2023MNRAS.520..917D

The advent of large-scale stellar spectroscopic surveys naturally leads to the implementation of machine learning techniques to isolate, for example, small sub-samples of potentially interesting stars from the full data set. A recent example is the application of the t-SNE statistical method to ∼600 000 stellar spectra from the GALAH survey in order to identify a sample of candidate extremely metal-poor (EMP; [Fe/H] ≤ -3) stars. We report the outcome of low-resolution spectroscopic follow-up of 83 GALAH EMP candidates that lack any previous metallicity estimates. Overall, the statistical selection is found to be efficient (∼one-third of the candidates have [Fe/H] ≤ -2.75) with low contamination (<10 per cent have [Fe/H] > -2), and with a metallicity distribution function that is consistent with previous work. Five stars are found to have [Fe/H] ≤ -3.0, one of which is a main sequence turnoff star. Two other stars are revealed as likely carbon-enhanced metal-poor (CEMP) stars of type CEMP-s, and a known carbon star is re-identified. The results indicate that the statistical selection approach employed was successful, and therefore it can be applied to forthcoming even larger stellar spectroscopic surveys with the expectation of similar positive outcomes.