Deep Learning Search for Gravitational Waves from Compact Binary Coalescence

TL;DR

This work investigates a hybrid approach that combines matched-filtering concepts with Convolutional Neural Networks, enabling efficient signal searches without relying on the usual $\chi^2$ rejection test, and shows that deep learning assisted searches can support sustainable gravitational-wave data analysis in future detector eras.

Abstract

Gravitational wave searches rely on a combination of methods, including matched filtering, coherent analyses, and more recent machine learning based pipelines. For compact binary coalescences, where signals originate from the relativistic dynamics of compact objects, matched filtering remains a central element, but its computational cost will increase substantially with the data volumes and parameter-space coverage required by next-generation interferometers such as the Einstein Telescope. Developing complementary strategies that reduce computational load while preserving detection performance is therefore essential. We investigate a hybrid approach that combines matched-filtering concepts with Convolutional Neural Networks, enabling efficient signal searches without relying on the usual $χ^2$ rejection test. Using simulated data sets that include injected signals in Gaussian noise, transient noise, and physical effects not represented in template bank, such as eccentricity, precession and higher-order modes, we show that the method achieves a detection efficiency comparable to a standard matched-filtering search while offering a more resource efficient pipeline. These results indicate that deep learning assisted searches can support sustainable gravitational-wave data analysis in future detector eras.

Figures (13)

• Figure 1: Examples of the TT-SNR-Map obtained by piling up the $\rho(t)$ after the normalization procedure. The left picture corresponds to Gaussian Noise solely, on the center to an injected signal and on the right the TT-SNR Map resulting from an injected signal with a glitch superimposed.
• Figure 2: EasyResNet architecture. A convolutional stem (Conv2D-BN-ReLU) processes the input and feeds three residual blocks (ResBlock1-3). The final feature maps are reduced by global average pooling (GAP) to a feature vector, regularized with dropout, and passed to a fully connected (FC) layer for classification. Arrows indicate data flow; numbers below the blocks denote channel counts (and, where shown, spatial dimensions)
• Figure 3: Residual block. The lower arrow represents the input passing through a sequence of convolutions and normalizations. The upper one carries the identity. The green $\oplus$ denotes element-wise addition. Arrow directions indicate data flow through the layers (Conv2D-BN-ReLU-Conv2D-BN-Conv2D in this example).
• Figure 4: Results for purely Gaussian Noise and BNS injections.
• Figure 5: Receiver Operating Characteristic Curve considering injections in pure Gaussian noise.