Inferred stellar parameters from 220 million XP spectra using an empirical forward model


Gaia data release 3 (GDR3) provides a trove of 220 million low-resolution BP/RP (XP) flux-calibrated spectra. In order to properly exploit the XP spectra, we need to accurately model them as a function of stellar type, extinction and distance. We build a machine-learning model based on the subset of XP stars with LAMOST LRS counterparts, whose stellar types are precisely determined by the higher-resolution spectroscopy. We apply our learned forward model to fit the XP low-resolution spectra of all 220 million sources, and determine their stellar types, extinctions and parallaxes. The typical uncertainty of our catalog is 90 K in T_eff, 0.15 dex in [Fe/H], 0.15 dex in log g and 0.03 mag in extinction. In the next version of our catalog (under development), we include APOGEE, LAMOST MRS and GALAH into the training set for better coverage of stellar parameter space, and take variations in the extinction curve (i.e., dust R_V) into account. We also expand our forward model to include latent variables, which encode unidentified stellar parameters and measurement systematics. Finally we present a catalog of (T_eff, [Fe/H], [alpha/Fe], log g, R_V, E, omega) for 220 million XP stars. Our catalogs build a foundation for the study of stellar populations, a 3D map of dust density and a 3D R_V map in the Milky Way.

Xiangyu Zhang (张翔宇), Max Planck Institute for Astronomy (MPIA)
KIAA 2nd meeting room
Friday, August 11, 2023 - 1:30PM to Friday, August 11, 2023 - 2:30PM