Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

High-resolution weak lensing mass mapping from DES-Y3 data using diffusion-based prior

Supranta S. Boruah, Michael Jacob, Bhuvnesh Jain, Riya Maiya, Raghav Venkataramanan

TL;DR

The paper tackles the challenge of reconstructing high-resolution dark matter mass maps from weak-lensing data, where traditional Gaussian priors fail to capture non-Gaussian late-time structure. It introduces diffusion-based mass-mapping using Diffusion Posterior Sampling (DPS) with a learned non-Gaussian prior from N-body simulations and a time-dependent likelihood scaling to yield unbiased, well-calibrated posteriors for the convergence field ${\boldsymbol{\kappa}}$ given DES-Y3 shear data ${\mathbf{D}}$. The approach achieves 1 arcmin resolution with enhanced small-scale detail and calibrated uncertainties, validated on simulations (achieving accurate cross-correlations up to $\ell \lesssim 1000$) and applied to DES-Y3 patches where reconstructed maps correlate with known clusters. This diffusion-based framework enables detailed astrophysical studies of filaments, voids, and galaxy formation environments and provides a robust, extensible method for upcoming Stage-IV surveys and non-Gaussian likelihoods.

Abstract

High-resolution mapping of cosmic mass distribution is essential for a variety of astrophysical applications including understanding cosmic structure formation, and galaxy formation and evolution. However dark matter is not directly observed and therefore we need advanced methods for solving inverse problems to reconstruct the underlying cosmic matter distribution. Here, we train a generative diffusion model and use it in the Diffusion Posterior Sampling (DPS) framework to reconstruct mass maps from Dark Energy Survey-Year 3 (DES-Y3) weak gravitational lensing data at high (1 arcminute) resolution. We show that the standard DPS results are biased, but they can be easily corrected by scaling the log-likelihood score during the diffusion process, yielding unbiased results with proper uncertainty quantification. The resulting mass maps reveal cosmic structures with enhanced detail, opening the door for improved astrophysical studies using the obtained mass maps.

High-resolution weak lensing mass mapping from DES-Y3 data using diffusion-based prior

TL;DR

The paper tackles the challenge of reconstructing high-resolution dark matter mass maps from weak-lensing data, where traditional Gaussian priors fail to capture non-Gaussian late-time structure. It introduces diffusion-based mass-mapping using Diffusion Posterior Sampling (DPS) with a learned non-Gaussian prior from N-body simulations and a time-dependent likelihood scaling to yield unbiased, well-calibrated posteriors for the convergence field given DES-Y3 shear data . The approach achieves 1 arcmin resolution with enhanced small-scale detail and calibrated uncertainties, validated on simulations (achieving accurate cross-correlations up to ) and applied to DES-Y3 patches where reconstructed maps correlate with known clusters. This diffusion-based framework enables detailed astrophysical studies of filaments, voids, and galaxy formation environments and provides a robust, extensible method for upcoming Stage-IV surveys and non-Gaussian likelihoods.

Abstract

High-resolution mapping of cosmic mass distribution is essential for a variety of astrophysical applications including understanding cosmic structure formation, and galaxy formation and evolution. However dark matter is not directly observed and therefore we need advanced methods for solving inverse problems to reconstruct the underlying cosmic matter distribution. Here, we train a generative diffusion model and use it in the Diffusion Posterior Sampling (DPS) framework to reconstruct mass maps from Dark Energy Survey-Year 3 (DES-Y3) weak gravitational lensing data at high (1 arcminute) resolution. We show that the standard DPS results are biased, but they can be easily corrected by scaling the log-likelihood score during the diffusion process, yielding unbiased results with proper uncertainty quantification. The resulting mass maps reveal cosmic structures with enhanced detail, opening the door for improved astrophysical studies using the obtained mass maps.

Paper Structure

This paper contains 8 sections, 3 equations, 3 figures.

Table of Contents

  1. Introduction
  2. Weak lensing mass mapping using diffusion posterior sampling
  3. Diffusion Posterior Sampling and likelihood scaling
  4. Training data and likelihood
  5. Results
  6. Validation on simulation
  7. Mass map reconstruction of DES-Y3 data
  8. Conclusion

Figures (3)

  • Figure 1: Left: Standard DPS (red) makes the latent maps go out-of-distribution. We resolve this with our likelihood scaling method (green). Right: TARP coverage test showing well-calibrated uncertainty quantification (following the $x=y$ line) with likelihood scaling, whereas the standard DPS gives biased results.
  • Figure 2: Validation of our method on simulations. Top row: True underlying mass map (left), mean of reconstructed mass map samples (center left), standard deviation of reconstructed samples (center right), and survey mask (right). Bottom row: Cross-correlation between reconstructed samples and true map (left), followed by comparisons of power spectrum, PDF, and peak counts between reconstructed samples (blue) and true map (black).
  • Figure 3: Comparison of the reconstructed maps from the DES-Y3 data in a $4.26^{\circ} \times 4.26^{\circ}$ representative patch. The left panel shows the mean of the DPS reconstructions, the center left panel shows the Weiner filtered reconstruction, and the third panel shows the Null-B reconstruction. The right panel shows the standard deviations computed from the DPS reconstructions. DES-Y3 RedMapper clusters are overplotted with red circles.