Mining Big Data Over the Entire IR Sky: Improved Photometric Classification of YSOs, AGB and Post-AGB Stars, Mira Variables, and Biogenic Ice Candidates through Machine Learning
Knowledge of the IR sky is crucial for understanding a wide array astrophysical phenomena including star formation and the late stages in the evolution of intermediate- to high-mass stars. It also provides a large potential for discovery because many sources detected in IR surveys such as protostars, evolved dusty stars, and young stars bearing biogenic ices in their circumstellar environments have not yet been fully characterized. To maximize the scientific potential of large IR surveys, we must be able to classify sources that often overlap in color-color diagrams. Here we aim to produce an improved census of IR sources in the galactic plane by employing machine learning (ML) techniques rather than traditional color cuts to source classification, which allows us to assign probabilistic rather than deterministic labels. Specifically, we aim to increase the number of classified YSOs, AGB and post-AGB stars, Mira variables, and biogenic ice candidates. The latter are particularly important to the study of planet formation and constitute excellent targets for follow-up with missions such as the proposed NASA Medium Explorer mission SPHEREx. We have constructed a robust training set of spectroscopically confirmed sources and used their AllWISE and 2MASS photometry to train three ML classifiers: Support Vector Machine, Random Forest, and Multi-layer Perceptron. When classifying a test set, all three optimal classifiers perform with scores >0.97. We apply the optimal RF algorithm to a science target set of WISE- and Gaia-selected sources in the galactic plane. As a preliminary result, we have classified over 60,000 of these sources, among which we find over 1,000 likely biogenic ice candidates. We will continue to expand our training set so that we can apply these techniques to even larger sets of unclassified sources in the future. The SAO REU program is funded in part by the National Science Foundation REU and Department of Defense ASSURE programs under NSF Grant no. AST-1659473, and by the Smithsonian Institution.
American Astronomical Society Meeting Abstracts #233
- Pub Date:
- January 2019