2025ApJS..278...41Z

This paper proposes a spectral transformer encoder (SpecTE) model based on a denoising pretraining technique, and applies it to estimating radial velocity (RV) and three stellar atmospheric physical parameters together with 15 chemical element abundances from low-resolution spectra from the Large Sky Area Multi-Object Fiber Spectroscopic Telescope (LAMOST). SpecTE enhances its sensitivity to spectral features and improves its parameter estimation accuracy by prelearning a mapping from low-quality spectra to high-quality spectra. We systematically evaluate the performance of SpecTE on the test set. On the cases with a signal-to-noise ratio of ​​​​​g ≥ 5, SpecTE achieved mean absolute error of 45 K for T_eff, 0.08 dex for $\mathrm{log}\,g$, 0.037-0.05 dex for Fe, Mg, Si, Ni, Ca, and C, 0.05-0.077 dex for Al, Mn, O, S, and K, 0.09-0.12 dex for Ti, N, and Cr, 0.15 dex for V, 0.21 dex for Na, and 3.87 km s^–1 for RV, respectively. Compared with StarNet, StarGRUNet, and SHBoost, SpecTE demonstrates significant advantages in both estimation accuracy and precision. Nevertheless, SpecTE also exhibits certain limitations, including relatively lower accuracy in the abundances of Na, V, and Cr. In addition, it shows some systematic biases when applied to metal-poor stars and stars with extreme elemental abundances. Finally, we estimated the parameters and their uncertainties for 9.80 million LAMOST DR11 low-resolution spectra, thereby compiling the LAMOST-SpecTE catalog. Compared with the DD-Payne and StarGRUNet catalogs, LAMOST-SpecTE exhibits higher accuracy and robustness. The computed catalog demonstrates good consistency with APOGEE and GALAH. The estimated stellar catalog, code, pretrained model, and experimental data are released online for astronomical exploration and algorithm research.