GalaxyNet: Connecting galaxies and dark matter haloes with deep neural networks and reinforcement learning in large volumes

TL;DR

GalaxyNet is a deep neural network trained with reinforcement learning that accurately models galaxy properties and their relation to dark matter haloes, enabling predictions of large-scale structure and galaxy clustering up to high redshift.

Contribution

This work introduces GalaxyNet, a novel neural network architecture trained on observed galaxy data using reinforcement learning, to connect galaxy properties with dark matter haloes.

Findings

GalaxyNet reproduces observed galaxy statistics with high accuracy (reduced chi-squared 1.05).

It predicts a stellar-to-halo mass relation with redshift-dependent features.

GalaxyNet enables large-volume galaxy population simulations for future surveys.

Abstract

We present the novel wide & deep neural network GalaxyNet, which connects the properties of galaxies and dark matter haloes, and is directly trained on observed galaxy statistics using reinforcement learning. The most important halo properties to predict stellar mass and star formation rate (SFR) are halo mass, growth rate, and scale factor at the time the mass peaks, which results from a feature importance analysis with random forests. We train different models with supervised learning to find the optimal network architecture. GalaxyNet is then trained with a reinforcement learning approach: for a fixed set of weights and biases, we compute the galaxy properties for all haloes and then derive mock statistics (stellar mass functions, cosmic and specific SFRs, quenched fractions, and clustering). Comparing these statistics to observations we get the model loss, which is minimised with particle swarm optimisation. GalaxyNet reproduces the observed data very accurately ($\chi_\mathrm{red}=1.05$), and predicts a stellar-to-halo mass relation with a lower normalisation and shallower low-mass slope at high redshift than empirical models. We find that at low mass, the galaxies with the highest SFRs are satellites, although most satellites are quenched. The normalisation of the instantaneous conversion efficiency increases with redshift, but stays constant above $z\gtrsim0.7$. Finally, we use GalaxyNet to populate a cosmic volume of $(5.9~\mathrm{Gpc})^3$ with galaxies and predict the BAO signal, the bias, and the clustering of active and passive galaxies up to $z=4$, which can be tested with next-generation surveys, such as LSST and Euclid.