Exoplanet atmospheres: AI failed to analyse the actual JWST spectra because it did not know the planet’s radius

A simulation-based inference method, which had proved reliable with synthetic data, failed in the case of WASP-39b. The problem lay not with the physics of the simulator but with an overlooked variable: MIRAGE incorporates this variable and reduces the χ²/N from 301 to 0.06.

Simulation-damped inference, where a grating trained on radiative transfer models deduces the composition of an atmosphere from a spectrum, works very well on simulated spectra from the James Webb Telescope. On real observations, however, it can collapse completely: on the exoplanet WASP-39b, the effective sample size dropped to 1 and the fit reached a χ²/N of 301. Physics is usually missing in the simulator.

Two researchers verified this hypothesis by nested sampling: adding a temperature gradient, the opacity of the SO₂ or a finer set of opacities did not change anything. The real culprit was a latent variable absent from the inference, the radius of the planet. Their method, MIRAGE, integrates this radius into a flow matching model, calibrated to an independent reference, and obtains for WASP-39b a solution that is physically consistent with the literature, with a χ²/N of 0.06.

The method can be applied without modification to two instruments and three real JWST targets, including a spectrum reduced from end to end by the authors from the raw data in the MAST archive. Their conclusion goes beyond astronomy: a variable that the simulator takes into account but that the inference ignores can be mistakenly mistaken for a modeling error.