Stellar parameters from very low resolution spectra and medium band filters: Teff, logg and [M/H] using neural networks

Abstract

Large scale, deep survey missions such as GAIA will collect enormous amounts of data on a significant fraction of the stellar content of our Galaxy. These missions will require a careful optimisation of their observational systems in order to maximise their scientific return, and will require reliable and automated techniques for parametrizing the very large number of stars detected. To address these two problems, I investigate the precision to which the three principal stellar parameters (Teff, logg, [M/H]) can be determined as a function of spectral resolution and signal-to-noise (SNR) ratio, using a large grid of synthetic spectra. The parametrization technique is a neural network, which is shown to provide an accurate three-dimensional physical parametrization of stellar spectra across a wide range of parameters. It is found that even at low resolution (50-100 AA FWHM) and SNR (5-10 per resolution element), Teff and \met can be determined to 1% and 0.2 dex respectively across a large range of temperatures (4000-30000 K) and metallicities (-3.0 to +1.0 dex), and that logg is measurable to +/- 0.2 dex for stars earlier than solar. The accuracy of the results is probably limited by the finite parameter sampling of the data grid. The ability of medium band filter systems (with 10-15 filters) for determining stellar parameters is also investigated. Although easier to implement in a unpointed survey, it is found that they are only competitive at higher SNRs (> 50).