Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Searching for binary black hole mergers with deep learning in Advanced LIGO's third observing run

Damon Beveridge, Alistair McLeod, Linqing Wen, Weichangfeng Guo, Andreas Wicenec

TL;DR

The paper addresses detecting binary-black-hole mergers in gravitational-wave data using a hybrid approach that couples matched filtering with deep learning. It builds a pipeline that produces SNR time series from a BBH template bank, then scores triggers with a pretrained DL model to yield a ranking statistic, enabling robust FAR and $p_{ m astro}$ estimates. On O3 data, the method demonstrates sensitivity comparable to LVK pipelines for high-mass systems ($m_{ m chirp}\gtrsim 25\,M_{\

Abstract

The detection of gravitational waves from compact binary coalescences has provided significant insights into our Universe, and the discovery of new and unique gravitational wave candidates from independent searches remains an ongoing field of research. In this work, we built a hybrid search pipeline that combines matched filtering and deep learning to identify stellar-mass binary black hole candidates from detector strain data. We first present results from a targeted injection study to benchmark the sensitivity of our method and compare it with existing search pipelines. We demonstrate that our hybrid approach has comparable sensitivity for injections with a source-frame chirp mass greater than 25$\,$M$_{\odot}$, and below this threshold our sensitivity drops off for signals with a network SNR less than 15. We also observe that our search method can identify a significant population of unique candidates. Furthermore, we conduct an offline search for gravitational wave candidates in the third observing run of the LIGO-Virgo-KAGRA Collaboration (LVK), yielding 31 candidates previously reported by the LVK with a probability of astrophysical origin $p_{\rm astro}\geq0.5$. We identify two other candidates: one previously reported only in a search conducted by the Institute for Advanced Study, and one previously unreported promising new candidate with a $p_{\rm astro}$ of 0.63. This unique candidate has a high chirp mass and a high probability that the primary black hole is an intermediate-mass black hole.

Searching for binary black hole mergers with deep learning in Advanced LIGO's third observing run

TL;DR

The paper addresses detecting binary-black-hole mergers in gravitational-wave data using a hybrid approach that couples matched filtering with deep learning. It builds a pipeline that produces SNR time series from a BBH template bank, then scores triggers with a pretrained DL model to yield a ranking statistic, enabling robust FAR and estimates. On O3 data, the method demonstrates sensitivity comparable to LVK pipelines for high-mass systems ($m_{ m chirp}\gtrsim 25\,M_{\

Abstract

The detection of gravitational waves from compact binary coalescences has provided significant insights into our Universe, and the discovery of new and unique gravitational wave candidates from independent searches remains an ongoing field of research. In this work, we built a hybrid search pipeline that combines matched filtering and deep learning to identify stellar-mass binary black hole candidates from detector strain data. We first present results from a targeted injection study to benchmark the sensitivity of our method and compare it with existing search pipelines. We demonstrate that our hybrid approach has comparable sensitivity for injections with a source-frame chirp mass greater than 25M, and below this threshold our sensitivity drops off for signals with a network SNR less than 15. We also observe that our search method can identify a significant population of unique candidates. Furthermore, we conduct an offline search for gravitational wave candidates in the third observing run of the LIGO-Virgo-KAGRA Collaboration (LVK), yielding 31 candidates previously reported by the LVK with a probability of astrophysical origin . We identify two other candidates: one previously reported only in a search conducted by the Institute for Advanced Study, and one previously unreported promising new candidate with a of 0.63. This unique candidate has a high chirp mass and a high probability that the primary black hole is an intermediate-mass black hole.

Paper Structure

This paper contains 9 sections, 4 equations, 3 figures.

Table of Contents

  1. Introduction
  2. Method
  3. Data
  4. Template bank
  5. Training dataset
  6. Model implementation
  7. Search statistics
  8. False alarm rate
  9. Probability of astrophysical origin

Figures (3)

  • Figure 1: A flowchart of the search pipeline presented in this paper. The pipeline takes in public O3 data from the LIGO-Hanford (H1) and LIGO-Livingston (L1) detectors, as well as pre-computed PSDs of this data (Sec. \ref{['sec:data']}). The data is downsampled to 2048$\,$Hz, and we perform matched filtering using a BBH template bank (Sec. \ref{['sec:template_bank']}). A peak-finding algorithm then produces triggers based on SNR $>$ 4 peaks in at least one detector in the SNR time series, and identifies a coincident peak in the other detector within the light-travel time between detectors ($\sim$10$\,$ms). For each second of analyzed data, triggers are generated from the 10 highest network SNR templates, and the SNR time series is sent to our pre-trained deep learning model (Sec. \ref{['sec:training_dataset']} and Sec. \ref{['sec:model_implementation']}) for predictions. We perform 16 predictions by stepping through the SNR time series in increments of 1/16th of a second for each trigger. These stepwise predictions start when the trigger time first enters the deep learning model's 1-second viewing window and continue until it exits the window; the average of these predictions is then taken. We retain one of the 10 template triggers per second that has the highest average model prediction, and this quantity becomes the search pipeline's ranking statistic. The false alarm rate and probability of astrophysical origin can be estimated (Sec. \ref{['sec:search-stats']}) by comparing the ranking statistic to empirical results from our search pipeline's analyses on time-shifted data and injections.
  • Figure 2: Source frame chirp mass and network signal-to-noise ratio (SNR) parameter distributions for the training dataset injection samples. The top histogram illustrates the bias towards low-mass signals, with twice as many samples in the lowest mass bin as in the highest mass bin. The right histogram shows the power-law distribution of network SNR. We chose these distributions to account for low injection counts in the high-mass and high-SNR regions when sampling from an astrophysical distribution.
  • Figure 3: Foreground (orange) and background (blue) trigger density distribution for our search pipeline in O3a. We tune our $p_{\rm astro}$ model such that the FAR at the intersection of the two distributions corresponds to a $p_{\rm astro}$ of 0.5.