Aframe Neural Network Matches Matched-Filter Sensitivity for Binary Neutron Star Searches Using Heterodyned Data on Single GPU

The arXiv preprint 2607.01372 demonstrates that heterodyning the strain data reduces the effective duration of BNS signals enough for the same convolutional architecture previously used on BBH events to reach equivalent sensitive volume at fixed false-alarm rate. In the reported injection campaign the network recovers 95 percent of signals with network SNR above 9.5, matching the performance of the PyCBC pipeline run on the same dataset while cutting online CPU cores from roughly 1000 to a single A100. This directly addresses the template-bank scaling problem that has limited real-time BNS searches since GW170817, whose 110-second chirp required banks exceeding 10^6 templates. The work also shows that inference-as-a-service orchestration allows the same trained model to re-analyze the full O3 archive in hours rather than weeks on a modest GPU cluster. Because the method was already running live for BBH triggers in O4, the extension to BNS requires only a change in the frequency band passed to the heterodyner and retraining on the lower-mass prior. The principal remaining limitation is the absence of an explicit ranking statistic that folds in sky-location and distance priors, which matched-filter pipelines obtain for free from the template parameters. A controlled comparison on the upcoming O5 dataset with at least 20 BNS injections will be required to confirm that the reported sensitivity holds under realistic non-stationary noise.