CAMELS Simulations Overview

• CAMELS Simulations are a coordinated family of cosmological runs that vary key parameters like Ωm and σ8 in controlled (25 h⁻¹ Mpc) volumes to probe galaxy formation and baryonic effects.
• They incorporate multiple galaxy formation models (IllustrisTNG, SIMBA, ASTRID, and Swift-EAGLE) to explore non-linear interactions between cosmology and astrophysics.
• The suite generates vast ML-ready datasets—from snapshots to 2D/3D maps—enabling advanced inference techniques and comparative studies of cosmic structure.

CAMELS Simulations, the “Cosmology and Astrophysics with MachinE Learning Simulations,” are a coordinated family of cosmological simulation suites built to sample cosmological and galaxy-formation parameter space at scale and to support machine-learning-based inference. The original CAMELS project consists of 4,233 cosmological simulations of $(25~h^{-1}{\rm Mpc})^3$ volume each: 2,184 state-of-the-art (magneto-)hydrodynamic simulations run with the AREPO and GIZMO codes and 2,049 N-body simulations, with the stated goal of providing theory predictions for different observables as a function of cosmology and astrophysics and of training machine-learning algorithms (Villaescusa-Navarro et al., 2020). Subsequent expansions added CAMELS-ASTRID and 28-parameter extensions of the TNG and SIMBA suites, and a second-generation IllustrisTNG-based program with 1,192 simulations in $(50\,{\rm Mpc}/h)^3$ boxes varying 35 cosmological, astrophysical, and numerical parameters (Ni et al., 2023, Genel et al., 8 Jun 2026).

1. Foundational design

The original CAMELS design couples relatively small simulation volumes to very large ensemble size. Each first-generation run evolves a periodic box of $(25\,h^{-1}\,\mathrm{Mpc})^3$; the project as a whole follows the evolution of more than 100 billion particles and fluid elements over a combined volume of $(400~h^{-1}{\rm Mpc})^3$ while varying $\Omega_m$, $\sigma_8$, and four parameters controlling stellar and AGN feedback (Villaescusa-Navarro et al., 2020). In the four-model comparisons later enabled by CAMELS, the common numerical setup is a $(25\,h^{-1}\,{\rm Mpc})^3$ box with $256^3$ dark-matter particles and $256^3$ gas resolution elements, allowing direct comparisons among galaxy-formation models at fixed volume and resolution (Gebhardt et al., 9 Jan 2026).

A central organizational feature is the separation into complementary simulation sets. In the six-parameter CAMELS framework, the sampled parameter vector is $$(\Omega_m,\sigma_8,A_{\rm SN1},A_{\rm SN2},A_{\rm AGN1},A_{\rm AGN2})$$. The CV set provides 27 hydrodynamical and 27 dark-matter-only runs per model at fixed fiducial parameters but with different initial phases; the 1P set varies one parameter at a time; and the LH set contains 1000 runs per model sampling the six-dimensional parameter space plus initial-condition variance (Gebhardt et al., 9 Jan 2026). This structure makes CAMELS simultaneously useful for controlled response studies, cosmic-variance estimates, and high-dimensional interpolation.

The original project already emphasized that IllustrisTNG and SIMBA produce roughly similar distributions of galaxy properties over the full parameter space but significantly different halo baryon fractions and baryonic effects on the matter power spectrum, a result that directly motivates the use of CAMELS for marginalizing baryonic effects in cosmological inference (Villaescusa-Navarro et al., 2020).

2. Suite families and expansion of model space

CAMELS has evolved from a dual-model program into a broader multi-model, multi-generation framework. The hydrodynamical branches now include at least SIMBA, IllustrisTNG, ASTRID, and Swift-EAGLE in comparative studies, implemented respectively with GIZMO, AREPO, MP-Gadget SPH, and SWIFT/SPHENIX (Gebhardt et al., 9 Jan 2026). This diversification is not cosmetic: CAMELS-ASTRID was introduced specifically to broaden the galaxy-formation model space, while the new TNG and SIMBA extensions were designed to expose the “enormity” of the overall model space and the complex non-linear interplay between cosmology and astrophysical processes (Ni et al., 2023).

Suite or expansion	Scope	Distinguishing feature
Original CAMELS	4,233 simulations; 2,184 hydrodynamic and 2,049 N-body; $(50\,{\rm Mpc}/h)^3$0 each	Core variation of $(50\,{\rm Mpc}/h)^3$1, $(50\,{\rm Mpc}/h)^3$2, and four stellar/AGN feedback parameters
CAMELS-ASTRID and 28-parameter extensions	CAMELS-ASTRID has 2,124 hydrodynamic runs; new TNG/SIMBA sets explore 28 parameters	Broader galaxy populations and broader baryonic impact on the matter power spectrum
Second-generation CAMELS	1,192 IllustrisTNG simulations in $(50\,{\rm Mpc}/h)^3$3 boxes varying 35 parameters	Lower sample variance, more massive halos, more diverse environments

CAMELS-ASTRID employs the galaxy-formation model following the ASTRID simulation and varies three cosmological parameters, $(50\,{\rm Mpc}/h)^3$4, $(50\,{\rm Mpc}/h)^3$5, and $(50\,{\rm Mpc}/h)^3$6, plus four parameters controlling stellar and AGN feedback (Ni et al., 2023). Its fiducial model is described as having the mildest AGN feedback and predicting the least baryonic effect on the matter power spectrum among the main CAMELS hydrodynamical branches (Ni et al., 2023). The extended TNG and SIMBA suites widen the parameter space to 28 parameters, and the second-generation IllustrisTNG program widens it again to 35 parameters, including four parameters that control the amplitude and timing of the ionizing background radiation (Genel et al., 8 Jun 2026).

This progressive broadening of model space has a methodological consequence stated explicitly in the CAMELS literature: building robust machine-learning models favors training and testing on the largest possible diversity of galaxy formation models (Ni et al., 2023).

3. Public data products and machine-learning orientation

CAMELS was designed not only as a simulation archive but as an ML-oriented data system. The public data release describes 4,233 cosmological simulations, 2,049 N-body and 2,184 hydrodynamic, and distributes more than 350 terabytes of data containing 143,922 snapshots, millions of halos and galaxies, and summary statistics including halo, subhalo, galaxy, and void catalogues, power spectra, bispectra, Lyman-$(50\,{\rm Mpc}/h)^3$7 spectra, probability distribution functions, halo radial profiles, and X-rays photon lists (Villaescusa-Navarro et al., 2022). The same release also provides over one thousand catalogues containing billions of galaxies from CAMELS-SAM (Villaescusa-Navarro et al., 2022).

The CAMELS Multifield Dataset extends this orientation toward field-level learning. It consists of hundreds of thousands of 2D maps and 3D grids containing many different properties of cosmic gas, dark matter, and stars from 2,000 distinct simulated universes at several cosmic times, and totals more than 70 terabytes (Villaescusa-Navarro et al., 2021). Its stated purpose is to train machine-learning models on spatially resolved simulation outputs rather than only on summary statistics (Villaescusa-Navarro et al., 2021).

A growing body of CAMELS work uses this structure directly. A diffusion generative model trained on CAMELS reconstructs dark matter fields from stellar mass fields while marginalizing over uncertainties in cosmological and astrophysical models; it also generalizes to simulation volumes approximately 500 times larger than those used in training and across different galaxy-formation models (Ono et al., 2024). A separate CycleGAN-based program learns bidirectional mappings among projected dark-matter density, neutral-hydrogen density, and magnetic-field-magnitude maps, with generated maps whose probability distribution functions and 2D power spectra agree well with those of the target fields (Andrianomena et al., 2023). These studies illustrate a distinctive CAMELS role: the suites are not merely reference simulations but supervised training distributions for field-to-field translation, inverse problems, and simulation-based inference.

4. Baryons, halos, and matter clustering

One of the most developed scientific uses of CAMELS is the quantification of baryonic effects on nonlinear structure. In “Cosmological baryon spread and impact on matter clustering in CAMELS,” the spread of baryons relative to their initial neighboring dark matter is shown to depend strongly on feedback implementation: the fiducial SIMBA model spreads $(50\,{\rm Mpc}/h)^3$8 of baryons more than $(50\,{\rm Mpc}/h)^3$9 away, compared to $(25\,h^{-1}\,\mathrm{Mpc})^3$0 for IllustrisTNG and ASTRID (Gebhardt et al., 2023). The same study reports that increasing the efficiency of AGN-driven outflows greatly increases baryon spread, while increasing the strength of SNe-driven winds can decrease spreading because of non-linear coupling between stellar and AGN feedback (Gebhardt et al., 2023). This makes CAMELS a controlled setting for relating matter-power suppression to physically interpretable summaries of baryon redistribution.

CAMELS has also been used to isolate baryonic back-reaction on dark matter by matching hydrodynamical and dark-matter-only simulations halo by halo. In that analysis, virial masses decrease owing to the ejection of baryons by feedback; halo profiles show increased dark matter density in the center and decreased density farther out; and the effect is strongest in SIMBA, with more than a 450% increase in central dark matter density at $(25\,h^{-1}\,\mathrm{Mpc})^3$1 (Gebhardt et al., 9 Jan 2026). Across the four-model comparison, dark-matter power spectra in some simulations from each model show as much as 20% suppression or increase in power at $(25\,h^{-1}\,\mathrm{Mpc})^3$2 relative to N-body simulations, and the back-reaction depends intrinsically on cosmology, specifically $(25\,h^{-1}\,\mathrm{Mpc})^3$3 and $(25\,h^{-1}\,\mathrm{Mpc})^3$4, at fixed baryonic physics (Gebhardt et al., 9 Jan 2026).

At the halo-structure level, CAMELS has been used to construct a simulation-informed nonlinear model for the concentration–mass relation $(25\,h^{-1}\,\mathrm{Mpc})^3$5. That model spans $(25\,h^{-1}\,\mathrm{Mpc})^3$6, $(25\,h^{-1}\,\mathrm{Mpc})^3$7, and six cosmological/astrophysical parameters, and shows that astrophysical model choices cause significant differences in the mass and redshift dependence of baryon imprints (Shao et al., 2022). A plausible implication is that CAMELS does not merely sample nuisance variation around a single halo model; it reveals that halo response functions themselves are model-dependent objects.

5. Observable predictions: CGM, FRBs, SZ, alignments, and environment

CAMELS has become a laboratory for forward-modeling observables that directly probe diffuse baryons. Fast radio burst studies provide one example. Using the IllustrisTNG, SIMBA, and Astrid 1P suites, a CAMELS-based FRB analysis finds that SIMBA exhibits the strongest feedback, leading to the smoothest distribution of baryons and reducing the sightline-to-sightline variance in dispersion measures between $(25\,h^{-1}\,\mathrm{Mpc})^3$8 and 1, whereas Astrid has the weakest feedback and the largest variance (Medlock et al., 2024). Within each suite, the largest differences are due to varying AGN feedback, and the apparent sensitivity of IllustrisTNG to supernova feedback is interpreted as a consequence of changing AGN feedback strengths, implying that black holes rather than stars are most capable of redistributing baryons in the IGM and CGM (Medlock et al., 2024). A later neural-network emulator for the cosmic dispersion-measure distribution $(25\,h^{-1}\,\mathrm{Mpc})^3$9 models feedback effects at arbitrary redshifts for $(400~h^{-1}{\rm Mpc})^3$0 and finds that $(400~h^{-1}{\rm Mpc})^3$1 does not depend monotonically on every feedback parameter; even the largest $(400~h^{-1}{\rm Mpc})^3$2 values in the explored parameter space remain small compared to current observational constraints, which the authors attribute to the limitations imposed by CAMELS’ small box sizes (Guo et al., 28 Jan 2025).

Sunyaev–Zeldovich and CGM studies use CAMELS in a parallel way. One emulator-based forecast concludes that, for a DESI-like galaxy sample observed by the Simons Observatory, all four feedback parameters are constrainable from SZ radial profiles, some within the 10% level, and that the inner SZ profiles contribute more to the constraining power than the outer profiles (Moser et al., 2022). The same study reports that, despite the wide range of AGN feedback parameter variation in CAMELS, the suite cannot reproduce the tSZ signal of BOSS-selected galaxies as measured by ACT (Moser et al., 2022). A related CAMELS analysis of the warm-hot circumgalactic medium finds that IllustrisTNG shows higher cumulative feedback energy across all halos, while SIMBA demonstrates a greater spread of baryons, quantified by the closure radius and CGM gas fraction, suggesting that feedback in SIMBA couples more effectively to baryons and drives them more efficiently within the host halo (Medlock et al., 2024). It also finds that parameters controlling stellar feedback efficiency significantly impact AGN feedback, underscoring the interdependence of feedback modes (Medlock et al., 2024).

CAMELS has likewise been extended to intrinsically anisotropic observables and environmental statistics. Intrinsic galaxy alignments have been detected in the CAMELS suite, with alignment amplitude depending significantly on $(400~h^{-1}{\rm Mpc})^3$3, $(400~h^{-1}{\rm Mpc})^3$4, $(400~h^{-1}{\rm Mpc})^3$5, and $(400~h^{-1}{\rm Mpc})^3$6, while no dependence on AGN feedback is observed because of the limited simulation volume; quiescent galaxies show alignment amplitudes exceeding those of star-forming galaxies by an order of magnitude (Bilsborrow et al., 26 Nov 2025). Environmental analyses across SIMBA, IllustrisTNG, ASTRID, and Swift-EAGLE show that satellite galaxies are significantly affected by environment in all models, with gas fraction and star-formation rate suppressed in overdense regions at $(400~h^{-1}{\rm Mpc})^3$7, whereas central galaxies are less sensitive and can transition from lower gas fraction and SFR in overdense regions at low stellar mass to higher gas fraction and SFR for massive galaxies in higher-density environments (Sims et al., 9 Jan 2026). Halo baryon fraction and CGM mass fraction also show clear but model-dependent environmental trends, with opposite signs in SIMBA and Swift-EAGLE at low halo mass (Sims et al., 9 Jan 2026).

6. Second-generation CAMELS and current trajectory

The second-generation CAMELS program expands the original design in both physical volume and parameter dimensionality. It introduces 1,192 IllustrisTNG-based simulations in $(400~h^{-1}{\rm Mpc})^3$8 boxes, eight times larger than previous CAMELS volumes, and explores 35 cosmological, astrophysical, and numerical parameters (Genel et al., 8 Jun 2026). The authors emphasize two consequences of this enlargement: lower sample variance and access to more massive halos and more diverse environments (Genel et al., 8 Jun 2026). They construct training sets from matter power spectra, projected maps, graphs representing galaxy spatial distributions, and thermodynamical properties of massive halos, and they pair these inputs respectively with multilayer perceptrons, convolutional neural networks, graph neural networks, and Gaussian processes for parameter inference (Genel et al., 8 Jun 2026).

The central methodological result of this second-generation study is that larger volumes do generally produce tighter marginal constraints, but the improvements scale more weakly than the square root of the increase in physical volume (Genel et al., 8 Jun 2026). The stated interpretation is either information loss due to mode coupling or complex degeneracies in parameter space (Genel et al., 8 Jun 2026). This suggests that CAMELS’ future development is not reducible to a simple “larger box” prescription: gains depend on the observable, the representation, and the structure of the parameter manifold. At the same time, the addition of four new parameters controlling the amplitude and timing of the ionizing background radiation extends CAMELS into the thermal history of the intergalactic medium, broadening the suite’s relevance for Ly$(400~h^{-1}{\rm Mpc})^3$9-forest and IGM-temperature studies (Genel et al., 8 Jun 2026).

Taken together, the CAMELS literature defines a simulation ecosystem rather than a single suite: a first-generation core optimized for breadth in $\Omega_m$0 boxes, public releases centered on ML-ready data products, expanded model families spanning IllustrisTNG, SIMBA, ASTRID, and Swift-EAGLE, and a second generation that pushes to $\Omega_m$1 and 35 parameters (Villaescusa-Navarro et al., 2020, Genel et al., 8 Jun 2026). The recurring conclusion across these stages is that cosmology and baryonic physics must be varied jointly, because the relevant observables—from matter clustering to FRB dispersion measures, SZ profiles, intrinsic alignments, and environmental trends—respond to both in model-dependent and often non-monotonic ways (Ni et al., 2023).