AAS Meeting #194 - Chicago, Illinois, May/June 1999
Session 84. Studying the Anatomy of the Milky Way
Display, Thursday, June 3, 1999, 9:20am-4:00pm, Southwest Exhibit Hall

[Previous] | [Session 84] | [Next]

[84.11] Automatic Identification, Classification, and Abundance Estimation for Metal-Poor Stars in the Galaxy from Objective-Prism Spectroscopy via Artificial Neural Network Analysis

J. Rhee, T.C. Beers (Michigan State U.), M.J. Irwin (IoA)

The HK prism survey of Beers and collaborators has been extremely successful in the identification of large numbers of metal-deficient stars in the thick disk and halo of the Galaxy. Such stars provide vital clues for unraveling the chemical and dynamical history of the Milky Way, and large spiral galaxies in general.

The original selection of candidate metal-poor stars from the HK prism plates was carried out using visual inspection, which introduces a number of (avoidable) biases in the resulting target lists (in particular a tendency to overlook metal-poor stars of low temperature). We are in the process of selecting new candidate metal-poor stars based on automated scans of the HK survey plates with the APM facility in Cambridge. Here we present the results of an artificial neural network analysis of this data, which enables us to objectively select, to classify by color and metallicity class, and to predict the metallicities of stars on the prism plates directly from the extracted spectra.

The training set consists of about 370 stars with abundances obtained from previous HK survey follow-up efforts, chosen from some of the 320,000 stars in the ``digital'' HK survey to date (over 1,500,000 stars are expected in the final sample). For first-pass classification, external estimates of the broadband color index, (B-V)o, and equivalent widths of the CaII H and K lines from the extracted prism spectra are used as input variables to separate the prism spectra into regions of similar (B-V)o and [Fe/H]. Currently, a correct classification rate is achieved for more than 70% of the stars. In the prediction step, these same quantities are used as input variables to predict stellar [Fe/H]. We presently obtain correlation coefficients between the predicted and known [Fe/H] for stars in our test sample of greater than 0.75, with an rms error of 0.1 dex, which is extremely encouraging. We discuss steps that are underway to improve on these results, primarily by obtaining improved color estimates for the stars.

This work received partial support from grant AST 95-29454 awarded by the National Science Foundation.

If the author provided an email address or URL for general inquiries, it is a s follows:


[Previous] | [Session 84] | [Next]