Abstract DGP2026-32

Planets form from the same dust and gas cloud as their host star. Consequently, the refractory composition of rocky planets is expected, in principle, to resemble that of the host star, albeit modified by formation and migration processes. Measurements of the stellar Fe/H ratio can therefore help, in combination with planet mass and radius, to significantly constrain planetary interior structures. However, conventional interior retrieval methods are typically computationally expensive and time consuming, which limits their applicability to large-scale population studies.

We here employ our machine-learning-based framework ExoMDN (Baumeister & Tosi 2023) to construct an interior retrieval model that accounts for planet mass, radius, orbital distance, and stellar Fe/H ratio. ExoMDN requires a large training dataset of synthetic planets, which we generate by combining a condensation model to estimate the chemical compositions of stars and planets (Carone et al., 2025) with a rocky planet interior structure model (Noack & Lasbleis, 2020). In total, we produce more than one million synthetic planets with randomly sampled stellar and planetary masses, orbital distances, and stellar metallicities. The trained model rapidly infers probability distributions of possible interior structures within fractions of a second, enabling the large-scale characterization of rocky exoplanet interiors.