Finding voids billions of years in advance with machine learning

Abstract

The $\Lambda$CDM model is the standard model of cosmology, able to describe the large-scale structure and evolution of the Universe with a handful of empirically determined and constrained free parameters. Two mysterious components of the model, a gravitationally interacting dark matter and an expansive dark energy, which contribute nearly all of the Universe’s energy density in the model, have yet to be explained. In addition, the driving force behind the formation of galaxies and large-scale structure is likely found in dark matter and dark energy. Cosmic voids, underdense regions within the large-scale structure of the Universe, could give us insight into these unknowns. Voids fill most of the volume of the Universe and to date have been studied significantly less than dense cosmic halos. Their size and low density allow them to preserve memory of the Universe’s initial conditions with greater fidelity and make them more sensitive to key unresolved cosmological puzzles including structure growth, dark energy, modified gravity, the sum of neutrino masses, and galaxy formation. In order to extract the cosmological information embedded in voids, we first have to identify them, which is a computationally expensive task that involves running a complex simulation from the relatively simple initial conditions to the desired redshift, e.g. z=0, and running a void finder. In this research, we therefore apply machine learning algorithms ranging from random forests to neural networks to previously conducted cosmological simulations in an attempt to predict the formation of voids without having to run new, expensive simulations. We find that, given density features, the random forest and multilayer perceptron neural network algorithms are able to predict the final "in-void" or "not-in-void" class label of dark matter particles to a high degree of accuracy. Physical implications of the finding should be pursued in further research.