Predicting metallicities and carbon abundances from Gaia XP spectra for (carbon-enhanced) metal-poor stars

Carbon-rich (C-rich) stars can be found at all metallicities and evolutionary stages. They are often the result of mass-transfer from a companion, but some of the most metal-poor C-rich objects are likely carrying the imprint of the metal-free First Stars from birth. In this work, we employ a neural network to predict metallicities and carbon abundances for over 10 million stars with Gaia low-resolution XP spectra, down to [Fe/H] = -3.0 and up to [C/Fe] $\approx$ +2. We identify ~2000 high-confidence bright (G<16) carbon-enhanced metal-poor (CEMP) stars with [Fe/H] < -2.0 and [C/Fe] > +0.7. The majority of our C-rich candidates have [Fe/H] > -2.0 and are expected to be binary mass-transfer products, supported by high barium abundances in GALAH and/or their Gaia RUWE and radial velocity variations. We confirm previous findings of an increase in C-rich stars with decreasing metallicity, adopting a definition of $3\sigma$ outliers from the [C/Fe] distribution, although our frequency appears to flatten for -3.0 < [Fe/H] < -2.0 at a level of 6-7%. We also find that the fraction of C-rich stars is low among globular cluster stars (connected to their lower binary fraction), and that it decreases for field stars more tightly bound to the Milky Way. We interpret these last results as evidence that disrupted globular clusters contribute more in the inner Galaxy, supporting previous work. Homogeneous samples like these are key to understanding the full population properties of C-rich stars, and this is just the beginning.