A graph-based Neural Network surrogate model for accelerating semi-analytical model of galaxy formation and evolution

Xuejie Li, Zhongxu Zhai, Xiaohu Yang, Andrew Benson, Yun Wang · 2026

Read A graph-based Neural Network surrogate model for accelerating semi-analytical model of galaxy formation and evolution by Xuejie Li, Zhongxu Zhai, Xiaohu Yang, Andrew Benson, Yun Wang (2026) 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-2604-23147.txt.

Introduction

Li and four colleagues build a graph neural network surrogate predicting galaxy properties from dark matter merger trees plus semi-analytic model parameters. Trained on Uchuu trees with the Galacticus model, it predicts stellar mass, luminosity, angular momentum, gas metal mass, and star formation rate to redshift 5. Stellar mass predictions to redshift 3 carry scatter near one-quarter dex with determination scores close to one, showing strong agreement with truth. One graph model thus reproduces galaxy traits across many model runs, trees, and redshifts, opening wider parameter searches.

Catalog record

CREATORXuejie Li, Zhongxu Zhai, Xiaohu Yang, Andrew Benson, Yun Wang
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE64417
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.23147

Browse the full library · About the Tower of Babel

Full text

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

A graph-based Neural Network surrogate model for accelerating semi-analytical model of galaxy formation and evolution Xuejie Li, Zhongxu Zhai, Xiaohu Yang, Andrew Benson, Yun Wang Abstract Understanding how galaxy populations emerge and evolve from the growth of dark matter structure is a central challenge in galaxy formation theory. Semi-analytic models (SAMs) provide an efficient framework to address this problem, but exploring large ensembles of merger trees across broad parameter spaces remains computationally demanding. We develop a conditional graph neural network surrogate model that combines merger tree information with SAM parameters to predict galaxy properties across cosmic time. Using merger trees of dark matter halos from the Uchuu simulation and the Galacticus SAM, the model predicts stellar mass, luminosity, angular momentum, gas metal mass, and specific star formation rate across the wide redshift range of 0 ≤ z ≲ 5 0\leq z\lesssim 5 . For instance, the model can predict stellar mass at 0 ≤ z ≲ 3 0\leq z\lesssim 3 with a scatter of 0.19-0.28 dex and coefficient of determination R 2 R^{2} of 0.946-0.973 ( R 2 R^{2} close to 1 indicates prediction closely matching the truth). The results show that a single graph based model can reproduce these galaxy properties with good accuracy over multiple SAM realizations, merger trees and redshifts. This catalog-level model provides a practical route for accelerating SAM based studies of galaxy formation to enable a more detailed investigation of the model parameter space. The inference code, trained models, and example data products are publicly available at https://github.com/MutongCat/sam2galaxy-gnn . I Introduction In our current understanding of galaxy formation and evolution, galaxy properties are shaped by the coupled effects of halo assembly and baryonic evolution across cosmic time. Gas accretion, cooling, star formation, feedback, mergers, and chemical enrichment all contribute to this connection, so a useful forward model must link halo growth to galaxy populations in a physically meaningful way ( 5 ; 46 ) . The current model of galaxy formation is built on this principle and has achieved substantial success in reproducing many observed trends ( 16 ; 45 ; 41 ) , but it remains far from complete in terms of theoretical description ( 5 ; 46 ) and detailed comparisons with observational data ( 21 ; 34 ; 48 ; 43 ) . More broadly, the relation between galaxies and dark matter halos has long been studied with a variety of empirical and statistical models, including halo occupation distributions, conditional luminosity functions, abundance matching, and related empirical galaxy–halo connection frameworks; see 50 for a review, and 55 ; 53 ; 35 ; 44 for representative examples. Among physically motivated forward models, two of the main approaches are hydrodynamical simulations and semi-analytic models (SAMs). Hydrodynamical simulations such as Illustris ( 16 ) , EAGLE ( 45 ) , and IllustrisTNG ( 41 ) evolve a tremendous number of particles of dark matter, gas, and stars self-consistently and have substantially improved comparisons between theory and observation. SAMs instead evolve galaxies in the background of dark matter halos and their merger trees using parameterized prescriptions for cooling, star formation, feedback, enrichment, and satellite evolution, and therefore offer greater flexibility for parameter studies and mock catalog construction ( 51 ; 8 ; 6 ; 46 ) . Although SAMs are flexible, many applications still require repeated model evaluations over broad parameter spaces for calibration and inference ( 21 ; 34 ; 48 ; 43 ) . A common used approach for calibrating SAM parameters relied on Monte Carlo merger trees generated from extended Press–Schechter-type models, instead of using the full merger trees obtained from cosmological simulations. This is because such Monte Carlo trees are inexpensive to produce and enable a fast exploration of the parameter space ( 39 ; 34 ; 43 ) . This approach has been widely used to calibrate SAM parameters by comparing predicted galaxy abundance statistics with observations, for example, through the stellar mass function and luminosity function ( 21 ; 34 ; 48 ) . However, such merger trees are not taken directly from simulations; they are less accurate for applications that require simulation-based assembly histories of dark matter halos or information about their spatial clustering ( 54 ) . For such problems, SAMs constructed from a small set of merger trees derived from cosmological simulations are often the more suitable option. Nevertheless, thoroughly probing the broad SAM parameter space via repeated forward modeling remains computationally costly. This is the regime in which surrogate models are particularly well motivated. Machine learning has become an increasingly important tool for accelerating expensive forward models in galaxy formation and cosmology. This includes applications to simulation emulation, fast nonlinear prediction, and data-driven approximations to complex physical evolution ( 19 ; 10 ; 9 ) . Early studies showed that supervised machine learning models could recover galaxy properties from dark matter halo information and from semi-analytic catalogs with reasonable accuracy ( 52 ; 30 ) . Machine learning has also been used to model the galaxy–halo connection in a more flexible way, beyond prescriptions based solely on halo mass, by incorporating secondary halo properties such as concentration, assembly history, and environment ( 12 ) . Subsequent works used machine learning to populate halos with galaxies, construct fast emulators, and accelerate downstream inference tasks ( 1 ; 29 ; 14 ; 33 ; 19 ; 10 ; 37 ) . These developments have established that learned models can capture nontrivial mappings between dark matter structure, baryonic physics, and observable galaxy properties, and can therefore serve as practical surrogates in settings where repeated model evaluations are required. Recent CAMELS-based work has further highlighted the breadth of galaxy formation parameter space and the nonlinear coupling among astrophysical parameters, cosmology, and galaxy populations ( 36 ) . Together, these developments show that learned surrogates can play a practical role in galaxy modeling pipelines built on large samples. For SAM emulation, the way in which the input is represented can affect the information that the model can use ( 17 ; 28 ; 7 ) . Galaxy properties at a given redshift depend not only on the instantaneous halo state but also on the hierarchical assembly history of the host halo, encoded by the merger tree ( 4 ) . Using only a tabulated set of halo properties or summary statistics may therefore lose information that is relevant to the prediction task. Since merger trees are non-Euclidean objects with explicit progenitor–descendant relations, graph neural networks (GNNs) can be a natural candidate for learning on such structured inputs. GNNs operate by passing and aggregating information along graph edges, allowing the representation of each node to depend on its local neighborhood, through repeated message passing, and on the broader graph structure ( 17 ) . This makes them well-suited to merger trees, where the relevant information is distributed across the full assembly history rather than concentrated on a small set of summary variables. Recent work has shown that GNNs can learn directly from merger trees and predict galaxy properties from halo assembly histories ( 28 ; 7 ) . More broadly, related graph-based generative models also suggest that graph representations are useful for galaxy population modeling beyond standard regression settings ( 26 ;

Read the complete text