Real-Time Multi-Mode Post-Merger Gravitational Wave Detection using Convolutional Neural Networks: Methodology Development for Third-Generation Detectors

Abstract

The detection and characterization of post-merger gravitational wave signals from binary neutron star mergers remains challenging with current ground-based detectors. We present a convolutional neural network framework designed for real-time detection and multi-mode frequency extraction of post-merger signals, achieving an inference latency of 3.0 ms and a frequency accuracy of 48.6 Hz on the direct-comparison subsets (53 Hz on the comprehensive test set). The framework is validated on realistic LIGO O4 detector noise including authentic GravitySpy glitch morphologies, demonstrating ROC AUC of 0.999999 and 99.998% detection efficiency at 1% false alarm rate. These exceptional performance metrics arise from an aggressive training augmentation strategy that exposes the network to artificially challenging conditions, enabling robust generalization to our synthetic O4 detector noise model. We compare performance against a simplified matched filtering baseline using Lorentzian templates (23x more accurate despite a 4.7x computational overhead) and Bayesian parameter estimation (1.2 million times faster), establishing complementary trade-offs in the analysis landscape. While current O4 sensitivity limits post-merger detections to ~20 Mpc (~1 detection per century), this methodology provides essential infrastructure for third-generation detectors (Einstein Telescope, Cosmic Explorer) where post-merger detection will become routine with annual detection rates exceeding 100 events. Our validation framework identifies expected behavior in uncertainty scaling that reflects realistic training constraints rather than idealized Fisher information limits, demonstrating honest assessment practices for machine learning applications in gravitational wave astronomy.