Dark Energy Survey Year 3 results: Simulation-based $w$CDM inference from weak lensing and galaxy clustering maps with deep learning. I. Analysis design

TL;DR

The paper tackles the challenge of extracting non-Gaussian information from cosmological large-scale structure by introducing a simulation-based inference pipeline that jointly analyzes DES Y3 weak lensing and galaxy clustering maps. It combines a forward model (CosmoGridV1) with map-level neural compression (DeepSphere) and neural likelihood estimation (normalizing flows) to infer a 10-parameter space including $w$CDM, intrinsic alignment, and linear galaxy bias while marginalising nuisance parameters. The authors validate the approach on extensive synthetic mocks (CosmoGridV1 and Buzzard), establish conservative scale cuts, and demonstrate substantial improvements in cosmological parameter constraints—up to ~189% FoM increase in certain planes—over baseline two-point statistics, highlighting the method’s potential for DES Y3 and future Stage-IV surveys. The work also provides a robust framework for handling baryonic effects, redshift errors, and shear biases within SBI and emphasizes the value of map-level compression in breaking degeneracies and extracting non-Gaussian information from LSS data.

Abstract

Data-driven approaches using deep learning are emerging as powerful techniques to extract non-Gaussian information from cosmological large-scale structure. This work presents the first simulation-based inference (SBI) pipeline that combines weak lensing and galaxy clustering maps in a realistic Dark Energy Survey Year 3 (DES Y3) configuration and serves as preparation for a forthcoming analysis of the survey data. We develop a scalable forward model based on the CosmoGridV1 suite of N-body simulations to generate over one million self-consistent mock realizations of DES Y3 at the map level. Leveraging this large dataset, we train deep graph convolutional neural networks on the full survey footprint in spherical geometry to learn low-dimensional features that approximately maximize mutual information with target parameters. These learned compressions enable neural density estimation of the implicit likelihood via normalizing flows in a ten-dimensional parameter space spanning cosmological $w$CDM, intrinsic alignment, and linear galaxy bias parameters, while marginalizing over baryonic, photometric redshift, and shear bias nuisances. To ensure robustness, we extensively validate our inference pipeline using synthetic observations derived from both systematic contaminations in our forward model and independent Buzzard galaxy catalogs. Our forecasts yield significant improvements in cosmological parameter constraints, achieving $2-3\times$ higher figures of merit in the $Ω_m - S_8$ plane relative to our implementation of baseline two-point statistics and effectively breaking parameter degeneracies through probe combination. These results demonstrate the potential of SBI analyses powered by deep learning for upcoming Stage-IV wide-field imaging surveys.

Figures (24)

• Figure 1: Normalized redshift distributions of (a) the Metacalibration source galaxy sample used for weak lensing and (b) the Maglim lens galaxy sample used for galaxy clustering. The colored lines indicate the base distributions, while the partially transparent overlapping gray lines illustrate the redshift uncertainty via fifty draws from the photo-$z$ distributions parametrized in \ref{['eq:z_error_metacal', 'eq:z_error_maglim']} according to \ref{['tab:prior_cosmogrid']}.
• Figure 2: Schematic overview of the processing pipeline we apply to transform (baryonified) particle shells from the CosmoGridV1 simulations into mock weak lensing and galaxy clustering maps matching selected DES Y3 properties. Sharp-cornered boxes represent processing steps, while blue rounded boxes denote (full- or partial-sky) HEALPix maps. Green and red ellipses indicate constrained and marginalized parameters, respectively. The tomographic bin index $i \in \{1,2,3,4\}$ for source (subscript $s$) and lens (subscript $l$) galaxy samples is omitted from map names for simplicity. The abbreviation "WL" indicates weak lensing signal, "IA" intrinsic alignment, and "SN" shape noise. We denote the scrambled shear catalog of randomly rotated source galaxies used to generate shape noise maps as $\tilde{\gamma}_\textsc{Metacal}\xspace$, and omit the $\kappa_\mathrm{IA,TA}$ (see \ref{['eq:kappa_sum']}) map for clarity.
• Figure 3: Projection of the fiducial and $2 \, 500$ grid cosmologies included in the CosmoGridV1 to the $\Omega_m\xspace - \sigma_8\xspace$ plane. Dashed and solid lines represent the wide and narrow priors, respectively, with the corresponding points colored in blue and orange. The black star marks the fiducial cosmology.
• Figure 4: Full-sky mollweide projection of (a) how the original DES Y3 footprint (blue) is rotated to a position (red) that allows for (b) four non-overlapping cutouts. These distinctly colored patches are related by HEALPix symmetries such that there is a perfect one-to-one correspondence between the pixels. The black padding with zeros along patch 1 is determined by the lowest $n_\textrm{side}$ used within the DeepSphere networks; $n_\textrm{side}\xspace = 16$ here. It is not part of the original survey area.
• Figure 5: Scatterplot depicting the source galaxy bias $b_{g,s}\xspace$ we find for the $2 \, 500$ unique cosmologies of the CosmoGridV1 by fitting the source galaxy number count histogram to a reference Buzzard simulation of fixed cosmology.