Radius Omission, Not Simulator Gaps, Collapses SBI on Real JWST Spectra

The arXiv preprint from Angshuman Chakravertty demonstrates that amortized simulation-based inference trained on radiative-transfer models fails on actual reduced JWST spectra because the planet radius latent is omitted from the inference network even though the simulator encodes it. Nested sampling first ruled out temperature gradients, SO2 opacity, and opacity tables as causes, leaving the posterior collapse (ESS=1) unchanged until radius augmentation was introduced. The fix used optimal-transport maps and importance sampling to calibrate the flow against an independent nested-sampling reference, enabling transfer across instruments and three real targets including end-to-end MAST reductions.

This result exposes a broader pattern: when a simulator contains a variable the inference step silently fixes or ignores, downstream diagnostics misattribute failure to missing physics. Similar latent-omission artifacts have appeared in cosmological SBI applications and in particle-physics unfolding, where fixing an apparently inconsequential parameter restored calibration without altering the forward model. The paper therefore reframes the usual “simulator improvement” narrative as an inference-design problem.

Future pipelines will need to treat radius, and potentially other boundary-condition latents such as stellar heterogeneity or instrumental throughput offsets, as jointly inferred rather than pre-marginalized. Validation on additional self-reduced spectra and blind tests against ground-truth injection-recovery campaigns will determine whether the method generalizes beyond the three targets shown.