TAU: A neural network based telluric correction framework

Abstract

Context. Telluric correction is one of the critically important outstanding issues for extreme precision radial velocities and exoplanet atmosphere observations. Thorough removal of small so-called micro tellurics across the entire wavelength range of optical spectro-graphs is necessary in order to reach the extreme radial velocity precision required to detect Earth-analog exoplanets orbiting in the habitable zone of solar-type stars. Likewise, proper treatment of telluric absorption will be important for exoplanetary atmosphere observations with high-resolution spectrographs on future extremely large telescopes (ELTs). Aims. In this work, we introduce the Telluric AUtoencoder (TAU). TAU is an accurate high-speed telluric correction framework built to extract the telluric spectrum with previously unobtained precision in a computationally efficient manner. Methods. TAU is built on a neural network autoencoder trained to extract a highly detailed telluric transmission spectrum from a large set of high-precision observed solar spectra. We accomplished this by reducing the data into a compressed representation, allowing us to unveil the underlying solar spectrum and simultaneously uncover the different modes of variation in the observed spectra relating to the absorption from H2O and O2 in the atmosphere of Earth. Results. We demonstrate the approach on data from the HARPS-N spectrograph and show how the extracted components can be scaled to remove H2O and O2 telluric contamination with improved accuracy and at a significantly lower computational expense than the current state of the art synthetic approach molecfit. We also demonstrate the capabilities of TAU to remove telluric contamination from observations of the ultra-hot Jupiter HAT-P-70b allowing for the retrieval of the atmospheric signal. We publish the extracted components and an open-source code base allowing scholars to apply TAU on their own data.