The Payne: self-consistent ab initio fitting of stellar spectra

We present The Payne, a general method for the precise and simultaneous determination of numerous stellar labels from observed spectra, based on fitting physical spectral models. The Payne combines a number of important methodological aspects: it exploits the information from much of the available spectral range; it fits all labels (stellar parameters and element abundances) simultaneously; it uses spectral models, where the atmosphere structure and the radiative transport are consistently calculated to reflect the stellar labels. At its core The Payne has an approach to accurate and precise interpolation and prediction of the spectrum in high-dimensional label-space, which is flexible and robust, yet based on only a moderate number of ab initio models (O(1000) for 25 labels). With a simple neural-net-like functional form and a suitable choice of training labels, this interpolation yields a spectral flux prediction good to $10^{-3}$ rms across a wide range of $T_{\rm eff}$ and log g (including dwarfs and giants). We illustrate the power of this approach by applying it to the APOGEE DR14 data set, drawing on Kurucz models with recently improved line lists: without recalibration, we obtain physically sensible stellar parameters as well as 15 element abundances that appear to be more precise than the published APOGEE DR14 values. In short, The Payne is an approach that for the first time combines all these key ingredients, necessary for progress towards optimal modelling of survey spectra; and it leads to both precise and accurate estimates of stellar labels, based on physical models and without re-calibration. Both the codes and catalog are made publicly available online.