dex. Applying this framework to the high-quality DESI-HQ-DATA dataset and applying a strict temperature cut ( T eff ≥ 4500 K) to avoid extrapolation in the cool dwarf regime, we constructed a highly purified catalog of 2569 high-confidence VMP candidates. Internal validation demonstrates a significant improvement in consistency between DESI pipeline measurements, and external crossmatching with Gaia XP spectra confirms the reliability of our metallicity estimates. Finally, compared to existing compilations, this work contributes 1377 new VMP candidates, providing a robust and statistically significant sample for future studies of the Galactic halo and ancient stellar populations.