Demonstrate cosmology-grade line-flux control for learned slitless-deblending models

Demonstrate that variational autoencoders, deep generative galaxy models, normalizing flows, score-based models, and diffusion models can control line-flux bias sufficiently for cosmology-grade slitless spectroscopy while satisfying the proposed detector-level injection-recovery requirements.

Background

The proposed telescope must reconstruct spectra in crowded fields where neighboring galaxy traces overlap. The baseline therefore uses a calibrated detector forward model with multiple dispersion orientations, while learned models are restricted to morphology priors, source masks, and quality-control tests.

The paper notes that existing generative-deblending studies have not demonstrated the line-flux accuracy needed for the emission-line-galaxy BAO and RSD sample. Establishing whether these methods can meet the required bias and recovery limits remains unresolved.