Jiutian-1G Cosmological Simulation

• Jiutian-1G Cosmological Simulation is a state-of-the-art dark matter-only N-body simulation that enables precision studies of galaxy formation and large-scale cosmic structure.
• It leverages high particle counts and advanced subhalo tracking techniques to achieve robust phase-space accuracy (5–10%) at sub-megaparsec scales.
• The modular workflow integrates emulation, Bayesian reconstruction, and hydrodynamic enhancements to support CSST survey calibration and theoretical extensions.

The Jiutian-1G Cosmological Simulation is a flagship dark matter–only N-body run designed to underpin extra-galactic survey science, precision cosmology, and methodological developments for the China Space Survey Telescope (CSST) and related major survey programs. It features a combination of high particle count, large volume, and modular workflow, enabling robust predictions for galaxy formation, large-scale structure analyses, and joint survey strategies—while providing tractable benchmarks for subhalo physics, reconstruction, and cross-correlation experiments.

1. Simulation Architecture and Technical Specifications

Jiutian-1G is part of the Jiutian simulation suite’s “primary runs” module (Han et al., 27 Mar 2025, &&&1&&&). It evolves 6144³ dark matter particles (total particle count ≈ 2.32×10¹¹) in a periodic cube of side L = 1 h⁻¹ Gpc. The nominal particle mass is 3.72×10⁸ h⁻¹ M_⊙, and the initial condition redshift is z_ini = 127. The run comprises 128 snapshots spaced roughly every 100 Myr from z = 127 to z = 0, enabling high time resolution for merger tree and halo tracking analyses. The mass and spatial resolution allow resolution of galaxy hosts with ≥100 particles while controlling cosmic variance over the full CSST area.

The fiducial cosmology matches Planck-2018 parameters: h = 0.6766, Ωm = 0.3111, Ω_b = 0.0490, ΩΛ = 0.6899, σ₈ = 0.8102, n_s = 0.9665 (Song et al., 16 Aug 2024). Calibration and validation include tests of halo mass function universality, light cone construction for survey projections, and subhalo statistical convergence.

2. Subhalo Identification and Phase-Space Correction Methodologies

Subhalos in Jiutian-1G are identified using the time-domain halo tracker hbt+ (Xu, 11 Aug 2025). This algorithm tracks self-bound particle groups across snapshots, allowing robust identification of both host halos and their subhalos (including those that become “orphans” after disruption). Orphan subhalos—whose bound mass drops below the identification threshold—are tracked via the orbit of their most bound particle.

To address numerical convergence, especially in the inner regions of host halos, Jiutian-1G employs correction models for subhalo survival, notably the Jiang08 merger timescale model. This assigns a merger timescale $T_\text{merge}$ as a function of infall orbital properties, mass ratio, and host mass (see Eq. 1 above). Inclusion of orphan subhalos according to their life expectancy enables recovery of the Surviving subhalo Peak Mass Function (SPMF) and phase-space properties down to subhalo masses of $\sim$5000 particle masses. Real-space density and velocity statistics are accurate to 5–10% at 0.1–0.2 h⁻¹ Mpc scales. Below 0.1 h⁻¹ Mpc, further sophistication is needed due to strong tidal disruption and baryonic effects.

In redshift space, small-scale inaccuracies propagate as Fingers-of-God distortions, making recovery of multipole correlation functions (ξ₀ˢ(s), ξ₂ˢ(s), ξ₄ˢ(s)) more difficult; convergence at the monopole is achieved only at scales >0.2 h⁻¹ Mpc. The preferred measures for subhalo-based studies involve imposing a cutoff in projected separation or using alternative statistics less sensitive to small-scale uncertainties.

3. Modular Workflows and Data Products

The Jiutian-1G simulation supports a broad range of scientific workflows (Han et al., 27 Mar 2025), organized as follows:

• Primary runs: High-resolution N-body evolution for standard ΛCDM cosmology, outputting snapshot data, merger trees (SubFind/LHaloTrees, hbt+), and subhalo/lightcone catalogs.
• Emulator runs (“Kun”): 129 medium-resolution simulations sampling an 8D cosmological parameter space (Ω_b, Ω_cb, H₀, n_s, A_s, w₀, w_a, Σν) using Sobol sequences, feeding a Gaussian Process Regression emulator with 1% prediction accuracy for $P_\text{mm}(k)$ at $k\leq 10\,h\,\text{Mpc}^{-1}$ and $z\leq 2$.
• Reconstruction runs: Hydrodynamical “zoom-in” simulations matched to SDSS-constrained initial conditions (ELUCID), with baryonic models for local universe analogues.
• Extension runs: Non-standard cosmologies including Warm Dark Matter (parameterized $T(k)$), $f(R)$ gravity (Hu–Sawicki model), Interacting Dark Energy, and massive neutrino fluid modeling.

Data products include four sets of mock galaxy catalogs (Gaea, LGalaxies-Line, LGalaxies-Lensing, SHAM), merger trees, strong/weak lensing maps, subhalo abundance/peak mass statistics, SEDs, and mock survey images. The simulation provides on-the-fly lightcone projections in various geometries for direct comparison with survey data.

4. Cosmological Probes: Void Statistics, Galaxy Clustering, Lensing, and HI Mapping

Jiutian-1G serves as the basis for several precision cosmological probe analyses:

• Void statistics: Galaxy mocks post-processed via watershed algorithms (VIDE/ZOBOV) yield void catalogs with volume-equivalent radii and ellipticity selection (Song et al., 8 Feb 2024, Song et al., 16 Aug 2024). The void size function (VSF) is modeled using a redshift-dependent linear underdensity threshold δ_v (ranging −0.4 to −0.1 across z = 0.5 to 1.1), in stark contrast to the theoretical δ_v ≈ −2.7 for spherical DM voids. Theoretical modeling accounts for the conversion from Lagrangian to Eulerian void radii,

$R_L \approx R_v [1 - \delta_v / c_v ]^{-c_v / 3}$

and the first-crossing distribution $𝒜(\nu, \delta_v, \delta_c)$ for VSF predictions. Joint MCMC analyses using VSF data enable cosmological parameter constraints at a few percent level.

• Galaxy and void power spectra: Auto and cross spectra for galaxies and voids are modeled via bias parameters and shot noise terms,

$$P_{vv}(k, z) \simeq [b_v(z) \, u_v(k, z)]^2 P_{mm}(k, z) + N_v(z)$$

Theoretical profiles use the Hamaus–Sutter–Wandelt (HSW) model, fitted with the mean void effective radius per redshift slice. Joint fitting with galaxy spectra improves cosmological parameter constraints by ~20%.

• Lensing and clustering: Lightcone construction for sky patches (up to 100 deg², extendable to 17,500 deg²) enables multi-lens-plane weak lensing simulation, tomographic shear measurement, and convergence map production via Kaiser–Squires inversion. Mock galaxy catalogs with realistic luminosity assignments and photometric redshifts allow 3×2pt cosmological joint analyses. Covariance matrices include non-Gaussian terms and super-sample covariance.
• 21cm HI mapping and cross-correlation: The simulation enables modeling of MeerKAT HI intensity maps—incorporating instrument beam pattern, polarization leakage, and foregrounds—and CSST spectroscopic galaxy catalogs (Jiang et al., 27 Sep 2025). Cross-power spectra are modeled as

$$P_{HI,g}(k) = T_0 \Omega_{HI} b_{HI} b_g r_{HI,g} P_m(k) \exp[-(1-\mu^2) k^2 R_{beam}^2/2] + P_{HI,g}^{SN}$$

Post-PCA foreground subtraction and signal compensation yield joint constraints on $\Omega_{HI}b_{HI}b_g r_{HI,g}$ and $\Omega_{HI}b_{HI} r_{HI,g}$ with 6–8% relative accuracy—a factor of 3–4 improvement over contemporary measurements.

5. Field Reconstruction and Bayesian Inference

The Jiutian-1G simulation enables probabilistic reconstruction of cosmological initial conditions via simulation-based Bayesian inference:

• Bayesian SBI (with autoregressive modeling): The posterior for initial field $𝒛$ given late-time field $𝒙$ uses neural ratio estimation and local Gaussian approximations,

$$\ln p(𝒙|𝒛) = -\frac{1}{2} 𝒛^T Q_{like} 𝒛 + b^T 𝒛 + C(𝒙)$$

Autoregressive decomposition sequentially factors the likelihood, facilitating scalability to 3D high-resolution data (List et al., 2023).

• Mean-field inference (Fourier-diagonal Gaussian): The posterior $p(z|x)$ is modeled as a diagonal Gaussian in Fourier space,

$$p(z|x) = \mathcal{N}(z; \hat{z}(x), (Q^P+Q^L)^{-1})$$

Training utilizes simulation pairs (z_i, x_i), with nonlinear corrections provided by a U-Net model. Sampling is rapid (thousands of samples in seconds for 128³), and sample statistics (power spectrum, transfer function, cross-correlation) are consistent with ground truth to 1–2% accuracy (Savchenko et al., 21 Oct 2024). Both approaches are fully compatible with non-differentiable N-body codes (e.g., Gadget-III).

6. Advanced Physics Modules: Scalar Field Dynamics and Hydrodynamics

Jiutian-1G is compatible with extensions beyond CDM N-body:

• Cosmological scalar fields: UltraDark.jl numerically solves the Gross-Pitaevskii–Poisson system,

$$i \hbar \frac{\partial \psi}{\partial t} = -\frac{\hbar^2}{2 m a(t)^2} \nabla^2 \psi + m \Phi \psi \ \nabla^2 \Phi = \frac{4\pi G}{a(t)} m |\psi|^2$$

using pseudo-spectral, symmetrized split-step integration, adaptive time-stepping, and MPI/thread parallelism. The approach enables simulations of axion dark matter, inflation, and tidal phenomena for large cosmological volumes (Musoke, 7 May 2024).

• Hydrodynamic enhancement: Gigaparsec-scale IGM hydrodynamics is achieved by training a deep-learning generative enhancement (conditional GAN; TSIT-style encoder-generator) on pairs of high-res and low-res Nyx hydrodynamical runs. The GAN injects small-scale (kpc) features into low-res large-volume simulations, resulting in 960 Mpc/h volumes with 6144³ cells and 11 TB memory requirement (Jacobus et al., 25 Nov 2024). Effective multi-scale spectral loss functions enable matching 1D/3D power spectra and PDF statistics (e.g., Lyman-α forest) to high-resolution references at the 10% level.

7. Scientific Impact and Applications

Jiutian-1G (and the greater Jiutian suite) forms a comprehensive and open simulation library tailored for the CSST era. Its capacity for large volume, multi-module extensibility, and high-fidelity mock generation supports:

• CSST survey support: Calibration, pre-survey design, cosmological inference pipelines, lensing prediction, emission line/spectral forecast, and strong/weak lensing studies.
• Method benchmarking: Numerical convergence analysis (SPMF, orphan corrections), Bayesian inference validation, emulator training, and cross-survey comparison.
• Fundamental physics: Testing alternate cosmologies (WDM, $f(R)$, IDE, massive neutrino cosmology), assessing the universality of subhalo mass functions, reconstructing observed Universe regions with constrained initial conditions, and enabling survey-scale hydrodynamic–N-body synergy.

A plausible implication is that the Jiutian-1G simulation provides both a tractable benchmark for precision cosmological analyses and a flexible base for integrating advanced physics, enabling robust inference in next-generation cosmological surveys.

Table: Key Jiutian-1G Simulation Specifications

Property	Value/Methodology	Significance
Volume	1 h⁻¹ Gpc cube	Controls cosmic variance
Particle count	6144³ ≈ 2.32×10¹¹	High resolution
Mass resolution	3.72×10⁸ h⁻¹ M_⊙	Resolves galaxy hosts (≥100 particles)
IC redshift	z_ini = 127	Early Universe evolution
Time resolution	128 snapshots	Accurate merger histories
Subhalo finder	hbt+	Robust phase-space tracking
Cosmology	Planck-2018 ΛCDM	Survey-comparable input
Extension modules	WDM, $f(R)$, IDE, ν masses	Probes theoretical diversity
Emulation accuracy	1% (k ≤ 10 h Mpc⁻¹, z ≤ 2)	Precision cosmology
Public release	Y (hydro/N-body/halo)	Enables community benchmarking

This summary delineates the architecture, methodology, probe applications, correction models, reconstruction strategies, and survey support functions of the Jiutian-1G Cosmological Simulation as documented in the cited arXiv papers.