For decades, cosmologists have treated the universe like a slightly wrinkled sheet. On very large scales, the distribution of matter is "linear"—the wrinkles are small, and their growth is governed by simple, predictable equations. However, as gravity pulls matter into dense filaments and clusters, these wrinkles fold and tangle into a "nonlinear" mess.
While the large-scale patterns tell us about the geometry of the universe, the small-scale, nonlinear clustering contains a wealth of untapped information about dark energy, neutrino masses, and the laws of gravity. Traditionally, researchers have been forced to discard this high-resolution data because our mathematical models simply break down when gravity becomes chaotic.
A new research direction, exemplified by the SimBIG framework, seeks to reclaim this lost information. By moving away from analytic equations and toward Simulation-Based Inference (SBI), researchers are now "forward modeling" the entire history of the universe to extract cosmological parameters directly from the complex, nonlinear web of galaxies.
The Analytic Wall
To understand why we need simulations, we must first understand why equations fail. Standard cosmological analysis, such as the foundational work by Wayne Hu and Daniel J. Eisenstein in Cosmic Complementarity, relies on Perturbation Theory. This approach treats the universe as a smooth background with tiny fluctuations. As long as those fluctuations remain small, we can predict their evolution with high precision.
However, as noted in Measuring Cosmological Parameters with Galaxy Surveys, gravity eventually creates "mildly nonlinear" regimes. In these regions, matter density is no longer a small correction; it is a massive departure from the average. Robert M. Wald argued in New framework for analyzing the effects of small scale inhomogeneities in cosmology that while these small-scale clumps don't mimic dark energy, they are mathematically grueling to model because the derivatives of the spacetime metric become large.
Because we cannot write a simple "likelihood function" (a formula that tells us how probable a galaxy map is given a set of parameters) for these chaotic scales, traditional surveys like BOSS (Lado Samushia et al., The clustering of galaxies in the SDSS-III Baryon Oscillation Spectroscopic Survey) often truncate their data, ignoring everything below a certain spatial scale to avoid theoretical bias.
Forward Modeling: The Simulation Shift
The core intuition of the new direction is simple: If you can't derive the answer, simulate the process.
Instead of trying to "undo" gravity to see the initial conditions, researchers like ChangHoon Hahn and Michael Eickenberg use a "forward modeling" approach. In SimBIG: A Forward Modeling Approach To Analyzing Galaxy Clustering, they start with a set of cosmological parameters, run a dark matter simulation, populate it with galaxies using models like Halo Occupation Distribution (HOD), and then "observe" it through a virtual telescope that mimics real-world errors and fiber collisions.
By repeating this thousands of times with different parameters, they build a library of "possible universes." Simulation-Based Inference then uses machine learning to learn the relationship between the observed patterns and the underlying parameters. This bypasses the need for an analytic likelihood entirely.
Field-Level Inference and Neural Emulators
The power of this approach isn't just in handling nonlinearity; it's in seeing the "hidden" patterns. Traditional methods compress galaxy maps into summary statistics like the "power spectrum." But as Pablo Lemos and Liam Parker demonstrate in SimBIG: Field-level Simulation-Based Inference of Galaxy Clustering, using 3D Convolutional Neural Networks (CNNs) to analyze the raw "field" of galaxies can yield constraints on the growth of structure ($\sigma_8$) that are over 2.6 times tighter than standard methods.
The bottleneck for this method is computational cost. Running thousands of full N-body simulations is expensive. To solve this, researchers are developing "emulators"—deep learning models that can mimic the output of a full simulation in a fraction of the time.
- Elena Giusarma and Shirley Ho developed D3M, which predicts neutrino effects 10,000 times faster than traditional methods (Learning neutrino effects in Cosmology with Convolutional Neural Networks).
- Drew Jamieson and Yin Li introduced field-level emulators that capture the chaotic, virialized motion of particles within dense halos (Field Level Neural Network Emulator for Cosmological N-body Simulations).
The Future: Voids and Wavelets
As we push deeper into the nonlinear regime, new types of structures become valuable. Leander Thiele and Elena Massara have shown that cosmic voids—the empty spaces between filaments—are particularly sensitive to neutrino masses when analyzed via SBI (Neutrino mass constraint from an Implicit Likelihood Analysis of BOSS voids).
Furthermore, techniques like the Wavelet Scattering Transform (WST) are being used to capture non-Gaussian information that the power spectrum misses (Galaxy Clustering Analysis with SimBIG and the Wavelet Scattering Transform).
The transition from analytic theory to simulation-based inference represents a fundamental shift in how we "read" the universe. By embracing the complexity of nonlinear gravity rather than avoiding it, cosmologists are turning the chaotic "noise" of small-scale structures into the most precise signal we have for understanding the fundamental nature of the cosmos.