GW-FALCON: A Novel Feature-Driven Deep Learning Approach for Early Warning Alerts of BNS and NSBH Inspirals in Next-Generation GW Observatories

TL;DR

GW‑FALCON introduces a feature‑driven deep‑learning approach for early redetection of BNS and NSBH inspirals in next‑generation GW observatories. By converting 60 s observational time windows into fixed‑length TSFEL feature vectors and training shallow feed‑forward ANNs, the method achieves high discrimination between signal+noise and noise‑only data across ET and CE PSDs, with CE delivering near‑perfect accuracy. The study delineates PI SNR–dependent detection efficiency, demonstrates sliding‑window triggering with low false‑alarm rates in Gaussian noise, and discusses practical deployment considerations, including multi‑detector networks and integration with low‑latency pipelines. Overall, GW‑FALCON offers a robust, computationally efficient framework for rapid premerger alerts that can enhance multimessenger follow‑up with next‑generation GW networks.

Abstract

Next-generation GW observatories such as the ET and CE will detect BNS and NSBH inspirals with high SNRs and long in-band durations, making systematic early-warning alerts both feasible and scientifically valuable. Such triggers are essential for coordinating rapid electromagnetic follow-up. In this work, we introduce GW-FALCON, a novel feature-driven DL framework for early-time detection between GW signal+noise and noise-only data in next-generation detectors. Instead of feeding raw time series to CNN or more complex neural network architectures, we first extract a large set of statistical, temporal, and spectral quantities from short observational time windows using the TSFEL library. The resulting fixed-length feature vectors are then used as input to feed-forward ANNs suitable for low-latency operation. We demonstrate the method using simulated BNS and NSBH inspiral waveforms injected into colored Gaussian noise generated from the ET and CE design PSDs. We train separate ANNs on feature sets extracted from partial-inspiral windows characterized by different maximum instantaneous frequencies, enabling early-warning triggers from tens to hundreds of seconds before merger. Across all detector configurations and datasets, the resulting classifiers achieve high accuracy and detection efficiency, with ET-like networks typically reaching test accuracies of order 90% and CE-like ones exceeding 97% at low false-alarm probability. To the best of our knowledge, this work presents the first comprehensive feature-based DL detection framework for Next-generation GW observatories, connecting feature extraction from strain time series data to robust signal-noise classification within a setup that can be extended to real data and to more advanced neural network architectures.

Figures (23)

• Figure 1: Left panel: Indicative plus polarization $h_{+}$ for a BNS GW template with component masses $m_{1} = 1.4\,M_{\odot}$ and $m_{2} = 1.0\,M_{\odot}$. Right panel: Same as left panel for an NSBH GW template with component masses $m_{1} = 4.0\,M_{\odot}$ and $m_{2} = 1.5\,M_{\odot}$. In both panels, the shaded region marks the time interval during which early-warning alerts are expected with the methodology deployed in this work. The merger time is set to $t_{c} = 0 \ \mathrm{s}$. The corresponding instantaneous GW frequency, shown as a function of time relative to the merger, is also displayed in each panel.
• Figure 2: Design-sensitivity (analytical) power spectral densities (PSDs) for next-generation LVK detectors and third-generation observatories. In this panel, we present the Advanced LIGO analytical PSD (O4 target), the Advanced LIGO A+, the Advanced Virgo (O4 target), the KAGRA (design), the Einstein Telescope (ET-D), and, finally, the Cosmic Explorer (CE, wideband) PSD as a function of frequency within the range $f\in[5,100] \ \mathrm{Hz}$. In addition, we show coloured vertical lines indicating benchmark frequencies that are useful for our investigation, as discussed in the following sections. In any case, for the chosen frequency band, the indicative noise curves illustrate the anticipated frequency-dependent improvements in strain sensitivity.
• Figure 3: Top panel: source frame $m_1-m_2$ masses for BNS and NSBH waveforms employed in this work, color-coded by the source frame chirp mass $\mathcal{M}_c$. Bottom panel: Luminosity-distance histogram distributions for the associated CBC populations considered. For each case, we also provide the median luminosity distance and the associated $90\%$ credible interval for the distributions illustrated.
• Figure 4: CE1 strain data together with an injection with component source masses similar to the GW170817 BNS event ($m_1 = 1.46 \ M_{\odot}$, $m_2 = 1.27 \ M_{\odot}$), but located at a different luminosity distance, $d_L = 440\ \mathrm{Mpc}$. Colored vertical lines mark the time instants relative to the merger when the waveform’s instantaneous frequency reaches the corresponding $f_{\mathrm{max}}$ values employed. For each reference value of instantaneous maximum frequency, we also show the corresponding PI SNR, illustrating how the accumulated SNR increases in the detector's band as the system evolves toward merger. The time of coalescence corresponds to $t_c = 0 \ \mathrm{s}$.
• Figure 5: CE1 Detector: PI SNR distributions for BNS (left panel) and NSBH (right panel) injections embedded in the CE1 detector's strain data. Each histogram is color-coded by the annotated maximum instantaneous frequency $f_{\mathrm{max}}$. For each histogram, we report the median PI SNR, along with the corresponding $90\%$ credible interval, which indicates the central value and spread of the distribution, enabling a direct comparison of the typical signal strength and its variability across the different cases.