On the cosmological performance of photometrically classified supernovae with machine learning
Aug 12, 201918 pages
Published in:
- Mon.Not.Roy.Astron.Soc. 497 (2020) 3, 2974-2991
- Published: Sep 21, 2020
e-Print:
- 1908.04210 [astro-ph.CO]
DOI:
- 10.1093/mnras/staa1968 (publication)
View in:
Citations per year
Abstract: (Oxford University Press)
The efficient classification of different types of supernovae is one of the most important problems for observational cosmology. However, spectroscopic confirmation of most objects in upcoming photometric surveys, such as the the Rubin Observatory Legacy Survey of Space and Time, will be unfeasible. The development of automated classification processes based on photometry has thus become crucial. In this paper, we investigate the performance of machine learning (ML) classification on the final cosmological constraints using simulated light-curves from the Supernova Photometric Classification Challenge, released in 2010. We study the use of different feature sets for the light-curves and many different ML pipelines based on either decision-tree ensembles or automated search processes. To construct the final catalogues we propose a threshold selection method, by employing a bias-variance tradeoff. This is a very robust and efficient way to minimize the mean squared error. With this method, we were able to obtain very strong cosmological constraints, which allowed us to keep || of the total information in the Type Ia supernovae when using the SALT2 feature set, and || for the other cases (based either on the Newling model or on standard wavelet decomposition).Note:
- 18 pages, 13 figures, 7 tables. Replaced to match the published version
- methods: data analysis
- techniques: photometric
- supernovae: general
- cosmology: observations
- astronomical data bases: miscellaneous
- surveys
References(70)
Figures(15)