Armillary: 32 labelled stars are sufficient to estimate the parameters of 200,000 others based on their spectra

Without deep learning, the method constructs a coordinate system based solely on the stellar spectra from APOGEE and DESI, and then applies the labels from a few stars to all the others.

A stellar spectrum depends on a small number of physical parameters, such as actual temperature, surface gravity or metallicity. A set of spectra thus forms a small variety whose parameters are the natural coordinates. Yuan-Sen Ting and Serat Saad present Armillary, a method that reconstructs these coordinates from spectra alone, by generalizing the sequencer algorithm of Baron and Ménard.

Distances along the variety

Two stars are compared by Wasserstein's distance between their absorption profiles, then the shortest paths in a graph of closest neighbours measure distance along the variety. The same graph is used to transfer labels from a few known stars to all others. The coordinates obtained follow temperature and gravity more linearly than the CPA, T-SNE, UMAP or autoencoder.

On 202,932 stars of the APOGEE DR17 survey, 32 driving stars are sufficient to recover temperature, gravity and metallicity with errors of about 100 K, 0.18 and 0.083 dex; with 2,048, they fall to 33 K, 0.06 and 0.026 dex. On 104,799 DESI stars, the method remains effective even for very noisy spectra. The code is public.