Machine learning as a tool to determine exoplanet properties

Characterization of exoplanetary atmospheres and interiors requires information on planetary properties such as mass, radius, and density. However, precise measurements of these fundamental properties are not always possible, as a result of which the masses and radii of these planets are often unknown. We propose a data driven machine learning method to estimate missing exoplanet properties by applying clustering algorithms to a subset of the currently known exoplanet populations, containing approximately 4000 exoplanets. We applied the Uniform Manifold Approximation Projection (UMAP, McInnes et al., 2018) algorithm, together with a range of clustering techniques, such as Gaussian Mixture Models, K-means and HDBSCAN, to a custom data set, created by merging the NASA exoplanet archive and the exoplanet.eu catalogue. This enables us to create the most complete parameter set, that current observations allow. We show that a combination of different clustering algorithms and trained UMAP models is able to infer estimates on planetary radii and masses, when provided with a large enough training sample. Based on this method, we can provide estimates for radii and masses of Hot Jupiters. Our results deviate from ground truth by 0.06 Jupiter radii and 0.23 Jupiter masses on average. This is achieved for a test data set containing three host star properties (mass, radius, effective temperature) and only the orbital period of the planet. The results show that data-driven methods are a promising approach for parameter estimation in exoplanets, which can be very useful in the target characterization of upcoming missions such as PLATO and Ariel.