🔭 astrophysics

SE3D: Building a radiative transfer emulator to fit panchromatic resolved galaxy observations with 3D models of dust and stars

The paper introduces SE3D, a computationally efficient machine learning framework that utilizes a Bayesian neural network trained on 3D radiative transfer simulations to simultaneously and self-consistently model the spectral energy distributions and spatially resolved structural properties of galaxies across the panchromatic spectrum.

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

Published 2026-05-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Steven Ramnichal, Junkai Zhang, 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 understand the inside of a complex, glowing machine just by looking at it from the outside. You can see how bright it is, what color it glows, how big it looks, and how round or flat it appears. But you can't see the gears, the fuel, or the dust clogging the pipes inside. This is exactly the challenge astronomers face when studying distant galaxies. They see the light, but they need to figure out the hidden physics: how much dust is there, how old the stars are, and how the stars and dust are arranged in 3D space.

For decades, scientists have used "toy models" to guess the answer. They build a simplified digital version of a galaxy, run a massive, slow computer simulation to see how light travels through it (a process called radiative transfer), and compare the result to what they see in the sky. The problem? These simulations are incredibly slow. Running one takes hours or even days. If you want to test millions of different possibilities to find the best match, you'd need a supercomputer the size of a planet and centuries of time.

Enter SE3D: The "Galaxy Translator"

This paper introduces a new tool called SE3D (Spectral Energy Distribution and 3D). Think of SE3D as a super-smart, lightning-fast translator that has learned to speak "Galaxy" without needing to run the slow, heavy simulations every single time.

Here is how it works, broken down into simple steps:

1. The Training Camp (The Library)

First, the scientists built a massive library of 80,000 "toy galaxies." These aren't real galaxies; they are digital constructions with adjustable knobs. They can change the amount of dust, the size of the galaxy, the age of the stars, and how the stars and dust are mixed together.

The Analogy: Imagine a chef baking 80,000 different cakes. Some have lots of chocolate, some have little; some are tall, some are flat. For every cake, they carefully measure exactly how the light reflects off the frosting and the crumbs. This is the "slow simulation" part.

2. The Machine Learning Emulator (The Shortcut)

Once they had the data from all those slow simulations, they trained a Machine Learning (ML) emulator. This is a type of artificial intelligence (specifically a Bayesian Neural Network).

The Analogy: Instead of baking a new cake from scratch every time you ask for one, the chef has memorized the patterns. Now, if you say, "I want a cake with 20% more chocolate and it needs to be taller," the emulator instantly predicts exactly how that cake will look and taste, based on what it learned from the 80,000 real cakes.
The Speed: The original simulations took about 4.5 minutes each on a powerful computer. The emulator does the same job in 0.018 seconds. That's a speed-up of over 100,000 times!

3. The Fitting Process (Solving the Puzzle)

Now, astronomers can use this fast emulator to study real galaxies. They take the light, size, and shape measurements from a real galaxy and ask the emulator: "Which combination of knobs (dust amount, star age, 3D shape) creates a match?"

The Analogy: It's like a detective trying to solve a crime. Instead of testing every suspect one by one (which takes forever), the detective has a super-fast computer that can instantly simulate the outcome of every suspect's alibi. The computer quickly narrows it down to the most likely culprit.

What Did They Discover?

The paper doesn't just build the tool; it tests how well it works and what it teaches us about galaxies:

It's Accurate: The emulator is incredibly precise. When it predicts the brightness, size, or shape of a galaxy across different colors of light (from ultraviolet to infrared), it is usually within 5% of the "real" slow simulation.
It Learned the Rules: The scientists checked if the AI actually understood the physics or just memorized numbers. They found that the emulator correctly learned complex relationships. For example, it knows that if you tilt a galaxy edge-on (like a coin seen from the side), it looks redder and dimmer because you are looking through more dust.
The "Dust Density" Rule: One of the key findings is about something called the attenuation ratio (how much dust dims light vs. how much it changes the color). The paper found that the most important factor determining this is the projected dust surface mass density.
- The Analogy: Think of dust like a fog. If the fog is thick and dense (high surface density), it doesn't just make things dim; it changes the color of the light in a specific way. The paper shows that how "thick" the dust layer looks from our viewpoint is the main driver of these color changes, more so than the specific shape of the galaxy.

Why Does This Matter?

Before SE3D, studying the detailed 3D structure of distant galaxies was too slow to do for large numbers of objects. Now, astronomers can take a telescope image of a galaxy, run it through SE3D, and quickly learn:

How the stars and dust are arranged in 3D.
If the galaxy is growing from the inside out or the outside in.
How the dust is distributed relative to the stars.

In short, SE3D acts as a bridge. It connects the slow, heavy physics of how light travels through dust with the fast, data-rich observations from modern telescopes, allowing scientists to "see" the hidden 3D structure of the universe much more clearly.

Technical Summary of SE3D: Building a Radiative Transfer Emulator to Fit Panchromatic Resolved Galaxy Observations

Problem Statement
Traditional galaxy Spectral Energy Distribution (SED) modeling relies on integrated light and often imposes simplifying assumptions, such as universal attenuation laws, uniform foreground dust screens, and isotropic emission. These approximations break down when analyzing galaxies in detail, particularly at high redshifts where spatially resolved observations across UV-to-submm wavelengths are becoming available. Current methods struggle to simultaneously model integrated fluxes and wavelength-dependent structural parameters (size, Sérsic index, axis ratio) while accounting for complex 3D star-dust geometries and radial stellar population gradients. Furthermore, full 3D dust radiative transfer (RT) calculations are computationally expensive, rendering them impractical for Bayesian inference within Monte Carlo Markov Chain (MCMC) fitting frameworks.

Methodology
The authors introduce SE3D, a framework designed to simultaneously and self-consistently model a galaxy's SED and its spectral distributions of global structural parameters (wavelength-dependent size $R_e(\lambda)$ , Sérsic index $n(\lambda)$ , and projected axis ratio $q(\lambda)$ ). The core of SE3D is a Machine Learning (ML) emulator trained to reproduce the outputs of 3D dust RT calculations.

Toy Model Definition: The framework utilizes a library of parametrized "toy model" galaxies. These models feature:
- Stellar Components: Axisymmetric oblate ellipsoids with log-normal star formation histories (SFHs) that allow for radial gradients in peak formation time and width.
- Dust Components: Diffuse dust distributions and compact "birth clouds" surrounding young stars. The geometry (size, thickness, Sérsic index) of dust is independent of the stellar distribution.
- Parameters: 16 input parameters define stellar mass, size, structure, SFH, metallicity, dust mass, dust geometry, birth cloud properties, and observer inclination.
Radiative Transfer Library: The authors generated a library of 61,261 physical toy models using the SKIRT code. Each model was processed with 3D dust RT to produce 3D data cubes at 512 wavelengths. From these, four spectral distributions (SDs) were extracted: Flux ( $F_\lambda$ ), Size ( $R_e$ ), Sérsic index ( $n$ ), and Axis ratio ( $q$ ).
Emulator Construction:
- Dimensionality Reduction: Principal Component Analysis (PCA) was applied to the SDs to reduce the 512 wavelength elements into a small set of basis coefficients (4–12 components depending on the SD) plus a zero-point.
- Architecture: A Bayesian Neural Network (BNN) was trained to predict these PCA coefficients and zero-points from the 16 input physical parameters. The BNN architecture allows for the quantification of both aleatoric and epistemic uncertainty.
- Training: The network was trained on 84% of the library, validated on 8%, and tested on 8%. Hyperparameters were tuned using Optuna to minimize the Normalized Median Absolute Deviation (NMAD) between predictions and SKIRT ground truth.
Fitting Framework: The trained emulator is integrated into an MCMC fitting algorithm (emcee). During the fitting process, the emulator predicts rest-frame SDs, which are then redshifted, corrected for Intergalactic Medium (IGM) absorption, and convolved with filter throughput curves to compare with observational data.

Key Contributions

Simultaneous Modeling: SE3D is the first framework to jointly model integrated SEDs and structural spectral distributions (size, profile, shape) within a single 3D RT-based inference pipeline.
Computational Efficiency: The ML emulator achieves a speed-up of over $10^5$ compared to full SKIRT RT calculations. A single prediction takes ~0.018s on a CPU, enabling thousands of calls within an MCMC chain, which was previously infeasible.
High Accuracy: The emulator reproduces SKIRT outputs with high fidelity across the dynamic range of input parameters.
Sensitivity Analysis: The paper provides a comprehensive mapping between physical inputs and observables, demonstrating that the emulator has successfully learned the complex, non-linear relationships between star-dust geometry and observed properties.

Results

Performance Metrics: The emulator predicts the spectral distributions with the following typical accuracies (NMAD in dex):
- Flux (SED): 0.043
- Size (SRD): 0.030
- Sérsic Index (SND): 0.056
- Axis Ratio (SQD): 0.004
- Accuracy is stable across the rest-frame UVJ color space of observed galaxies and the full range of input parameters, with only minor degradation at edge-on inclinations for axis ratios (still < 0.01 dex error).
Physical Mappings: The study confirms that the emulator correctly captures known degeneracies and sensitivities. For instance, it identifies that projected dust surface mass density is the dominant driver of the total-to-selective attenuation ratio ( $R_V$ ), followed by the dust-to-star size ratio ( $R_{dust}/R_{star}$ ), inclination, and stellar Sérsic index.
Training Requirements: The authors demonstrate that emulator performance scales exponentially with the dimensionality of the toy model; doubling the number of input parameters roughly requires doubling the training sample size to maintain accuracy.

Significance and Claims
The paper positions SE3D as a bridge between computationally expensive, high-fidelity 3D radiative transfer and the need for robust statistical analysis of large, multi-wavelength, spatially resolved galaxy datasets (e.g., from JWST and ALMA).

Scientific Utility: By accounting for non-uniform attenuation and 3D geometries, SE3D aims to break degeneracies in deriving intrinsic galaxy properties, such as stellar age gradients and star-dust distributions, which are inaccessible via standard integrated SED fitting.
Methodological Advancement: The work validates the use of BNN-based emulators for complex astrophysical forward modeling, showing that they can learn intricate physical mappings without the need for analytical approximations of dust attenuation.
Limitations and Scope: The authors acknowledge that their toy models simplify real galaxy structures (omitting bars and spiral arms) and assume smooth SFHs. They note that while SE3D fills the gap between zero-dimensional SED fitting and high-resolution individual object modeling, it is not a replacement for the latter in cases requiring detailed sub-galactic structural analysis. The paper explicitly states that further testing on simulated cosmological galaxies (TNG) is presented in a companion paper (Zhang et al. 2025) to assess recovery of intrinsic properties under realistic observational conditions.

1. The Training Camp (The Library)

2. The Machine Learning Emulator (The Shortcut)

3. The Fitting Process (Solving the Puzzle)

What Did They Discover?

Why Does This Matter?

More like this