Close the sim-to-real gap for JWST TNO spectra

Improve the agreement between TNFlow inferences and real JWST Trans-Neptunian Object spectra by determining how training-set coverage of trace abundances and simulator fidelity contribute to the failure to recover observed materials and the overprediction of water.

Background

TNFlow is trained exclusively on synthetic spectra generated with the Shkuratov radiative transfer model. When applied qualitatively to JWST observations, it fails to fully recover clearly visible CO2, CO, and CH3OH signatures and assigns H2O as the dominant component even for objects classified by the survey as water-poor.

The paper identifies the sim-to-real discrepancy as its main unresolved challenge. The authors suggest that insufficient representation of trace-abundance components in the training corpus and missing physical effects in the Shkuratov simulator may both contribute, motivating improved training-set coverage and greater simulator fidelity.

References

In summary, TNFlow demonstrates that amortized posterior inference over TNO surface compositions is practical: a sub-second, fully offline surrogate for the Shkuratov RTM that replaces the manual hours-per-object fitting loop and makes the inherent degeneracy of the inverse problem explicit. On synthetic data, composition accuracy is comparable to or better than the disagreements typically reported between independent manual refits. The main open challenge is closing the sim-to-real gap on real JWST spectra, which will require improvements to both training-set coverage and simulator fidelity.

— TNFlow: Amortized Posterior Inference for Trans-Neptunian Object Surface Composition  (2609.04305 - Gaur et al., 3 Sep 2026) in Section 6, Limitations and Future Work; concluding paragraph