Machine learning applications in astrophysics: Reduced-order modelling for chemical kinetics and galaxy merger reconstruction with graph neural network

Tang, Kwok Sun

Permalink

https://hdl.handle.net/2142/124391 Copy

Description

Title

Machine learning applications in astrophysics: Reduced-order modelling for chemical kinetics and galaxy merger reconstruction with graph neural network

Author(s)

Tang, Kwok Sun

Issue Date

2024-04-26

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

Turk, Matthew

Doctoral Committee Chair(s)

Ricker, Paul

Committee Member(s)

• Fields, Brian
• Narayan, Gautham

Department of Study

Astronomy

Discipline

Astronomy

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Astronomy
• Machine Learning
• Astrophysics
• Deep Learning

Abstract

This thesis explores the application of deep learning techniques to two astrophysical problems: simplifying chemical kinetics calculations and reconstructing galaxy merger histories. A framework called Dengo is presented that can automatically generate chemical kinetics solvers from user-specified networks, enabling simplified integration of customized chemistry in simulations. The combination of neural ordinary differential equations and autoencoders is shown to be a promising approach for reducing the complexity of chemical kinetics simulations. Specifically, autoencoders identify the reduced reaction subspace while the neural ODE learn the latent space dynamics. This demonstrates the potential of using deep learning for reduced order modeling of complex chemical networks in astrophysics. For studying galaxy evolution, a conditional graph generative model is developed that can reconstruct the merger histories of observed galaxies from cosmological simulations. This allows identifying progenitors and formation pathways of galaxies across cosmic time. The model captures statistical properties of high-redshift progenitors and enables outlier detection and correlation identification. An overview of deep learning methods with a focus on techniques used in this thesis is also provided. The works demonstrate the potential of using deep learning for the selected problems in computational astrophysics and cosmic structure formation.

Graduation Semester

2024-05

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/124391

Copyright and License Information

© 2024 Kwok Sun Tang

Owning Collections