Analysis of stellar spectra with machine learning

Abstract: Researchers in the field of Galactic Archaeology have entered the era of industrial revolution. Upcoming surveys are planning on observing tens of millions of stars and high precision and accuracy must be ensured when deriving their stellar parameters and elemental abundances. Unconventional data-driven techniques hold the promise of efficiently dealing with these vast collections of data while still rendering results of astrophysical value. The Cannon is a supervised machine learning algorithm implemented to transfer stellar properties or labels from a dataset of reference to any desired collection of stars. In this thesis, The Cannon is trained on a set of synthetic spectra generated ab initio and applied to a sub-set of 1410 FGK-type stars from the Gaia-ESO Survey for a label space of high dimensionality (Teff , log g, vmic, v sin i and 16 [X/H] abundances, where X is Mg, Na, Ca, Sc, Si, V, Ti, Mn, Fe, Ni, Cr, Co, Ba, Eu, O and Al). The aforementioned synthetic training set does not represent a grid of synthetic spectra or a sub-sample of stars with well studied properties. Instead, we have designed a sophisticated training set predominantly based on the Bensby catalogue of 714 stars with well measured stellar parameters and elemental abundances. The Cannon is indeed very fast, taking an average time of 15 seconds to simultaneously fit 20 labels on one single spectrum after having trained on the model. It succeeds in recovering the Teff , log g and [Fe/H] stellar parameters with typical deviations of sigma_[Fe/H] = 0.08 dex, sigma_Teff = 88 K and sigma_log g = 0.14 dex in the label offsets with respect to the GES values, as well as determine 15 elemental abundances within a SNR range spanning from 10 to 300.