A method based on Generative Adversarial Networks for disentangling physical and chemical properties of stars in astronomical spectra

Abstract

This work presents the design of an autoencoder architecture that uses adversarial training in the context of astrophysical spectral analysis. We aim to develop a middle representation of stellar spectra in which the influence of the most prominent physical properties, such as surface temperature and gravity, is effectively removed. This allows the variance within the representation to primarily reflect the effects of the star’s chemical composition on the spectrum. We apply a scheme of deep learning to unravel in the latent space the desired parameters of the rest of the information contained in the data. This work proposes a version of adversarial training that uses one discriminator per parameter to be disentangled, avoiding the exponential combination that occurs when using a single discriminator. Synthetic astronomical data from the APOGEE and Gaia surveys were used to test the method’s effectiveness. Our approach demonstrates a marked improvement in disentangling, reflected in an improvement in the $R^2$ score of up to 0.7. Additionally, we introduce an ad-hoc framework, GANDALF, designed to facilitate visualization and adaptation of the methodology to other domains in astronomical spectroscopy.