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

TL;DR

SE3D presents a forward-modeling framework that jointly analyzes panchromatic and spatially resolved galaxy observations by coupling 3D dust radiative transfer with a Bayesian neural network emulator. A large library of toy galaxies processed with SKIRT is used to train the emulator to predict wavelength-dependent spectral distributions and structural properties with about $0.05$ dex accuracy, enabling rapid Bayesian inference of stellar and dust content and geometry. The approach reveals how observables map to physical properties and identifies the main drivers of attenuation (notably the projected dust surface density) while offering orders-of-magnitude speedups over full radiative transfer. This framework enables robust, joint constraints on star-dust geometries from diverse data and provides a scalable path for exploiting upcoming panchromatic, resolved galaxy surveys.

Abstract

We present a framework for analysing panchromatic and spatially resolved galaxy observations, dubbed SE3D. SE3D simultaneously and self-consistently models a galaxy's spectral energy distribution and its spectral distributions of global structural parameters: the wavelength-dependent galaxy size, light profile and projected axis ratio. To this end, it employs a machine learning emulator trained on a large library of toy model galaxies processed with 3D dust radiative transfer and mock-observed under a range of viewing angles. The toy models vary in their stellar and dust geometries, and include radial stellar population gradients. The computationally efficient machine learning emulator uses a Bayesian neural network architecture, and reproduces the spectral distributions at an accuracy of ~ 0.05 dex or less across the dynamic range of input parameters, and across the rest-frame UVJ colour space spanned by observed galaxies. We carry out a sensitivity analysis demonstrating that the emulator has successfully learned the intricate mappings between galaxy physical properties and direct observables (fluxes, colours, sizes, size ratios between different wavebands, ...). We further discuss the physical conditions giving rise to a range of total-to-selective attenuation ratios, Rv, with among them most prominently the projected dust surface mass density.

Figures (14)

• Figure 1: A visual schematic displaying the workflow of SE3D modelling. Top: radiative transfer is applied to toy model galaxies to produce 3D mock data cubes, which are distilled into 4 spectral distributions (flux, size, Sérsic index and axis ratio as a function of wavelength). Middle: a Bayesian Neural Network (BNN) emulator is trained to improve computational efficiency. Bottom: the flexible and efficient emulator, along with a post-processing step, is placed within a fitting framework to extract physical properties from observed galaxies. The post-processing step accounts for redshift, IGM absorption and filter band convolution.
• Figure 2: Effect of varying a small subset of model parameters on the predicted SEDs and SRDs. Each panel shows the result of modifying one parameter while keeping others fixed to the reference model (shown in black and described in Appendix \ref{['app:oneparam']}). From left to right, we vary the dust content, the scale of the galaxy, the radial gradient of the time of peak star formation, and the inclination. Note: because of how we defined parameters #3, #12, and #13 in Table \ref{['tab:param_library']}, varying $R_{\rm star}$ effectively scales the entire toy model galaxy up/down, preserving geometric thickness and the relative distribution of dust and stars. Emulator predictions and uncertainties are plotted in dashed lines and shaded regions, whereas ground truth SKIRT outputs are plotted in solid lines.
• Figure 3: Validation of the emulator. Residuals, in dex, between 75,000 (unseen) SKIRT ground truth spectral distributions and the corresponding emulator predictions are displayed. The blue shaded regions showcase the central 68 percentile with the running median shown as the blue solid line. Integrating over wavelengths, we further mark the 16 (red dashed), 50 (red solid) and 84 (red dashed) percentiles for each spectral distribution. For reference, 1% (5%) errors are marked in dark (light) grey. The NMADs for Flux, Size, Sérsic index and projected axis ratio $q$ are 0.043, 0.030, 0.056, and 0.004 dex, respectively.
• Figure 4: Emulator performance (NMAD) as evaluated on our testing set of 75,000 toy models, for varying training library sizes. The coloured polygons display the range of computed NMADs for a given library size, where for sizes smaller than the full library size, multiple iterations are computed by sampling disjoint subsets of the corresponding size from the remaining data. The red dashed line indicates a 0.05 dex error for reference. Emulator accuracy improves with increasing training sample size.
• Figure 5: Number of training samples required to reach the NMAD performance of 0.043, 0.030, 0.056, and 0.004 dex for Flux, Size, Sérsic, and axial ratio ($q$), as a function of the number of input dimensions defining the toy model. The solid lines correspond to the running mean of the coloured polygons, which display the range of training samples required to reach the target NMAD, depending on the order in which the input dimensions are built up. Required training rows to reach our defined threshold accuracy, on average, double with each added input dimension.