The problem
TESS has measured the frequency of maximum oscillation power for about 158,000 red giants across the sky, and asteroseismology turns that into masses and ages. Ages put the events of Galactic history in sequence: the formation of the thick and thin discs, the merger of accreted systems such as Gaia-Enceladus, the migration of stars through the disc. Reading that history also needs each star's chemistry, and the high-resolution surveys cover few of these stars: about one in ten has APOGEE DR19 values that pass the quality cuts.
What JANUS does
JANUS (Joint Abundance Network with Uncertainty and Stratification) is an ensemble of neural networks built on conditional normalising flows, trained on APOGEE DR19. From the low-resolution Gaia BP/RP spectrum of a red giant, with its astrometry, photometry and extinction, it predicts a joint probability distribution over T_eff, log g, [M/H], [Fe/H], [α/M] and individual abundances. Each star also receives a novelty score and P(H), a calibrated probability that its four atmospheric parameters are jointly off the APOGEE values, so that a user can choose which stars to trust.
What the tests showed
On stars withheld from training the robust scatter about APOGEE is 32 K, 0.072 dex, 0.051 dex and 0.021 dex in the four parameters above, with calibrated uncertainties. Repeating every test with a model trained without the spectrum shows that the XP spectra carry information on [M/H], [α/M], carbon and nitrogen, and that JANUS separates the two α sequences of the disc, which the model without the spectrum merges.
Where it runs
Two FCT advanced-computing projects carry the work, both with me as principal investigator: 299,520 ARM core-hours and 1,500 A100 GPU-hours on Deucalion to run JANUS on the Gaia DR3 spectra of about 220 million stars, and 1,002,400 core-hours and 1,000 H100 GPU-hours on MareNostrum 5 for Gaia DR4. The catalogue is the chemistry layer of ArqueoGal and will be released through the project.