Large surveys now measure millions of stars, too many to analyse one at a time. In supervised learning, a model such as a neural network learns from stars whose properties are already known from careful measurements, then predicts those properties for the rest. Unsupervised methods such as clustering search a catalogue for groups of similar stars without being told what the groups are.
Leung and Bovy (2019) trained a neural network on APOGEE spectra that returns abundances of 18 elements to about 0.03 dex (about 7 per cent), with uncertainties that grow for stars unlike its training set; it analysed about 250,000 spectra in ten minutes on one GPU. Hunt and Reffert (2023) ran the clustering algorithm HDBSCAN on 729 million Gaia DR3 sources and recovered 7,167 star clusters, 2,387 of them new candidates; a neural network that classifies colour-magnitude diagrams helped them validate the clusters.
Where I come in
My 2025 paper introduced a method based on UMAP and HDBSCAN for separating the Milky Way's stellar populations and tested it on a synthetic galaxy where the answer is known. Starfold, my open-source Python package, extends it by refitting each grouping on 32 random subsets to measure how often every group recurs; it matched or beat the library defaults on 8 of 10 benchmark tables, but on only 2 of 4 held-out tables. I am also developing JANUS, a neural network that predicts APOGEE-like parameters and abundances from Gaia XP spectra, with a reliability score for each star.