Machine-Learning Enabled Stellar Classification
and the Prediction of Fundamental Atmospheric
Parameters From Photometric Light Curves

Adam A. Miller
Jet Propulsion Laboratory
California Institute of Technology

Abstract. 

The falling costs of computing and CCD detectors has led to
a great boom in wide-field time-domain surveys during the past decade,
with several new surveys expected prior to the arrival of the Large Synoptic
Survey Telescope (LSST). This observational boon, however, comes
with a catch: the data rates from these surveys are so large that discovery
techniques heavily dependent on human intervention are becoming unviable.
In this talk I will detail new methods, which utilize semi-supervised
machine-learning algorithms, to automatically classify the light curves of
time-variable sources. Using these methods, we have produced a datadriven
probabilistic catalog of variables found in the All Sky Automated
Survey (ASAS). I will also present a new machine-learning-based framework
for the prediction of the fundamental stellar parameters, Teff, log
g, and [Fe/H], based on the photometric light curves of variable stellar
sources. The method was developed following a systematic spectroscopic
survey of stellar variability. I will demonstrate that, for variable sources,
the machine-learning model can determine Teff, log g, and [Fe/H] with
a typical scatter of 130 K, 0.38 dex, and 0.26 dex, respectively, without
obtaining a spectrum. Instead, the random-forest-regression model uses
SDSS color information and light-curve features to infer stellar properties.
The precision of this method is competitive with what can be
achieved with low-resolution spectra. These results are an important
step on the path to the efficient and optimal extraction of information
from future time-domain experiments, such as LSST. We argue that this
machine-learning framework, for which we outline future possible improvements,
will enable the construction of the most detailed maps of
the Milky Way ever created.