Galactic Alchemy: Deep Learning Map-to-Map Translation in Hydrodynamical Simulations

TL;DR

This work pioneers multi-domain map-to-map translation in galaxy formation simulations by leveraging high-resolution IllustrisTNG data to train conditional GANs and diffusion models across seven astrophysical domains. It demonstrates that GANs can rival diffusion models in fidelity while offering substantially lower training and inference costs, especially for tightly coupled mappings like gas to dark matter or HI content, though mappings involving stellar mass remain highly challenging due to intrinsic non-locality. The authors introduce physics-aware evaluation metrics that capture structural realism and mass distribution fidelity, providing a robust framework beyond standard computer-vision metrics. The study offers practical guidance for model selection and design in forward modelling and observational reconstruction, with implications for interpreting SKA-era 21-cm and HI observations and enabling scalable, physics-conscious generative surrogates for galaxy formation.

Abstract

We present the first systematic study of multi-domain map-to-map translation in galaxy formation simulations, leveraging deep generative models to predict diverse galactic properties. Using high-resolution magneto-hydrodynamical simulation data, we compare conditional generative adversarial networks and diffusion models under unified preprocessing and evaluation, optimizing architectures and attention mechanisms for physical fidelity on galactic scales. Our approach jointly addresses seven astrophysical domains - including dark matter, gas, neutral hydrogen, stellar mass, temperature, and magnetic field strength - while introducing physics-aware evaluation metrics that quantify structural realism beyond standard computer vision measures. We demonstrate that translation difficulty correlates with physical coupling, achieving near-perfect fidelity for mappings from gas to dark matter and mappings involving astro-chemical components such as total gas to HI content, while identifying fundamental challenges in weakly constrained tasks such as gas to stellar mass mappings. Our results establish GAN-based models as competitive counterparts to state-of-the-art diffusion approaches at a fraction of the computational cost (in training and inference), paving the way for scalable, physics-aware generative frameworks for forward modelling and observational reconstruction in the SKA era.

Figures (8)

• Figure 1: Samples from various models and tasks. Each panel shows a model input map on the left, the corresponding ground truth in the middle, and prediction on the right. Qualitative comparison confirms the alignment of astrophysical plausibility and human perception with astrophysical metrics and FID (see Tables \ref{['tab:map_to_map_results_pixelwise']} and \ref{['tab:map_to_map_results_astroph']}).
• Figure 2: Examples of normalized asymmetry error maps for the mappings gas$\rightarrow$21cm (top) and gas$\rightarrow$stars (bottom) in the test set, inferred by GANs. The overall mean error is around an order of magnitude larger for gas$\rightarrow$stars and relatively uniform but exhibits a slight chequerboard pattern, indicating the difficulty to model the fine-grained structure of the stellar mass distribution. gas$\rightarrow$21cm exhibits smaller irregular errors which are noticeable due to overall lower average error.
• Figure 3: Examples of normalized clumpiness error maps for the mappings gas$\rightarrow$dm (top) and dm$\rightarrow$gas (bottom) in the test set, inferred by GANs. To keep numerical stability, the inner regions of the error maps have been masked to 5% of the map's respective half-mass radius. The overall mean error is around an order of magnitude larger for dm$\rightarrow$gas, indicating the increased difficulty of predicting baryonic components from DM compared to the inverse mapping. Moreover, due to the collisionless nature of DM, its distributions tend to be smoother, which also contributes to the lower mean error. For gas$\rightarrow$dm, errors mainly arise due to the wrong estimate of DM substructure in the haloes, whereas errors for dm$\rightarrow$gas indicate unrealistic fragmentation in small-scale structures.
• Figure 4: Examples of the angular distribution of COM drifts for the mappings gas$\rightarrow$hi (top) and gas$\rightarrow$stars (bottom) in the test set, inferred by DDPMs. While the upper wind rose diagram shows a uniform distribution for gas$\rightarrow$hi COM drifts, gas$\rightarrow$stars exhibits an angular bias towards 0$\textdegree$. The concentration of these errors in lower offset bins (in units of R$_{50}$), as shown for gas$\rightarrow$hi, indicates low overall drift and typically good agreement with the ground truth.
• Figure 5: Global statistics of inferred vs true integrated quantities. The colour scheme qualitatively indicates histogram density and matches the task assignment analogous to Figure \ref{['fig:samples']}. In general, GANs and DDPMs show no biases and minimal scatter of integrated quantities (except for gas$\rightarrow$stars).