AAS Meeting #193 - Austin, Texas, January 1999
Session 44. Modeling Stellar Characteristics
Display, Thursday, January 7, 1999, 9:20am-6:30pm, Exhibits Hall 1

## [44.09] Neural Network Techniques Applied to Low Resolution Spectra of Halo Stars

Y. Qu, S. Snider (U. Texas), T. von Hippel (Gemini), C. Sneden, D.L. Lambert (U. Texas), T.C. Beers (Michigan State U.), S. Rossi (IAG, U. Sau Paulo)

Recent large surveys of Galactic halo stars have uncovered kinematically and chemically diverse substructures that contain vital clues to the early evolution of the Galaxy. As extant spectroscopic sample sizes grow into the thousands, traditional star-by-star chemical composition analyses simply will not be able to keep pace. New analytical tools must be found that can attack the large spectroscopic databases in a partially or fully automated fashion (thus providing useful astrophysical data nearly in real-time") without sacrificing information content. Here, for a variety of halo-population stars, we present preliminary results of applying an artificial neural network code to low resolution (R\equiv\lambda/\delta\lambda~2000) spectra in the wavelength range 3800--5000Å. We have adapted the back-propogation neural network technique originally devised for stellar spectral classification (von Hippel et al. 1994, MNRAS, 269, 97) to predict effective temperatures, gravities, and overall metallicities from spectra that are being gathered by Beers and colleagues (e.g. 1992, AJ, 103, 1987). First results demonstrate that with properly trained neural networks, the T\rm eff and [Fe/H] values may be predicted to typically ±50K and ±0.2~dex, respectively, for stars with well-determined parameters from the literature. We will discuss these trends, as well as additional studies of the application to log~g predictions, and experiments that focus in on individual abundance ratios (principally [C/Fe], [Ba/Fe], and [Sr/Fe]).