2nd Place Solution: Bayesian modeling of exoplanet transit-depth measurements

Understanding exoplanet transit depths is essential for characterizing distant worlds and their atmospheres. The NeurIPS Ariel Data Challenge asked competitors to find wavelength-dependent exoplanet transit depths from simulated sensor measurements of the future Ariel satellite. The main challenge was the extremely low signal-to-noise ratio, along with several different types of measurement errors.

In this talk we present our second-place solution, based on a fully Bayesian model of the measurement. Our prior describes the distribution of the various signal and noise elements that make up the signal, such as shot noise, sensor drift, and the actual transit. Using Bayes’ law we then obtain the posterior for a given transit measurement, which indicates how this measurement is described by the signal and noise elements. We can then simply read out the desired transit depth, as well as its uncertainty.

Such a Bayesian model brings several advantages apart from good performance on the competition metric. It is straightforward to include new physics, it provides accurate error estimates including covariance matrices, and is fully explainable.