Super-resolving galaxy spectra, from JWST to Roman
Low-resolution spectra blend the emission lines used to measure dust, metallicity, and ionization, and high-resolution follow-up is expensive. I built a physics-informed deep-learning pipeline that sharpens JWST/NIRSpec prism spectra by roughly a factor of ten in resolving power. Trained on 2,858 paired prism and grating spectra from JADES, it infers a redshift and then refines emission lines with attention across line tokens, separating features such as the [OIII]λλ4959,5007 doublet that the prism cannot resolve.
To test how much of that sharpening is real, I benchmarked the network against seven classical deconvolution methods on 572 held-out spectra. It appears to win on mean absolute error by 30%, but the margin comes from shrinking line amplitudes; on a scale-free metric all methods lie within 1.2% of one another. The network recovers line widths far better than any classical method, but it also places confident lines where the reference detects none, in 30% of spectra for Hβ and 44% for [OII].
That failure mode shaped the next step: adapting the framework to slitless grism spectra from the Nancy Grace Roman Space Telescope. Trained on 36,404 simulated OpenUniverse2024 spectrum pairs, the Roman model is penalized for drawing lines the data could not have revealed. It recovers 85% of the flux of strong lines while leaving undetectable ones near zero, and its redshift head, which also reads three Roman imaging bands, reaches a 5.1% catastrophic outlier rate.
Advised by Bahram Mobasher & Shoubaneh Hemmati — UC Riverside
