Testing Theories About Our Universe’s Evolution with Convolutional Neural Networks
Using CNNs to predict galaxy formation by leveraging dark matter properties
It is rare for most people to step back and consider their own small place in the universe — except perhaps in the field of cosmology. From the Big Bang to the present and beyond, cosmologists attempt to understand the universe and its development. A new research paper describing a novel approach to simulating cosmological evolution via convolutional networks that rely chiefly on dark matter. Researchers involved in the publication include CDS’ Xinyue Zhang, Yanfang Wang, Wei Zhang, Yueqiu Sun, along with Siyu He, Center for Computational Astrophysics, Flatiron Institute, Gabriella Contardo and Francisco Villaescusa-Navarro, both of Flatiron Institute Center for Computational Astrophysics, and Shirley Ho, Flatiron Institute Center for Computational Astrophysics.
A key aim of cosmology is to “understand and define the physical rules and parameters that led to our actual universe.” For this, cosmologists require computer simulations to model their theories and determine whether the observed universe could indeed be produced by the suggested conditions. Predictably, fast forwarding through the advent of galaxies over light years of evolution is not quite as simple as hitting a button on a remote. In fact, it turns out that simulating even a tiny fraction of the universe with the relevant physics presently takes about 2,000 years on a single CPU–and this is a state of the art, fully gravo-hydrodynamical cosmological simulation.
Since it is a time intensive, difficult task to model the layout of the universe from the beginning of time to our current, observed universe, researchers propose a solution offered by the standard cosmological model. Matter in the universe is primarily dark matter. Because of this, it is possible to model the structure of the universe quite accurately simply by evolving dark matter through time. This system only factors in the physics of gravity. Gas collects in pockets formed by dark matter (“halos”) and cools down, creating galaxies and stars. Researchers logically conclude that there is a strong relationship to leverage between dark matter halos and the emergence of galaxies. Dark matter halos enable researchers to learn about “growth, internal properties, and spatial distribution of galaxies.”
The preferred approach to model dark matter distribution is a gravity-only N-body simulation. The approach is less computationally expensive than those that include more physics properties. It also predicts dark matter position and velocity. Unfortunately, this simulation does not include galaxy distribution. Thus, the problem becomes how to best link the 3D dark matter field from N-body simulations to the 3D galaxy distribution from hydrodynamic simulations. Researchers propose a CNN to map “the 3D matter field in an N-body simulation to galaxies in a full hydrodynamic simulation.” The CNN combats issues introduced by the spatial nature of the data, and the two-phase architecture and learning process contends with the high sparsity of the galaxy distribution. The two-phase approach refers to breaking training into separate processes. This is necessary to prevent artificially high accuracy caused by the imbalanced distribution of input and output.
Researchers introduce existing network architectures and describe modifications that make them better suited to cosmology. Specifically, researchers studied U-Net, a fully convolutional neural network, and R2U-Net (Recurrent Residual U-Net), which was proposed as an upgrade of U-Net and consists of two stacked RCNNs. Additionally, Zhang et al. experimented with inception networks. After experimentation, researchers concluded that the model with the best performance involved two phases; first, an inception network, and then, R2U-Net. This model outperforms or equals the results produced by HOD (Halo Occupation Distribution), the computationally expensive benchmark against which researchers measured their proposed method. The two phase model also offers better scaling and generalization abilities. These findings open up opportunities for developing and training the model further, as well as establishing a link between cosmology and astronomy. This research represents a very exciting step forward in cosmology as it circumvents the need for HODs, and reduces the time needed to answer critical research questions.
By Sabrina de Silva
