Machine learning for galactic archaeology: A chemistry-based neural network method for identification of accreted disc stars

We develop a method ('Galactic Archaeology Neural Network', gann) based on neural network models (NNMs) to identify accreted stars in galactic discs by only their chemical fingerprint and age, using a suite of simulated galaxies from the Auriga Project. We train the network on the target galaxy's own local environment defined by the stellar halo and the surviving satellites. We demonstrate that this approach allows the detection of accreted stars that are spatially mixed into the disc. Two performance measures are defined - recovery fraction of accreted stars, f(recov) and the probability that a star with a positive (accreted) classification is a true-positive result, P(TP). As the NNM output is akin to an assigned probability (P-a), we are able to determine positivity based on flexible threshold values that can be adjusted easily to refine the selection of presumed-accreted stars. We find that gann identifies accreted disc stars within simulated galaxies, with high f(recov) and/or high P(TP). We also find that stars in Gaia-Enceladus-Sausage (GES) mass systems are over 50 per cent recovered by our NNMs in the majority (18/24) of cases. Additionally, nearly every individual source of accreted stars is detected at 10 per cent or more of its peak stellar mass in the disc. We also demonstrate that a conglomerated NNM, trained on the halo and satellite stars from all of the Auriga galaxies provides the most consistent results, and could prove to be an intriguing future approach as our observational capabilities expand.