StarEvolve: Multiscale Stellar & AI Evolution

• StarEvolve is a suite of time-resolved modeling frameworks that simulate the evolution of stars, clusters, and planetary systems.
• It employs hybrid techniques—combining N-body, hydrodynamic, and probabilistic methods—to capture complex multi-physics dynamics.
• Beyond astrophysical modeling, StarEvolve extends to AI with a hierarchical framework for strategic decision-making in StarCraft II.

StarEvolve denotes several distinct but conceptually related frameworks in the literature represented here. In astrophysics, the label is applied to multiscale star-cluster dynamics, stellar-evolution emulation, habitability and planetary-survival calculations, and precision analyses of rapidly evolving post-AGB and other evolved stars; in a separate AI context, it names a hierarchical large-language-model framework for StarCraft II. This suggests that “StarEvolve” is best treated not as a single canonical codebase but as a family of time-resolved modeling programs centered on the evolution of stars, clusters, and their environments, with representative implementations including Enzo-N for self-consistent cluster–galaxy evolution (Jo et al., 2024), EMACSS for long-term cluster evolution [(Alexander et al., 2012); (Alexander et al., 2014)], the Dartmouth Stellar Evolution Emulator integrated into CONF1DENCE (Jiaqi et al., 7 Apr 2026), and the SC2Arena StarEvolve agent (Shen et al., 14 Aug 2025).

1. Scope, terminology, and major usages

In this corpus, “StarEvolve” spans several technical domains rather than a single narrowly defined method. The common thread is explicit temporal evolution under coupled physics, whether the evolving object is a star cluster embedded in gas, a stellar track conditioned on uncertain microphysics, a post-AGB star observed in real time, or an agent operating in a sequential decision process.

Usage domain	Representative framework	Core object of evolution
Cluster–galaxy co-evolution	Enzo-N	Star clusters inside parent galaxies
Embedded-cluster formation	Torch in AMUSE	Young gas-rich clusters
Long-term secular cluster modeling	EMACSS	Bound clusters in tidal fields
Stellar track emulation	DSEE + CONF1DENCE	Tracks, isochrones, ages
Habitability and planetary fate	TYCHO/CHAD, STAREVOL, MESA	HZ histories, engulfment, star-lifting
Evolved-star diagnostics	HST/COS, TMAP, PoWR, interferometry	Rapid post-AGB and late-stage evolution
AI decision making	SC2Arena StarEvolve	Full StarCraft II gameplay

A further astrophysical extension places StarEvolve at galactic scale: gas density and star formation rate are modeled as a stochastic Hopf oscillator, implemented with the “Stochastic Hopf Engine,” with Euler–Maruyama integration, a radial Fokker–Planck reduction, differential shear, and AGN-driven quenching (Kumar et al., 26 Feb 2026). That usage remains astrophysical but is structurally different from stellar-track or cluster-dynamics applications.

2. Multiscale star-cluster evolution in gas-rich and galactic environments

A major StarEvolve usage concerns the simultaneous evolution of star clusters and their larger-scale environments. Enzo-N is a GPU-accelerated hybrid hydro/N-body code that couples the cosmological (magneto-)hydrodynamic code Enzo to the direct N-body code Nbody6++GPU through a semi-stationary background acceleration approximation (Jo et al., 2024). The motivation is the large scale separation between galaxies, of order $\sim 0.1\,\mathrm{Mpc}$, and star clusters, of order $\sim \mathrm{pc}$, which makes fully direct treatment computationally prohibitive. Enzo-N solves internal cluster dynamics with a direct N-body solver, including regularization for few-body interactions, while evolving dark matter, gas, and stars outside the cluster with particle-mesh gravity and hydrodynamic methods. The reported demonstrations include core collapse and tidal stripping due to galactic tides, establishing a genuinely self-consistent co-evolution problem rather than a cluster evolved in a fixed external potential.

At earlier evolutionary stages, Torch provides a different StarEvolve implementation for embedded clusters (Cournoyer-Cloutier et al., 2023). Built within AMUSE, it couples FLASH self-gravitating AMR hydrodynamics, PH4 direct stellar dynamics, the Multiples module for stable hierarchies and resonant few-body encounters, sink-particle star and binary formation, SeBa stellar evolution, and radiation and wind feedback through Fervent and Vink wind rates. The simulations follow $10^4\,M_\odot$ turbulent clouds at $0.0683\,\mathrm{pc}$ resolution for approximately $2\,\mathrm{Myr}$ after the onset of star formation, exploring virial parameters $\alpha=0.8, 2.0, 4.0$, multiple IMF realizations, and several primordial-binary prescriptions.

Those simulations emphasize that embedded-cluster growth is not monotonic. Clusters can lose up to half of their mass while still embedded, and morphology can vary on $\sim 0.01\,\mathrm{Myr}$ timescales (Cournoyer-Cloutier et al., 2023). The 3D structure is quantified with reduced-inertia ellipsoids enclosing fixed Lagrangian mass fractions, characteristic radii $\tilde r = (abc)^{1/3}$, and ellipticity

$\epsilon = \frac{a-c}{a}.$

Within this framework, cluster mass is not correlated with $r_{50}$ or with $\sim \mathrm{pc}$0, and for embedded clusters with $\sim \mathrm{pc}$1 the dominant dynamical agent is the global gravitational potential of the star-forming region rather than internal two- or few-body relaxation. A common misconception is therefore that dense, compact young clusters are necessarily relaxation-dominated; these calculations show that in the embedded regime the external gas-plus-stars potential can dominate even when densities are high.

3. Secular, tidal, and cosmological cluster evolution

On longer timescales, StarEvolve is represented by reduced-order but physically calibrated prescriptions. EMACSS, “Evolve Me A Cluster of StarS,” models post-core-collapse cluster evolution in static tidal fields through global energy flow rather than explicit resolution of core few-body dynamics (Alexander et al., 2012). Its central assumption is Hénon-like balanced evolution,

$\sim \mathrm{pc}$2

with relaxation time

$\sim \mathrm{pc}$3

and coupled rates

$\sim \mathrm{pc}$4

In its equal-mass form, EMACSS reproduces direct N-body results to within $\sim \mathrm{pc}$5 over the entire post-collapse evolution for $\sim \mathrm{pc}$6 in isolated and tidal cases (Alexander et al., 2012). The code is explicitly limited by its equal-mass, no-stellar-evolution assumptions and by a static point-mass galactic tide.

The unequal-mass extension adds stellar evolution, an evolving mean stellar mass, a modified relaxation time $\sim \mathrm{pc}$7, induced escape near Roche-volume filling, and preferential ejection of low-mass stars (Alexander et al., 2014). The formalism separates unbalanced early evolution, driven by stellar-evolution mass loss, from balanced evolution reached after about $\sim \mathrm{pc}$8 modified relaxation times. In this regime, the half-mass radius can either expand or contract depending on the Roche filling factor, while preferential low-mass escape increases the mean stellar mass. This version is calibrated against N-body suites spanning initial star number, mass, half-mass radius, and tidal environment, and is intended for rapid population studies that would be prohibitively expensive with direct N-body.

A different tidal StarEvolve program embeds live clusters in cosmological structure formation (Rieder et al., 2013). In that work AMUSE is coupled to the dark-matter-only CosmoGrid $\sim \mathrm{pc}$9CDM simulation, following $10^4\,M_\odot$0 equal-mass clusters from $10^4\,M_\odot$1 to $10^4\,M_\odot$2 in two Milky-Way-mass haloes. The tidal field is represented by the local tidal tensor $10^4\,M_\odot$3, applied through Bridge with interpolation between cosmological snapshots. The principal result is that cluster mass loss is continuous irrespective of the host halo’s tidal history, but major mergers increase the instantaneous rate; “native” clusters, accreted into the main halo before its final major merger, evaporate faster than “immigrant” clusters, which spend more time in weaker pre-accretion environments (Rieder et al., 2013). The study also shows that using a static $10^4\,M_\odot$4 potential overestimates disruption for accreted clusters. This directly qualifies the common simplification that present-day halo structure is an adequate proxy for a cluster’s full tidal history.

4. Stellar tracks, probabilistic emulation, and habitable-zone histories

At the level of individual stars, StarEvolve includes both deterministic track grids and probabilistic emulators. The Dartmouth Stellar Evolution Emulator is a flow-based generative model trained on over eight million evolutionary tracks across twenty input-physics dimensions, with broad coverage in mass and composition (Jiaqi et al., 7 Apr 2026). DSEE learns phase-conditioned stellar state snapshots, treats tracks and isochrones as marginals of one model, and provides continuous interpolation across high-dimensional physics together with probabilistic predictions and calibrated credible intervals. Integrated into the open-source CONF1DENCE package, it supports end-to-end creation of tracks and isochrones and uncertainty-aware age determinations for clusters. Relative to fixed-physics sparse grids, the key conceptual shift is explicit marginalization over uncertain stellar-physics hyperparameters rather than interpolation within a small set of frozen assumptions.

A complementary deterministic StarEvolve resource is the catalog of Truitt et al., built from TYCHO stellar-evolution tracks and CHAD habitable-zone post-processing (Truitt et al., 2015). The grid covers $10^4\,M_\odot$5–$10^4\,M_\odot$6, scaled metallicities $10^4\,M_\odot$7–$10^4\,M_\odot$8, and $10^4\,M_\odot$9–$0.0683\,\mathrm{pc}$0 times solar. Habitable-zone boundaries are computed from stellar luminosity and effective temperature using the Kopparapu et al. parameterization

$0.0683\,\mathrm{pc}$1

and

$0.0683\,\mathrm{pc}$2

The catalog defines both the continuously habitable zone over the full main sequence and a CHZ2, continuously habitable for at least $0.0683\,\mathrm{pc}$3. It shows that metallicity and detailed abundance ratios materially alter luminosity evolution, effective temperature, and habitable-zone migration; in particular, varying $0.0683\,\mathrm{pc}$4 can produce effects comparable to or larger than a $0.0683\,\mathrm{pc}$5 change in bulk metallicity (Truitt et al., 2015). This is an important corrective to workflows that condition only on $0.0683\,\mathrm{pc}$6.

5. Precision evolved-star analysis and real-time stellar evolution

Another StarEvolve axis is the high-precision analysis of evolved stars. The white paper “Precision Analysis of Evolved Stars” argues that late stellar evolution is intrinsically 3D, time-variable, and often asymmetric, so progress depends on combining Gaia astrometry, pulsations and asteroseismology, optical/IR/radio interferometry, ALMA molecular-line imaging, high-resolution spectroscopy, and 3D radiation–hydrodynamics (Ridgway et al., 2019). The observational program is tied to standard relations such as

$0.0683\,\mathrm{pc}$7

and to empirical mass-loss prescriptions including Reimers’ law and Blöcker’s law. The paper’s central claim is methodological: low-resolution, effectively 1D diagnostics systematically miss the clumps, spirals, shocks, disks, and velocity substructure now seen in objects such as Antares, Betelgeuse, R Dor, W Hya, and R Scl.

The Stingray Nebula central star SAO244567, also identified as V839 Ara, provides a benchmark case of StarEvolve on human timescales [(Reindl et al., 2014); (Reindl et al., 2016); (Schaefer et al., 2020)]. Between 1988 and 2002 its effective temperature rose from $0.0683\,\mathrm{pc}$8 to $0.0683\,\mathrm{pc}$9, surface gravity increased from $2\,\mathrm{Myr}$0 to $2\,\mathrm{Myr}$1, the mass-loss rate declined from $2\,\mathrm{Myr}$2 to $2\,\mathrm{Myr}$3, and the terminal wind velocity increased from $2\,\mathrm{Myr}$4 to $2\,\mathrm{Myr}$5 (Reindl et al., 2014). By 2015, HST/COS data showed cooling to $2\,\mathrm{Myr}$6, $2\,\mathrm{Myr}$7, $2\,\mathrm{Myr}$8, and $2\,\mathrm{Myr}$9, implying re-expansion of the envelope and strongly supporting a late thermal pulse rather than a very late thermal pulse (Reindl et al., 2016). A central controversy is that state-of-the-art LTP calculations do not simultaneously reproduce the observed $\alpha=0.8, 2.0, 4.0$0, $\alpha=0.8, 2.0, 4.0$1, $\alpha=0.8, 2.0, 4.0$2, and $\alpha=0.8, 2.0, 4.0$3, so the source remains a challenge to stellar-evolution theory.

The nebular side of the Stingray system is equally rapid. HST imaging from 1996 to 2016 found no evidence for ongoing massive, high-velocity outflows or new shock-brightened structures, while emission lines faded with markedly different half-lives: typically $\alpha=0.8, 2.0, 4.0$4 for [O III], down to $\alpha=0.8, 2.0, 4.0$5 in some outer-shell regions, and $\alpha=0.8, 2.0, 4.0$6 for H$\alpha=0.8, 2.0, 4.0$7 (Schaefer et al., 2020). The preferred interpretation is that a modest fast-wind episode in the 1980s collisionally ionized a pre-existing neutral shell, after which the nebula entered recombination-driven decline. This case demonstrates that post-AGB stellar and nebular evolution can proceed on decade scales, and that broadband photometry alone can be misleading when ionization structure, extinction, and stellar variability evolve simultaneously.

6. Planetary-system consequences and engineered control of stellar evolution

StarEvolve also encompasses the effect of stellar evolution on planetary systems. In STAREVOL-based models of post-main-sequence evolution, orbital response to isotropic adiabatic mass loss obeys

$\alpha=0.8, 2.0, 4.0$8

while RGB and AGB tides, drag, engulfment, and PN photoevaporation compete to determine survival (Villaver, 2011). The equilibrium-tide term scales steeply as $\alpha=0.8, 2.0, 4.0$9, so capture occurs even when the orbit lies several stellar radii outside the photosphere. The paper gives explicit RGB survival thresholds: for a $\sim 0.01\,\mathrm{Myr}$0 star, a Jupiter-mass planet requires $\sim 0.01\,\mathrm{Myr}$1 to avoid engulfment, and a $\sim 0.01\,\mathrm{Myr}$2 planet requires $\sim 0.01\,\mathrm{Myr}$3 (Villaver, 2011). During the PN phase, intense XUV irradiation can cause severe atmospheric escape; for low-mass white-dwarf remnants, Jupiter-like planets at $\sim 0.01\,\mathrm{Myr}$4–$\sim 0.01\,\mathrm{Myr}$5 can lose at least half their mass, and “Jupiter in the Solar System is barely expected to survive” in the modeled scenario (Villaver, 2011).

A more interventionist StarEvolve usage is the “Lazarus stars” study, which investigates star-lifting as a life-extension strategy with MESA (Scoggins et al., 2022). Two control objectives are distinguished. In isoluminosity, lifted mass remains interior to the planet’s orbit, so $\sim 0.01\,\mathrm{Myr}$6 stays fixed and constant insolation requires $\sim 0.01\,\mathrm{Myr}$7. In isoirradiance, lifted mass moves beyond the planetary orbit, angular momentum is conserved, and constant insolation is obtained by maintaining $\sim 0.01\,\mathrm{Myr}$8. The control condition is written as

$\sim 0.01\,\mathrm{Myr}$9

with $\tilde r = (abc)^{1/3}$0 found numerically from MESA tracks. For initial masses below about $\tilde r = (abc)^{1/3}$1, star-lifting can extend main-sequence lifetimes by up to $\tilde r = (abc)^{1/3}$2 until the hydrogen-burning limit is approached; for a Sun-like star, the main-sequence lifetime can be increased by up to $\tilde r = (abc)^{1/3}$3, with a representative isoluminosity rate of about $\tilde r = (abc)^{1/3}$4 (Scoggins et al., 2022). The paper explicitly frames this as a controlled, external mass-loss prescription rather than a revision of internal stellar microphysics.

7. Distinct non-stellar and decision-theoretic usage

A separate usage of the name is entirely non-astrophysical. In SC2Arena, StarEvolve is a hierarchical LLM framework for full StarCraft II gameplay, designed for complex decision-making under partial observability and low-level action constraints (Shen et al., 14 Aug 2025). Its architecture is Planner–Executor–Verifier: the Planner emits natural-language commands, a Planner Verifier checks them against resources, tech, supply, and other rules, the Executor converts valid commands into JSON actions, and an Executor Verifier enforces syntactic and semantic legality. Each stage supports up to three rounds of self-correction.

The framework is tightly coupled to SC2Arena’s text-based observations and full low-level action space. Observations include action history, positions as integer grid coordinates, and proximity-based ordering of units to improve spatial reasoning; actions are JSON objects with keys such as "action", "units", "target_unit", and "target_position" (Shen et al., 14 Aug 2025). Continuous improvement is driven by trajectory logging, an RL-inspired scoring rule

$\tilde r = (abc)^{1/3}$5

and supervised fine-tuning on verified plan, execution, and verifier data. Reported evaluation shows that the SFT version based on Qwen2.5-7B-Instruct reached a LV7 built-in-AI win rate of $\tilde r = (abc)^{1/3}$6 with a Valid Action Ratio of $\tilde r = (abc)^{1/3}$7, while removing the scoring function reduced win rate to $\tilde r = (abc)^{1/3}$8 and VAR to $\tilde r = (abc)^{1/3}$9 (Shen et al., 14 Aug 2025). Although unrelated to stellar evolution, this usage preserves the general StarEvolve theme of hierarchical state evolution, constraint checking, and iterative refinement.

Taken together, these usages define StarEvolve as a broad computational and observational paradigm rather than a single artifact. In astrophysics it covers hybrid hydro/N-body cluster dynamics, cosmological tidal evolution, secular cluster prescriptions, probabilistic stellar tracks, composition-dependent habitable-zone histories, evolved-star diagnostics, and planetary-system outcomes. In a separate AI context it denotes a self-correcting decision architecture. The unifying methodological pattern is explicit evolution in time under coupled constraints, with model choice determined by scale: direct or hybrid simulation for strongly multiscale cluster environments, calibrated reduced-order dynamics for secular evolution, generative emulation for high-dimensional stellar-physics uncertainty, and multimodal observational inversion for evolved stars whose changes are now resolvable on human timescales.