Characterizing the interior structure of exoplanets is an essential part in understanding the diversity of observed exoplanets, their formation processes and their evolution. As the interior of an exoplanet is inaccessible to observations, an inverse problem must be solved, where numerical structure models need to conform to observed parameters such as mass and radius. Since the relative proportions of iron, silicates, water ice, and volatile elements are not known, this is a highly degenerate problem whose solution often relies on computationally-expensive and time-consuming inference methods such as Markov Chain Monte Carlo.
We present here ExoMDN, a new machine-learning-based approach to the interior characterization of observed exoplanets using Mixture Density Networks that improves upon our previous work (Baumeister et al., ApJ, 2020). This improved model, trained on a large database of 5.6 million synthetic interior structures, can make a complete probabilistic inference about possible planetary interior structures within a fraction of a second, without the need for extensive modeling of each exoplanet's interior. We can demonstrate how the model, trained on different sets of (potentially) observable parameters including the received irradiation at the planet’s orbit and the fluid Love number, can help to further constrain the interior of a large number of exoplanets. In particular, we can show how precisely these parameters need to be measured to well constrain the interior.