| dc.description.abstract | Understanding how structures in the universe form and grow under the influence of gravity is fundamental to our understanding of how the universe we observe came to be. However, simulations of dark matter, which are one of the most common testbeds for probing structure formation, are incredibly computationally expensive to run, acting as a potential barrier to some studies. The rapidly developing field of machine learning has offered many powerful tools for mapping complex relationships between inputs and outputs. These tools can be used for a variety of astrophysical studies to model highly complex behaviors, saving time and in some cases leading to increased accuracy. In this research, we use different machine learning algorithms to make predictions for how cosmological structures grow and evolve in simulations of dark matter, on both the small scale of halo merging, and on the much larger scale of the cosmic web as a whole.
The research begins by predicting the survivability, mass loss, final position, and merge time of dark matter subhalos as they fall into larger host halos, using the decision tree based methods of random forests and gradient boosting regressors. This identifies only 5 physically-motivated parameters as being necessary to predict the evolution of a subhalo within its host, but additionally points to a degree of stochasticity in the merging process that cannot be explained by only the features from the start of the interaction. Next, this research investigates the use of convolutional neural networks to predict the full redshift zero results of a dark-matter only simulation from its initial conditions. A U-Net that uses the mean squared error as the loss function and one that uses a discriminator network with a generative adversarial network structure are both trained on this problem and their respective generated cosmic webs are compared. The models developed in this research have use cases where they could significantly reduce the computational cost associated with running complete high-resolution simulations and point to promising avenues for future development of AI simulators of cosmological behaviors. | |
