Hunting for C-rich long-period variable stars in the Milky Way's bar-bulge using unsupervised classification of Gaia BP/RP spectra
The separation of oxygen- and carbon-rich AGB sources is crucial for their accurate use as local and cosmological distance and age/metallicity indicators. We investigate the use of unsupervised learning algorithms for classifying the chemistry of long-period variables from Gaia DR3's BP/RP spectra. Even in the presence of significant interstellar dust, the spectra separate into two groups attributable to O-rich and C-rich sources. Given these classifications, we utilise a supervised approach to separate O-rich and C-rich sources without BP/RP spectra but instead given broadband optical and infrared photometry finding a purity of our C-rich classifications of around $95$ per cent. We test and validate the classifications against other advocated colour-colour separations based on photometry. Furthermore, we demonstrate the potential of BP/RP spectra for finding S-type stars or those possibly symbiotic sources with strong emission lines. Although our classification suggests the Galactic bar-bulge is host to very few C-rich long-period variable stars, we do find a small fraction of C-rich stars with periods $>250\,\mathrm{day}$ that are spatially and kinematically consistent with bar-bulge membership. We argue the combination of the observed number, the spatial alignment, the kinematics and the period distribution disfavour young metal-poor star formation scenarios either in situ or in an accreted host, and instead, these stars are highly likely to be the result of binary evolution and the evolved versions of blue straggler stars already observed in the bar-bulge.