⚛️ general relativity

A Fast and Scalable Transformer Pipeline for Binary Black Hole Detection

This paper introduces \castor, a fast and scalable transformer-based pipeline for detecting binary black hole gravitational-wave signals that achieves high sensitivity, outperforms adapted audio foundation models, and significantly reduces the computational cost of background estimation.

Original authors: Chayan Chatterjee, Abigail Petulante, Haowei Fu, Yang Hu, Roy Lau, Karan Jani

Published 2026-09-02

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: Chayan Chatterjee, Abigail Petulante, Haowei Fu, Yang Hu, Roy Lau, Karan Jani

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

The universe is filled with invisible ripples in space and time, known as gravitational waves. These ripples are created when massive objects, such as black holes, crash into one another. To hear these whispers from the cosmos, scientists use giant laser instruments called interferometers, which can detect changes in distance smaller than the width of an atom. However, the data these instruments collect is often drowned out by noise from the Earth itself, from traffic to distant earthquakes. Finding a real cosmic signal in this static is like trying to hear a single violin in a crowded stadium. For decades, the best way to find these signals has been to compare the data against a massive library of predicted wave patterns, a method that works well but requires immense computing power. As the number of detected collisions grows, scientists need faster and more efficient ways to sort through the noise to find the next big event.

A team of researchers has developed a new tool called Castor to solve this problem. Castor is a type of artificial intelligence designed specifically to listen for the collision of two black holes. Unlike older methods that try to match the data to a library of pre-made patterns, Castor learns to recognize the shape of a black hole collision directly from the raw data. The researchers built this system to be incredibly fast, allowing it to process data from two different detectors independently and then combine the results later. This approach is crucial because it allows the system to estimate how often random noise might trick it into thinking it found a signal, without having to re-run the entire complex computer program millions of times. By caching the initial results and simply rearranging them to simulate different time shifts, the team can generate a massive amount of background data to test their findings, a task that would be prohibitively expensive with other methods.

The researchers tested Castor using both simulated data and real recordings from the third observing run of the LIGO detectors. They compared their new system against another artificial intelligence model that had been adapted from a program originally designed to understand human speech. The results showed that Castor was significantly more sensitive than the speech-based model, capable of detecting fainter signals from much farther away. More importantly, Castor was about twenty times faster at processing the data. This speed advantage means that scientists can now run much more rigorous tests on their findings, ensuring that when they announce a new discovery, they are certain it is not just a glitch in the machine. The system successfully identified the majority of the confident black hole collisions from a recent catalog of known events, provided those events fell within the mass range the system was trained to recognize.

The study also highlighted a specific weakness in using models adapted from other fields, such as speech recognition, for astronomy. While those models can learn quickly, they sometimes struggle with the loudest and most significant signals because they rely on a way of processing sound that discards absolute volume information. Castor, by contrast, looks at the raw vibration of the detectors, preserving the full strength of the signal. This allowed it to maintain its accuracy even for the most powerful collisions. The researchers found that the few events Castor missed were either too light or too heavy for its training, or they were simply too faint to be distinguished from the background noise. This suggests that while the tool is highly effective, it is not a replacement for all existing methods, but rather a powerful, specialized addition to the toolkit.

Ultimately, this work demonstrates a practical path forward for the future of gravitational wave astronomy. As observatories become more sensitive and detect more events every day, the ability to process data quickly and reliably will become the bottleneck for discovery. Castor offers a way to keep up with this growing flood of information, providing a fast, scalable method to find black hole collisions while keeping the risk of false alarms under strict control. The researchers have made their code available to the public, allowing the wider scientific community to build upon this foundation. By balancing speed with accuracy, this new approach ensures that as we listen deeper into the universe, we will not miss the most important stories it has to tell.

Technical Summary: Castor – A Fast and Scalable Transformer Pipeline for Binary Black Hole Detection

Problem Statement
With the projected increase in gravitational-wave (GW) detection rates from compact-binary coalescences (CBCs), there is a critical need for search pipelines that are faster, robust, and more scalable than traditional matched filtering. While deep learning offers a promising alternative by amortizing computational costs into a one-time training stage, a significant bottleneck remains in multi-detector searches: the estimation of false-alarm rates (FAR).

Traditional coherent searches analyze data from multiple detectors jointly to maximize sensitivity but require re-evaluating the entire neural network for every time-slide shift used to generate background statistics. This makes large-scale background estimation computationally prohibitive for deep learning models. Conversely, coincident searches analyze detectors independently and combine outputs post-hoc, which is computationally efficient but traditionally lacks the representational power of coherent analysis. The challenge is to retain the sensitivity of coherent cross-detector information while preserving the computational efficiency of separable, post-hoc ranking.

Methodology: The Castor Pipeline
The authors introduce Castor (Coincident Analysis Siamese TransfORmer), a transformer-based pipeline designed to overcome the computational limitations of background estimation.

Architecture: Castor is a compact time-domain model (~2.9 million parameters) trained from scratch. It employs a Siamese architecture with shared weights, processing the whitened, calibrated time-domain strain data from each interferometer (H1 and L1) independently.
Input Processing: The network takes 1-second segments of strain data (sampled at 2048 Hz). A stack of three strided 1D convolutional layers compresses the data into a sequence of 64 contextual tokens.
Dual-Head Output:
1. Global Classification: A single-query attention pooling layer aggregates the token sequence into a global vector, feeding into a classifier head that outputs signal-versus-noise log-odds (Unbounded Softmax Replacement or USR).
2. Local Localization: A lightweight "frame" head processes the unpooled token sequence to predict a per-token merger localization profile (a Gaussian centered at the merger time).
Coincident Ranking Statistic: The final detection statistic is a closed-form combination of the single-detector outputs:
$R = \text{USR}_{HL} + w_c C(f_H, f_L)$
Where $\text{USR}_{HL}$ is the network log-odds derived from independent detectors, and $C(f_H, f_L)$ is a cross-correlation term measuring temporal coherence between the localization profiles of the two detectors.
Efficient Background Estimation: Because the neural network is evaluated only once per detector, the single-detector outputs (log-odds and frame profiles) can be cached. Time slides are generated by simply recomputing the closed-form coincidence statistic on the cached, shifted arrays. This eliminates the need for repeated neural network forward passes during background estimation.
Training: The model is trained using a hybrid curriculum combining real-time signal generation (using the ml4gw framework) and pre-generated injections from the MLGWSC-1 benchmark. The training objective combines binary cross-entropy for global classification and a Gaussian-profile loss for the frame head.

Comparison Baseline
Castor is benchmarked against GW-Whisper, a domain-adapted version of OpenAI's Whisper audio foundation model. GW-Whisper operates on time-frequency spectrograms (Q-transforms) and uses low-rank adaptation (DoRA) to fine-tune a pre-trained audio model (~39 million parameters) for GW detection.

Key Results

Sensitivity: On the MLGWSC-1 benchmark datasets (Dataset 3: Gaussian noise; Dataset 4: Real O3a noise), Castor ranks among the most sensitive machine-learning pipelines. At a FAR of 1 month $^{-1}$ , Castor achieves sensitive distances of ~1482 Mpc (Dataset 3) and ~1350 Mpc (Dataset 4). This represents an improvement in sensitive distance by factors of ~1.66 and ~1.49, respectively, over GW-Whisper.
Computational Efficiency: Castor is significantly faster. Inference on one month of MLGWSC-1 data takes ~0.9 hours on a single A100 GPU, compared to ~18 hours for GW-Whisper. The time-slide background estimation for Castor is ~20 times faster than what would be required for a jointly-evaluated network like GW-Whisper.
Real Data Application (O3b): Applied to five months of O3b observing strain, Castor successfully recovered 14 out of 31 confident BBH events from the GWTC-3 catalog with a FAR < 1 month $^{-1}$ . The majority of missed events were either outside the training mass range ([10, 50] $M_\odot$ ), had low matched-filter SNR, or were excluded due to data quality issues.
Background Validation: The authors validated that the time-slide background generated from cached outputs statistically matches the distribution of real coincident noise, allowing for empirically calibrated FARs without extrapolation.

Significance and Claims
The paper claims that Castor demonstrates a practical, scalable approach for deep-learning GW searches. Its primary significance lies in resolving the trade-off between sensitivity and computational cost in background estimation:

Scalability: By decoupling neural network inference from time-slide recombination, Castor enables the generation of multi-year background exposures at a negligible arithmetic cost, allowing for fully empirical FARs below 1 year $^{-1}$ without exponential extrapolation.
Architecture Design: The results suggest that for BBH detection at fixed FAR, a compact, purpose-built time-domain transformer trained from scratch substantially outperforms a much larger, repurposed audio foundation model. The authors attribute GW-Whisper's lower performance to the loss of absolute amplitude information during spectrogram normalization and the mismatch between acoustic pre-training and GW morphology.
Modesty: The authors explicitly state they do not claim Castor matches the sensitivity of state-of-the-art coherent matched-filter analyses (which remain more sensitive but computationally expensive for background estimation). Instead, Castor offers a "lightweight" solution that provides a practical route to trustworthy, cheap background estimation for coincident deep-learning searches.

The study concludes that while Castor is not yet production-ready for all scenarios, it validates a design pattern where a modest sensitivity trade-off yields a massive gain in the efficiency of statistical validation, a critical requirement for future high-rate observing runs.

Technical Summary: Castor – A Fast and Scalable Transformer Pipeline for Binary Black Hole Detection

More like this