A Consensus-driven and Physically Validated Catalog of Metallic-line A-star Candidates from LAMOST DR12 Based on Physics-constrained Generative Augmentation with a High-confidence Core of 18,408 Objects

Abstract We present a consensus-driven and physically validated catalog of 36,226 metallic-line A-star (Am) candidates from LAMOST DR12, constructed by a multimethod framework with physics-constrained generative augmentation, and containing a high-confidence core of 18,408 objects. A high-purity benchmark of 12,352 spectra is first built through unanimous consensus among multiple classifiers, MKCLASS , and an empirical separation curve. Given the scarcity of high-quality Am spectra, we develop the Astronomical Contrastive Self-Attention Residual Generative Adversarial Network (ACSR-GAN) integrating 1D–2D spectral representations and self-attention, with contrastive learning to strengthen the learning of weak Fe-group features. A closed-loop “generation–physical filtering–retraining” scheme iteratively refines synthetic spectra under the guidance of the empirical separation curve. Classifiers trained with this augmented dataset show consistent performance improvement (random forest accuracy: 0.9784 → 0.9859; ensemble regression model accuracy: 0.8209 → 0.8944). Applying the enhanced classifiers to DR12 yields 36,836 initial candidates. After requiring at least two classifiers in agreement, we apply a physical confidence parameter C phys defined in the differential line-index plane. Objects with C phys > 0.2 constitute the final catalog of 36,226 candidates, within which 18,408 objects with C phys ≥ 0.8 form a high-confidence core exhibiting strong Ca depletion and Fe enhancement. This catalog provides a robust foundation for studying atomic diffusion and Galactic disk evolution.