Subgrid modeling of general relativistic magnetohydrodynamic turbulence with physics informed deep learning

Rosofsky, Shawn Garrett

Permalink

https://hdl.handle.net/2142/120193 Copy

Description

Title

Subgrid modeling of general relativistic magnetohydrodynamic turbulence with physics informed deep learning

Author(s)

Rosofsky, Shawn Garrett

Issue Date

2023-01-04

Director of Research (if dissertation) or Advisor (if thesis)

• Huerta, Eliu
• Seidel, Edward

Doctoral Committee Chair(s)

Allen, Gabrielle

Committee Member(s)

• Zhao, Zhizhen
• Cooper, Lance

Department of Study

Physics

Discipline

Physics

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Gravitational Waves
• Numerical Relativity
• Deep Learning
• Neutron Stars
• General Relativity
• Magnetohydrodynamics

Abstract

The recent detections of gravitational waves (GWs) from mergers of binary black holes (BBH), binary neutron stars (BNS), and black hole neutron stars (BHNS) by LIGO have allowed us to view the universe like never before. BNS mergers in particular allow us to combine our GW observations with the electromagnetic (EM) counterparts to provide even more insight into these extreme events. To better understand the inner workings of neutron stars and BNS mergers, we would like to simulate BNS mergers from start to finish and obtain their GW signal along with their EM counterpart. We employ numerical relativity (NR) simulations to model the dynamic evolution of spacetime for this purpose. To model neutron stars, we add general relativistic magnetohydrodynamics (GRMHD) to these NR simulations to evolve the neutron stars and their magnetic fields in a curved spacetime. However, at the merger, the Kelvin-Helmholz instability (KHI) creates GRMHD turbulence that amplifies the magnetic field at a smaller scale than BNS simulations are capable of resolving. This prevents us from accurately modeling the EM counterparts of BNS mergers in NR simulation. This thesis seeks to address such GRMHD turbulence by using artificial neural networks (ANNs) as subgrid-scale (SGS) models for this turbulent amplification. We begin by examining these ANN based SGS models within the large eddy simulation (LES) formulation. We study how these ANN models perform on precomputed data and demonstrate that ANN models are superior to the state of the art traditional models of Newtonian magnetohydrodynamic (MHD) turbulence designed for this purpose. Then, we implemented these networks in actual simulations, illustrating their effect on the fluid evolution and their limitations. To overcome such limitations, we looked towards physics informed machine learning and physics informed neural operators (PINO) in particular. We tested PINOs on a wide variety of problems ranging from the wave equation to the nonlinear shallow water equations to assess their strengths and weaknesses. Encouraged by the flexibility and accuracy of PINOs, we applied PINOs to the Newtonian incompressible MHD equations and assessed their performance as we increased the kinetic and magnetic Reynolds numbers. The results of this thesis provides a stepping stone for future work in resolving GRMHD turbulence associated with the KHI. In addition, this thesis presents our direct contributions to the field of NR that are not directly related to the GRMHD turbulence problem. The first of these works on the necessary resolutions for NR neutron star simulations to compute the theoretically accurate frequencies and damping times of f-mode oscillations. We provide our contribution to two studies involving NR simulations of eccentric BBH mergers. One involves creating a large catalog of non-spinning eccentric BBH mergers. The other work focuses on running and closely examining a smaller set of spinning eccentric BBH simulations.

Graduation Semester

2023-05

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/120193

Copyright and License Information

Copyright 2022 Shawn Rosofsky

Owning Collections