CoChat

Kraken reveals itself – the merger history of the Milky Way reconstructed with the E-MOSAICS simulations

Open access

The DOI matches its Crossref record; 125 of the 138 references it cites were found in a registry or an index; 41 of the 79 sources its claims cite were checked for support.

Abstract

Globular clusters (GCs) formed when the Milky Way experienced a phase of rapid assembly. We use the wealth of information contained in the Galactic GC population to quantify the properties of the satellite galaxies from which the Milky Way assembled. To achieve this, we train an artificial neural network on the E-MOSAICS cosmological simulations of the co-formation and co-evolution of GCs and their host galaxies. The network uses the ages, metallicities, and orbital properties of GCs that formed in the same progenitor galaxies to predict the stellar masses and accretion redshifts of these progenitors. We apply the network to Galactic GCs associated with five progenitors: Gaia-Enceladus, the Helmi streams, Sequoia, Sagittarius, and the recently discovered ‘low-energy’ GCs, which provide an excellent match to the predicted properties of the enigmatic galaxy ‘Kraken’. The five galaxies cover a narrow stellar mass range [M⋆ = (0.6–4.6) × 108 M⊙], but have widely different accretion redshifts ($\mbox{$z_{\rm acc}$}=0.57\!-\!2.65$). All accretion events represent minor mergers, but Kraken likely represents the most major merger ever experienced by the Milky Way, with stellar and virial mass ratios of $\mbox{$r_{M_\star }$}=1$:$31^{+34}_{-16}$ and $\mbox{$r_{M_{\rm h}}$}=1$:$7^{+4}_{-2}$, respectively. The progenitors match the z = 0 relation between GC number and halo virial mass, but have elevated specific frequencies, suggesting an evolution with redshift. Even though these progenitors likely were the Milky Way’s most massive accretion events, they contributed a total mass of only log (M⋆, tot/M⊙) = 9.0 ± 0.1, similar to the stellar halo. This implies that the Milky Way grew its stellar mass mostly by in-situ star formation. We conclude by organizing these accretion events into the most detailed reconstruction to date of the Milky Way’s merger tree.

Overview

CoChat Summary

TL;DR

Kruijssen et al. train an artificial neural network on the E-MOSAICS cosmological simulations to infer the stellar masses and accretion redshifts of five Milky Way satellite progenitors from globular cluster ages, metallicities, and orbital properties. They identify the low-energy GC group as Kraken, find that all five accretion events were minor mergers, and reconstruct the Milky Way’s merger tree.

Key findings

  • The neural network predicts satellite stellar masses with scatter \(\sigma[\log_{10}(M_{\star,\mathrm{pred}}/M_{\star,\mathrm{true}})] = 0.41^{+0.05}_{-0.04}\) and accretion redshifts with \(\sigma[\log_{10}(1+z_{\mathrm{acc,pred}}/1+z_{\mathrm{acc,true}})] = 0.13^{+0.01}_{-0.01}\), with validation score \(0.89^{+0.06}_{-0.06}\).
  • The five progenitors span \(M_{\star} = (0.6\text{–}4.6)\times 10^8\,M_\odot\) and \(z_{\mathrm{acc}} = 0.57\text{–}2.65\). Kraken has \(M_{\star}=1.9^{+1.0}_{-0.6}\times 10^8\,M_\odot\) and \(z_{\mathrm{acc}}=2.26^{+0.39}_{-0.45}\); Gaia-Enceladus has \(M_{\star}=2.7^{+1.1}_{-0.8}\times 10^8\,M_\odot\) and \(z_{\mathrm{acc}}=1.35^{+0.26}_{-0.23}\); Sagittarius has \(M_{\star}=2.8^{+1.8}_{-1.1}\times 10^8\,M_\odot\) and \(z_{\mathrm{acc}}=0.76^{+0.22}_{-0.19}\).
  • All accretion events are minor mergers. Kraken is likely the most major merger ever experienced by the Milky Way, with stellar mass ratio \(r_{M_\star}=1{:}31^{+34}_{-16}\) and halo mass ratio \(r_{M_h}=1{:}7^{+4}_{-2}\); its probability of being a major merger is 0.3 per cent.
  • The five progenitors contributed total stellar mass \(\log(M_{\star,\mathrm{tot}}/M_\odot)=9.0\pm0.1\), similar to the stellar halo, and total halo mass \(\log(M_{h,\mathrm{tot}}/M_\odot)=11.57\pm0.04\), or 25–45 per cent of the Milky Way’s virial mass. The paper estimates 35–50 per cent of Galactic GCs formed ex-situ.

Methods

  • The authors train a sequential artificial neural network on 25 E-MOSAICS zoom-in cosmological simulations, using 205 accretion events with \(\log(M_\star/M_\odot)>6.5\) that host GCs. The eight feature variables are the medians and interquartile ranges of GC age, metallicity, orbital apocentre radius, and eccentricity; the two target variables are accretion redshift and host stellar mass at accretion.
  • They run 10,000 Monte Carlo realisations with scikit-learn and Keras, using dense hidden layers, ReLU activation, Adam optimizer, mean squared error loss, and early stopping. The network is applied to Galactic GCs associated with five progenitors following Massari et al. (2019), with small changes, using ages and metallicities from Kruijssen et al. (2019b) and orbits from Baumgardt et al. (2019).
  • They consider all GC membership permutations, combine predictions with the Milky Way stellar mass growth history from Kruijssen et al. (2019b), and use the Moster et al. (2013) stellar-mass–halo-mass relation to infer halo masses and merger halo mass ratios.

Limitations

  • Stated: E-MOSAICS provides only 25 Milky Way-mass galaxies, and the Milky Way’s accretion history is atypical, so the training sample is small; the authors note they do not find strong variations across the suite but cannot guarantee the Milky Way’s particular history is captured.
  • Stated: quoted error bars reflect random uncertainties; comparison to the network scatter suggests additional systematic uncertainties of up to 0.3 dex in stellar mass and up to 0.25 in redshift. GC membership ambiguities affect Sequoia and the Helmi streams more strongly, and GC counts are lower limits.
  • Evident from the text: the method depends on the E-MOSAICS sub-grid GC formation/disruption model and on the simulation definition of accretion, and the application is to a single galaxy, so broader generality remains to be tested with larger simulation volumes.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1093/mnras/staa2452

Cites

125 of 125 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 Validation of the neural network. Left: PDF of validation scores across all 10,000 Monte Carlo realisations of the neural network, for models including di erent subsets of the feature variables as indicated by the legend. The data point with error bar and the annotation in the top right indicate the median and 16th-to-84th percentiles for the experiment including all feature variables. Middle: PDF of the logarithmic scatter of the predicted stellar masses of the satellite progenitors around their true stellar masses. Lines and annotation as in the left-hand panel. Right: PDF of the logarithmic scatter of the predicted accretion redshifts of the satellite progenitors around their true accretion redshifts. Lines and annotation as in the left-hand panel. This figure shows that the neural network performs well, predicting stellar masses ( M ? ) within a factor of 2.5 and accretion redshifts (1 + z acc ) within a factor of 1.35.
Figure 1 Validation of the neural network. Left: PDF of validation scores across all 10,000 Monte Carlo realisations of the neural network, for models including di erent subsets of the feature variables as indicated by the legend. The data point with error bar and the annotation in the top right indicate the median and 16th-to-84th percentiles for the experiment including all feature variables. Middle: PDF of the logarithmic scatter of the predicted stellar masses of the satellite progenitors around their true stellar masses. Lines and annotation as in the left-hand panel. Right: PDF of the logarithmic scatter of the predicted accretion redshifts of the satellite progenitors around their true accretion redshifts. Lines and annotation as in the left-hand panel. This figure shows that the neural network performs well, predicting stellar masses ( M ? ) within a factor of 2.5 and accretion redshifts (1 + z acc ) within a factor of 1.35.From the PDF
Figure 2 Orbital properties of the Galactic GC population, expressed in terms of their apocentre radii and eccentricities. In all panels, we distinguish GCs that formed in the Main progenitor and those that formed in satellites (largely following Massari et al. 2019, with changes as discussed in the text), as well as fossil streams that are likely relics of GCs (taken from Bonaca & Kruijssen 2020), which we do not associate to any specific accretion event here. In each of the panels, the unambiguous members of each satellite progenitor are highlighted, demonstrating that GCs associated with di erent satellite progenitors have di erent orbital properties.
Figure 2 Orbital properties of the Galactic GC population, expressed in terms of their apocentre radii and eccentricities. In all panels, we distinguish GCs that formed in the Main progenitor and those that formed in satellites (largely following Massari et al. 2019, with changes as discussed in the text), as well as fossil streams that are likely relics of GCs (taken from Bonaca & Kruijssen 2020), which we do not associate to any specific accretion event here. In each of the panels, the unambiguous members of each satellite progenitor are highlighted, demonstrating that GCs associated with di erent satellite progenitors have di erent orbital properties.From the PDF
Figure 3 Age-metallicity distribution of the Galactic GC population. In all panels, we distinguish GCs that formed in the Main progenitor and those that formed in satellites (largely following Massari et al. 2019, with changes as discussed in the text). In each of the panels, the unambiguous members of each satellite progenitor are highlighted, demonstrating that GCs associated with di erent satellite progenitors follow di erent tracks in age-metallicity space. Each panel also includes a vertical line indicating the time of accretion inferred in Section 3.2, with the shaded band representing the 1 uncertainty.
Figure 3 Age-metallicity distribution of the Galactic GC population. In all panels, we distinguish GCs that formed in the Main progenitor and those that formed in satellites (largely following Massari et al. 2019, with changes as discussed in the text). In each of the panels, the unambiguous members of each satellite progenitor are highlighted, demonstrating that GCs associated with di erent satellite progenitors follow di erent tracks in age-metallicity space. Each panel also includes a vertical line indicating the time of accretion inferred in Section 3.2, with the shaded band representing the 1 uncertainty.From the PDF
Image 4
From the PDF
Figure 5 PDFs of the predicted accretion redshifts of the satellite progenitors, inferred by applying a neural network trained on the E-MOSAICS simulations to their GC populations. Each PDF shows the distribution across all 10,000 Monte Carlo realisations of the neural network, applied to di erent GC membership permutations as indicated by the legend (also see Table 1). In each panel, the data point with error bar and the annotation in the top right indicate the median and 16th-to-84th percentiles for the experiment using the unambiguous member GCs (corresponding to the solid line in that panel). This figure shows that the five satellite progenitors accreted over a wide redshift range of z acc = 0 : 57 2 : 65, corresponding to lookback times of t acc = 5 : 7 11 : 3 Gyr.
Figure 5 PDFs of the predicted accretion redshifts of the satellite progenitors, inferred by applying a neural network trained on the E-MOSAICS simulations to their GC populations. Each PDF shows the distribution across all 10,000 Monte Carlo realisations of the neural network, applied to di erent GC membership permutations as indicated by the legend (also see Table 1). In each panel, the data point with error bar and the annotation in the top right indicate the median and 16th-to-84th percentiles for the experiment using the unambiguous member GCs (corresponding to the solid line in that panel). This figure shows that the five satellite progenitors accreted over a wide redshift range of z acc = 0 : 57 2 : 65, corresponding to lookback times of t acc = 5 : 7 11 : 3 Gyr.From the PDF
Image 6
From the PDF
Image 7
From the PDF
Image 8
From the PDF
Figure 9 Galaxy merger tree of the Milky Way, inferred by applying the insights gained from the E-MOSAICS simulations to the Galactic GC population. This figure summarises many of the results presented in this paper. The main progenitor is denoted by the trunk of the tree, coloured by stellar mass (based on the reconstruction from Kruijssen et al. 2019b; see the colour bar). Black lines indicate the five identified (and likely most massive) satellites, with the shaded areas visualising the PDFs of the accretion redshifts from Figure 5. The coloured circles indicate the stellar masses at the time of accretion (both with their colours and sizes) and the subtly di erent colours of the outer rings and central dots visualise the uncertainties on the stellar masses. The annotations list the minimum number of GCs brought in by each satellite. Light grey lines illustrate the global statistics of the Milky Way's merger tree inferred by Kruijssen et al. (2019b), but have no absolute meaning. Thin lines mark tiny mergers (with mass ratio rM ? < 1:100) and thick lines denote minor (or possibly major) mergers (with mass ratio rM ? > 1:100). This merger tree is consistent with the stellar mass growth history of the Milky Way, the total number of mergers ( N br), the number of high-redshift mergers ( N br ; z > 2), and the numbers of tiny and minor mergers ( N < 1:100 and N 1:100 1:4) from Kruijssen et al. (2019b), as well as with the five identified satellite progenitors discussed in this work, including their stellar masses, accretion redshifts, and GC populations. Note that only progenitors with masses M ? > 4 : 5 10 6 M are included. From left to right, the six images along the top of the figure indicate the identified progenitors, i.e. Sagittarius, Sequoia, Kraken, the Milky Way's Main progenitor, the progenitor of the Helmi streams, and Gaia -Enceladus.
Figure 9 Galaxy merger tree of the Milky Way, inferred by applying the insights gained from the E-MOSAICS simulations to the Galactic GC population. This figure summarises many of the results presented in this paper. The main progenitor is denoted by the trunk of the tree, coloured by stellar mass (based on the reconstruction from Kruijssen et al. 2019b; see the colour bar). Black lines indicate the five identified (and likely most massive) satellites, with the shaded areas visualising the PDFs of the accretion redshifts from Figure 5. The coloured circles indicate the stellar masses at the time of accretion (both with their colours and sizes) and the subtly di erent colours of the outer rings and central dots visualise the uncertainties on the stellar masses. The annotations list the minimum number of GCs brought in by each satellite. Light grey lines illustrate the global statistics of the Milky Way's merger tree inferred by Kruijssen et al. (2019b), but have no absolute meaning. Thin lines mark tiny mergers (with mass ratio rM ? < 1:100) and thick lines denote minor (or possibly major) mergers (with mass ratio rM ? > 1:100). This merger tree is consistent with the stellar mass growth history of the Milky Way, the total number of mergers ( N br), the number of high-redshift mergers ( N br ; z > 2), and the numbers of tiny and minor mergers ( N < 1:100 and N 1:100 1:4) from Kruijssen et al. (2019b), as well as with the five identified satellite progenitors discussed in this work, including their stellar masses, accretion redshifts, and GC populations. Note that only progenitors with masses M ? > 4 : 5 10 6 M are included. From left to right, the six images along the top of the figure indicate the identified progenitors, i.e. Sagittarius, Sequoia, Kraken, the Milky Way's Main progenitor, the progenitor of the Helmi streams, and Gaia -Enceladus.From the PDF

Verification

Claims the paper makes

29 Valid · 6 Not valid · 29 Not checked · 152 Own results and interpretations

130 more claims · 11 with their cited sources checked

Create a free account to see all 216 claims and to verify the 4 references still open.

See all claims — free account
  • Probably not a substantial claim. 16
  • It may state two separate claims as one. 66
  • The quote only partly states this claim. 3

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

1 of 1

Citations

Cites

Research works printed but not linked: 8

  1. Bastian, 2020, MNRAS
  2. Bonaca, 2020, ApJ
  3. Buitinck, 2013, ECML PKDD Workshop: Languages for Data Mining and Machine Learning, 108
  4. Chollet, 2015, Keras
  5. Horta, 2020, MNRAS
  6. Pedregosa, 2011, J. Mach. Learn. Res., 12, 2825
  7. Pfeffer, 2020, MNRAS
  8. Trujillo-Gomez, 2020, MNRAS

Cited by

See all 252

Metadata

Authors
Kruijssen, J. M. D., Pfeffer, J. L., Chevance, M., Bonaca, A., Trujillo-Gomez, S., Bastian, N., Reina-Campos, M., Crain, R. A., & Hughes, M. E.
Affiliations
  • Harvard University (US)
  • Liverpool John Moores University (GB)
  • Heidelberg University (DE)
Corresponding author
Kruijssen, J M Diederik
Venue
Monthly Notices of the Royal Astronomical Society
Published
2020-08-17
Volume
498
Issue
2
Pages
2472-2491
ISSN
  • 0035-8711 (ISSN-L)
  • 1365-2966
OpenAlex
W3009673653
Cited by
252
References
138
Citations by year
2026: 25 · 2025: 32 · 2024: 42 · 2023: 56 · 2022: 41 · 2021: 37 · 2020: 19
FWCI
17.05
Type
journal article
Language
en
Open access
Locations
  • Monthly Notices of the Royal Astronomical Society (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • Liverpool John Moores University (submittedVersion)
  • arXiv (Cornell University)
Funders
  • California Department of Fish and Game (KR4801/1-1, KR4801/2-1)
  • H2020 European Research Council (714907)
Topics
  • Physical Sciences › Physics and Astronomy › Astronomy and Astrophysics › Galaxies: Formation, Evolution, Phenomena
  • Physical Sciences › Physics and Astronomy › Astronomy and Astrophysics › Stellar, planetary, and galactic studies
  • Physical Sciences › Physics and Astronomy › Astronomy and Astrophysics › Astrophysical Phenomena and Observations
Keywords
globular clusters, satellite galaxies, Gaia-Enceladus, Helmi streams, Sagittarius dwarf galaxy, Sequoia, minor mergers, major mergers, stellar halo, in situ star formation, artificial neural networks

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol and qwen3.8-omni-flash).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.