Extending SOMPZ beyond Spectroscopic Coverage for Roman Weak Lensing

Precision weak lensing cosmology requires accurate galaxy redshift distributions. However, the spectroscopic samples used to characterize these distributions do not fully span the magnitude--color--redshift space of weak lensing source galaxies, necessitating extrapolation. This challenge will be critical at the depth of the Roman Space Telescope's High Latitude Imaging Survey (Roman HLIS). Using simulated LSST Year 4 and Roman HLIS photometry from the Cardinal simulations, we quantify how spectroscopic incompleteness biases photometric-redshift (photo-z) calibration and cosmological constraints. We find that the standard Self-Organizing Map p(z) (SOMPZ) method produces substantial tomographic bin overlaps and biased source redshift distributions. To address these limitations, we develop Roman-SOMPZ, which uses dedicated individual-galaxy photo-z estimators to supply redshift references for tomographic binning and redshift estimation. For the latter, we introduce UMAP-flow-z, an estimator that uses a UMAP embedding of deep-field galaxy colors to infer redshifts for phenotypes lacking spectroscopic representation. Compared with the standard pipeline, Roman-SOMPZ produces better-separated tomographic bins and lower absolute mean-redshift biases in eight of nine bins. Propagating the estimated redshift distributions into cosmological inference, we find that existing spectroscopy is adequate for early-stage weak lensing analyses in the Roman deep fields. Conversely, full medium-tier cosmology will require further improvements in redshift characterization.