Novel neural-network architecture for continuous gravitational waves

The high computational cost of wide-parameter-space searches for continuous gravitational waves (CWs) significantly limits the achievable sensitivity. This challenge has motivated exploration alternative search methods, such as deep neural networks (DNNs). Previous attempts [1,2] to apply convolutional image-classification DNN architectures all-sky and directed CW showed promise short, one-day durations, but proved ineffective longer durations around ten days. In this paper, we offer a hypothesis limitation propose new design principles overcome it. As proof concept, show that our novel architecture attains matched-filtering sensitivity targeted (i.e., single sky-position frequency) in Gaussian data from two detectors spanning We illustrate performance different sky positions five frequencies 20--1000 Hz range, spectrum an ``easy'' ``hardest'' case. corresponding depths fall range $82--86/\text{ }\text{ }\sqrt{\mathrm{Hz}}$. same is trained each case, taking between 4--32 hours reach detection probability DNNs function signal amplitude varies consistently with matched filtering. Furthermore, statistic distributions can be approximately mapped those $\mathcal{F}$-statistic under simple monotonic function.