Disentangling stellar atmospheric parameters in astronomical spectra using generative adversarial neural networks

Context . The rapid expansion of large-scale spectroscopic surveys has highlighted the need to use automatic methods to extract information about the properties of stars with the greatest efficiency and accuracy, and also to optimise the use of computational resources. Aims . We developed a method based on generative adversarial networks (GANs) to disentangle the physical (effective temperature and gravity) and chemical (metallicity and overabundance of α elements with respect to iron) atmospheric properties in astronomical spectra. Using a projection of the stellar spectra, commonly called latent space, in which the contribution due to one or several main stellar physicochemical properties is minimised while others are enhanced, it was possible to maximise the information related to certain properties. This could then be extracted using artificial neural networks (ANNs) as regressors, with a higher accuracy than a reference method based on the use of ANNs that had been trained with the original spectra. Methods . Our model utilises auto-encoders, comprising two ANNs: an encoder and a decoder that transform input data into a low-dimensional representation known as latent space. It also uses discriminators, which are additional neural networks aimed at transforming the traditional auto-encoder training into an adversarial approach. This is done to reinforce the astrophysical parameters or disentangle them from the latent space. We describe our Generative Adversarial Networks for Disentangling and Learning Framework (GANDALF) tool in this article. It was developed to define, train, and test our GAN model with a web framework to show visually how the disentangling algorithm works. It is open to the community in Github. Results . We demonstrate the performance of our approach for retrieving atmospheric stellar properties from spectra using Gaia Radial Velocity Spectrograph (RVS) data from DR3. We used a data-driven perspective and obtained very competitive values, all within the literature errors, and with the advantage of an important dimensionality reduction of the data to be processed.