1st Place Solution :Bayesian inference for exoplanet transits

The NeurIPS Ariel Data Challenge tasked participants with inferring exoplanet spectral properties from simulated sensor measurements of the upcoming Ariel satellite. The 2025 edition built on the previous year’s challenge, introducing additional physical complexity such as stellar limb darkening and a wider diversity of planetary atmospheres.

In this talk, we present our first-place solution, which applies Bayesian modeling to both the underlying astrophysics and the sensor characteristics. We argue that Bayesian inference is particularly well suited to this problem: it directly incorporates domain knowledge, provides interpretable uncertainty estimates, and exposes the relationships among inferred parameters.

However, the competition also revealed key limitations of the Bayesian approach. Our model’s missing physics became apparent only through comparison with the synthetic ground truth—a luxury unavailable to the real mission. We close by discussing how hybrid strategies that integrate Bayesian and deep learning methods might help overcome these challenges in future work.