Simulation-Based Forward Modeling for Cosmological Inference

26 concepts25 papers2 proposals

To understand the history and composition of our universe, cosmologists look at the large-scale distribution of matter. By analyzing how galaxies cluster together across billions of light-years, we can infer fundamental parameters like the total matter density ($\Omega_m$) and the amplitude of cosmic fluctuations ($\sigma_8$).

Traditionally, this inference relies on analytical models. Scientists calculate expected statistical summaries—such as the two-point matter power spectrum—using perturbation theory, and then compare them directly to observations. However, this traditional pipeline hits a hard wall at small, non-linear scales. As gravity pulls matter into dense halos, filaments, and voids, the equations become intractable. Compounding this challenge is the fact that telescopes do not observe dark matter directly; they observe galaxies, which are complex, biased tracers affected by messy astrophysical processes like supernova explosions and black hole feedback.

Instead of trying to mathematically invert these complex observational effects, a powerful alternative paradigm has emerged: simulation-based forward modeling. Rather than simplifying the data to fit an analytical theory, forward modeling runs the physics forward—generating synthetic universes that replicate every stage of the observational pipeline.


The Core Mechanism: The Forward Modeling Pipeline

The goal of a forward model is to start with a set of cosmological parameters and produce a realistic mock catalog (or synthetic survey) that looks exactly like real telescope data. This pipeline typically operates in three distinct layers.

[ Cosmological Parameters ] 
           │
           ▼
 1. Matter Backbone (N-body / Perturbation Theory)
           │
           ▼
 2. Galaxy Population (HOD / Semi-Analytic Models)
           │
           ▼
 3. Survey Realism (Noise, Selection Masks, Geometry)
           │
           ▼
   [ Realistic Mock Catalog ]

1. Simulating the Matter Backbone

The first layer simulates the gravitational backbone of the universe: the distribution of dark matter. Ideally, this is achieved using high-resolution, three-dimensional N-body simulations. However, because running thousands of full N-body simulations is computationally expensive, researchers frequently employ clever approximations depending on the scale of interest.

For instance, Lauren Anderson et al. demonstrated the use of second-order Lagrangian perturbation theory to rapidly generate large ensembles of mock datasets for baryon acoustic oscillation measurements in The clustering of galaxies in the SDSS-III Baryon Oscillation Spectroscopic Survey: baryon acoustic oscillations in the Data Release 9 spectroscopic galaxy sample. On even broader statistical scales, Will J. Percival et al. showed in The Shape of the Sloan Digital Sky Survey Data Release 5 Galaxy Power Spectrum that log-normal distributions can serve as an efficient baseline for estimating survey covariances.

When precise non-linear scales are required, approximate halo catalogs must be carefully validated. To address this, Manuel Colavincenzo et al. performed systematic benchmarking of approximate halo-generation methods against full N-body simulations in Comparing approximate methods for mock catalogues and covariance matrices – III: bispectrum, proving that predictive approximate methods can accurately replicate complex higher-order statistics like the bispectrum.

2. Populating the Universe with Galaxies

Once the dark matter structure is established, the next step is determining where the visible galaxies reside. Simulating every star and gas cloud from first principles across gigaparsec scales is computationally impossible, so cosmologists use empirical shortcuts.

The most prominent method is the Halo Occupation Distribution (HOD), which probabilistically populates dark matter halos with galaxies based on the halo's mass. As shown by G. Weiglein et al. in The clustering of Lyman Alpha Emitting galaxies at z=2–3, HOD models can be calibrated to match the exact number densities and redshift distributions of highly specific, complex galaxy populations. Alternatively, researchers can couple semi-analytic galaxy formation models with N-body backbones to capture physical properties across cosmic time, an approach detailed by Yun Wang et al. in The High Latitude Spectroscopic Survey on the Nancy Grace Roman Space Telescope.

3. Injecting Survey Realism

A mock catalog is only useful if it shares the flaws of the real data. The final layer of the forward model applies instrumental noise, selection effects, and geometric masks unique to specific observatories. For example, H. T. Ihle et al. utilized fast, empirical simulations in Joint Power Spectrum and Voxel Intensity Distribution Forecast on the CO Luminosity Function with COMAP to generate thousands of mock realizations that accounted for line-intensity mapping instrument noise. Similarly, in the context of the cosmic microwave background, F. Finelli⋆ et al. demonstrated in Exploring cosmic origins with CORE: Inflation how synthetic datasets must incorporate realistic instrumental effects to accurately forecast the constraints an observatory can place on primordial physics.


Why It Works: Overcoming the Likelihood Bottleneck

Why invest massive computational power into generating fake universes? The answer lies in how we perform statistical inference.

Stochastic Covariance Estimation

To know whether an observed cosmic pattern is a profound discovery or a random statistical fluke, we must know how our errors are correlated across different scales. This requires a covariance matrix. Because non-linear gravitational collapse introduces complex, non-Gaussian correlations, calculating this matrix analytically is impossible.

By generating thousands of independent forward-modeled mock catalogs—using pipelines like the Quick Particle Mesh or ELEPHANT simulations described by Siddharth Satpathy et al. in Measurement of marked correlation functions in SDSS-III Baryon Oscillation Spectroscopic Survey using LOWZ galaxies in Data Release 12—cosmologists can numerically estimate these covariance matrices directly from the simulations. This provides a robust, empirical baseline to test theoretical consistency and safeguard against systematic errors.

Simulation-Based Inference (SBI)

The true paradigm shift occurs when forward modeling is paired with machine learning to bypass analytical likelihoods entirely. This is known as Simulation-Based Inference (SBI) or Likelihood-Free Inference.

Traditional analysis compresses a galaxy map into a simple summary statistic (like the power spectrum), throwing away the rich, non-Gaussian information hidden in small-scale cosmic webs. Because we have a forward model that can generate realistic data for any set of parameters, we can train deep neural networks to learn the implicit mapping from complex data structures directly to cosmological parameters.

[ Cosmological Parameters ] ──► [ Forward Model ] ──► [ Synthetic Catalogs ]
                                                             │
                                                             ▼
[ Inferred Parameters ]     ◄── [ Trained Network ] ◄────────┘

A prime example of this is the SimBIG framework. As explained by ChangHoon Hahn et al. in A forward modeling approach to analyzing galaxy clustering with S im BIG, by training normalizing flows on forward-modeled simulations that capture non-linear scales and survey realism, SBI can extract significantly more precise constraints on parameters like $\sigma_8$—effectively providing a statistical gain equivalent to a 60% larger traditional galaxy survey.

To squeeze even more information out of these catalogs, researchers are exploring advanced data representations:


Taming Astrophysical Uncertainty

The primary vulnerability of simulation-based inference is model misspecification: if the simulation does not reflect reality, the trained neural network will yield biased results. This is particularly dangerous when modeling baryonic feedback, as the exact physics of how gas, stars, and black holes interact remains uncertain.

To solve this, modern simulation suites vary both cosmological parameters and astrophysical feedback parameters simultaneously. The CAMELS (Cosmic Afternoon & Evening Merger Experiment Large-scale Structure) suite provides thousands of hydrodynamic simulations spanning diverse feedback prescriptions.

By training models across this multi-dimensional parameter space, researchers can find solutions that are immune to astrophysical variations. For example, Digvijay Wadekar et al. utilized the CAMELS suite alongside symbolic regression in The SZ flux-mass (i>Y–M) relation at low-halo masses: improvements with symbolic regression and strong constraints on baryonic feedback to discover a novel cluster mass scaling relation that remains robust regardless of the underlying baryonic feedback model.

Furthermore, these varied suites enable generative models to learn the underlying physics. In 3D Reconstruction of Dark Matter Fields with Diffusion Models: Towards Application to Galaxy Surveys, Core Francisco Park et al. trained a 3D diffusion model on the CAMELS Multifield Dataset, demonstrating that generative models can successfully reconstruct underlying dark matter fields from sparse galaxy distributions even when subjected to highly varied astrophysical backgrounds.

By shifting the analytical burden from mathematical inversion to forward simulation, cosmologists are unlocking the full, non-linear potential of upcoming astronomical surveys, ensuring that our models can finally match the precision of our telescopes.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_camels_simulation_suite_77e2e3fb,
  title = {CAMELS simulation suite},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/camels-simulation-suite/rdr_77e2e3fb8d9b4fd1998afe412c557299},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_camels_simulation_suite_77e2e3fb_v1,
  title = {CAMELS simulation suite},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/camels-simulation-suite/rdr_77e2e3fb8d9b4fd1998afe412c557299/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake