On the robustness of atmospheric inferences. Estimating the predictive accuracy of atmospheric models for transmission spectra of transiting exoplanets.

Up until now, most exoplanet transit spectroscopy observations are comprised of relatively low signal to noise and sparse wavelength coverage data. As such, atmospheric inferences are currently derived from a limited number of spectral points and, in some cases, one data-point can significantly affect the interpretation of these observations. Furthermore, when interpreting the goodness-of-fit between the data and the atmospheric models, using different statistical metrics can lead to different conclusions. In some cases, while Bayesian evidence model comparisons suggest the detection of an atomic or molecular species, frequentist metrics like the p-value suggest that the chosen atmospheric model is not supported by the data. Can we trust a chemical detection if some metric suggests that the model does not fit the data? This conundrum highlights the need for methods that can assess the sensitivity of the resulting inferences to the data. Such need becomes more urgent with the imminent wave of spectroscopic information coming from the James Webb Space Telescope for which understanding the limits of our models and our data is paramount.

In this work we will present a complementary goodness-of-fit metric applied to transmission spectra of transiting exoplanets. Our work leverages the power of Bayesian atmospheric retrievals to understand which specific data points are responsible for an inferred atmospheric property (e.g., chemical detection), and is robust against considerations in the model (e.g., choice of priors) and noise properties of the data. We will present our computationally practical approach and demonstrate its applicability on recent transmission spectra with unsettled atmospheric detections. We will show how this metric can help us better determine the reliability of an inference and quantify its significance. This complementary approach to estimate the goodness of fit of atmospheric models to the data opens a new avenue to determine the reliability of a given result and inform our understanding of exoplanet atmospheres.