Investigating all-sky Frequency Hough performances for neutron stars

TL;DR

This work addresses the detection of continuous gravitational waves from isolated neutron stars by focusing on all-sky searches to bridge the gap between the predicted population of $\mathcal{O}(10^{8-9})$ and the observed $\mathcal{O}(10^{3})$. It evaluates the Frequency Hough (FH) pipeline and introduces a neural-network-based follow-up to tag Hough maps as signal-bearing or noise, enabling subthreshold candidates to be studied. Using real O3 data, the NN follow-up achieves high classification accuracy for subthreshold amplitudes, supported by feasible GPU runtimes such as $e=16.07\ \mathrm{s}$, $f=2.24\ \mathrm{s}$, and $e\rightarrow f=64.58\ \mathrm{s}$. The results suggest extending the datasets to include binary and weaker signals and further optimizing training and sky-mapping to improve generalization.

Abstract

Between the estimated population of Neutron Stars (NSs) and the actual number present in the catalogs, there is a huge gap: O(10$^{8-9}$) vs O(10$^3$). Among the different search techniques for Continuous gravitational waves (CWs), the all-sky could help to reduce the discrepancy. We focus on the all-sky CW pipeline Frequency Hough (FH), which operates without prior knowledge of the source parameters ($f,\dot{f}, λ, β$). Here, we present a Machine Learning strategy, diverging from the standard follow-up(FU) of the FH pipeline. We study the performance with real interferometer data, until reaching $h$ value subthreshold for the standard FU procedure ($CR_{thr}=5$), with encouraging classification results.

Figures (4)

• Figure 1: Frequency Hough pipeline standard workflow (purple), with the Machine Learning approach (green).
• Figure 2: Alternative FU workflow.
• Figure 3: Repartition of the HM for the three stages (Train, Validation, and Test).
• Figure 4: Test results of the two LIGO detectors displayed with a confusion matrix.