Improving Posterior Inference of Galaxy Properties with Image-Based Conditional Flow Matching

TL;DR

This work presents a conditional flow matching (CFM) framework that leverages pixel-level imaging alongside photometry to improve posterior inference of galaxy properties and highlights the potential of integrating morphology into photometric SED fitting pipelines.

Abstract

Estimating physical properties of galaxies from wide-field surveys remains a central challenge in astrophysics. While spectroscopy provides precise measurements, it is observationally expensive, and photometry discards morphological information that correlates with mass, star formation history, metallicity, and dust. We present a conditional flow matching (CFM) framework that leverages pixel-level imaging alongside photometry to improve posterior inference of galaxy properties. Using $\sim10^5$ SDSS galaxies, we compare models trained on photometry alone versus photometry plus images. The image+photometry model outperforms the photometry-only model in posterior inference and more reliably recovers known scaling relations. Morphological information also helps mitigate the dust--age degeneracy. Our results highlight the potential of integrating morphology into photometric SED fitting pipelines, opening a pathway towards more accurate and physically informed constraints on galaxy properties.

Figures (4)

• Figure 1: Left: distributions of $\Delta\log p(\theta_\ast;\mathcal{D})$ for the image model (purple, $\mu=2.17, \ \sigma=3.30$) and the photometry model (yellow, $\mu=1.26, \ \sigma=3.98$). Middle: distributions of $D_{\mathrm{KL}}\!\left[p(\theta\!\mid\!\mathcal{D})\,\|\,p(\theta)\right]$ for the image model (purple, $\mu=3.41, \ \sigma=0.95$) and the photometry model (yellow, $\mu=2.55, \ \sigma=0.97$). Right: per-object scatter plot of $\Delta\log p$ vs. $D_{\mathrm{KL}}$ (color indicates model). Panels show $N=1000$ galaxies from the test set. For these objects, the image model attains higher $\Delta\log p$ for 81.5% of objects and higher $D_{\mathrm{KL}}$ for 96.5% of objects compared to the photometry model.
• Figure 2: Image-model predictions (blue) recover known scaling relations in SDSS data (gray) more faithfully than the photometry model (green). Each row shows a different relation: $M_\star$--$Z_{\rm gas}$ (top), $M_\star$--SFR (middle), and SFR--$Z_{\rm gas}$ (bottom). Red boxes mark selections defined on the image-model predictions, with three representative galaxy cutouts shown on the right.
• Figure 3: Corner plots of $A_V$ (dust attenuation) versus $D_n(4000)$ (stellar age proxy) for two representative galaxies. Left: old, dust-poor galaxy. Right: young, dust-rich galaxy. Contours show posterior predictions for the photometry model (green) and image model (blue); red crosses mark spectroscopic targets. The image model moves closer to the spectroscopic target than the photometry model does in both cases, indicating potential to weaken the dust--age degeneracy.
• Figure A.1: Distributions of SDSS galaxy properties and $r$-band magnitudes. The 16th, 50th, and 84th percentiles of the univariate distributions are labeled with dashed lines.