The Cannon 2: A data-driven model of stellar spectra for detailed chemical abundance analyses

We have shown that data-driven models are effective for inferring physical attributes of stars (labels; Teff, logg, [M/H]) from spectra, even when the signal-to-noise ratio is low. Here we explore whether this is possible when the dimensionality of the label space is large (Teff, logg, and 15 abundances: C, N, O, Na, Mg, Al, Si, S, K, Ca, Ti, V, Mn, Fe, Ni) and the model is non-linear in its response to abundance and parameter changes. We adopt ideas from compressed sensing to limit overall model complexity while retaining model freedom. The model is trained with a set of 12,681 red-giant stars with high signal-to-noise spectroscopic observations and stellar parameters and abundances taken from the APOGEE Survey. We find that we can successfully train and use a model with 17 stellar labels. Validation shows that the model does a good job of inferring all 17 labels (typical abundance precision is 0.04 dex), even when we degrade the signal-to-noise by discarding ~50% of the observing time. The model dependencies make sense: the spectral derivatives with respect to abundances correlate with known atomic lines, and we identify elements belonging to atomic lines that were previously unknown. We recover (anti-)correlations in abundance labels for globular cluster stars, consistent with the literature. However we find the intrinsic spread in globular cluster abundances is 3--4 times smaller than previously reported. We deliver 17 labels with associated errors for 87,563 red giant stars, as well as open-source code to extend this work to other spectroscopic surveys.