Comparative study of 1D and 2D convolutional neural network models with attribution analysis for gravitational wave detection from compact binary coalescences

Recent advancements in gravitational wave astronomy have seen the application of convolutional neural networks (CNNs) in signal detection from compact binary coalescences. This study presents a comparative analysis of two CNN architectures: one-dimensional (1D) and two-dimensional (2D) along with an ensemble model combining both. We trained these models to detect gravitational wave signals from binary black hole (BBH) mergers, neutron star-black hole (NSBH) mergers, and binary neutron star (BNS) mergers within real detector noise. Our investigation entailed a comprehensive evaluation of the detection performance of each model type across different signal classes. To understand the models’ decision-making processes, we employed feature map visualization and attribution analysis. The findings revealed that while the 1D model showed superior performance in detecting BBH signals, the 2D model excelled in identifying NSBH and BNS signals. Notably, the ensemble model outperformed both individual models across all signal types, demonstrating enhanced detection capabilities. Additionally, input feature visualization indicated distinct areas of focus in the data for the 1D and 2D models, emphasizing the effectiveness of their combination.

Figure 1

Example strain data at H1 detector in the training set. The upper figures show the whitened time-series data used as input to the 1D model, and the lower figures show the time-frequency maps for the 2D model. The component masses of the BBH signal are $51.3M_{\odot }$ and $50.9M_{\odot }$, whereas in the NSBH sample, the respective masses are $33.3M_{\odot }$ and $1.83M_{\odot }$, and the masses of the BNS sample are $1.54M_{\odot }$ and $1.40M_{\odot }$. The single-detector SNR of each signal is 15, and the merger time is fixed at 3.9 s.

Figure 2

Illustration of the ensemble model. The final output is the weighted average of the outputs from the 1D CNN, 2D CNN, and ensemble network.

Figure 3

ROC curves of the three models for BBH, NSBH, and BNS signals at a fixed network SNR of 8. The SNRs are computed with four-second signals.

Figure 4

Sensitivity curves of the three models for BBH, NSBH, and BNS signals at a fixed false alarm probability of 0.001. The SNRs are computed with four-second signals.

Figure 5

Two-dimensional representations of feature maps of the 1D model by t-SNE. The marker size of signal sample corresponds to the SNR.

Figure 6

Two-dimensional representations of feature maps of the 2D model by t-SNE. The marker size of signal sample corresponds to the SNR.

Figure 7

Attribution maps of the 1D and 2D models for a BBH signal and corresponding input samples at H1 detector. The component masses are $37.7M_{\odot }$ and $6.94M_{\odot }$. The red dashed line shows the time of coalescence. The bottom plot shows the values of integrated gradients summed over all frequencies for each time bin.

Figure 8

Attribution maps of the 1D and 2D models for a NSBH signal and corresponding input samples at H1 detector. The component masses are $6.78M_{\odot }$ and $1.99M_{\odot }$. The red dashed line shows the time of coalescence. The bottom plot shows the values of integrated gradients summed over all frequencies for each time bin.

Figure 9

Attribution maps of the 1D and 2D models for a BNS signal and corresponding input samples at H1 detector. The component masses are $1.18M_{\odot }$ and $1.15M_{\odot }$. The red dashed line shows the time of coalescence. The bottom plot shows the values of integrated gradients summed over all frequencies for each time bin.