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

TL;DR

This work presents a hybrid approach to gravitational-wave detection that combines matched-filter preprocessing with a convolutional neural network operating on SNR maps derived from a template bank. The SNR map inputs are constructed from $ ho_I[j; r] = \frac{|z_I[j; r]|}{\sigma_I[r]}$, spanning $256$ templates with chirp masses $5$–$50\,M_\odot$ in a $(2048\times256)$ image, with time-axis smoothing and normalization. A CNN processes these images to predict signal presence, producing a detection statistic $\Lambda = x_1 - x_0$ after using an unbounded softmax replacement; training uses cross-entropy with the Adam optimizer and curriculum learning, plus data augmentation to cover a range of SNRs. The method achieves sensitivity comparable to PyCBC and ML competitors in the MLGWSC-1 challenge, e.g., a sensitive distance around $2428$ Mpc at FAR $=1$ per month, and demonstrates practical computational efficiency on contemporary CPU/GPU hardware. The results support extending this approach to real non-Gaussian noise, higher- and lower-mass BBH systems, and integration into existing GW search pipelines for faster candidate screening. $ $

Abstract

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.

Figures (6)

• Figure 1: Workflow of our algorithm
• Figure 2: Example of an SNR map for the strain data containing only Gaussian noise. The upper and bottom panels show the data assuming LIGO Hanford and Livingston, respectively. The values are normalized so that the highest and the lowest values must be 1 and 0, respectively (see Eq. \ref{['eq: normalization snr map']}).
• Figure 3: Example of an SNR map for the data containing the simulated Gaussian noise and the injected CBC signal. Upper and lower panels correspond to LIGO Hanford and Livingston, respectively. Injected signal has the chirp mass of 17.2, and its SNR are 28.4 for Hanford and 22.4 for Livingston. The normalization \ref{['eq: normalization snr map']} is applied. A characteristic pattern of the injected signal is clearly visible around 0.5s.
• Figure 4: Accuracy curve. The blue and orange lines represent the accuracy for the training data and validation data, respectively. The sudden decreases at the epochs of 15, 30, and 60 are when the SNR range of the training data is changed to include the low SNR data.
• Figure 5: Distributions of the statistics for the noise data set and the CBC datasets. The blue and the orange lines are the results for the noise data and the CBC data, respectively.