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

The Big Picture

Imagine trying to hear a whisper from a billion light-years away, lasting less than a second, buried beneath the constant rumble of traffic, wind, and seismic noise. That is more or less what physicists do every time they hunt for gravitational waves at LIGO. The signal from two colliding black holes arrives as a ripple in spacetime so faint that the detector mirrors move by a distance a thousand times smaller than a proton’s width.

For a decade, LIGO has relied on matched filtering, cross-checking incoming data against millions of pre-calculated signal templates. It works well, but it’s slow and computationally expensive. It’s also poorly suited for the rapid alerts astronomers need when two neutron stars collide and their light starts fading within hours.

Machine learning offers a faster path. Neural networks can be trained to recognize the characteristic “chirp” of a black hole merger (a rapid upward sweep in frequency as two objects spiral together) and classify new data in milliseconds rather than minutes.

Researchers Ethan Silver, Plamen Krastev, and Edo Berger at Harvard and IAIFI have built a machine-learning search pipeline that scans all three of LIGO’s first observing runs and recovers 57 out of 75 cataloged binary black hole mergers.

Key Insight: By training convolutional neural networks directly on real LIGO noise, not simulated Gaussian noise, this pipeline achieves competitive detection rates across three full observing runs while running orders of magnitude faster than traditional matched-filter methods.

How It Works

The pipeline starts with real data. The team pulled roughly 22 days of continuous output from LIGO’s second observing run (O2), from both the Livingston (L1) and Hanford (H1) detectors, sampled at 4096 Hz. They injected simulated black hole merger waveforms into half of these snippets, creating a labeled training set of “signal” and “noise” examples.

Getting the mass distribution right was critical. Real binary black hole events span a wide range of masses, but sampling each independently from a uniform distribution yields very few systems where both holes are low-mass, even though those systems produce longer, harder-to-detect signals. The team solved this with intentional oversampling: they first drew a maximum mass from a triangular distribution peaked at 5 solar masses, then sampled individual masses uniformly below that ceiling. This ensured the training set included plenty of challenging low-mass events.

Waveforms were generated using the SEOBNRv4 approximant, a state-of-the-art model for the time-domain gravitational wave signal from spinning, non-precessing black hole binaries, via the PyCBC library. Both signal and noise were then whitened: a preprocessing step that flattens the noise power spectrum so the network sees a statistically uniform background. Each 4-second template captures the inspiral (gradual approach), merger (collision), and ringdown (final settling) of the signal across both detectors simultaneously.

The team trained two distinct families of models:

• Time-series CNNs that process raw whitened strain data directly. (CNNs, or convolutional neural networks, are a class of neural network that detects patterns in sequential data.)
• Time-frequency CNNs that operate on spectrograms, visual representations of how signal power varies across frequency bands over time.

Both architectures learned to distinguish the characteristic chirp from noise transients. Running them in parallel and requiring agreement between them helped suppress false alarms.

Why It Matters

The pipeline scanned all of LIGO’s O1, O2, and O3 data, covering 75 binary black hole events with two-detector data available, and recovered 57 of those 75 events: a detection rate of 76%. That compares favorably to recent competing ML searches; AresGW found 40 of 65 O3 events, and AFrame found 38. The 57 false positives can largely be eliminated through quick parameter inference (fitting the signal to extract mass and spin) and human review, reducing the burden to a manageable vetting step.

To measure how often the network would flag non-existent signals, the team ran an extensive time-shift test. They offset the L1 and H1 data streams by amounts large enough to destroy any genuine coincident signal. A real gravitational wave arrives at both detectors within ~10 milliseconds; a 100-second shift eliminates all real coincidences. Detections in this shifted data measured the false alarm rate (FAR) as a function of detection threshold, which is the standard LIGO method for characterizing detection significance.

Speed is the central motivation. LIGO O4 has cataloged over 200 events in its first 23 months, and next-generation detectors like Einstein Telescope and Cosmic Explorer promise event rates orders of magnitude higher. The field needs tools that triage candidates instantly. A neural network processing 4 seconds of data in milliseconds can send alerts to optical telescopes while a neutron star merger’s kilonova is still brightening.

Binary neutron star and neutron star-black hole mergers produce electromagnetic counterparts: gamma-ray bursts, kilonovae, radio afterglows. These carry independent information about nuclear physics and cosmology. Catching them within minutes of the gravitational wave trigger is the difference between a full multi-messenger portrait and a missed opportunity.

Bottom Line: A CNN-based pipeline trained on real LIGO noise recovers 76% of known binary black hole mergers across O1–O3, with a false alarm rate low enough for practical use. Machine learning is ready to play a central role in the next era of gravitational wave astronomy.