Galactic Alchemy: Deep Learning Map-to-Map Translation in Hydrodynamical Simulations 2025 Galactic Alchemy: Deep Learning Map-to-Map Translation in Hydrodynamical Simulations – References

Philipp Denzel, Yann Billeter, Frank-Peter Schilling, Elena Gavagnin · 2025

Read Galactic Alchemy: Deep Learning Map-to-Map Translation in Hydrodynamical Simulations 2025 Galactic Alchemy: Deep Learning Map-to-Map Translation in Hydrodynamical Simulations – References by Philipp Denzel, Yann Billeter, Frank-Peter Schilling, Elena Gavagnin (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-23768.txt.

Introduction

Philipp Denzel and three co-authors test whether deep learning models can translate between maps of a simulated galaxy, predicting one physical property from another. Working with high-resolution simulation data, they compare conditional GANs and diffusion models under one shared pipeline. Seven domains are in play: dark matter, gas, neutral hydrogen, stellar mass, temperature, and magnetic field strength.

Beyond standard image scores, the authors add physics-aware metrics that judge structural realism. Difficulty tracked physical coupling: gas-to-dark-matter and gas-to-hydrogen maps came out nearly perfect, while weakly constrained jobs like gas-to-stellar-mass stayed hard. GANs matched the diffusion models' quality at a fraction of the training and inference cost. The authors frame the study as the first systematic multi-domain test of its kind, pointing toward cheap forward modeling for the Square Kilometre Array era.

Catalog record

CREATORPhilipp Denzel, Yann Billeter, Frank-Peter Schilling, Elena Gavagnin
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE90 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2510-23768

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Galactic Alchemy: Deep Learning Map-to-Map Translation in Hydrodynamical Simulations 2025 Galactic Alchemy: Deep Learning Map-to-Map Translation in Hydrodynamical Simulations – References Philipp Denzel, Yann Billeter, Frank-Peter Schilling, Elena Gavagnin 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 H i 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. 1 Introduction The spatial matter distribution of galaxies is the result of a complex and chaotic interaction between its individual components, such as dark matter (DM), stellar populations, (predominantly hydrogen) gas, super-massive black holes (SMBHs), and their surrounding environment. Interactions between these components are governed by their mutual gravitational and electromagnetic forces, and hydrodynamical processes which collectively shape the structure and properties of galaxies over cosmic time ( Binney & Tremaine 2011 ; Tinsley 2022 ; Conselice 2014 ; D’Onofrio et al. 2016 ) . These interactions imprint subtle signatures in the phase-space distribution of a galaxy, retaining the various feedback mechanisms that have influenced its formation and evolution ( Binney & Vasiliev 2023 ; Bassini et al. 2024 ) . Thus, the physical components encode distinct aspects of galaxy evolution: • DM haloes of galaxies dominate the gravitational potential into which baryonic matter flows and forms visible substructure ( White & Frenk 1991 ; Moore et al. 1999 ; Frenk & White 2012 ) . • Stellar mass reflects the cumulative outcome of star formation and feedback, but its distribution is temporally highly non-local and entropic ( McKee & Ostriker 2007 ; Kennicutt & Evans 2012 ; Hwang et al. 2019 ) , while star formation is spatially localized in H 2 clouds within the interstellar medium ( Colman et al. 2024 ; Schinnerer & Leroy 2024 , ISM;) . • Gas traces the baryonic backbone of galaxies, regulating cooling, heating, and star formation through radiative feedback cycles ( Gavagnin et al. 2017 ; Luisi et al. 2021 ) and turbulence induced by active galactic nuclei ( Biernacki & Teyssier 2018 ; Valentini et al. 2019 , AGNs;) , supernovae ( Fielding et al. 2017 ; Ibrahim & Kobayashi 2023 ) , stellar winds ( Krumholz et al. 2014 ; Bally 2016 ) , galaxy-galaxy mergers ( Hopkins et al. 2006 ; Cibinel et al. 2019 ) , or interaction with the intergalactic medium ( Muratov et al. 2017 ; Poggianti et al. 2019 , IGM;) . • Neutral hydrogen and 21cm brightness are key observational tracers of the ISM for low-redshift galaxies, critical for radio surveys with MeerKAT , ASKAP , or the upcoming SKA-Mid ( Maccagni & Blok 2024 ; O’Beirne et al. 2025 ; Maddox et al. 2021 ; de Blok et al. 2024 ; Maccagni & Serra 2025 , such as WALLABY, MIGHTEE- H i , MHONGOOSE, or the MeerKAT Fornax Survey;) . • Temperature captures thermodynamic states shaped by shocks, cooling, and AGN-driven outflows ( Zubovas et al. 2024 ; Ward et al. 2024 ) . • Magnetic fields emerge from turbulent amplification and influence gas dynamics ( Beck 2015 ; Rieder & Teyssier 2017 , e.g.,) , yet remain poorly constrained observationally. Recovering these domains from limited information is challenging; observationally due to technical limitations: for most instruments, signals beyond the local Universe – especially from HI – become too faint due to intrinsic dimming ( Messias et al. 2024 ) , foreground contamination ( Collaboration et al. 2025 ) , and low-frequency RFI ( Harper & Dickinson 2018 ; Engelbrecht et al. 2024 ) . Conversely, theoretical inference of these domains has computational challenges: the understanding of the distribution of matter in the Universe remains largely driven by numerical simulations. Among these, (magneto-)hydrodynamical simulations present the most principled approach to model and capture the non-linear co-evolution of dark and baryonic matter fields across cosmological and astrophysical scales ( Crain & van de Voort 2023 , for recent reviews, see) . However, this quality comes at steep computational costs or forces detrimental trade-offs between resolution and volume. To mitigate these challenges, simpler alternatives such as dark-matter-only (DMO) simulations ( Potter et al. 2017 ; Ishiyama et al. 2021 ; Cheng et al. 2020 , e.g.,) reproduce large-scale structure and halo statistics at reduced cost but omit baryonic physics. Semi-analytical models (SAMs) attempt to compensate by applying post de facto prescriptions to approximate baryonic effects on top of DMO outputs ( Berlind et al. 2003 ; Somerville et al. 2008 ; Schneider et al. 2019 ; Obuljen et al. 2023 , e.g., ) . While these methods enable exploration of cosmological parameter space, they lack the fidelity needed to capture the full complexity of galaxy-scale feedback and morphology. In particular, their intrinsic post-hoc nature often ignores the gravitational back-reaction caused by the redistribution of baryons on the DM field. With the proliferation of deep generative models, a complementary line of research has emerged that seeks to emulate aspects of these simulations rather than compute them from first principles. Recent efforts have explored enhancing simulations and augmenting galaxy models through scalable deep learning techniques in various ways. For instance, Perraudin et al. 2019 use scalable GANs (for details see Section 2.3.1 ) to produce entire N-body 3D cubes of the cosmic DM distribution in a multi-scale approach. Still, techniques aiming for the full 3D reconstruction of cosmological simulations often face challenges in scaling to resolutions where individual galaxies can be resolved. Alternatively, Bernardini et al. 2021 employ Wasserstein-GANs to paint baryons onto thin slices of simulation boxes from the FIRE simulation suite. Li et al. 2021 ; Schanz et al. 2024 use StyleGAN and denoising diffusion models, respectively, to super-resolve cosmic large-scale structure predictions. Thiele et al. 2020 applied a U-Net architecture (for details see Section 2.4 ) to infer observable thermal and kinematic Sunyaev-Zel’dovich maps of haloes from DMO simulations, explicitly linking theory to observations. Similarly, Chadayammuri et al. 2023 use a U-Net for image-to-image translation of IllustrisTNG galaxy cluster haloes to the corresponding baryonic fields. Most studies focus on a single aspect of a simulation’s galaxy formation or feedback model and do not fully reproduce (or harness) all physical modes of simulated galaxies (for details see Section 2 , Equation 2 ). In this paper, we introduce a novel application of deep generative models for map-to-map translation across multiple astrophysical domains in cosmological simulations on the

Read the complete text