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

Technical Summary: Searching for Binary Black Hole Mergers with Deep Learning in Advanced LIGO's Third Observing Run

Problem Statement
The detection of gravitational waves (GWs) from compact binary coalescences has become routine, yet the discovery of new and unique candidates remains a critical area of research. While the LIGO-Virgo-KAGRA (LVK) Collaboration utilizes established matched-filtering pipelines (e.g., PyCBC, GstLAL, MBTA) and unmodeled burst searches (cWB), independent offline searches continue to identify additional candidates and unique events. Existing deep-learning approaches (e.g., AresGW, Aframe) have shown promise but often overlap significantly with standard catalogs. The challenge addressed in this work is to develop a hybrid search pipeline that combines the robustness of matched filtering with the pattern recognition capabilities of deep learning to identify stellar-mass binary black hole (BBH) candidates from detector strain data, specifically aiming to recover unique events and benchmark sensitivity against existing LVK pipelines during the third observing run (O3).

Methodology
The authors developed a hybrid search pipeline that processes data from the LIGO Hanford (H1) and Livingston (L1) detectors. The workflow proceeds as follows:

1. Matched Filtering and Trigger Generation: The pipeline performs matched filtering using a template bank of 97,802 BBH waveforms (SEOBNRv4 ROM approximant) with component masses $>2 M_\odot$ and a low-frequency cutoff of 15 Hz. A peak-finding algorithm identifies triggers where the signal-to-noise ratio (SNR) exceeds 4 in at least one detector, requiring a coincident peak within the light-travel time ($\sim10$ ms) in the other detector.
2. Deep Learning Classification: For each second of data, the ten templates with the highest network SNR are selected. The SNR time series for these templates is fed into a pre-trained deep learning model (based on a ResNet architecture with convolutional and fully connected layers). The model processes the time series in 16 steps (increments of 1/16 s) across a 1-second window.
3. Ranking Statistic: The model outputs a prediction for each step. The pipeline averages these stepwise predictions for the ten templates and selects the single template with the highest average prediction as the ranking statistic for that second.
4. Training Strategy: The model was trained on 1-second SNR time series samples derived from O3 data. The dataset balanced pure noise, glitch-contaminated noise (identified via Omicron), and injected BBH signals. Crucially, the injection distribution was manually controlled to ensure sufficient representation of high-mass systems, which are often underrepresented in astrophysical priors. Glitches were aligned with signal frequencies to maximize interaction during training.
5. Significance Estimation:
   • False Alarm Rate (FAR): Computed by analyzing time-shifted data (decoupling detectors by $>10$ ms) to create a background distribution of noise triggers.
   • Probability of Astrophysical Origin ($p_{astro}$): Calculated using a signal-versus-noise Bayes factor, modeling foreground (signal) and background (noise) as independent Poisson processes. Parameters were tuned using injection campaigns to ensure a $p_{astro} \ge 0.5$ threshold corresponds to a specific intersection of background and foreground trigger densities.

Key Contributions and Results

• Sensitivity Benchmarking: The pipeline was tested against 83,815 injected BBH signals from the GWTC-3 injection set. The hybrid approach demonstrated comparable sensitivity to existing pipelines (PyCBC-BBH, GstLAL, etc.) for source-frame chirp masses greater than $25 M_\odot$. However, sensitivity dropped off for signals with chirp masses below this threshold and network SNRs less than 15.
• Unique Detections: The search yielded 33 candidate events in O3 data meeting the criteria of $p_{astro} \ge 0.5$ and FAR $< 2$ per day.
  • Recovery: 31 of these candidates matched those in the LVK's GWTC-3 catalog.
  • New/Unique Candidates: The search identified two additional candidates:
    1. GW190605 025957: Previously reported only by the Institute for Advanced Study (IAS) higher-order modes search. The authors note potential data quality issues (a transient noise event) but confirm the candidate's significance in their pipeline.
    2. GW190929 091722: A previously unreported candidate with a $p_{astro}$ of 0.63. Parameter estimation suggests a high total mass ($171^{+172}_{-58} M_\odot$), with a significant probability that the primary component is an intermediate-mass black hole (IMBH) potentially residing in the pair-instability mass gap.
• Deep Learning vs. Template Bank: A comparative study using the PyCBC pipeline with the same template bank as the deep learning search confirmed that the unique detection capability stems from the deep learning ranking statistic, not the choice of template bank. The deep learning approach identified significantly more unique events (348 vs. 23 in a specific test window) than the standard PyCBC statistic.
• Missed Candidates: The authors analyzed 75 missed candidates (events detected by other pipelines but not by this search). The majority of misses were attributed to the pipeline's reduced sensitivity to low-mass systems (chirp mass $< 25 M_\odot$) and low SNR events. Other misses were caused by design constraints, such as data segment cuts (first 200s and last 24s) and glitch interference with long-duration templates.

Significance and Claims
The paper claims that the proposed hybrid search pipeline successfully identifies a significant population of unique gravitational wave candidates that are missed by traditional matched-filtering pipelines using standard ranking statistics. While the method currently exhibits lower sensitivity for low-mass binaries compared to optimized pipelines, its ability to detect high-mass BBH mergers and unique events demonstrates the value of integrating deep learning classifiers into GW search pipelines.

The authors emphasize that the unique detections are a direct result of the deep learning model's ability to distinguish signals from noise in the SNR time series differently than traditional statistical methods. They conclude that while the current implementation requires optimization for low-mass sensitivity and robustness against glitches, the hybrid approach offers a complementary search strategy that enhances the overall discovery potential of the LVK observing runs. The identification of GW190929 091722 as a potential IMBH candidate highlights the pipeline's utility in exploring extreme mass regimes, even if further verification (e.g., dedicated IMBH searches) is required.