Using deep learning to denoise and detect gravitational waves

We have upgraded the MSNRnet framework to MSNRnet-2 by refining the training strategy, drawing inspiration from generative adversarial networks for data generation. The astrophysical discrimination network enforces constraints on the denoising output, ensuring that the ``$\mathrm{signal}+\mathrm{noise}$'' case conforms to an astrophysical-origin shape, while limiting the denoised output of the ``noise'' case to a nonastrophysical origin shape. This improvement has led to enhanced denoising and astrophysical discrimination performance in most scenarios. For the analysis of confident events, we observed a 1.3% increase in the detection rate. Remarkably, no false triggers were generated during the analysis of the data jointly processed by the Hanford and Livingston interferometers in O3b (about five months of data). These findings suggest that the denoising-discrimination-matched filtering framework holds promising potential for gravitational wave searches in the future.