Model-driven Stellar Spectroscopy with the Total Differential Spectral Model ( TDSM ): Framework and Its Application to 11 Million LAMOST Low-resolution Spectra

Abstract Deriving stellar astrophysical parameters from large-scale spectroscopic surveys requires accurate modeling of spectra across a wide wavelength coverage and a broad range of stellar parameter space, which remains a great challenge. In this work, we present the Total Differential Spectral Model ( TDSM ), a physically motivated neural network architecture that self-consistently emulates both the flux and its first-order derivatives of the ab initio stellar spectra. The method enables accurate and precise emulation of Kurucz synthetic spectra with severely blended features at R ≃ 2000 in a high-dimensional parameter space, achieving a typical pixel-wise flux precision of 0.05% in the wavelength range 3700 < λ < 9000 Å over the parameter ranges of 3000 < T eff < 45,000 K, − 1 < log g < 5 , and −4.0 < [Fe/H] < 0.5. The model is applied to LAMOST DR10 to derive stellar parameters for 11 million spectra spanning the full model parameter range. We conduct detailed refinement and calibration efforts to mitigate systematic errors, which are revealed to be significant and complex due to imperfections in the ab initio model atmosphere. The resultant catalog provides eight stellar labels— T eff , log g , v mic , [Fe/H], [ α /Fe], [C/Fe], [N/Fe], and [Mg/Fe]—for eight million unique stars (11 million spectra including repeat observations) across the full parameter space. This adds a valuable dataset for studies of Galactic structure and archaeology, stellar populations, and related areas of astrophysics.