Peculiar Velocity Reconstruction From Simulations and Observations Using Deep Learning Algorithms

TL;DR

This paper presents a deep learning Unet model for reconstructing 3D peculiar velocity fields, demonstrating improved accuracy over traditional methods in simulations and successfully applying it to SDSS observational data.

Contribution

The study introduces a novel Unet deep learning approach for velocity reconstruction, outperforming analytical methods and effectively handling realistic observational effects.

Findings

16% improvement in precision over analytical methods

Effective capture of velocity features in non-linear regions

Successful application to SDSS DR7 data

Abstract

In this paper, we introduce a Unet model of deep learning algorithms for reconstructions of the 3D peculiar velocity field, which simplifies the reconstruction process with enhanced precision. We test the adaptability of the Unet model with simulation data under more realistic conditions, including the redshift space distortion (RSD) effect and halo mass threshold. Our results show that the Unet model outperforms the analytical method that runs under ideal conditions, with a 16% improvement in precision, 13% in residuals, 18% in correlation coefficient and 27% in average coherence. The deep learning algorithm exhibits exceptional capacities to capture velocity features in non-linear regions and substantially improve reconstruction precision in boundary regions. We then apply the Unet model trained under SDSS observational conditions to the SDSS DR7 data for observational 3D peculiar velocity reconstructions.