Astronomy Seminar: Mining Big Data Over the Entire IR Sky: Improved Photometric Classification of YSOs, AGB and Post-AGB Stars, Mira Variables, and Bi
| Date/Time: | Friday, 07 Sep 2018 from 4:10 pm to 5:10 pm |
|---|---|
| Location: | Rm. 38, Physics Bldg. |
| Cost: | Free |
| Contact: | Curt Struck |
| Phone: | 515-294-3666 |
| Channel: | College of Liberal Arts and Sciences |
| Actions: | Download iCal/vCal | Email Reminder |
The study of the infrared (IR) sky is of paramount importance for understanding astrophysical phenomena such as star formation in the Milky Way or the late stages in the evolution of intermediate- to high-mass stars. Studying intrinsically red sources also provides a large potential for discovery, as many sources detected in IR surveys have not yet been fully characterized, including many protostars, evolved dusty stars, and young stars bearing biogenic ices in their circumstellar environments. To maximize the scientific potential of large IR surveys, we must be able to identify sources belonging to different classes 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-color cuts to source classification. These techniques allow us to assign probabilistic classifications to sources rather than deterministic labels. Specifically, we aim to increase the number of classified YSOs, AGB, 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 (SVM), Random Forest (RF), and Multi-layer Perceptron (MLP). When classifying a test set of these five source types, we find that all three optimized classifiers perform with F1 scores >0.97. By applying the optimal RF algorithm to a science target set of sources in the galactic plane, we have classified over 60,000 of these sources, among which we find over 1,000 likely biogenic ice candidates.