Topics at the interface of machine learning and high-performance computing for gravitational wave astrophysics

Khan, Asad

Permalink

https://hdl.handle.net/2142/115341 Copy

Description

Title

Topics at the interface of machine learning and high-performance computing for gravitational wave astrophysics

Author(s)

Khan, Asad

Issue Date

2022-03-16

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

• Huerta, Eliu A.
• Seidel, Harry Edward

Doctoral Committee Chair(s)

Allen, Gabrielle D.

Committee Member(s)

• Zhao, Zhizhen
• Cooper, S. 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)

• General Relativity
• Gravitational Waves
• LIGO
• Multimessenger Astrophysics
• Artificial Intelligence
• High-Performance Computing

Abstract

Multi-Messenger Astrophysics (MMA) aims to provide key insights into the origin, evolution and large scale structure of the universe by probing and combining information from different cosmic messengers; gravitational waves (GW), electromagnetic waves (EM), cosmic rays, and neutrinos. The recent direct detections by LIGO have firmly added gravitational waves (GW) to this list of cosmic messengers and thus a new era of GW astrophysics has started. However, the full potential for scientific discovery of current and future MMA surveys in the context of gravitational waves is limited by the key challenges involved in; (i) real time detection of GW signals, (ii) rapid and accurate parameter estimation from those signals, (iii) rapid modeling of numerical relativity waveforms covering the entire range of parameters of interest, and (iv) extreme computational costs and high latency of analyzing rapidly increasing volumes of observational data. In this thesis, we present a framework that combines deep learning (DL) and high performance computing (HPC) to contend with many of the existing challenges of real-time MMA. In particular; (i) we present an ensemble of deep learning models that can detect gravitational wave events from compact binary coalescences in real-time. When deployed on archival data from LIGO's O2 and O3 runs, this framework is able to identify all events with previously confirmed detections, and has a false alarm rate of 1 misclassification for every 2.7 days of searched data. We then develop a workflow to demonstrate the scaling of such an analysis on an HPC platform to reduce time to insight. (ii) We also study the efficacy of deep learning models to estimate parameters from waveforms of spinning, non-precessing bninary black hole mergers in the absence of noise, the ability of these models to handle parameter space degeneracies, and the effects of including higher order modes on the accuracy of parameter estimations. (iii) Finally, we test the ability of deep learning to model full numerical relativity waveforms, as an alternative to surrogate modeling and Gaussian emulation methods. In the last chapter of this thesis, we present a deep learning model that can efficiently and accurately forecast the late-inspiral, merger and ringdown of numerical relativity waveforms that describe quasi-circular, spinning, non-precessing binary black hole mergers. These results indicate that combining deep learning algorithms and high performance computing is a promising new paradigm that can help accelerate scientific discovery in the era of multi-messenger astrophysics.

Graduation Semester

2022-05

Type of Resource

Thesis

Copyright and License Information

Copyright 2022 Asad Khan

Owning Collections