Complete Multi-Band Structural Characterization for 10 Million Galaxies in the Hyper Suprime-Cam Wide Survey using Deep Learning

Abstract

Galaxy structure is wavelength-dependent, as different photometric bands trace distinct stellar populations and structural components. We present per-galaxy Bayesian posterior distributions of the bulge-to-total light ratio ($L_B/L_T$), effective radius ($R_e$), and flux ($F$) in the $g$, $r$, and $i$ bands for nearly $10$ million galaxies across $\sim$1400\,deg$^2$ of the Hyper Suprime-Cam Subaru Strategic Program (HSC-SSP) Wide survey at $z \leq 0.75$ and $m < 23$. To our knowledge, this is the largest multi-band catalog of bulge--disk structural parameters with per-galaxy posteriors. $\sim$3.9 million of the above galaxies have full multi-band coverage. Measurements are generated using the Galaxy Morphology Posterior Estimation Network (\gampen{}), a Bayesian deep-learning framework. We demonstrate that we can adapt \gampen{} from previously-trained single-band models to the new bands by using transfer learning on only $\sim$0.5\% of our dataset. Despite this extremely small training set, \gampen{} successfully reproduces light-profile fitting measurements and its output posteriors are extremely well-calibrated with $\lesssim6\%$ mean deviation. This approach offers a scalable pathway to structural parameter measurements with robust uncertainties for Rubin, Euclid, and Roman data, which we will pursue in forthcoming work. While structural measurements remain robust across most bands and redshift ranges, performance degrades significantly for the $g$-band at $z \gtrsim 0.5$, where it increasingly probes rest-frame ultraviolet emission; these measurements are therefore excluded from the public release. This public catalog will enable large-scale statistical studies of morphological color gradients, wavelength-dependent galaxy structure, and galaxy evolution.