Description
Ground-based astronomy observations are contaminated by the emission and absorption of light by molecules in the atmosphere. This atmospheric emission is particularly a problem when observing at Near-InfraRed (NIR) wavelengths, where the atmospheric emission can be orders of magnitude brighter than the object being observed. Furthermore, the atmospheric emission spectrum varies both spatially and temporally and the physical processes which drive this variability are poorly understood and difficult to model accurately. Accurate atmospheric correction (sky subtraction) techniques will be essential to exploit the high sensitivity of the ELT’s spectrograph instruments, particularly at NIR wavelengths. Current sky subtraction methods require either simultaneous observation of the sky near the target, resulting in a loss of efficiency (particularly for multi-object spectrographs e.g. MOSAIC-ELT), or line modelling which lacks the required accuracy due to uncertainty in the sky model and/or instrumental effects. In this work, we explore data driven models for sky subtraction, focusing on spectrographs operating at NIR. Data driven methods use the large number of sky spectra observed by an instrument in normal operation to create an instrument-specific sky model which can then be used for sky subtraction. This reduces the need for simultaneous sky observation and better corrects for instrumental effects. In particular, we explore the use of neural network autoencoders to model and subtract the sky spectrum. We present an implementation of such a method for use in NIR instruments for the ELT and and report on results from testing with simulated spectra from the MOONS-VLT instrument.
