In this project I investigate the use of artificial neural networks and principal components analysis (PCA) in the analysis of stellar spectra. In particular, I examine the application of these techniques to the automation of MK spectral classification.
For the purposes of this project, I scanned and reduced 100 plates from the Michigan Spectral Survey. This yielded a spectral database of over 15,000 spectra with an approximate wavelength range of 3800-5200 Angstroms at a resolution of approximately 3 Angstronms. The plate reduction and the extraction and processing of the spectra is discussed. From this database, a subset of over 5,000 spectra (with two-dimensional MK classifications listed in the Michigan Henry Draper Catalogue) was used to develop supervised neural network classifiers. I show that neural networks can give accurate spectral type classifications (mean 1 sigma error = 0.82 subtypes, RMS error = 1.09 subtypes) across the full range of spectral types present in the database (B2-M7); I show also that the networks yield correct luminosity classes for over 95% of both dwarfs and giants with a high degree of confidence. The high level of reproducibility of neural network classifications is demonstrated. The analysis includes a comparison of the performance of networks of differing degrees of complexity and modes of application.
Stellar spectra generally contain a large amount of redundant (correlated) information. I investigate the application of PCA to the optimal compression of spectra. This investigation shows that PCA can compress the spectra by a factor of over 30 whilst retaining more than 95% of the variance in the data set. Furthermore, this compression leads to no decrease in classifier performance, indicating that the PCA compression from 820 to 25 components results in no significant loss of relevant information. I also demonstrate how PCA acts as a filter of noise and bogus features in a spectrum and can be used to identify unusual spectra.
The ultimate goal of stellar classification should be a physical parameterization of the stars. I examine the application of neural networks to the problem of obtaining physical stellar parameters (T_eff, log g, etc.) directly from an observed spectrum, by training a neural network on synthetic spectra. It is demonstrated that this approach is capable of yielding physical parameters independently of the MK system, thus avoiding the assumptions and limitations of any such intermediate classification scheme.
Return to my thesis page.