Accounting for Planet Radius Resolves Simulation-Based Atmospheric Retrieval Collapse on Real JWST Spectra
What Was Observed (Empirical Data)
Researchers analyzed exoplanet transmission spectra from the James Webb Space Telescope (JWST) across two instruments for three targets, including archival WASP-39b data from the Mikulski Archive for Space Telescopes. Initial amortized simulation-based inference collapsed on real reduced WASP-39b spectra, showing an effective sample size of 1 and a reduced chi-squared of 301 despite succeeding on synthetic benchmarks. Adding planetary radius as an explicit latent parameter recovered physical atmospheric retrievals consistent with nested sampling, improving the reduced chi-squared to 0.06.
The Reality Check (Anti-Hype Analysis)
This study is an algorithmic correction to neural retrieval pipelines rather than a discovery of new physical phenomena or atmospheric constituents on WASP-39b. The initial model failure was caused by omitting an existing geometric parameter during training rather than incomplete radiative-transfer physics. In addition, the method has currently been demonstrated on only three targets across two JWST instruments, requiring broader validation on diverse exoplanet types and cloudy atmospheres.
PRIMARY SOURCE:
arXiv Astrophysics (astro-ph)