🔭 astrophysics

SE3D: Testing the recovery of stellar population, dust and structural properties on mock-observed toy model and simulated galaxies

This paper introduces and validates SE3D, a novel radiative transfer-based modeling framework that successfully recovers key physical properties of toy and TNG50 simulated galaxies from mock observations with high accuracy, while also identifying model mismatches in star formation histories as a primary limitation and revealing discrepancies between simulated and observed dust-to-stellar mass evolution.

Original authors: Junkai Zhang, Steven Ramnichal, Stijn Wuyts, Cheng Li

Published 2026-05-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Junkai Zhang, Steven Ramnichal, Stijn Wuyts, Cheng Li

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to figure out what a car is made of, how fast it's going, and how much fuel it has, but you can't open the hood or look inside. You can only look at the car from a distance, see how bright its headlights are, how big it looks, and how its color changes depending on the angle you view it from.

This is essentially what astronomers face when studying distant galaxies. They can't touch them or see inside them; they only see the light that reaches Earth. This light is a mix of starlight and dust, and the dust acts like a foggy window, distorting the view and changing the colors.

The Problem: A Foggy Window
In this paper, the authors are testing a new tool called SE3D. Think of SE3D as a super-smart detective that tries to reverse-engineer a galaxy's secrets. It looks at the light coming from a galaxy (its "Spectral Energy Distribution" or SED) and its shape (how big it is and how round or flat it looks) across many different colors of light, from ultraviolet to radio waves.

The tricky part is that dust doesn't just block light; it scatters it and re-emits it. Also, stars aren't all the same age, and they aren't spread out evenly. This creates a complex puzzle where the "mass-to-light ratio" (how much actual stuff is there for how much light we see) changes from the center of the galaxy to the edges.

The Solution: A Virtual Training School
To train their detective (SE3D), the authors created a massive library of "toy models." Imagine these as thousands of virtual, simplified galaxies built in a computer. They know exactly what these toy galaxies are made of: how much dust, how many stars, how old the stars are, and how everything is arranged. They then used a high-tech physics simulator (called SKIRT) to calculate exactly what these toy galaxies would look like if we observed them from Earth.

They then trained a Machine Learning "emulator" on this library. Think of the emulator as a student who has studied millions of these virtual galaxies and learned to recognize patterns: "Oh, if the galaxy looks this red and this flat, it probably has this much dust and these old stars."

The Test: Two Types of Challenges
The authors put SE3D through two types of tests:

The "Toy" Test: They took the virtual toy galaxies, simulated observations of them (adding a little bit of "noise" to mimic real telescope errors), and asked SE3D to guess the properties.
- Result: SE3D did an excellent job. It could guess the total mass of the stars, the amount of dust, and how fast new stars are being born with very high accuracy (within about 10-20% error). It could also guess the size of the galaxy and how thick the disk of stars and dust is. Even without high-resolution images, just looking at the total light across all colors, it could still guess the size reasonably well.
The "Realistic" Test: They then tried SE3D on galaxies from TNG50, a famous, complex cosmological simulation that tries to mimic the real universe. These galaxies are messy, with complex histories and irregular shapes, unlike the neat "toy" models.
- Result: SE3D still did a great job guessing the big numbers: the total mass of stars, the total dust, and the size of the galaxy. However, it struggled a bit more with the finer details, like the exact age of the stars in different parts of the galaxy.
- Why? The authors found that the main reason for the struggle was a "mismatch" in the story of how the stars were born. The toy models assumed a smooth, simple history of star formation, while the TNG galaxies had a chaotic, bumpy history. It's like trying to guess the plot of a complex, twist-filled movie by only knowing the plot of a simple fairy tale. The biggest source of error was the difference in the "Star Formation History" (SFH).

The Reality Check: Simulations vs. Real Life
The authors also noticed something interesting about the TNG simulations themselves. When they looked at the simulated galaxies, they didn't look quite like the real galaxies we see in the sky.

The Color Problem: Real galaxies can get very red (dusty), but the simulated ones tended to stay a bit too blue.
The Dust Problem: The simulations didn't seem to produce enough dust relative to the stars as the universe got older.
The Conclusion: The simulations might be missing some physics about how dust forms or how it is distributed. The "toy models" the authors built actually covered a wider range of colors and dust levels than the simulations did, which is why the toy models were a better testbed for the tool.

The Takeaway
The paper concludes that SE3D is a powerful new tool. It can successfully take a blurry, multi-colored picture of a galaxy and tell us how much stuff is in it, how big it is, and how thick it is, even if we don't have a perfect, high-definition image.

However, the tool is only as good as the "toy models" it was trained on. If the real universe is much more complex than the simple models (like having a chaotic history of star formation), the tool's guesses become less precise. The authors suggest that to get even better results in the future, we need to make the "toy models" more complex to match the messy reality of the universe.

Technical Summary: SE3D: Testing the recovery of stellar population, dust and structural properties on mock-observed toy model and simulated galaxies

Problem Statement
Reconstructing the evolutionary histories of galaxies requires translating direct observables (photometry and structure) into physical properties (stellar mass, dust mass, star formation history, and geometry). This process is complicated by the fact that star-dust geometry and inhomogeneous stellar populations induce spatial variations in the mass-to-light ratio. Conventional Spectral Energy Distribution (SED) modeling often treats galaxies as single points or uses simplified attenuation laws that do not self-consistently account for 3D radiative transfer (RT). While 3D RT approaches exist, they are computationally expensive, hindering their use in Bayesian fitting routines required for large datasets. The challenge is to develop a framework that can efficiently model the complex interplay between stellar populations and dust distribution to recover intrinsic properties from multi-wavelength, spatially resolved observations.

Methodology
The authors evaluate SE3D, a novel modeling framework introduced in a companion paper (Ramnichal et al. 2026, R26). SE3D utilizes a machine learning (ML) emulator trained on a library of 3D radiative transfer calculations (performed with the SKIRT code) to rapidly predict spectral distributions.

Framework: SE3D jointly fits galaxy-integrated SEDs and "Spectral Structural Distributions" (SSDs)—wavelength-dependent measurements of effective radii, Sérsic indices, and projected axis ratios.
Test Objects: The study employs two distinct datasets for recovery tests:
1. Toy Model Galaxies: Parametrized galaxies with log-normal star formation histories (SFHs), specific dust geometries, and metallicity distributions. These form the basis of the emulator's training library.
2. TNG50 Simulated Galaxies: Massive star-forming galaxies (SFGs) from the TNG50 cosmological simulation ( $z=0, 1, 2$ ). These possess complex, self-consistent SFHs and spatial distributions. Dust is assigned to gas particles in post-processing based on temperature, density, and metallicity (following Torrey et al. 2012 and De Vis et al. 2019).
Mock Observations: Synthetic observations are generated using SKIRT RT calculations. These are converted into 1D spectral distributions (flux, size, Sérsic index, axis ratio) across a wide wavelength range (UV to 870 $\mu$ m) covering HST, JWST, Spitzer, Herschel, and ALMA bands. Gaussian noise is added to simulate measurement uncertainties.
Fitting: An affine-invariant Markov Chain Monte Carlo (MCMC) sampler (emcee) is used to fit the mock data. The emulator predicts observables for a given set of physical parameters (stellar/dust mass, size, thickness, SFH gradients, viewing angle, metallicity, and birth cloud covering fraction).
Ground Truth Definition: For TNG galaxies, "truth" is defined by projecting the particle data onto the best-fitting toy model geometry (axisymmetric structure) to derive comparable parameters (e.g., fitting a Sérsic profile to the TNG mass distribution to define the "true" $n$ ).

Key Contributions and Results

Recovery Performance on Toy Models:
- SE3D demonstrates encouraging performance in recovering intrinsic properties from mock toy models, even with finite wavelength sampling and measurement errors.
- Bulk Properties: Stellar mass ( $M_*$ ), dust mass ( $M_{dust}$ ), and Star Formation Rate (SFR) are recovered with high accuracy ( $\lesssim 0.1$ dex scatter) and negligible systematic offsets.
- Structural Properties: Half-light radii for stars, dust, and SFR are recovered to within $\sim 0.1$ dex.
- Complex Parameters: Mass-weighted ages and age gradients show larger scatter but significant correlation. Stellar metallicity and birth cloud covering fraction are more difficult to constrain but remain statistically significant.
- Data Requirements: The study finds that even without spatially resolved data (using only integrated SEDs), structural parameters like radii can be recovered to $\sim 0.2-0.3$ dex precision, provided the physical makeup matches the toy model library. This is attributed to the sensitivity of the far-IR Rayleigh-Jeans tail to dust mass and the sensitivity of short-wavelength reddening to dust column density (which depends on size).
Recovery Performance on TNG50 Galaxies:
- When applied to the more complex TNG50 galaxies, SE3D maintains good recovery for bulk properties ( $M_*$ , $M_{dust}$ , SFR, and radii) with similar accuracy ( $\sim 0.1$ dex) to the toy models.
- Model Mismatch: Recovery of parameters sensitive to the specific shape of the SFH and density profiles (e.g., Sérsic indices, age gradients) shows increased scatter and systematic offsets. This is attributed to "template mismatch" between the simplified toy model assumptions and the complex reality of TNG galaxies.
- Mismatch Analysis: The authors quantify four sources of mismatch: SFH, metallicity, radial profile shape, and azimuthal structure. They find that SFH mismatch is the dominant contributor to recovery residuals. Radial and azimuthal mismatches play a secondary role, while metallicity mismatch (assuming mono-metallicity in toy models vs. gradients in TNG) has a negligible impact on the recovery of other parameters.
Intrinsic Properties of TNG50 SFGs:
- Evolution: TNG50 SFGs show a trend of "inside-out" growth, with age gradients becoming more negative from $z=2$ to $z=0$ .
- Geometry: Dust disks are consistently more radially extended than stellar disks ( $R_{dust}/R_{star} \approx 1.6-2.0$ ) and geometrically thinner ( $C_{dust}/R_{dust} < C_{star}/R_{star}$ ).
- Observational Discrepancies: When placed in diagnostic diagrams ($UVJ$ and IRX- $\beta$ ), TNG50 galaxies occupy a narrower dynamic range than observed galaxies. Specifically, they lack the reddest $V-J$ colors and the full spread in IRX- $\beta$ seen in observations. The authors attribute this partly to a lack of evolution in the dust-to-stellar mass ratio ( $M_{dust}/M_*$ ) in TNG compared to observations and potentially inefficient dust reddening due to star-dust geometries.

Significance and Limitations
The paper claims that SE3D offers a computationally efficient framework for jointly analyzing panchromatic photometry and structural data to infer 3D dust and stellar distributions. The method successfully recovers key physical quantities for both idealized and complex simulated galaxies, validating its utility for future large-scale surveys (e.g., Euclid, Rubin-LSST, Roman).

However, the authors modestly highlight several limitations:

Template Mismatch: The recovery accuracy for complex parameters (like detailed SFHs) is limited by the simplifications inherent in the toy model library (e.g., log-normal SFHs, axisymmetry).
Simulation Realism: The TNG50 mock observations do not fully reproduce the color distributions of real observed galaxies (specifically the reddest $V-J$ colors), suggesting that current cosmological simulations may still struggle with dust physics (formation, destruction, and spatial distribution relative to stars).
Data Constraints: While structural parameters can be inferred from SEDs alone in the context of the toy models, this relies on the assumption that real galaxies share the same physical correlations. The authors caution that without spatially resolved data, constraints on viewing angles and Sérsic indices become weak or unconstrained.

In conclusion, SE3D provides a robust tool for extracting physical properties from multi-wavelength data, but its application to real galaxies requires careful consideration of model mismatch and the limitations of current cosmological simulations in reproducing the full diversity of observed galaxy properties.

More like this