Using deep learning to predict matched signal-to-noise ratio of gravitational waves

The existing matched filtering method for gravitational wave (GW) search relies on a template bank. The computational efficiency of this method scales with the sizes of the templates within the bank. Higher-order modes and eccentricity will play an important role when third-generation detectors operate in the future. In this case, traditional GW search methods will hit computational limits. To speed up the computational efficiency of GW searches, we propose the utilization of a deep learning (DL) model bank as a substitute for the template bank. This model bank predicts the latent templates embedded in the strain data. Combining an envelope extraction network and an astrophysical origin discrimination network, we realize a novel GW search framework. The framework can predict the GW signal's matched filtering signal-to-noise ratio (SNR). Unlike the end-to-end DL-based GW search method, our statistical SNR holds stronger physical interpretability than the $p_{\text{score}}$ metric. Moreover, the intermediate results generated by our approach, including the predicted template, offer valuable assistance in subsequent GW data processing tasks such as parameter estimation and source localization. Compared to the traditional matched filtering method, the proposed method can realize real-time analysis.