Abstract
Neural networks are widely used in astronomical spectroscopy to analyse stellar spectra, which encode rich information about the physical and chemical properties of stars. In tasks like identifying stellar chemical abundance patterns (chemical tagging), performance improves when the parameters of interest are isolated and the influence of others is reduced. With this objective, we present GANDALF (Generative Adversarial Networks for Disentangling And Learning Framework), a software framework specifically designed to tackle this problem through the use of generative adversarial networks. By projecting the stellar spectra into a latent space, where the influence of one or several physicochemical properties is reduced while others are emphasised, it becomes possible to amplify the information associated with specific properties. Moreover, our software can be used to reconstruct new spectra by modifying the disentangled parameters, working as a generative model. Additionally, it includes an interactive visualizer that allows users to explore the data and manually modify its disentangled parameters. The effectiveness of GANDALF has been demonstrated using data from two broadly used astronomical surveys, the Apache Point Observatory Galactic Evolution Experiment (APOGEE) and Gaia's Radial Velocity Spectrometer (RVS), where it yielded low coefficients of determination (R2 = 0.2) for the parameters that were suppressed, while preserving high values for those that were retained.

This publication has 6 references indexed in Scilit: