Modeling Galaxy Bias with Artificial Neural Networks

Abstract

We implement a feedforward neural network to determine the galaxy bias parameters between the underlying dark matter distribution and the distribution of galaxies. We map the density of dark matter to the dark matter halo density, using halo density as a proxy for galaxy density. The performance of transfer functions, for which the bias parameters are functions of wavenumber \(textit{k}\), is the baseline we use to compare the neural network performance. We find that the neural network is able to match the performance of the transfer functions and exceed it with statistical significance at lower resolution. Both models achieve over 90% accuracy up to a maximum frequency \(textit{k}\)\(_{max}\). The transfer functions and neural network work equally well at predicting the large-scale structure at both resolutions. Predicting the extent of the density at lower resolution using the power spectra is where the neural network outperforms the transfer functions above \(textit{k}\)=0.1. At higher resolution, the neural network exhibits slightly higher average performance at high \(textit{k}\), but it is not statistically significant. For both resolutions, the neural network exhibits much higher variability. Because of this variability as well as the time to train the network, transfer functions are the better method. This research shows that machine learning methods should be more physically-motivated in modeling galaxy bias.