Can a Cosmology-Aware Generative Model Super-Resolve N-Body Simulations?

The model outputs 6D phase space, not just density

The paper's target is not a density map but particle phase space. Each snapshot is represented as a six-channel field on the regular Lagrangian grid: three displacement components and three velocity components. If q is a particle's Lagrangian grid coordinate, its Eulerian position is x = q + s(q), with positions wrapped periodically on the box. The generator maps low-resolution displacement and velocity fields to a super-resolution field with the same six-channel representation as the high-resolution target, so the output can be fed into the same halo-finding pipeline used on the high-resolution simulations.

That distinction matters for what the results can support. Because displacements and velocities are generated jointly, the model can be tested on collapsed structures and substructure, not only on two-point clustering. The paper validates with the matter power spectrum, halo mass function, subhalo mass function, and mean satellite occupation, and it treats the satellite statistics as information complementary to P(k) about how particles are reorganized inside dense environments.

One model, five cosmological parameters, no retraining

The generator and discriminator are both conditioned on the same style vector s_style = (a, Ωm, Ωb, h, ns, σ8), passed to a learned linear mapping at each modulated convolution. All training and evaluation use z = 0, so a = 1 throughout. The architecture follows the StyleGAN2-adapted generator and discriminator of the authors' earlier super-resolution work, with StyleGAN2 weight modulation and demodulation, Gaussian noise injection for stochastic variation, and a single H-block that doubles grid resolution along each spatial axis.

Training data are 100 matched low–high resolution pairs from the Quijote Latin-hypercube suite, one per distinct LH cosmology, each with its own random seed at z = 0. The LH set varies all five parameters simultaneously over Ωm ∈ [0.10, 0.50], Ωb ∈ [0.03, 0.07], h ∈ [0.50, 0.90], ns ∈ [0.80, 1.20], and σ8 ∈ [0.60, 1.00], with neutrino mass fixed at Mν = 0 and dark-energy equation of state fixed at w = −1. Boxes are 1 h⁻¹ Gpc on a side, pairing 512³ LR against 1024³ HR particles, a fixed super-resolution factor of 2 per axis, or 8 in particle count.

Evaluation uses ten further LH cosmologies excluded from super-resolution training: a near-fiducial reference plus nine additional points labeled by their dominant parameter offset (Ωm±, Ωb±, σ8±, ns±, h−). In every non-reference test cosmology all five parameters differ from the reference, so the labels identify cases rather than isolate single-parameter variations.

Super-resolution is trained as a stochastic mapping

The low-resolution simulations do not specify the additional short-wavelength modes present in the high-resolution runs. The paper therefore models super-resolution as a conditional stochastic mapping, with noise inputs allowing different outputs for the same LR realization and cosmology, and it deliberately omits a per-voxel reconstruction loss from the optimization objective. Training uses Wasserstein-GAN losses with a gradient penalty (λ_GP = 10), evaluated once every 16 discriminator updates with the coefficient multiplied by 16 on those updates. The paired mean-squared error is recorded only as a diagnostic and contributes no gradients.

This is a different task from deterministic emulation, where high-resolution initial conditions are supplied to identify the target realization. The consequence for readers is that agreement should be judged statistically across realizations, not by whether a specific particle lands where its high-resolution counterpart did. The paper's own framing is that the LR simulation provides the large-scale realization and its non-linear evolution, while the model learns to add smaller-scale structure to that realization.

Power spectra agree at a few percent, with a shared bias

Matter power spectra are computed with nbodykit by assigning particles in the 1 h⁻¹ Gpc box to a 1024-cell-per-side mesh using Cloud-in-Cell mass assignment, with CIC window compensation and interlacing disabled. All HR targets and SR outputs use the same mesh, mass-assignment scheme, and k-binning. The mesh spans from the fundamental mode k_f ≈ 6.3 × 10⁻³ h Mpc⁻¹ to the mesh Nyquist wavenumber k_Nyq^mesh ≈ 3.2 h Mpc⁻¹.

At the fiducial cosmology, the maximum of |ε_abs| is 3.1% over 0.1 < k < 3 h Mpc⁻¹, with a mean signed error of +0.9%, where ε_abs ≡ A² − 1 and A is the amplitude ratio √(P_SR/P_HR). Across all ten cosmologies including the fiducial case, the mean of the per-cosmology maximum absolute power errors over the same interval is 3.6%. The deviations from unity are small but scale dependent: the spectra agree closely on large scales, show a common excess of SR power at intermediate wavenumbers, and spread more among test cases at high k.

The paper separates this shared model bias from cosmology-to-cosmology differences by defining a scale-dependent fiducial bias b_fid(k) = P_SR(k; θ_fid)/P_HR(k; θ_fid) and applying the same correction factor 1/b_fid(k) to each SR spectrum. The calibrated ratio Q stays within about 2% of unity across the nine non-fiducial cosmologies. The agreement persists through the LR particle Nyquist frequency k_Nyq^LR ≈ 1.6 h Mpc⁻¹ and into the interval between the LR and HR particle Nyquist frequencies, where the LR field no longer contains independently resolved Fourier modes and the model must generate unresolved small-scale structure from the resolved LR field, the stochastic input, and the cosmological conditioning. None of this agreement comes from an explicit power-spectrum term in the training loss.

Halo abundances hold; satellite statistics do not

Because the model outputs full particle phase space, its predictions can be processed with the same halo-finding pipeline as the HR simulations. Friends-of-friends halo abundances agree with HR to within 15% in the analyzed mass bins below 10^14.8 h⁻¹ M⊙. Subhalo abundances have larger mass-dependent errors, and satellite occupation is increasingly suppressed in massive hosts, with a similar host-mass dependence across the tested cosmologies.

The paper connects this deficit to a related one found in the super-resolution baseline of its earlier work, where the division into particle chunks was investigated as a limitation. That suggests varying the spatial context of the training chunks, which would require GPUs with more memory, as a concrete next step, or involving a halo-finding step during training via reinforcement learning to send non-differentiable feedback to the model. The fiducial corrections account for a substantial shared component of these statistical errors, but residual satellite discrepancies remain at the level of tens of percent.

The practical reading is that the statistic determines how far the model can be trusted. Clustering and host-halo abundance are reproduced at the few-percent and 15% levels respectively, while the substructure-sensitive statistics remain the clear weak point, and the paper treats them as the target for the next round of model improvement.

Small spectral residuals still move inferred parameters

To connect spectral accuracy to parameter estimation, the paper passes matched HR and SR spectra through the same neural posterior estimator trained on HR spectra, keeping the estimator fixed so that the measured shift reflects only the substitution of SR for HR input. For each matched pair i and parameter j, the normalized posterior-mean shift is Δ = (θ̂_X − θ̂_HR)/σ_HR, using the HR posterior standard deviation as the common unit.

For raw SR spectra, the RMS normalized shifts are 2.1 in Ωm, 1.7 in ns, and 2.7 in σ8, while Ωb and h shift by approximately 0.1. The σ8 shifts are positive in all ten cosmologies, whereas the Ωm and ns shifts vary in sign. Applying the same fiducial power-spectrum correction reduces the RMS shift in σ8 from 2.7 to 0.7 and in Ωm from 2.1 to 1.7, while the ns shift remains approximately 1.7. The correction therefore removes a substantial component of the positive σ8 shift, and the remaining Ωm and ns shifts measure the parameter impact of spectral differences left after calibration.

The paper's explanation is that the impact of a spectral residual depends on its scale dependence and on the direction in which the spectrum changes with the cosmological parameters, so a small but coherent difference across bins can produce a substantial parameter displacement. This comparison uses k_max = 2 h Mpc⁻¹, which extends into the nonlinear regime and beyond the LR particle-grid Nyquist wavenumber. The fiducial reference LH 1128 is itself used to determine b_fid(k), so its corrected spectrum matches the HR reference by construction; the other nine pairs test whether the correction transfers.

Calibration transfers, but only for some statistics

The improvement after fiducial calibration indicates that a substantial part of the SR-to-HR discrepancy is shared across the tested cosmologies. For each statistic, a scale- or mass-dependent correction is derived from the SR-to-HR ratio at the near-fiducial cosmology and applied to the other nine test cosmologies. This reduces the power-spectrum offset and improves agreement of FoF halo abundances and satellite occupation in massive hosts. Near k ≈ 0.4 h Mpc⁻¹, an SR power excess of approximately 2–4% is reduced to fractional residuals of order 1% after calibration.

Two boundaries qualify this. First, the corrections are applied to the measured statistics only and leave the actual generated particle field unchanged, so they do not fix the underlying substructure prediction. Second, larger residuals remain in the satellite statistics, meaning the effectiveness of calibration depends on both the statistic and the scale or mass considered. The paper states that improvements to the model's understanding of structure within massive hosts are needed to fulfill its goal of field-level simulation-based inference, where parameters are inferred from information learned directly from the 3D field without summary statistics.

Judge this model by statistic, not by average accuracy

The paper's own summary of its contribution is that a cosmology-aware super-resolution model can reproduce matter clustering and halo abundance across the tested cosmologies while quantifying the remaining errors in satellite statistics and parameter inference. The single most decision-relevant split in the evidence is between statistics: power spectra at the few-percent level and FoF halo abundances within 15% below 10^14.8 h⁻¹ M⊙ on one side, and mass-dependent subhalo and satellite-occupation errors on the other.

The second split is between raw and calibrated use. A single fiducial power-spectrum calibration transfers to the other nine cosmologies and cuts the σ8 RMS normalized shift from 2.7 to 0.7, but it leaves Ωm and ns shifts near 1.7 and does not alter the generated field. For anyone using these realizations, the operative question is therefore which statistic and which parameter the downstream analysis depends on: σ8 inference from the power spectrum is the case the calibration most clearly rescues, while substructure-sensitive measurements and residual Ωm and ns shifts are where the model's remaining error still has to be carried explicitly.

The paper also notes that extending the model jointly across cosmology and redshift, improving substructure predictions, and employing other advanced approaches remain open, and that the generated matter and halo fields would additionally enter a galaxy-population and survey-selection model in forward-modeling applications. Those are stated directions rather than demonstrated results here.

Sources used in this answer