Synthetic JWST galaxy images in the TNG50 simulation - I. Model validation and comparison to observations

TL;DR

This work investigates how well forward-modelled, dust-aware JWST-like images of galaxies from the high-resolution TNG50 simulation reproduce key observational inferences. By coupling SKIRT radiative transfer with realistic instrument realism and applying standard SED fitting to both synthetic and real JWST data, the authors quantify the fidelity of photometric redshifts, stellar masses, UVJ diagnostics, and quiescent fractions in the $3 \le z \le 6$ regime. They find robust redshift recovery up to $z \le 5$ but increasing challenges at $z=6$, and a systematic underestimation of stellar masses that grows with $M_*$, largely driven by dust attenuation. The results support the viability of forward modelling for validating SED-fitting pipelines against JWST data, while also highlighting the need for improved dust treatments and template sets to reconcile color offsets, particularly in $V-J$ and the dusty star-forming regime.

Abstract

We use the TNG50 cosmological simulation and three-dimensional radiative transfer post-processing to generate dust-aware synthetic observations of galaxies at $ 3 \leq z \leq 6 $ and $ \log_{10}(M_\ast/\mathrm{M}_\odot) \geq 8.5 $, tailored to match the depth and resolution of current deep JWST surveys (NGDEEP and JADES). We analyse the performance of spectral energy distribution (SED) fitting on the simulated sample, focusing on the recovery of photometric redshift and stellar mass. At $ z \leq 5 $, we find that 90 per cent of redshifts are recovered within $ \pm0.2 $, but performance declines at $ z = 6 $. Stellar masses are generally well-recovered within a factor of 2, but are systematically underestimated regardless of redshift, a trend that is more pronounced at the high-mass end $ ( \log_{10}(M_\ast/\mathrm{M}_\odot) \geq 10 ) $. In addition, we study the observer-frame colours of galaxies in this redshift range as well as the SED-inferred $UVJ$ diagram. We find that TNG50 galaxies broadly follow the tendencies marked by observations, but tend to be slightly redder at lower masses and bluer at higher masses, regardless of redshift. Finally, using a colour-based definition of quiescence, we determine the fraction of quiescent galaxies as a function of stellar mass at $ 3 \leq z \leq 6 $, which we find to be broadly consistent with observations.

Figures (13)

• Figure 1: First four rows: example TNG50 NGDEEP-like synthetic images in F356W at $z=3, 4, 5$ and $6$. Each row shows galaxies at the same redshift, with objects sorted from left to right by increasing stellar mass (subfind values). We only include close companions in the same host halo as the central object. Last four rows: example NGDEEP galaxies in F356W, ordered from top to bottom by decreasing photometric redshift. As opposed to the synthetic images, these contain a contribution from foreground and background objects that, in most cases, are not physically associated with the main galaxy.
• Figure 2: (a) Violin plots of the photometric redshift distributions of TNG50 galaxies at $z_\text{true} = 6, 5, 4$ and $3$. Each panel shows mirrored kernel density estimates of the predicted redshifts. The horizontal line marks the true redshift; the inner box-and-whisker plot indicates with a thick bar the interquartile range (IQR; 25th to 75th percentile), with thin vertical lines extending to the most extreme point within $1.5 \times \text{IQR}$, and with a white line the median. Summary statistics are listed in each panel. Photometric redshifts are well recovered for $z_\text{true} \leqslant 5$ with near-zero bias and a low outlier rate. At $z_\text{true} = 6$, performance declines, with larger underestimations and more outliers. (b) Residuals of the estimated photometric redshifts versus the true stellar mass (subfind values). Purple triangles mark medians in bins of up to 40;120;240;480 objects for $z_\text{true} = 6, 5, 4$ and $3$, respectively; horizontal bars show the bin widths. Dashed lines indicate the outlier threshold, $\left|\Delta z\right| / (1+z_\text{true}) > 0.15$, with the number of outliers shown in each panel. Residuals remain largely flat with stellar mass, except at the low-mass end $\left( \log_{10} \left( M_\ast / \unit{\msun} \right) \lesssim 9\text{\textendash}10) \right)$, where they are the largest at all redshifts.
• Figure 3: Distribution of stellar mass residuals, $\Delta M_\ast$, for the TNG50 sample at $z_\text{true} = 6, 5, 4$ and $3$. The violin plots show the kernel density estimates of the residuals, while the box-and-whisker plots inside represent the interquartile range (IQR), with the white line indicating the median. The horizontal solid line marks zero residual, and the dashed lines indicate ± 0.3dex (about a factor of 2). Stellar masses are mostly underestimated, but are generally recovered within 0.3dex. The population mean, $\mu$, becomes less negative and the distributions show increased scatter at higher redshifts.
• Figure 4: (a) Fitted versus true stellar masses for the TNG50 subsamples at $z_\text{true} = 6, 5, 4$ and $3$. The colour scale shows the predicted value of V-band attenuation for objects in each bin. The solid line indicates the one-to-one relation, while the dashed lines mark offsets by a factor of 2. (b) Mass residuals as a function of true stellar mass. Purple triangles mark median values, with error bars spanning the 16th to 84th percentiles; horizontal dashed lines indicate ±0.3dex. For low-mass galaxies $\left( \log_{10}\left( M_\ast / \unit{\msun} \right) \lesssim 9 \right)$, residuals are flat with underestimations up to a factor of 2.0. At higher masses $\left( \log_{10}\left( M_\ast / \unit{\msun} \right) \gtrsim 10 \right)$, offsets become increasingly negative even exceeding -1dex. These high-mass galaxies with the largest offsets, exhibit the highest predicted V-band attenuation values. The mass-binned mean $\bar{\mu}$ peaks at $z = 3$, while the scatter $\bar{\sigma}$ decreases with lower redshift. In contrast, using dustless synthetic photometry (blue line) nearly eliminates the systematics, with stellar masses recovered within 0.2dex or less, indicating that dust attenuation is the main driver of the discrepancies. This suggests that the strength of dust attenuation is underestimated in the SED fitting, especially at high masses.
• Figure 5: The UVJ diagram for the TNG50 sample at $z = 3$. The plot shows best-fit predicted $U - V$ and $V - J$ for the sample, colour-coded (from left to right) by the specific star formation rate (sSFR), stellar mass, mass-weighted age $(t_\text{MW})$, V-band attenuation, and metallicity. The dashed and solid lines represent the Williams2009 and Whitaker2011 boundaries between quiescent and star-forming galaxies, respectively. According to these criteria, no objects fall into the quiescent region. Clear trends are seen: $U - V$ colours are inversely related to sSFR (redder colours correspond to lower sSFRs); both stellar mass and mass-weighted age correlate with ${{V - J}}$ colour, with more massive and older galaxies being redder; and V-band attenuation increases with redder UVJ, with the bluest objects having the lowest $A_V$ values, while those more attenuated are located closer to the top-right corner. These trends are consistent with previous studies at lower redshifts. The final panel shows the distribution of objects in this plane, with green contours indicating the 68.0 and 95 levels. Colours are expressed in the AB system.