CosmoGEMS: Cosmological GC Stream Modeling

• CosmoGEMS is a framework that self-consistently models globular cluster streams by integrating detailed cluster evolution with a time-dependent galactic potential.
• It combines collisional dynamics from CMC and orbit integration in a FIRE-derived host to generate realistic stream features emerging from tidal stripping.
• Demonstration cases reveal diverse stream morphologies, where episodic stripping and orbital phase changes lead to both long thin streams and diffuse shell-like structures.

CosmoGEMS, short for Cosmological Globular cluster streams Exploration and Modeling with Simulations, is a post-processing framework for generating globular-cluster stellar streams star by star inside a fully cosmological, Milky Way-like environment. Its defining aim is to self-consistently link small-scale cluster physics with large-scale Galactic dynamics by combining cluster formation and collisional evolution with orbit integration of escaped stars in a time-dependent Galactic potential extracted from the same cosmological simulation. In this formulation, stream morphology is not imposed through idealized particle-spray prescriptions alone; it emerges from the coupling of GC evolution, stripping history, and hierarchical host-galaxy assembly (Panithanpaisal et al., 3 Sep 2025).

1. Definition and astrophysical setting

CosmoGEMS was introduced to address a specific modeling gap in GC-stream theory: conventional stream methods often simplify either the collisional internal dynamics of globular clusters or the time-dependent structure of the host galaxy. The framework is designed to include both. On the cluster side, the relevant processes include two-body relaxation, strong encounters, stellar and binary evolution, and tidal stripping. On the galaxy side, the host is not static: its halo is triaxial, its disk evolves, satellites perturb it, and a progenitor’s orbit can change over time (Panithanpaisal et al., 3 Sep 2025).

The observational motivation is explicit. More than a hundred stellar streams are already known around the Milky Way, with roughly $\sim 80$ thought to come from globular clusters, and Euclid, Rubin/LSST, and Roman are expected to reveal many more, potentially thousands of Galactic and extragalactic GC streams. This makes a cosmological stream framework operationally important: if interpretation continues to rely on static or idealized potentials, theoretical inference may lag behind the incoming survey data (Panithanpaisal et al., 3 Sep 2025).

The host system used in the proof-of-concept implementation is the FIRE-2 cosmological zoom-in halo m12i, simulated with the magnetohydrodynamic version of GIZMO. At $z=0$, the simulated host has stellar mass

$M_\star = 6.7\times 10^{10}\,M_\odot$

and halo mass

$M_{200m} = 1.2\times 10^{12}\,M_\odot,$

values stated to be broadly consistent with Milky Way mass estimates. The baryonic particle mass is about

$m_b \sim 7000\,M_\odot,$

which is too coarse to form ordinary globular clusters directly, so the GC population is introduced through post-processing rather than through direct formation in the simulation (Panithanpaisal et al., 3 Sep 2025).

2. Formation of the cluster population and collisional evolution

CosmoGEMS builds on the two earlier Great Balls of FIRE stages. In GBoF1, a cluster formation model is constructed from the giant molecular cloud population in m12i. Because FIRE resolves GMC bulk properties only down to roughly $\gtrsim 10^5\,M_\odot$, a calibrated statistical prescription maps each GMC’s mass, size, and metallicity to a GC population. In GBoF2, the resulting clusters are evolved to $z=0$ with the Cluster Monte Carlo (CMC) code (Panithanpaisal et al., 3 Sep 2025).

The GBoF2 realization contains 895 clusters evolved with CMC. The code uses Hénon’s Monte Carlo method, assuming spherical symmetry for the cluster while statistically treating many two-body encounters. The evolution includes a time-varying external tidal field, stellar evolution and binary evolution through COSMIC, a Kroupa IMF from $0.08$ to $150\,M_\odot$, and an initial 10% binary fraction, with binary mass ratios uniform between $0.1$ and $z=0$0. A single star particle associated with the parent GMC is used as a tracer to follow the cluster’s orbit through the galaxy across all snapshots (Panithanpaisal et al., 3 Sep 2025).

A central point is that GBoF2 already tracks cluster mass loss, but in that earlier setup escaped stars were simply removed. CosmoGEMS converts those escapers into the initial conditions of stellar streams. This makes the framework a bridge between collisional GC evolution and resolved stream construction rather than a separate stream toy model layered on top of an arbitrary orbit prescription (Panithanpaisal et al., 3 Sep 2025).

The paper also notes an important caveat: the cluster formation model in this realization does not reproduce the age-metallicity distribution of the oldest Milky Way GCs well, likely because the early star-formation history of m12i differs from that of the real Milky Way. This suggests that the framework is physically rich, while the specific cluster population should not be read as an exact Milky Way analog (Panithanpaisal et al., 3 Sep 2025).

3. Pipeline and dynamical formalism

The CosmoGEMS workflow is described as a three-stage pipeline. First, clusters are formed in the cosmological simulation through the GMC-based statistical model. Second, each cluster is evolved collisonally with CMC to $z=0$1. Third, each escaped star is orbit-integrated from its removal time to the present in the combined potential of the host galaxy and the evolving cluster (Panithanpaisal et al., 3 Sep 2025).

The host-galaxy potential is reconstructed directly from FIRE snapshots through a basis-function expansion. For dark matter plus hot gas ($z=0$2), the potential is modeled with a spherical harmonic expansion up to $z=0$3: $z=0$4 For stars plus cold gas ($z=0$5), the potential is represented in cylindrical coordinates with azimuthal modes up to $z=0$6: $z=0$7 These potential models are constructed using AGAMA. The force model is snapshot-based and linearly interpolated in time; over the last $z=0$8 Gyr this corresponds to about 100 snapshots, or a spacing of roughly 20 Myr (Panithanpaisal et al., 3 Sep 2025).

Because the host-centered frame is non-inertial, CosmoGEMS spline-fits the host center-of-mass acceleration in each Cartesian direction and includes that frame acceleration. The coordinate-frame orientation is held fixed so that the present-day disk normal defines the $z=0$9-axis. The paper states that the actual disk tilts by less than $M_\star = 6.7\times 10^{10}\,M_\odot$0 over the period relevant to the stream integrations. Orbit-validation tests show that the replayed orbits recover the correct $M_\star = 6.7\times 10^{10}\,M_\odot$1 and $M_\star = 6.7\times 10^{10}\,M_\odot$2, with some phase drift and an apocenter mismatch of $M_\star = 6.7\times 10^{10}\,M_\odot$3 kpc even for the harder case (Panithanpaisal et al., 3 Sep 2025).

The cluster self-gravity is represented by an evolving Plummer sphere,

$M_\star = 6.7\times 10^{10}\,M_\odot$4

where $M_\star = 6.7\times 10^{10}\,M_\odot$5 is the cluster mass from CMC and $M_\star = 6.7\times 10^{10}\,M_\odot$6 is the time-dependent Plummer scale radius. This is an approximation to the real cluster density profile, adopted because CMC assumes spherical symmetry and because a smooth model is sufficient to allow stars to leave through the Lagrange-point geometry (Panithanpaisal et al., 3 Sep 2025).

A key distinction is made between ejection time and escape time. In CMC, a star is marked stripped when

$M_\star = 6.7\times 10^{10}\,M_\odot$7

and that defines the ejection time $M_\star = 6.7\times 10^{10}\,M_\odot$8. However, the star is often still physically inside the cluster at that moment. CosmoGEMS therefore defines a more observationally meaningful escape time $M_\star = 6.7\times 10^{10}\,M_\odot$9 as the first time the star reaches a fixed cluster-centric threshold,

$M_{200m} = 1.2\times 10^{12}\,M_\odot,$0

For the example clusters, this exceeds the actual tidal radius at all times, so a star crossing 100 pc is treated as genuinely escaped (Panithanpaisal et al., 3 Sep 2025).

The tidal radius itself is estimated from the largest eigenvalue of the tidal tensor,

$M_{200m} = 1.2\times 10^{12}\,M_\odot,$1

For stars stripped in the last 2 Gyr, CMC provides the speed relative to the cluster and the distance from the cluster at injection. Because CMC does not provide a preferred direction in a non-spherical escaping geometry, CosmoGEMS assigns angular position and velocity direction from isotropic distributions before evolving the star under the combined cluster and host potentials (Panithanpaisal et al., 3 Sep 2025).

A plausible implication is that CosmoGEMS should be read as a compromise between dynamical realism and practical replay accuracy: it imports detailed collisional stripping histories from CMC, but the launch geometry of individual escapers is still approximated.

4. Demonstration cases: GC1 and GC2

The proof-of-concept analysis follows two accreted clusters, GC1 and GC2, both surviving to $M_{200m} = 1.2\times 10^{12}\,M_\odot,$2 and both on eccentric orbits. Their properties are summarized below (Panithanpaisal et al., 3 Sep 2025).

Cluster	Key properties	Stream outcome
GC1	$M_{200m} = 1.2\times 10^{12}\,M_\odot,$3, $M_{200m} = 1.2\times 10^{12}\,M_\odot,$4, age $M_{200m} = 1.2\times 10^{12}\,M_\odot,$5 Gyr, $M_{200m} = 1.2\times 10^{12}\,M_\odot,$6 kpc, $M_{200m} = 1.2\times 10^{12}\,M_\odot,$7, $M_{200m} = 1.2\times 10^{12}\,M_\odot,$8 kpc	Long, thin stream
GC2	$M_{200m} = 1.2\times 10^{12}\,M_\odot,$9, $m_b \sim 7000\,M_\odot,$0, age $m_b \sim 7000\,M_\odot,$1 Gyr, $m_b \sim 7000\,M_\odot,$2 kpc, $m_b \sim 7000\,M_\odot,$3, $m_b \sim 7000\,M_\odot,$4 kpc	Thin component plus diffuse shell-like component

The orbital eccentricity proxy used throughout is

$m_b \sim 7000\,M_\odot,$5

For both clusters, the escape histories are strongly episodic and peak near pericenter, indicating that tidal stripping, rather than steady evaporation, dominates stream production in these examples (Panithanpaisal et al., 3 Sep 2025).

GC1

The present-day GC1 stream is built from 44,045 stars and forms a long, thin stream extending about

$m_b \sim 7000\,M_\odot,$6

in the stream-aligned coordinates used in the paper. Its velocity dispersion is broadly consistent with observed Milky Way GC streams: it is generally below $m_b \sim 7000\,M_\odot,$7, with a minimum near the progenitor of

$m_b \sim 7000\,M_\odot,$8

The trailing arm is more extended than the leading arm, and the stream also shows feather-like substructures near the progenitor (Panithanpaisal et al., 3 Sep 2025).

A particularly prominent feature is a clump in the trailing arm near

$m_b \sim 7000\,M_\odot,$9

By decomposing the stream by stripping episode, the paper shows that this overdensity contains stars from multiple different episodes, mostly ejected $\gtrsim 10^5\,M_\odot$0–$\gtrsim 10^5\,M_\odot$1 Gyr ago, rather than from a single release event. The proposed interpretation is that it may be associated with orbit-specific interactions, perhaps involving Galactic substructure or the disk, and that its visibility is enhanced because it lies near orbital apocenter (Panithanpaisal et al., 3 Sep 2025).

GC2

The present-day GC2 stream is built from 19,206 stars and is morphologically different. It forms a thin stream-like component plus a dominant diffuse shell-like component, a combination described as similar to features observed in streams like Jhelum. Only about

$\gtrsim 10^5\,M_\odot$2

of the escaped stars belong to the thin component; most reside in the shell-like structure (Panithanpaisal et al., 3 Sep 2025).

The local kinematics are correspondingly hotter than in GC1. For the thin component, the typical local velocity dispersion is

$\gtrsim 10^5\,M_\odot$3

while for the shell-like component the overall dispersion is

$\gtrsim 10^5\,M_\odot$4

The paper interprets this as an orbital-phase effect: at the present day the progenitor is near its slow apocenter, the thin tail is near fast pericenter, and the stars have slightly longer orbital timescales than the progenitor, allowing phase offsets to accumulate (Panithanpaisal et al., 3 Sep 2025).

5. Cosmological stream dynamics and methodological consequences

One of the central results of CosmoGEMS is that time dependence of the Galactic potential qualitatively changes streams. The paper quantifies orbital-plane evolution through the tilt angle

$\gtrsim 10^5\,M_\odot$5

where $\gtrsim 10^5\,M_\odot$6 is the unit orbital angular momentum vector at a given time and $\gtrsim 10^5\,M_\odot$7 is the present-day direction. Over the integration interval, GC1 changes its orbital plane by more than $\gtrsim 10^5\,M_\odot$8 and GC2 by more than $\gtrsim 10^5\,M_\odot$9 (Panithanpaisal et al., 3 Sep 2025).

This produces orbital-phase-dependent stream-track misalignment. Near the progenitor, the GC1 stream track is fairly well aligned across stripping times when the progenitor is near pericenter, but around apocenter different stripping episodes visibly diverge, with the effect strongest exactly at apocenter. The interpretation offered is that recently stripped stars track the progenitor’s current orbit reasonably well, whereas stars stripped $z=0$0 Gyr earlier evolved on a somewhat different orbit because the progenitor’s orbital plane, pericenter, and apocenter changed over time. Near apocenter, orbital compression enhances the apparent separation between streamlets; near pericenter, orbital stretching suppresses it (Panithanpaisal et al., 3 Sep 2025).

This result directly challenges several standard assumptions in stream analysis. The paper argues that in a fully cosmological setting, real streams need not trace a single orbit, and the discrepancy can depend on both orbital phase and escape time. It therefore identifies limitations in methods such as STREAMFINDER, pole-count / great-circle methods, and Hough Stream Spotter, all of which rely on more restrictive geometric or dynamical expectations. The authors suggest that more agnostic machine-learning approaches such as VIA MACHINAE may perform better in some cases because they do not hard-code the assumption that stream stars delineate the progenitor orbit (Panithanpaisal et al., 3 Sep 2025).

Because the streams are generated star by star, CosmoGEMS also allows mass-based selections to be tested explicitly. For GC1, selecting only higher-mass stars sharply reduces the number of visible stream stars, but the normalized spatial PDFs along and across the stream remain nearly identical. This suggests that, under perfect membership selection, inferred stream width and length would not strongly depend on survey depth, although measured velocity dispersions could become noisier or biased in shallower samples (Panithanpaisal et al., 3 Sep 2025).

6. Scope, limitations, and nomenclatural ambiguity

CosmoGEMS is presented as an advance over both particle-spray models and static analytic Milky Way potentials, but it is not a fully direct cosmological star-cluster calculation. It remains a post-processing framework. Several approximations are explicit: CMC assumes spherical symmetry; escaper directions at injection are isotropized; the progenitor is represented by a single Plummer model; and stream integrations are restricted to the last 2 Gyr because orbit-reconstruction errors grow with lookback time. The proof-of-concept also studies only two example clusters, so its morphological conclusions are suggestive rather than population-level (Panithanpaisal et al., 3 Sep 2025).

In this respect, CosmoGEMS differs from another GEMS-related framework, the Globular cluster Extra-tidal Mock Star catalogue. That catalogue is a simulated library of extra-tidal stars and recoil binaries ejected from GC cores by three-body dynamical encounters, sampled as 50,000 single extra-tidal stars per cluster for 159 Galactic globular clusters and integrated in seven Milky Way potential models. By contrast, CosmoGEMS derives escapers from collisional GC evolution with CMC and places them in a time-evolving cosmological host potential extracted from FIRE (Grondin et al., 2023).

A recurrent source of confusion is nomenclature. In some usage, “CosmoGEMS” is applied informally to the Cosmic Gems Arc, the strongly lensed high-redshift galaxy behind SPT-CL J0615-5746. That object is distinct: it is a spectroscopically confirmed galaxy at $z=0$1 with ultra-compact star clusters resolved by JWST, not a GC-stream framework. The literature itself notes that if someone says “CosmoGEMS” in that context, they are referring to the Cosmic Gems system rather than to a distinct source. The two topics therefore share a similar label but belong to different research programs (Bradley et al., 2024, Messa et al., 24 Jul 2025).

Within the stream-modeling literature, the significance of CosmoGEMS lies in a narrower but technically consequential claim. It pushes GC-stream theory away from “idealized tracers in a chosen Milky Way potential” toward a framework that links cluster evolution, star-by-star stripping histories, and host-galaxy assembly in a single cosmological environment. This suggests that clumps, asymmetries, stream-track divergence, and thin-plus-diffuse multi-component morphologies are not necessarily anomalies; in a time-dependent cosmological host, they may be natural outcomes of GC evolution (Panithanpaisal et al., 3 Sep 2025).