A Search for Binary Black Hole Mergers in LIGO O1-O3 Data with Convolutional Neural Networks

TL;DR

This work demonstrates that a convolutional neural network–based pipeline, trained on real LIGO noise with injected BBH waveforms and applied to O1–O3 data, can detect 57 of 75 cataloged BBH events using a two-detector, multi-resolution approach that includes time-series and time-frequency representations. By incorporating false-positive templates and a chirp-mass regression constraint, the method achieves a calibrated false alarm rate through extensive time-shift analyses, supporting low-latency, multi-messenger follow-up potential. The study shows competitive performance relative to other ML-based searches, highlighting the benefits of an ensemble of models and a two-stage refinement strategy. It also outlines clear paths for future enhancements, including extension to additional detectors and source classes, and faster parameter estimation for real-time alerts.

Abstract

Since the first detection of gravitational waves in 2015 by LIGO from the binary black hole merger GW150914, gravitational wave astronomy has developed significantly, with over 200 compact binary merger events cataloged. The use of neural networks has the potential to significantly speed up the detection, classification, and especially parameter estimation for gravitational wave events, compared to current techniques, quite important for electromagnetic follow-up of events. In this work, we present a machine learning pipeline using neural networks to detect gravitational wave events. We generate training data using real LIGO data to train and refine neural networks that can detect binary black hole (BBH) mergers, and apply these models to search through LIGO's first three observing runs. We detect 57 out of the 75 total cataloged BBH events with two detectors of data in O1, O2, and O3, with 57 false positives that can mostly be ruled out with parameter inference and human inspection. Finally, we extensively test this pipeline on time-shifted data to characterize its False Alarm Rate (FAR). These results are an important step in developing machine learning-based GW searches, enabling low-latency detection and multi-messenger astronomy.

Figures (12)

• Figure 1: Comparison of the various BBH mean mass distributions of the training data for models we have tested. We first show sampling each mass independently from a uniform distribution. Second, is doing the same except forcing 20% of those to be in the range (5-10M$_\odot$). Third, is the mass distribution we use in the datasets for the models shown in this paper, as described in the text.
• Figure 2: Left: Example of a BBH template injection in LIGO noise from our training data. The template has $M_1=25$ M$_\odot$, $M_2=20$ M$_\odot$, and a network ${\rm SNR}=10$. The red lines are the BBH waveforms alone, and the gray lines are the combined BBH waveform and noise. Right: The time-frequency data corresponding to the same template. We perform the Q-transform of a 0.5 s window around the signal peak of the time-series data to convert it to a $300x300$ time-frequency image. The signal chirp and merger are clearly seen.
• Figure 3: Schematic diagram of our residual neural network architecture for two-detector time-series ResNet models. This model contains 3,862,756 parameters.
• Figure 4: Flow-chart of our GW search process, starting from the initial LIGO files, data whitening and splitting into 4-s segments, evaluation through the low- and high-resolution time-series and time-frequency models, and generation of a final list of candidates.
• Figure 5: The number of false positives and true positives for our model pipeline as the final threshold changes, in all of O1-O3, and O3 specifically.