Simulation-based inference of galaxy properties from JWST pixels

Abstract

We present an efficient Bayesian SED-fitting framework tailored to multiwavelength pixel photometry from the JWST Advanced Deep Extragalactic Survey (JADES). Our method employs simulation-based inference to enable rapid posterior sampling across galaxy pixels, leveraging the unprecedented spatial resolution, wavelength coverage, and depth provided by the survey. It is trained on synthetic photometry generated from MILES stellar population models, incorporating both parametric and non-parametric SFHs, realistic noise, and JADES-like filter sensitivity thresholds. We validate this amortised inference approach on mock datasets, achieving robust and well-calibrated posterior distributions, with an $R^2$ score of 0.99 for stellar mass. Applying our pipeline to real observations, we derive spatially resolved maps of stellar population properties down to $\mathrm{S/N}_{\rm{pixel}}=5$ (averaged over F277W, F356W, F444W) for 1083 JADES galaxies and ~2 million pixels with spectroscopic redshifts. These maps enable the identification of dusty or starburst regions and offer insights into mass growth and the structural assembly. We assess the outshining phenomenon by comparing pixel-based and integrated stellar mass estimates, finding limited impact only in low-mass galaxies ($<10^8M_{\odot}$) but systematic differences of ~0.20 dex linked to SFH priors. With an average posterior sampling speed of $10^{-4}$ seconds per pixel and a total inference time of ~1 CPU-day for the full dataset, our model offers a scalable solution for extracting high-fidelity stellar population properties from HST+JWST datasets, opening the way for statistical studies at sub-galactic scales.

Figures (17)

• Figure 1: Residuals of properties compared to the medians of the posterior distributions obtained with the $\tau$-delayed model, for $\log_{10} (M_{*}/\rm{M}_{\odot})$, the $\log_{10}(\rm{age}/{yr})$, referring to the mass-weighted age, and $A_V$ [mag]. We split the simulated test sample in bins of mean S/N in the filters F277W, F356W, and F444W, 1-5 (left), 5-10 (middle) and $>10$ (right). We colour-coded each simulation with the standard deviation of the posterior distribution for the three properties and included dashed and dotted lines corresponding to the average one and two standard deviations of the posterior distributions respectively. We also plotted kernel density distribution contours with black solid lines for clarity.
• Figure 2: Same as Fig \ref{['true_vs_pred']} for the Dirichlet model. We show $\log_{10} (M_{*}/\rm{M}_{\odot})$, log$_{10}(t_{50\%})$, referring to the lookback time in Gyr at which 50% of the total stellar mass was formed, and $A_V$ [mag].
• Figure 3: Distribution of $1.083$ galaxies from the JADES survey in the plane of integrated stellar mass versus spectroscopic redshift, with colours indicating the fraction of pixels within two effective radii (in F444W) that have a mean S/N of 5 or greater in the filters F277W, F356W, and F444W. The lines represent the parameter-space limits used to set thresholds at pixel fractions of 0.20 (red dashed line) and 0.80 (red dotted line) averaged across the full sample. These thresholds indicate, for the ranges of integrated masses and redshifts, the fraction of pixels within the galaxies that can be effectively fitted with our model.
• Figure 4: Probability distributions for $\log_{10}(M_{*}/\rm{M}_{\odot})$, the mass-weighted age [Gyr], and $A_{V}$ [mag] for the pixel with highest mean S/N in the filters F277W, F356W, and F444W (in green), and a random pixel out of $1R_{\rm{eff}}$ (in red) for the six galaxies. Step histograms (solid lines) represent the prior distributions (in black) and our posterior distributions (in green and red). The posterior distributions from Prospector are shown with lower opacity (in green and red). The dotted lines correspond to the minimum $\chi^2$ parameters obtained with the optimisation code Synthesizer.
• Figure 5: Galaxy property maps. We include only the pixels with a S/N (averaged between F277W, F356W, and F444W) higher than 5. We use a $\tau$-delayed prior for the SFHs. From left to right: RGB images (built using the method described in lupton04 with the filters specified above each image to enhance the different structures), and the inferred properties per pixel $\log _{10}\left(\Sigma_{\star} /\left(M_{\odot} \mathrm{kpc}^{-2}\right)\right)$, the mass-weighted age [Gyr], and $A_V$ [mag]. In the lower-left corner of the first column, we show circles with diameters equal to the FWHM of the F444W PSF. Scale bars in the first column indicate angular (arcseconds) and physical (kiloparsecs) sizes.