Date of Award
2022
Document Type
Open Access Dissertation
Degree Name
Doctor of Philosophy in Applied Physics (PhD)
Administrative Home Department
Department of Physics
Advisor 1
Elena Giusarma
Committee Member 1
Robert Nemiroff
Committee Member 2
Francisco Villaescusa-Navarro
Committee Member 3
Mauricio Reyes
Committee Member 4
Petra Huentemeyer
Abstract
Multi-billion dollar cosmological surveys are being conducted almost every decade in today’s era of precision cosmology. These surveys scan vast swaths of sky and generate tons of observational data. In order to extract meaningful information from this data and test these observations against theory, rigorous theoretical predictions are needed. In the absence of an analytic method, cosmological simulations become the most widely used tool to provide these predictions in order to test against the observations. They can be used to study covariance matrices, generate mock galaxy catalogs and provide ready-to-use snapshots for detailed redshift analyses. But cosmological simulations of matter formation in the universe are one of the most computationally intensive tasks. Faster but equally reliable tools that could approximate these simulations are thus desperately needed. Recently, deep learning has come up as an innovative and novel tool that can generate numerous cosmological simulations orders of magnitude faster than traditional simulations. Deep learning models of structure formation and evolution in the universe are unimaginably fast and retain most of the accuracy of conventional simulations, thus providing a fast, reliable, efficient, and accurate method to study the evolution of the universe and reducing the computational burden of current simulation methods.
In this dissertation, we will focus on deep learning-based models that could mimic the process of structure formation in the universe. In particular, we focus on developing deep convolutional neural network models that could learn the present 3D distribution of the cold dark matter and generate 2D dark matter cosmic mass maps. We employ summary statistics most commonly employed in cosmology and computer vision to quantify the robustness of our models.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Kaushal, Neerav, "TOWARD DEEP LEARNING EMULATORS FOR MODELING THE LARGE-SCALE STRUCTURE OF THE UNIVERSE", Open Access Dissertation, Michigan Technological University, 2022.