| Abstract: |
I present a new non-parametric machine-learning method for predicting stellar metallicity ([Fe/H]) based on photometric colors from the Sloan Digital Sky Survey (SDSS). The method is trained using a large sample of 〜170k stars with SDSS spectra and atmospheric parameter estimates (Teff, log g, and [Fe/H]). For bright stars (g < 19.5 mag), the method is capable of predicting [Fe/H] with a typical scatter of 〜0.29 dex. This scatter is similar to the typical uncertainty associated with [Fe/H] measurements from low-resolution spectra. Following minor adjustments to the model, the method is suitable for the discovery of extremely metal poor (EMP) stars ([Fe/H] < -3). I further show that the inclusion of light curve information improves the overall performance of the model, suggesting that future wide-field time-domain surveys, such as the Zwicky Transient Facility and LSST, will act as pseudo-spectrographic engines. I will conclude by arguing that these methods, when applied to LSST and similar surveys will enable the construction of extremely detailed maps of the Milky Way, its structure, and history. |
