⚛️ general relativity

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

This paper presents a machine learning pipeline using convolutional neural networks trained on real LIGO data to detect binary black hole mergers, successfully identifying 57 out of 75 cataloged events across observing runs O1–O3 while characterizing its false alarm rate to enable low-latency gravitational wave detection.

Original authors: Ethan Silver, Plamen Krastev, Edo Berger

Published 2026-01-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ethan Silver, Plamen Krastev, Edo Berger

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the universe is a giant, dark ocean. For a long time, we could only see the surface. But in 2015, scientists built a special kind of "underwater microphone" called LIGO that could hear the ripples caused by massive objects crashing together, like two black holes merging. These ripples are called gravitational waves.

Since then, scientists have found over 200 of these crashes. But listening to the ocean is hard. The microphones are so sensitive that they hear everything: the distant rumble of a storm, the squeak of a door hinge, and the actual crash of black holes. It's like trying to hear a specific person whisper in a crowded, noisy stadium.

This paper is about teaching a computer to be a super-smart listener who can instantly spot the whisper of a black hole crash amidst all that noise.

The Problem: Too Much Noise

Traditionally, scientists find these crashes by comparing the sound in the microphone to a massive library of "expected sounds" (templates). It's like having a dictionary of every possible whisper and checking the recording against every single word. This works, but it's slow and computationally heavy. If a black hole crashes, we want to know fast so we can point our telescopes at the spot in the sky to see if there's a flash of light (an "electromagnetic follow-up").

The Solution: A Neural Network "Detective"

The authors built a Neural Network, which is a type of computer brain designed to learn patterns, much like how a child learns to recognize a dog by seeing many pictures of dogs.

Here is how they trained this detective:

The Training Camp: They took real recordings of LIGO's "silence" (which is actually full of noise) and secretly planted fake black hole crash sounds into half of them.
The Lesson: They showed the computer brain these recordings, telling it, "This one has a crash, this one doesn't."
The Twist: To make the detective smarter, they didn't just show it one microphone's recording. They showed it two microphones (LIGO has two detectors, one in Washington and one in Louisiana) at the same time. They taught the brain that a real crash must happen in both places at almost the exact same time. If only one microphone hears a "crash," it's probably just a local glitch.

The Search: Scanning the Ocean

The team took this trained detective and let it scan three years' worth of data (called O1, O2, and O3) from the LIGO detectors.

The Speed Run: First, the detective did a quick, low-resolution scan of the whole ocean to find potential suspects.
The Close-Up: For the suspects it found, it did a slow, high-resolution look to make sure they were real.

The Results: A Good Catch, Some False Alarms

Out of 75 known black hole crashes that the official catalog says happened during those three years, this new AI pipeline found 57 of them.

Why did it miss 18? The ones it missed were usually very quiet (low volume) or involved very small black holes. It's like the detective missed the whispers because they were too faint or the person was too far away.
The False Alarms: The pipeline also flagged 57 "false alarms." These were moments where the noise in the two microphones happened to line up perfectly by accident, sounding like a crash.
- How they fixed it: The authors showed that if you run a quick math check (called "parameter inference") on these false alarms, you can prove they aren't real. It's like checking the ID of the person who whispered; if the ID doesn't match the voice, you know it's a fake.

The "Time-Shift" Test

To prove their detective wasn't just guessing, they did a clever trick. They took the data from the two microphones and shifted them in time so they were completely out of sync. Since a real black hole crash can't happen in two places at different times, every signal found in this shifted data had to be a false alarm.

They ran this test over 100 years' worth of shifted data. This allowed them to calculate exactly how often their detective cries "Wolf!" when there is no wolf. They found that their system is very reliable, with a very low rate of false alarms for the loudest signals.

How It Compares to Others

The paper compares their AI detective to two other similar AI systems (AresGW and AFrame).

The other systems found fewer black holes (around 38 to 40 out of the 65 available in the third year of data).
This new pipeline found 49 of those same events.
While the other systems were slightly better at ignoring the faintest false alarms, this new pipeline was better at catching the most events overall.

The Bottom Line

This paper shows that we can use machine learning to find black hole crashes much faster and more efficiently than before. While it didn't find every single crash (especially the quiet ones), it found the vast majority of them and proved it can do so with a very low rate of false alarms.

The authors conclude that this is a major step forward. In the future, this kind of AI could be used to scan data in real-time, telling astronomers immediately when a black hole crashes so they can point their telescopes at the sky before the light fades away. They also plan to teach this AI to listen for other types of cosmic crashes, like neutron stars colliding.

Technical Summary: A Search for Binary Black Hole Mergers in LIGO O1–O3 Data with Convolutional Neural Networks

Problem Statement
Since the first detection of gravitational waves (GW) in 2015, the number of cataloged compact binary coalescence (CBC) events has exceeded 200. Current detection algorithms rely on matched-filtering, which compares data against pre-calculated template banks. While effective, this method can be computationally intensive. Machine learning (ML), specifically Convolutional Neural Networks (CNNs), offers the potential to accelerate detection and parameter estimation, which is critical for rapid electromagnetic follow-up of multi-messenger events. However, early ML applications were often limited to evaluating models on known signal times, failing to properly account for false positives (FPs) or calculate False Alarm Rates (FARs) across large, continuous data streams.

Methodology
The authors present a "blind" search pipeline designed to detect Binary Black Hole (BBH) mergers across the entirety of LIGO's first three observing runs (O1–O3). The methodology involves the following key components:

Dataset Construction: Training data was generated using approximately 22 days of continuous, real LIGO noise from the Livingston (L1) and Hanford (H1) detectors. BBH waveforms were injected into half of these noise segments using the LALSuite library (SEOB-NRv4 approximant). To ensure robustness at lower masses, the authors employed an over-sampling strategy for low-mass BBHs (sampling maximum mass from a triangular distribution and individual masses from a uniform distribution).
False Positive Training: To teach the network to reject non-astrophysical signals, the training data for two-detector models included four specific classes of injected false positives: single-detector signals, signals with inconsistent parameters between detectors, time-offset signals (>50 ms), and coincident noise spikes (specifically for time-frequency data).
Model Architecture: The pipeline utilizes an ensemble of 10 neural network models, including both time-series (1D) and time-frequency (2D Q-transform) inputs. The architectures primarily consist of ResNet and standard CNN structures with varying numbers of convolutional layers, filters, and dense layers.
Search Pipeline Strategy: The search employs a two-stage resolution approach to balance speed and sensitivity:
1. Low-Resolution Search: Applied to the entire dataset in 0.1 s time-steps using two detector models (one time-series, one time-frequency) to generate an initial list of candidates.
2. High-Resolution Search: The 0.2 s regions surrounding initial candidates are re-evaluated in 0.01 s time-steps using a suite of high-resolution models (including time-series, time-frequency, and a chirp mass regression model).
Coincidence and Validation: A candidate is considered detected only if it is identified by the low-resolution models and subsequently confirmed by all high-resolution models. A chirp mass regression model is used to calculate a "mass similarity" metric; candidates with significantly different predicted chirp masses between detectors are ruled out.
FAR Characterization: To determine the False Alarm Rate, the pipeline was evaluated on 100 time-shifted datasets (totaling ~100 years of data), where all events are known to be false positives.

Key Contributions

Blind Search Implementation: Unlike previous studies that tested ML on known signal times, this work performs a full blind search across O1, O2, and O3, enabling the first proper calculation of FARs for a CNN-based BBH search in this context.
Ensemble Approach: The paper demonstrates the efficacy of combining multiple models with different architectures (time-series vs. time-frequency) and resolutions to optimize the trade-off between detection efficiency and false positive rates.
Comprehensive False Positive Handling: The inclusion of specific false positive classes in the training data and the use of a chirp mass regression model to filter candidates significantly improve the pipeline's ability to distinguish real signals from noise transients.

Results

Detection Efficiency: The pipeline successfully detected 57 out of the 75 total cataloged BBH events in the GWTC-3 catalog that had data from two detectors. The 18 missed events were primarily characterized by low network Signal-to-Noise Ratios (SNR < 10), low chirp masses, or low SNR in a single detector (e.g., GW170818).
False Positives: The search yielded 57 false positives with significance levels at least as high as the least significant true positive. These FPs were largely attributed to transient noise aligning in both detectors. The authors note that parameter inference can rule out most of these candidates, as they often fail to converge to realistic solutions or exhibit physical inconsistencies (e.g., duration vs. mass).
False Alarm Rate: In the time-shifted analysis, 6,273 false positives were found above the significance threshold ( $S=0.54$ ). The highest significance observed in the time-shifted data was $S=3.43$ . The authors extrapolated the FAR distribution using a power-law fit to estimate rates for events more significant than the time-shifted background.
Comparison: When compared to other ML-based searches (AresGW and AFrame) on O3 data, this pipeline detected more total events (49 vs. 40 and 38, respectively) while achieving lower FARs for a comparable number of detections.

Significance and Claims
The authors position this work as an "important step" in developing machine learning-based gravitational wave searches. They claim that their pipeline enables low-latency detection and facilitates multi-messenger astronomy by significantly speeding up the identification of events compared to current techniques. The paper emphasizes that while the pipeline missed some low-SNR events, it successfully detected the vast majority of cataloged BBHs outside of those specific limitations. The authors conclude that this represents a crucial test of applying neural networks to multiple entire observing runs and suggest that future iterations could extend to Binary Neutron Star (BNS) and Neutron Star-Black Hole (NSBH) mergers, incorporate more detectors (Virgo, KAGRA), and provide direct parameter estimation.