In the context of large spectroscopic surveys of stars, data-driven methods are key in deducing physical parameters for millions of spectra in a short time. Convolutional neural networks (CNNs) enable us to connect observables (e.g. spectra, stellar magnitudes) to physical properties (atmospheric parameters, chemical abundances, or labels in general). We trained a CNN, adopting stellar atmospheric parameters and chemical abundances from APOGEE DR16 (resolution R=22500) data as training set labels. As input, we used parts of the intermediate-resolution RAVE DR6 spectra (R~7500) overlapping with the APOGEE DR16 data as well as broad-band ALL_WISE and 2MASS photometry, together with Gaia DR2 photometry and parallaxes. We derived precise atmospheric parameters Teff, log(g), and [M/H] along with the chemical abundances of [Fe/H], [alpha/M], [Mg/Fe], [Si/Fe], [Al/Fe], and [Ni/Fe] for 420165 RAVE spectra. The precision typically amounts to 60K in Teff, 0.06 in log(g) and 0.02-0.04 dex for individual chemical abundances. Incorporating photometry and astrometry as additional constraints substantially improves the results in terms of the accuracy and precision of the derived labels. We provide a catalogue of CNN-trained atmospheric parameters and abundances along with their uncertainties for 420165 stars in the RAVE survey. CNN-based methods provide a powerful way to combine spectroscopic, photometric, and astrometric data without the need to apply any priors in the form of stellar evolutionary models. The developed procedure can extend the scientific output of RAVE spectra beyond DR6 to ongoing and planned surveys such as Gaia RVS, 4MOST, and WEAVE. We call on the community to place a particular collective emphasis and on efforts to create unbiased training samples for such future spectroscopic surveys.
Download