Effective cosmic density field reconstruction with convolutional neural network

Abstract We present a cosmic density field reconstruction method that augments the traditional algorithms with convolutional neural network (CNN). Following previous work, key component of our is to use reconstructed as input network. extend this work by exploring how performance these ideas depends on algorithm, parameters, and shot noise field, well robustness method. build an eight-layer CNN train fields computed from Quijote suite simulations. The are generated both standard algorithm new iterative algorithm. In real space at z = 0, we find 90% correlated true initial out $k\sim 0.5\, h\rm {Mpc}^{-1}$, significant improvement over 0.2\, {Mpc}^{-1}$ achieved algorithms. similar improvements in redshift space, including improved removal distortions small scales. also robust across changes cosmology. Additionally, removes much variance choice different parameters. However, effectiveness decreases increasing noise, suggesting such approach best suited high samples. This highlights additional information beyond linear scales power complementing analysis approaches machine learning techniques.