Collaborators: M. Ntampaka, F. Dauphin, S. Ravindranath Project includes full funding starting January 12, 2022
Modern wide-field imaging surveys performed by Roman Space Telescope and Vera Rubin Observatory's Large Synoptic Survey will open a new window on the Universe, enabling the discovery and study of stars and galaxies too faint to be studied previously. To realize the full scientific potential of these surveys, we must examine those objects using spectroscopic techniques. Given the large numbers of objects to be studied – some individual science cases require observations of tens of millions of sources – we must develop an efficient process to carry out these surveys. Several massively multiplexed spectroscopic instruments from the ground are currently in the planning phase or nearing execution. High-quality calibration and stability are required to ensure that millions of fiber hours with 8-10-m class facilities. Most science cases for these massively multiplexed missions will rely on their ability to detect faint sources and/or measure the widths of narrow lines next to skylines. Both capabilities are dependent on our ability to subtract the contribution of the sky to the observed spectra.
This project aims to enhance ground-based NIR spectra of astrophysically interesting objects with ground-sky spectra, atmospheric data, HST spectra and images, and machine learning techniques proven to predict galaxy spectra from images.
The student hired for this project will focus on 600 high-resolution (R~30,000) NIR sky spectra taken contiguously (with a cadence of 5 and 10mn) over 6 nights with the SpectroPolarimètre Infra-Rouge (SPIRou) - a high precision spectropolarimeter and velocimeter on the Canada-France-Hawaii Telescope (CFHT) on Maunakea. The student will use atmospheric models employed by the ETCs of several facilities, atmospheric data including solar activity from the Maunakea Weather Center and use machine learning techniques to predict the evolution of sky continuum and lines between two sky observations separated by time scales. We will then seek to test this model on a set of archived spectra of non-variable point sources from Gemini’s Near-Infrared Spectrograph (GNIRS) also observed by HST (e.g. spectro-photometric calibrators) and develop ways to refine the observed ground based spectra when both space and sky data are available.
Deep learning is a powerful, efficient method to extract subtle signals from data by processing them through multiple unseen layers (e.g., LeCun et al. 1999). Autoencoders (Rumelhart, Hinton, and Williams 1986) are a class of deep learning algorithms that have been applied in astronomy for a range of tasks, including generating realistic galaxy images (Ravanbakhsh et al. 2016), classifying supernovae (Villar et al. 2020), and extracting new physical information from catalogs (Ntampaka & Vikhlinin, 2021 ). A traditional application of an autoencoder can be thought of as a flexible version of principal component analysis, summarizing a complicated signal in a few essential values and then reconstructing the input signal. Because autoencoders build both a flexible compression ("encoder") and decompression ("decoder") of data, they can be used as both a replacement for PCA and also as a data generation tool. The principal advantage of autoencoders over PCA is that auto-encoders are capable of modeling complex non-linear functions such as those associated with non-linear detector responses, fiber throughputs, and other instrumental features that compound the sky subtraction challenge. This research will explore the use of decoders for quickly generating many realistic mock sky spectra. We will make all codes with descriptions of the modifications available to the general community on github. Our goal is to optimize the sky subtraction for a fiber-fed spectrograph on a large aperture telescope. However, the strategies developed with this project will be useful for any spectroscopic observation, including improvements of archival NIR spectroscopic data.
References: Ellis & Bland-Hawthorn 2008, MNRAS, 386, 1¨ LeCun et al. 1999, Proceedings of the IEEE, 86(11):2278-2324¨ Ntampaka et al. 2019, ApJ, 880, 154N, ¨ Ntampaka & Vikhlinin, 2021 under-review, ¨ Ravanbakhsh, Schneider & Po ́czos, ICLR 2017, astroph:611.04500. ¨D. Rumelhart, G. Hinton, & R. Williams, 1986, Nature 323 ¨ Smette et al. 2015, A&A, 576, A77¨ Villar et al. 2020, ApJ, 905, 94
