Automated classification of stellar spectra - I. Initial results with artificial neural networks

We have initiated a project to classify stellar spectra automatically from high-dispersion objective prism plates. The automated technique presented here is a simple backpropagation neural network, and is based on the visual classification work of Houk. The plate material (Houk's) is currently being digitized, and contains ≈ 105 stars down to V ≈ 11 at ≈ 2-A resolution from ≈ 3850 to 5150 A. For this first paper in the series we report on the results of 575 stars digitized from 6 plates. We find that even with the limited data set now in hand we can determine the temperature classification to better than 1.7 spectral subtypes from B3 to M4. Our current sample size provides insufficient training set material to generate luminosity and metallicity classifications. Our eventual aims in this project are (1) to create a large and homogeneous digital stellar spectral library; (2) to create a well-understood and robust automatic classification algorithm which can determine temperatures, luminosities and metallicities for a wide variety of spectral types; (3) to use these data, supplemented by deeper plate material, for the study of Galactic structure and chemical evolution; and (4) to find unusual or new classes of objects.