Hybrid algorithm combining matched filtering and convolutional neural networks for searching gravitational waves from binary black hole mergers

Efficient searches for gravitational waves from compact binary coalescence are crucial for gravitational wave observations. We present a proof-of-concept for a method that utilizes a neural network taking an SNR map, a stack of SNR time series calculated by the matched filter, as input and predicting the presence or absence of gravitational waves in observational data. We demonstrate our algorithm by applying it to a dataset of gravitational-wave signals from stellar-mass black hole mergers injected into stationary Gaussian noise. Our algorithm exhibits comparable performance to the standard matched-filter pipeline and to the machine-learning algorithms that participated in the mock data challenge, MLGWSC-1. The demonstration also shows that our algorithm achieves reasonable sensitivity with practical computational resources.

💡 Research Summary

This paper presents a proof-of-concept for a novel hybrid algorithm designed to search for gravitational waves (GWs) from binary black hole (BBH) mergers. The core innovation lies in combining the well-established matched-filtering technique with a convolutional neural network (CNN), aiming to leverage the strengths of both approaches for efficient and sensitive detection.

The algorithm’s workflow is two-stage. First, the raw strain data from detectors (e.g., LIGO Hanford and Livingston) is processed using a standard matched filter against a bank of 256 non-spinning, equal-mass-ratio BBH waveform templates, covering chirp masses from 5 to 50 solar masses. This process generates a Signal-to-Noise Ratio (SNR) time series for each template. These time series are then stacked to form a 2D image called an “SNR map,” with one axis representing time and the other representing the template index. This map is preprocessed by smearing (averaging) over the time axis to reduce dimensionality and normalized before being fed into the CNN.

The CNN, which serves as the second stage, is tasked with classifying the presence or absence of a GW signal within the data segment. The network architecture comprises a series of convolutional layers with ReLU activations and max-pooling, followed by fully connected layers. During training, the model outputs a probability via a softmax layer, while during inference, an unbounded detection statistic (Λ = output for “signal” class minus output for “noise” class) is used for ranking candidates.

To train and evaluate the model, the authors generated datasets containing stationary Gaussian noise with the aLIGOZeroDetHighPower PSD. Signals from non-spinning BBH mergers were injected using the IMRPhenomXPHM waveform model. The training process utilized a unique data augmentation strategy, combining pools of pure-noise and pure-signal (noise-free waveform) SNR maps with randomized extrinsic parameters at each training iteration to enhance generalization.

The performance of the hybrid algorithm was assessed on a separate test dataset containing detectable signals within 15-150 Mpc. The results demonstrated that the algorithm achieves a detection performance—measured by the Area Under the Receiver Operating Characteristic Curve (AUC)—comparable to that of a standard matched-filter pipeline and to other machine learning algorithms that participated in the MLGWSC-1 mock data challenge. Crucially, the implementation leverages GPU-accelerated, vectorized matched-filter computations, allowing the entire process to achieve reasonable sensitivity with practical computational resources.

The paper concludes that the hybrid method, using an SNR map as an intermediate data representation, is a viable and promising approach. It retains the optimal theoretical baseline of matched filtering for Gaussian noise while delegating the complex task of handling residual non-Gaussian features to a fast neural network. Future work directions include applying the method to real observational data with non-stationary noise and glitches, extending the template bank to include spins and asymmetric mass ratios, and integrating the network into a full pipeline that includes parameter estimation.