Seminar: Machine-learning for stellar spectroscopy: past, present and future

In this colloquium, I will present past and recent developments in the field of machine-learning applied to stellar spectra in the context of large scale spectroscopic surveys, such as Gaia-ESO and RAVE surveys. I also focus on Gaia DR3, which provided the community with one million RVS spectra covering the CaII triplet region. One third of the spectra have a signal-to-noise ratio from 15 to 25 per pixel. I will demonstrate that precise parametrization can be achieved for such a type of dataset by using machine-learning and the full Gaia data product. I will present a new approach in the form of a hybrid Convolutional Neural-Network (CNN) to derive atmospheric parameters (Teff, log(g), and [M/H]) and chemical abundances ([Fe/H] and [α/M]). Our CNN is designed to effectively combine the Gaia DR3 RVS spectra, photometry (G, Bp, Rp), parallaxes, and XP coefficients and is able to extract formation from non-spectral inputs to supplement the limited spectral coverage of the RVS spectrum. We manage to characterize the [α/M] − [M/H] bimodality from the innerregions to the outer part of the Milky Way, which has never been characterized using RVS spectra or similar datasets. I will also discuss on the benefits to use CNNs for future large scale spectroscopic surveys such as 4MOST.

-------

We shall leave for lunch at around 12:15. There will be fika after the seminar.