Skip to main content
SearchLoginLogin or Signup

Unsupervised Learning and Feature Engineering of Exoplanet Transmission Spectra

Presentation #223.02 in the session Exoplanet Atmospheres (Poster + Lightning Talk)

Published onOct 23, 2023
Unsupervised Learning and Feature Engineering of Exoplanet Transmission Spectra

The vast majority of data in the real world (including the data from transit spectroscopy) is unlabeled, meaning nature does not provide us with a set of answers for the atmospheric parameters. We explore a number of unsupervised tasks (standardizing and transforming the data, correlation analysis, dimensionality reduction, manifold learning, clustering and anomaly detection), which can be implemented as part of inverision pipelines for large exoplanet transit surveys. None of those techniques requires any knowledge of the underlying physics and chemistry of the unobserved atmospheres and can serve as a useful pre-processing step. Advantages of this unsupervised approach include: improved ML performance; better understanding of the structures in data; proper planet classification and categorization.

Comments
0
comment
No comments here