Physical Parameterization of Stellar Spectra: The Neural Network Approach

We present a technique which employs artificial neural networks to produce physical parameters for stellar spectra. A neural network is trained on a set of synthetic optical stellar spectra to give physical parameters (e.g. Teff , log g, [M/H]). The network is then used to produce physical parameters for real, observed spectra. Our neural networks are trained on a set of 155 synthetic spectra, generated using the spectrum program written by Gray (Gray & Corbally 1994, Gray & Arlt 1996). Once trained, the neural network is used to yield Teff for over 5000 B–K spectra extracted from a set of photographic objective prism plates (Bailer-Jones, Irwin & von Hippel 1997a). Using the MK classifications for these spectra assigned by Houk (1975, 1978, 1982, 1988), we have produced a temperature calibration of the MK system based on this set of 5000 spectra. It is demonstrated through the metallicity dependence of the derived temperature calibration that the neural networks are sensitive to the metallicity signature in the real spectra. With further work it is likely that neural networks will be able to yield reliable metallicity measurements for stellar spectra.