AbacusSummit Suite: Precision Cosmology Simulations

Updated 7 December 2025

AbacusSummit Suite is a public-domain cosmological simulation resource that generates large-volume, high-resolution N-body datasets using nearly 60 trillion dark matter particles.
It employs GPU-accelerated Abacus code with subpercent gravitational force accuracy and a hybrid near-far field force method on Summit, ensuring efficient and precise computations.
The suite provides diverse data products—including full particle snapshots, halo catalogs, light-cone outputs, and weak lensing maps—supporting advanced analyses in galaxy clustering and mock survey generation.

The AbacusSummit Suite is a public-domain, large-volume, high-resolution cosmological $N$ -body simulation resource, designed specifically to meet the needs of current and next-generation large-scale structure surveys such as DESI, Euclid, and LSST. Generated with the GPU-accelerated Abacus code on the Oak Ridge Leadership Computing Facility's Summit supercomputer, the suite contains nearly 60 trillion dark matter particles distributed across a broad set of cosmologies and volume realizations. Its architecture is optimized for subpercent-level gravitational force accuracy, extensive cosmological parameter coverage, and rapid generation of synthetic survey observables, positioning it as a foundational platform for precision cosmology and large-scale structure analyses (Maksimova et al., 2021, Garrison et al., 2021).

1. Architecture and Core Simulation Design

AbacusSummit leverages the Abacus $N$ -body code, which achieves a median fractional force error of $\mathcal{O}(10^{-5})$ via an exact analytic near-far field force decomposition. Each simulation divides the volume into a uniform Cartesian mesh. Short-range (near-field) gravitational forces are computed via direct, GPU-accelerated summation out to two cell widths, while long-range (far-field) forces employ a Taylor expansion of multipole moments up to $p=8$ , utilizing FFT-based convolutions for CPU efficiency. This method eliminates the need for overlap, transition, or corrective force regions, ensuring both physical fidelity and computational efficiency (Garrison et al., 2021, Maksimova et al., 2021).

On the Summit supercomputer, each node (6 NVIDIA V100 GPUs + 2 POWER9 CPUs) delivers up to 70 million particle updates per second at early redshifts, and 45 million at late times. Abacus executes time integration using a global Kick–Drift–Kick scheme, pipelined as slabs to sustain high throughput and minimal memory overhead.

2. Cosmological and Volume Coverage

The AbacusSummit suite encompasses a wide range of cosmological models and box sizes:

Base-resolution runs: 25–139 realizations at $L_{\rm box} = 2000\,h^{-1}\mathrm{Mpc}$ per side, $N_p = 6912^3 \approx 3.3\times10^{11}$ particles, $m_p=2.1\times10^{9}\,h^{-1}M_\odot$ (Maksimova et al., 2021, Hadzhiyska et al., 2021).
Huge boxes: $L_{\rm box}=7500-7600\,h^{-1}{\rm Mpc}$ , $N_p=8640^3$ , $m_p\sim1.05\times10^{11}\,M_\odot/h$ for full-sky light-cone applications (Hadzhiyska et al., 2023, Hadzhiyska et al., 2021).
High-resolution small boxes: $L_{\rm box}=500\,h^{-1}\mathrm{Mpc}$ , up to $N_p=3072^3$ (Yuan et al., 2021).

Across the suite, 97 cosmologies are realized, including variations in $\Omega_m, \sigma_8, w_0, w_a, n_s,$ and the summed neutrino mass $\Sigma m_\nu$ (from 0.0 to 0.8~eV), with each cosmology sampled with 1–6 independent phases (Yuan et al., 2021). Secondary cosmologies enable the training of emulators and systematic forecasts.

The total simulated volume vastly exceeds DESI and Euclid requirements—with individual base boxes at $(2\,h^{-1}\mathrm{Gpc})^3$ and a cumulative volume $> 800\, (h^{-1}{\rm Gpc})^3$ for covariance analyses (Ding et al., 2022).

3. Data Products and Outputs

AbacusSummit delivers a wide array of data products, all accessible in highly efficient, compressed ASDF formats:

Full particle snapshots: Positions, velocities, and IDs at up to 47 discrete redshifts between $z=8$ and $z=0.1$ (Maksimova et al., 2021, Yuan et al., 2021).
Halo catalogs: On-the-fly spherical overdensity halo finding (CompaSO) at every output, including merger tree linkage, concentrations, and “cleaned” catalogs for robust HOD analysis (Yuan et al., 2021, Garrison et al., 2021).
Light-cone outputs: Observer-centered outputs with full particle and/or halo information, providing octant (base runs) or full-sky (huge runs) coverage out to $z \approx 2.2$ (Hadzhiyska et al., 2021, Hadzhiyska et al., 2023).
Mock catalogs: Synthetic galaxy, quasar, and Lyman- $\alpha$ forest catalogs for DESI, eBOSS, and weak lensing surveys, built with AbacusHOD and alternative prescriptions (Yuan et al., 2021, Hadzhiyska et al., 17 Mar 2025).
Weak lensing maps: Publicly released high-resolution HEALPix ( $N_{\rm side}=16384$ ; 0.21 arcmin pixels) shear, convergence, and deflection maps, as well as CMB lensing convergence, matched to the geometry of major surveys (Hadzhiyska et al., 2023).

These data products enable analyses of the matter and galaxy power spectrum, redshift-space clustering, weak lensing statistics, and cross-correlation measurements, with validation against theoretical predictions at the subpercent level.

4. Light-Cone Construction and Mock Survey Applications

AbacusSummit implements state-of-the-art light-cone catalog generation for both haloes and particles. For haloes, a merger-tree-based algorithm interpolates halo positions, velocities, and masses to their light-cone crossing times, with particle sub-sampling and duplication mitigation. For particles, on-the-fly outputs provide rarefied catalogs suitable for lensing and line-of-sight analyses (Hadzhiyska et al., 2021).

Mock catalogs of emission-line galaxies (ELGs), luminous red galaxies (LRGs), and quasars are assembled on the light cone using extended Halo Occupation Distribution (HOD) models (e.g., AbacusHOD). These mocks include redshift-space distortions and are validated using clustering multipoles, power spectra, and lensing maps, all of which show subpercent agreement with theoretical predictions and synthetic snapshot results (Yuan et al., 2021, Hadzhiyska et al., 2023).

Weak lensing catalogs employ Born-approximation ray-tracing over the full light-cone, matching source distributions of surveys such as KiDS, DES, and HSC, and producing convergence, shear, and deflection fields across $z=0.15$ –2.45, as well as CMB lensing at $z\sim1090$ (Hadzhiyska et al., 2023).

5. Statistical Techniques: Covariance Estimation and Variance Suppression

AbacusSummit enables precise estimation of covariance matrices via a dedicated suite of 1883 small-box ( $500\,h^{-1}$ Mpc) periodic simulations (Maksimova et al., 2021). For larger-scale suppression of sample variance, AbacusSummit pairs high-fidelity $N$ -body runs with many quasi- $N$ -body FastPM surrogates under the CARPool control variate methodology (Ding et al., 2022). This approach produces variance reductions by up to $100\times$ at $k<0.3\,h/$ Mpc for power spectra and bispectra, yielding an effective simulation volume two orders of magnitude larger than DESI for these statistics.

The CARPool estimator for an observable $x$ is:

$x = y - \beta(c - \mu_c)$

where $y$ is the measurement from AbacusSummit, $c$ is from FastPM using matched initial conditions, $\mu_c$ is the mean over many independent FastPM runs, and $\beta$ is an optimally estimated coefficient per $k$ -bin. This method is extensible to cross-cosmology applications (Ding et al., 2022).

6. Applications: Galaxy Clustering, BAO, Lyman- $\alpha$ Forest, and Weak Lensing

The suite supports joint analysis of cosmological probes with unprecedented realism and precision:

Baryon Acoustic Oscillations (BAO): AbacusSummit underpins the BAO shift and damping analyses for galaxies and Lyman- $\alpha$ forest. For Ly $\alpha$ , the suite supports both auto- and cross-correlation with quasars, noise-reduced using linear control variates (variance suppression by $\sim7$ –8 on BAO scales) (Hadzhiyska et al., 17 Mar 2025).
Effective Field Theory (EFT): Mock data from AbacusSummit have been used to test EFT-based full-shape modeling of the Ly $\alpha$ forest, demonstrating unbiased BAO parameter inference when one-loop EFT models are applied (Hadzhiyska et al., 17 Mar 2025).
Lensing and Cross-Correlation: High-resolution lensing maps on the light cone are validated against theory to $\sim$ 0.1% for auto- and $\sim$ 1.7% for galaxy-convergence cross-spectra. The calibration allows them to be employed in joint clustering-lensing cosmology pipelines, including DESI, Euclid, and CMB-S4 predictions (Hadzhiyska et al., 2023).

7. Data Access, Utility, and Limitations

All major AbacusSummit data products are publicly released and accessible via the OLCF Constellation portal and associated repositories. Software tools such as abacusutils enable efficient reading and manipulation of halo and particle catalogs, merger trees, and lensing maps (Hadzhiyska et al., 2021).

The suite is recommended for haloes with $M_{\rm halo}>2.1\times10^{11}M_\odot/h$ where mass and positional interpolation are robust. Below this regime, completeness and fidelity of merger trees and interpolated properties degrade. Special care is advised on box-edge haloes; a 10~Mpc/h exclusion is enforced to avoid artifacts from periodic images. For mapping substructure, some bias in central positions and velocities can be compensated within HOD modeling (Hadzhiyska et al., 2021).