NERV: Neural-network Enhanced Reconstruction of the UniVerse with Application to Baryon Acoustic Oscillations in the BOSS DR12 Galaxy Sample

Abstract

We present the first application of neural-network-based baryon acoustic oscillation (BAO) reconstruction to real galaxy survey data, restoring the acoustic signature damped by nonlinear structure growth. {\texttt{NERV}} ({\bf N}eural-network {\bf E}nhanced {\bf R}econstruction of the Uni{\bf V}erse) augments standard reconstruction with a convolutional neural network trained on cubic $N$-body simulations, and explicitly accounts for realistic observational effects including the curved-sky geometry, the redshift-dependent selection function, and finite survey boundaries, by tessellating the survey volume into local patches. We validate the method on the \textsc{MultiDark-Patchy} mock catalogs, recovering unbiased BAO dilation parameters. Applied to the BOSS DR12 galaxy sample, NERV improves the precision of the BAO distance measurements significantly. These results establish neural reconstruction as a practical component of BAO analyses for ongoing surveys such as DESI, with the potential to substantially tighten constraints on the cosmic expansion history and the nature of dark energy.

Abstract: arXiv descriptive metadata (CC0). Thank you to arXiv for use of its open access interoperability. This service was not reviewed or approved by, nor does it necessarily express or reflect the policies or opinions of, arXiv.