AI-Driven Reconstruction of Large-Scale Structure from Combined Photometric and Spectroscopic Surveys

Abstract

Galaxy surveys are crucial for studying large-scale structure (LSS) and cosmology, yet they face limitations--imaging surveys provide extensive sky coverage but suffer from photo-$z$ uncertainties, while spectroscopic surveys yield precise redshifts but are sample-limited. To take advantage of both photo-$z$ and spec-$z$ data while eliminating photo-$z$ errors, we propose a deep learning framework based on a dual UNet architecture that integrates these two datasets at the field level to reconstruct the 3D photo-$z$ density field. We train the network on mock samples representative of stage-IV spectroscopic surveys, utilizing CosmicGrowth simulations with a $z=0.59$ snapshot containing $2048^3$ particles in a $(1200~h^{-1}\rm Mpc)^3$ volume. Several metrics, including correlation coefficient, MAE, MSE, PSNR, and SSIM, validate the model's accuracy. Moreover, the reconstructed power spectrum closely matches the ground truth at small scales ($k \gtrsim 0.06~h/\rm Mpc$) within the $1σ$ confidence level, while the UNet model significantly improves the estimation of photo-$z$ power spectrum multipoles. This study demonstrates the potential of deep learning to enhance LSS reconstruction by using both spectroscopic and photometric data.