Exoplanet VAE
A one-class anomaly detector that looks for habitable worlds in the places an Earth-similarity score would never think to check.
Founded a student research group investigating whether machine-learning methods can identify potentially habitable exoplanets that conventional Earth-similarity metrics may overlook.
The argument is that the Earth Similarity Index is biased: it only measures how much a planet physically resembles Earth, and so ignores alternative planetary structures that could support life — massive Super-Earth ocean worlds among them. A standard classifier would simply inherit that bias, along with the dataset’s extreme class imbalance.
So the problem is framed as anomaly detection instead. A one-class variational autoencoder is trained only on standard, non-habitable planets, learning the characteristics of the broader exoplanet population rather than a handful of predefined habitable examples. Passing the candidate planets back through it, the ones flagged with high reconstruction loss but stable energy budgets are the hidden candidates the ESI fails to catch.
Current work focuses on developing and validating the model, and on assessing whether the identified anomalies correspond to physically meaningful planetary populations rather than artifacts of the dataset or the model itself.
Context
- Role — Founder and first author.
- Team — formed with the same people as the Beamline for Schools group.
- The question — whether anomaly detection is a better complementary framework than assuming Earth is the optimal template for life.
- Methods — one-class variational autoencoder, reconstruction-loss analysis, energy-budget filtering, ESI critique.