Flow-variational auto-encoder for the data augmentation of the Horizon-AGN dual-AGN dataset

Abstract

Dual and double active galactic nuclei (AGNs) are complex systems that encompass several astrophysical domains. Studying these objects helps address the formation and mass build-up of galaxies through hierarchical merging processes, emission of low-frequency gravitational waves, and AGN feedback on the star formation histories of galaxies. The next generation of large-scale wide-field imaging surveys will soon provide very large sets of observational data to look for AGN pairs, surpassing the computational capabilities of classical identification methods. Automatic selection algorithms have therefore been endorsed for such analyses, with the drawback that they necessitate extensive labelled datasets for training. As genuine dual-AGN observations are still scarce and numerical simulation tend to produce an insufficient number of complex systems over a reasonable computation time, it is necessary to introduce a data augmentation step to construct a viable training dataset for the efficient detection of dual-AGNs. Here, we present a deep learning data augmentation method aimed at emulating realistic dual-AGN images with a large range of physical properties. Our method combines a variational auto-encoder (VAE) with a normalising flow, trained and tested on images extracted from the catalogue of dual-AGNs resulting from the Horizon-AGN hydrodynamical simulation. We constructed our model in this astrophysical framework, developing a specific training loss function for the VAE in both the direct and Fourier spaces, which has enabled us to secure the construction of the small-scale and thin features associated with the gravitational interaction between merging galaxies like tidal tails. The normalising flow was used to map the distribution of the VAE latent space and sample it directly to emulate images similar to the original Horizon-AGN dataset. A generation score was defined to keep track of the quality of the emulated images. This score was computed using measures of the quality of the VAE reconstruction and estimations of the VAE latent space density based on the normalising flow's probability estimation. Overall, we estimate that the quality of the generated images is comparable to the initial dataset from Horizon-AGN and identify a possible quality cut as the 95th percentile of the distribution of generation scores. With this flow-VAE model, we were able to generate a large set of high-quality dual-AGN images, assessed from the generation score and quantitative measure of light profiles, at a rate of 200 images per second once the model had been properly trained. We note that this is a fraction of the time necessary to generate such images from a cosmological simulation. The images obtained from our data augmentation method can then be used as a training dataset for an automatic detection method, such a convolutional neural network (CNN), to analyse images from a catalogue of dual-AGN candidates. Our method could also be suited to studies of other astrophysical objects.