Deep learning methods for enabling real-time gravitational wave and multimessenger astrophysics

George, Daniel

Permalink

https://hdl.handle.net/2142/102945 Copy

Description

Title

Deep learning methods for enabling real-time gravitational wave and multimessenger astrophysics

Author(s)

George, Daniel

Issue Date

2018-12-07

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

Allen, Gabrielle D

Doctoral Committee Chair(s)

Allen, Gabrielle D

Committee Member(s)

• Huerta, Eliu A
• Seidel, Harry E
• Zhao, Zhizhen
• Fields, Brian D.
• Shutz, Bernard F

Department of Study

Astronomy

Discipline

Astronomy

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Deep Learning
• Convolutional Neural Networks
• Gravitational Waves
• LIGO
• Time-series Signal Processing
• Classification
• Regression
• Multimessenger Astrophysics
• Transfer Learning
• Anomaly Detection
• Unsupervised Clustering

Abstract

A new era of gravitational wave (GW) astronomy has begun with the recent detections by LIGO. However, we need real-time observations of GW signals and their electromagnetic (EM) and astro-particle counterparts to unlock its full potential for scientific discoveries. Extracting and classifying the wide range of modeled and unmodeled GWs, whose amplitudes are often much weaker than the background noise, and rapidly inferring accurate parameters of their source is crucial in enabling this scenario of real-time multimessenger astrophysics. Identifying and automatically clustering anomalous non-Gaussian transient noises (glitches) that frequently contaminate the data and separating them from true GW signals is yet another difficult challenge. Currently, the most sensitive data analysis pipelines are limited by the extreme computational costs of template-matching methods and thus are unable to scale to all types of GW sources and their full parameter space. Accurate numerical models of GW signals covering the entire range of parameters including eccentric and spin-precessing compact binaries, which are essential to infer the astrophysical parameters of an event, are not available. Searches for unmodeled and anomalous signals do not have sufficient sensitivity compared to the targeted searches. Furthermore, existing search pipelines are not optimal for dealing with the non-stationary, non-Gaussian noise in the detectors. This indicates that many critical events will go unnoticed. The primary objective of this thesis is to resolve these issues via deep learning, a state-of-the-art machine learning method based on artificial neural networks. In this thesis we develop robust GW analysis algorithms for analyzing real LIGO/Virgo data based on deep learning with neural networks, that overcomes many limitations of existing techniques, allowing real-time detection and parameter estimation modeled GW sources and unmodeled GW bursts as well as classification and unsupervised clustering of anomalies and glitches in the detectors. This pipeline is designed to be highly scalable, therefore it can be trained with template banks of any size to cover the entire parameter-space of eccentric and spin-precessing black hole binaries as well as other sources and also optimized based on the real-time characteristics of the complex noise in the GW detectors. This deep learning framework may also be extended for low-latency analysis of the raw big data collected across multiple observational instruments to further facilitate real-time multimessenger astrophysics, which promises groundbreaking scientific insights about the origin, evolution, and destiny of the universe. In addition, this work introduces a new paradigm to accelerate scientific discovery by using data derived from high-performance physics simulations on supercomputers to train artificial intelligence algorithms that exploit emerging hardware architectures.

Graduation Semester

2018-12

Type of Resource

text

Permalink

http://hdl.handle.net/2142/102945

Copyright and License Information

Copyright 2018 Daniel George

Owning Collections