Inferring redshift and galaxy properties via a multi-task neural net with probabilistic outputs

An application to simulated MOONS spectra

¹ Dipartimento di Fisica e Astronomia, Università di Firenze, Via G. Sansone 1, I-50019 Sesto F.no (Firenze), Italy
² INAF – Osservatorio Astrofisico di Arcetri, Largo E. Fermi 5, I-50125 Florence, Italy
³ School of Physics and Astronomy, University of St Andrews, North Haugh, St Andrews KY16 9SS UK
⁴ INAF-Osservatorio di Astrofisica e Scienza dello Spazio di Bologna, Via Piero Gobetti 93/3, 40129 Bologna, Italy
⁵ GEPI, Observatoire de Paris, PSL University, CNRS, Meudon, France
⁶ Cavendish Laboratory, University of Cambridge, 19 J. J. Thomson Ave., Cambridge CB3 0HE UK
⁷ Kavli Institute for Cosmology, University of Cambridge, Madingley Road, Cambridge CB3 0HA UK
⁸ Department of Physics and Astronomy, University College London, Gower Street, London WC1E 6BT, UK
⁹ European Southern Observatory, Karl-Schwarzschild-Strasse 2, D-85748 Garching bei Muenchen, Germany

^⋆ Corresponding author; michele.ginolfi@unifi.it

Received: 29 October 2024
Accepted: 25 November 2024

Abstract

The era of large-scale astronomical surveys demands innovative approaches for rapid and accurate analysis of extensive spectral data, and a promising direction in which to address this challenge is offered by machine learning. Here, we introduce a new pipeline, M-TOPnet (Multi-Task network Outputting Probabilities), which employs a convolutional neural network with residual learning to simultaneously derive redshift and other key physical properties of galaxies from their spectra. Our tool efficiently encodes spectral information into a latent space, employing distinct downstream branches for each physical quantity, thereby benefiting from multi-task learning. Notably, our method handles the redshift output as a probability distribution, allowing for a more refined and robust estimation of this critical parameter. We demonstrate preliminary results using simulated data from the MOONS instrument, which will soon be operating at the ESO/VLT. We highlight the effectiveness of our tool in accurately predicting the redshift, stellar mass, and star formation rate of galaxies at z ≳ 1 − 3, even for faint sources (m_H ∼ 24) for which traditional methods often struggle. Through analysis of the output probability distributions, we demonstrate that our pipeline enables robust quality screening of the results, achieving accuracy rates of up to 99% in redshift determination (defined as predictions within |Δz|< 0.01 relative to the true redshift) with 8 h exposure spectra, while automatically identifying potentially problematic cases. Our pipeline thus emerges as a powerful solution for the upcoming challenges in observational astronomy, combining precision, interpretability, and efficiency, all aspects that are crucial for analysing the massive datasets expected from next-generation instruments.