Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Convolutional Neural Networks for signal detection in real LIGO data

Ondřej Zelenka, Bernd Brügmann, Frank Ohme

TL;DR

This paper evaluates a CNN-based approach for gravitational-wave signal detection in real LIGO data within the MLGWSC-1 framework. It details the data generation, whitening, and two-channel coherent network using a ranking statistic $\Delta x$ to detect injections, along with a modified loss and training regime. The authors report an updated submission with improved robustness to non-Gaussian noise, competitive sensitivity relative to PyCBC on Gaussian data, and successful recovery of GWTC-3 events in O3b open data, demonstrating practical applicability to current catalogs and data releases. The work highlights the importance of training distribution and normalization choices for real-noise performance and notes non-monotonic training behavior that warrants further study.

Abstract

Searching the data of gravitational-wave detectors for signals from compact binary mergers is a computationally demanding task. Recently, machine learning algorithms have been proposed to address current and future challenges. However, the results of these publications often differ greatly due to differing choices in the evaluation procedure. The Machine Learning Gravitational-Wave Search Challenge was organized to resolve these issues and produce a unified framework for machine-learning search evaluation. Six teams submitted contributions, four of which are based on machine learning methods and two are state-of-the-art production analyses. This paper describes the submission from the team TPI FSU Jena and its updated variant. We also apply our algorithm to real O3b data and recover the relevant events of the GWTC-3 catalog.

Convolutional Neural Networks for signal detection in real LIGO data

TL;DR

This paper evaluates a CNN-based approach for gravitational-wave signal detection in real LIGO data within the MLGWSC-1 framework. It details the data generation, whitening, and two-channel coherent network using a ranking statistic to detect injections, along with a modified loss and training regime. The authors report an updated submission with improved robustness to non-Gaussian noise, competitive sensitivity relative to PyCBC on Gaussian data, and successful recovery of GWTC-3 events in O3b open data, demonstrating practical applicability to current catalogs and data releases. The work highlights the importance of training distribution and normalization choices for real-noise performance and notes non-monotonic training behavior that warrants further study.

Abstract

Searching the data of gravitational-wave detectors for signals from compact binary mergers is a computationally demanding task. Recently, machine learning algorithms have been proposed to address current and future challenges. However, the results of these publications often differ greatly due to differing choices in the evaluation procedure. The Machine Learning Gravitational-Wave Search Challenge was organized to resolve these issues and produce a unified framework for machine-learning search evaluation. Six teams submitted contributions, four of which are based on machine learning methods and two are state-of-the-art production analyses. This paper describes the submission from the team TPI FSU Jena and its updated variant. We also apply our algorithm to real O3b data and recover the relevant events of the GWTC-3 catalog.
Paper Structure (20 sections, 9 equations, 4 figures, 4 tables)

This paper contains 20 sections, 9 equations, 4 figures, 4 tables.

Table of Contents

  1. Introduction
  2. MLGWSC-1
  3. Test data
  4. Evaluation procedure
  5. Experimental setup
  6. Data processing
  7. Training and validation data
  8. Test data
  9. Machine Learning
  10. Model architecture
  11. Results
  12. MLGWSC-1 submission
  13. MLGWSC-1 results
  14. Updated submission
  15. Input normalization and network architecture
  16. ...and 5 more sections

Figures (4)

  • Figure 1: Evolution of the training and validation loss values throughout the training of the submission.
  • Figure 2: Sensitivity curves of the network at 3 minima of the validation loss highlighted in Fig. \ref{['fig:tpi_losses']} used to select the final network state for the submission.
  • Figure 3: Sensitivity curves of the submitted trained network using multiple $\Delta x$ thresholds to determine a suitable value. The datasets are generated by the script provided by the in the length of 1 day (86400 seconds), difficulty specified as datasets 3 and 4 in the top and bottom panel, respectively. Due to a large overlap between the sensitivity curves, rather than color-coding, their left ends are annotated with the threshold value. All curves reach the same point at the right end $\mathcal{F} = 1\,\mathrm{day}^{-1}$.
  • Figure 4: Sensitivity curves of 3 selected submissions, along with updated versions of 2 of them, on datasets 3 and 4 of the . Each panel contains the performance of the submissions on one test dataset. Dashed lines mark conventional analyses and solid lines mark -based search algorithms. In case of the TPI FSU Jena and Virgo-AUTh teams, the dotted lines mark the original submissions, while the solid lines mark the updated algorithms. The remaining submissions are shown in gray for illustration of overall challenge results.