Deep Generative Modeling of Periodic Variable Stars Using Physical Parameters

Abstract The ability to generate physically plausible ensembles of variable sources is critical the optimization time domain survey cadences and training classification models on data sets with few no labels. Traditional augmentation techniques expand by reenvisioning observed exemplars, seeking simulate observations specific under different (exogenous) conditions. Unlike fully theory-driven models, these approaches do not typically allow principled interpolation nor extrapolation. Moreover, principal drawback lies in prohibitive computational cost simulating source observables from ab initio parameters. In this work, we propose a computationally tractable machine learning approach realistic light curves periodic variables capable integrating physical parameters variability classes as inputs. Our deep generative model, inspired transparent latent space adversarial networks, uses variational autoencoder (VAE) architecture temporal convolutional network layers, trained using OGLE-III optical characteristics (e.g., effective temperature absolute magnitude) Gaia DR2. A test temperature–shape relationship RR Lyrae demonstrates efficacy our “physics-enhanced VAE” (PELS-VAE) model. Such serving nonlinear nonparametric emulators, present novel tool for astronomers create synthetic series over arbitrary cadences.