A hierarchical conditional generative model for planetary systems

To characterise the diversity of planetary systems, we need to go beyond the properties of individual planets to capture the statistical relationships between planets within a system, but to also include the properties of their host star and protoplanetary disk. Physics-based population synthesis frameworks, such as the Bern model, provide valuable insight into the architecture of the planetary systems and correlations therein, but their high computational cost restricts their use for large-scale statistical inference and direct comparison with the observed population of exoplanets.

We present a novel hierarchical conditional generative model designed to produce synthetic planetary systems with both high fidelity and high efficiency. The architecture explicitly conditions on host star and disk properties, including stellar metallicity, disk lifetime, and gas disk mass, enabling the model to reproduce the multi-scale dependencies that shape the formation of these planets and system architectures. By jointly modelling several system- and planet-level properties, it captures more subtle correlations and structural features than the previous model in (Alibert et al. 2025).

The model is trained on synthetic populations from the Bern model rather than on observed catalogues, avoiding the detection biases while preserving the underlying physical relationships. Once trained, it generates new planetary systems orders of magnitude faster than the original simulations, while remaining statistically consistent with the training population. We validate the model by recovering known trends linking stellar and disk properties to planetary system architecture, and by verifying that the generated systems satisfy expected dynamical stability criteria and training probability distributions.

Finally, we illustrate the potential of this framework by applying it in the context of the TESS survey, demonstrating how the model can be used to predict planetary properties and inform the interpretation of current and forthcoming exoplanet observations.