Neural-network-derived Stellar Parameters and Distances for Cool Giants

Abstract

In this work, we employ a multilayer perceptron to derive atmospheric parameters (effective temperature Teff, surface gravity logg , and metallicity [M/H]) for a large sample of cool giants from LAMOST DR9 low-resolution spectra. For spectra with a median signal-to-noise ratio ∼ 105, the achieved internal precisions are 36 K, 0.22 dex, and 0.16 dex for Teff, logg , and [M/H], respectively. By integrating these parameters with multiband photometry from the Two Micron All Sky Survey and the Wide-field Infrared Survey Explorer, we estimate stellar distances within a Bayesian framework incorporating 3D extinction maps. Our distance estimates exhibit a remarkably stable fractional internal uncertainty (σD/D) of ∼8.4% across the sampled parameter space. External validations against Gaia EDR3 Bayesian geometric distances show a median distance bias of only 0.1%, with a 16th–84th percentile dispersion of −21.8%–28.9%. Furthermore, comparisons with open cluster distances confirm that systematic deviations remain within 10% for stars with Teff ∼ 3700–4100 K. Applying these high-precision distances and velocities, we characterize the spatial distribution and kinematics of the Galactic disk. Our results demonstrate that cool giants are powerful tracers for the outer disk, effectively revealing its complex asymmetric structures and kinematic substructures.