Quantitatively rating galaxy simulations against real observations with anomaly detection

Zehao Jin, Andrea V. Macciò, Nicholas Faucher, Mario Pasquato, Tobias Buck, Keri L. Dixon, Nikhil Arora, Marvin Blank, Pavle Vulanovic

Research output: Contribution to journalArticlepeer-review


Cosmological galaxy formation simulations are powerful tools to understand the complex processes that govern the formation and evolution of galaxies. However, evaluating the realism of these simulations remains a challenge. The two common approaches for evaluating galaxy simulations is either through scaling relations based on a few key physical galaxy properties, or through a set of pre-defined morphological parameters based on galaxy images. This paper proposes a novel image-based method for evaluating the quality of galaxy simulations using unsupervised deep learning anomaly detection techniques. By comparing full galaxy images, our approach can identify and quantify discrepancies between simulated and observed galaxies. As a demonstration, we apply this method to SDSS imaging and NIHAO simulations with different physics models, parameters, and resolution. We further compare the metric of our method to scaling relations as well as morphological parameters. We show that anomaly detection is able to capture similarities and differences between real and simulated objects that scaling relations and morphological parameters are unable to cover, thus indeed providing a new point of view to validate and calibrate cosmological simulations against observed data.

Original languageEnglish (US)
Pages (from-to)3536-3549
Number of pages14
JournalMonthly Notices of the Royal Astronomical Society
Issue number4
StatePublished - Apr 1 2024


  • galaxies: evolution
  • galaxies: formation
  • methods: data analysis
  • methods: numerical
  • methods: statistical
  • quasars: supermassive black holes

ASJC Scopus subject areas

  • Astronomy and Astrophysics
  • Space and Planetary Science


Dive into the research topics of 'Quantitatively rating galaxy simulations against real observations with anomaly detection'. Together they form a unique fingerprint.

Cite this