Unsupervised discovery of exoplanet populations via deep generative models

A full exoplanetary classification takes into account a complete list of stellar and planetary physical and orbital parameters. This thesis investigates whether a deep generative model trained jointly on the full set of features can recover the known populations without supervision and reveal additional structure. We apply Variational Deep Embedding (VaDE) to the NASA Exoplanet Archive, benchmark it against classical baselines, study the geometry of the learned latent space through a β-VAE, and test a physics-informed variant that constrains the model with known scaling laws.