Contributed Talk - Splinter Learning

Wednesday, 23 September 2020, 09:45   (virtual room G)

The RAdial Velocity Experiment: Parametrisation of RAVE spectra based on Convolutional Neural Networ

G. Guiglion, G. Matijevic, A. Queiroz, M. Valentini, M. Steinmetz.
Leibniz-Institut für Astrophysik Potsdam (AIP), Germany

In recent years, data-driven methods have started to play an important role in the field of astrophysics. In the context of large Milky Way spectroscopic surveys such as GALAH, APOGEE or RAVE, these tools are key in parameterizing millions of spectra in a short time. We show that a Convolutional-Neural-Network-based approach (CN) offers a unique way of combining spectroscopic, photometric and astrometric data smoothly. We adopted atmospheric parameters and chemical abundances from APOGEE DR16 for the training set labels, and used part of the intermediate-resolution RAVE DR6 spectra set (R=7500) overlapping with APOGEE DR16 data set. We derived precise atmospheric parameters and chemical abundances for more than 400000 RAVE spectra. Incorporating broad-band WISE and 2MASS photometry and Gaia DR2 photometry and parallaxes as an extra set of constraints allows us to improve the results drastically. We obtain very good consistency of surface gravity with the results from the Kepler mission. Scientific validation has been performed and shows robustness of the CNN output. The developed procedure gives good insights for the planned surveys such as Gaia RVS and 4MOST.