Challenges in the detection of gases in exoplanet atmospheres

TL;DR

The paper challenges the reliability of gas detections in exoplanet atmospheres by showing that Bayesian evidence is highly sensitive to the chosen model space and reference model. It demonstrates, through a K2-18 b case study with JWST/MIRI data, that detections claimed in limited, degenerate model spaces can disappear when the hypothesis space is broadened, and that even modest Bayesian preferences may not signify a true atmospheric constituent. The authors argue for treating model comparisons as relative adequacy tests, supplemented by null-hypothesis tests and physically motivated models, to avoid overinterpretation in low-SNR data. They advocate broader spectral coverage, higher fidelity data, and a multi-framework inferential approach to advance reliable atmospheric characterization in the JWST era.

Abstract

Claims of detections of gases in exoplanet atmospheres often rely on comparisons between models including and excluding specific chemical species. However, the space of molecular combinations available for model construction is vast and highly degenerate. Only a limited subset of these combinations is typically explored for any given detection. As a result, apparent detections of trace gases risk being artifacts of incomplete modeling rather than robust identification of atmospheric constituents, especially in the low signal-to-noise regime. Using the sub-Neptune K2-18 b as a case study, we show that recent biosignature claims vanish when the model space is expanded, with numerous alternatives providing equally good or better fits. We demonstrate that the significance of a claimed detection relies on the choice of models being compared, and that model preference does not in itself imply the presence of a specific gas. We recommend treating model comparisons instead as relative adequacy tests, which should be supported by theoretical predictions and complementary metrics of statistical significance in order to attribute a signal to a particular gas.

Figures (9)

• Figure 1: Combinatorial growth of model space and dependence of apparent detection significance on reference model.A) Number of unique models and pairwise model comparisons that would be needed to exhaustively search the model space as a function of the number of molecules considered. Exploring any significant fraction of this space for typical models with $>10$ molecules is computationally infeasible. These numbers are lower limits because they do not account for other discrete atmospheric model choices such as cloud and pressure-temperature profile parameterizations. B) Schematic showing how the apparent detection significance from Bayesian model comparison depends on the chosen reference model.
• Figure 1: Posterior model and data realizations for MIRI observations of K2-18 b.A) Posterior data (gray) and model (orange) realizations generated from a flat-line model. Purple points indicate the median retrieved transit depth in each wavelength bin; vertical error bars represent the uncertainty in transit depth and horizontal bars indicate the bin width. B) Residuals between the data realizations and the best-fit flat-line model. C) As panel A, but for the model including a single-kernel Gaussian Process (GP) noise component. The GP identifies correlated structure near 7 µm and 8.5–9 µm, but this additional complexity is not statistically preferred over the flat-line model with white noise. D) Residuals between the data realizations and the best-fit GP model.
• Figure 2: Selected JWST transmission spectra of small exoplanets ($R_p<3R_{\oplus}$). Transmission spectra for seven exoplanets --- GJ 486 bMoran2023, L 98-59 bBello-Arufe2025, GJ 1214 bSchlawin2024Kempton2023b, GJ 1132 bMay2023, L 98-59 dGressier2024, K2-18 bMadhusudhan2023Madhusudhan2025, and LHS 1140 bDamiano2024---observed with JWST. Colors distinguish each planet. Offsets in normalized transit depth are applied for clarity. Vertical error bars show 1$\sigma$ uncertainties (standard deviations) on the measured transit depth; data points without visible bars have smaller errors than the plotted symbols. These spectra have been analysed using Bayesian model comparison to determine preference for an atmospheric model over a featureless spectrum.
• Figure 2: Normalized residuals of the JExoRES and JexoPipe spectra relative to flat-line models. Histograms of normalized residuals, defined as $([{\rm data} - {\rm model}]/\sigma)$, for the JExoRES and JexoPipe reductions compared with their best-fit featureless spectra. Data represent single measurements per wavelength bin (n=1); bars show the frequency distribution of normalized residuals. The black curve indicates the standard Normal (Gaussian) distribution used for comparison.
• Figure 3: JWST/MIRI transmission spectrum and model fits for K2-18 b.A Transmission spectrum of K2-18 b reduced with the JExoRES pipelineMadhusudhan2025. Black points show the mean transit depth per wavelength bin. Vertical error bars represent 1$\sigma$ standard deviations on the measured transit depth; horizontal bars indicate bin widths. The green line shows the median model including CH$_4$, CO$_2$ and C$_3$H$_4$ (propyne) with shaded bands denoting the 1$\sigma$ and 2$\sigma$ posterior intervals. B Normalized residuals ([data-model]/error) for parametric models including combinations of CH$_4$, CO$_2$, C$_3$H$_4$, DMS, and DMDS; self-consistent models considering radiative-convective-thermochemical equilibrium (RCTE) and radiative-convective-photochemical equilibrium (RCPE); a chemical equilibrium model where temperature structure is freely retrieved; and a constant transit depth (flat line) model. The black curve shows the expected normal distribution for comparison.