Exoplanet Habitability Models: A One-Class Variational Autoencoder for Identifying Non-Earth-like Habitability Candidates

The Earth Similarity Index (ESI) and related habitability metrics rank exoplanets by their closeness to Earth’s bulk physical properties. While these metrics are useful for rapid screening, they are fundamentally Earth-centric and may overlook planets that do not resemble Earth but could still support liquid water or other habitable conditions. We present a complementary screening method based on a one-class variational autoencoder (VAE). The VAE is trained only on ordinary planetary systems, defined as those outside a generous habitable insolation and radius range, using five observationally available parameters: planet radius, bulk density, insolation flux, stellar effective temperature, and orbital eccentricity. Anomalies are identified by combining reconstruction error with the Mahalanobis distance from the training latent centroid. We apply the method to 5,140 confirmed exoplanets from the NASA Exoplanet Archive. After a physical shortlist filter, the VAE top candidates have low ESI values but plausible densities and energy budgets. Within the shortlist, the VAE top 20 and ESI top 20 have zero overlap. Known potentially habitable planets have a median VAE percentile of 85.4 and a median ESI percentile of 92.4, indicating that the VAE identifies a different population of structural outliers. The method is intended as a complementary discovery tool, not as a replacement for existing habitability indices.