Technical Summary: AI-enabled Gravitational-Waves Searches for Binary Neutron Stars at Optimal Sensitivity

Problem Statement

Gravitational-wave (GW) searches for compact binary coalescences (CBCs) in the LIGO-Virgo-KAGRA (LVK) network face significant computational challenges, particularly for Binary Neutron Star (BNS) systems. Traditional matched-filter pipelines (e.g., GstLAL, PyCBC) require matching data streams against $\sim$millions of reference waveforms, consuming up to 1,000 CPU cores. While neural network-based approaches like Aframe have successfully detected Binary Black Hole (BBH) mergers in real-time, they struggle with BNS signals. BNS inspirals are long-duration signals (up to minutes in the detector band), whereas the original Aframe architecture relies on short, fixed time windows (1.5-second kernels) optimized for the shorter-duration BBH signals. Extending the input window to capture full BNS signals increases computational costs and latency, and standard sequence-based architectures (e.g., RNNs, Transformers) suffer from optimization instability and quadratic scaling with sequence length when handling such long inputs.

Methodology

The authors extend the Aframe framework to the BNS regime by introducing a heterodyning (demodulation) preprocessing step that compresses long-duration inspiral signals into shorter, compact representations without losing critical phase information.

1. Heterodyne Preprocessing

• Concept: The method demodulates the whitened GW strain data using a bank of reference waveforms based on the chirp mass ($M$). The phase evolution of the inspiral is dominated by the chirp mass term in the post-Newtonian approximation.
• Implementation:
  • A bank of $N=100$ reference phases is generated, spanning the BNS mass range (1–2.5 $M_\odot$) sampled uniformly in log mass.
  • The whitened data is heterodyned against these $N$ reference phases, producing $2N$ channels (one per detector, H1 and L1).
  • This process transforms the long-duration inspiral into a signal where power is concentrated near the merger time, effectively compressing the signal duration from minutes to seconds.
• Channel Selection: The heterodyned channels are averaged and combined in quadrature across detectors to enforce coherence. The top $k=10$ channels with the highest peak amplitude are selected.
• Input to Network: The selected top $k$ channels from both detectors are stacked to form $2k$ input channels. A 1.5-second segment ending at the peak time is extracted. This allows the use of the same ResNet-34 architecture previously used for BBHs, which is efficient for short temporal windows.

2. Training and Data Strategy

• Data: The model is trained on real detector data from the LVK's O3 observing run. Background is generated via time-sliding to ensure signal-free segments.
• Signal Injection: Simulated BNS waveforms (IMRPhenomPv2 approximant) are injected into the background. Intrinsic and extrinsic parameters follow astrophysical distributions consistent with GWTC-3.
• Augmentation: Techniques include time-reversal, detector swapping, random muting, and dynamic rescaling of injected signals to enforce learning of signal coincidence and coherence.
• Loss Function: Binary cross-entropy loss is used for classification (signal vs. background).

Key Contributions

1. Extension of AI to BNS Regime: The paper demonstrates that the Aframe neural network, previously limited to BBHs, can achieve sensitivity comparable to matched-filter pipelines for BNS systems by employing heterodyning.
2. Computational Efficiency: The approach reduces the input dimensionality of long-duration signals, allowing the use of a lightweight ResNet architecture. It requires only a single non-flagship GPU (e.g., NVIDIA A30) for online deployment, significantly lowering computational costs compared to CPU-intensive template banks.
3. Real-Time Viability: The pipeline achieves low-latency inference (median end-to-end latency of $\sim$4.7 seconds from data arrival to GraceDB registration), making it suitable for real-time multi-messenger alerts.
4. Robustness: The model is trained on real, non-stationary detector noise and demonstrates stability across different observing periods without requiring frequent retraining.

Results

• Sensitivity Comparison:
  • Vs. Matched-Filter Pipelines: Aframe achieves sensitive volumes comparable to state-of-the-art matched-filter searches (MBTA, GstLAL, PyCBC) across the BNS parameter space. For higher-mass BNS systems (e.g., $1.8-2.0 M_\odot$), Aframe exceeds the sensitive volume of matched-filter pipelines. For lower-mass systems ($1.4 M_\odot$), sensitivity is slightly lower at very low false alarm rates (FAR < 1/year), attributed to the 8-second kernel limit capturing less of the long inspiral SNR.
  • Vs. Previous ML Searches: Compared to the machine learning search by Schäfer et al. [23], Aframe shows a three-fold improvement in sensitive distance ($\sim$160 Mpc vs. $\sim$50 Mpc at FAR $\sim$1/month) when tested on idealized Gaussian noise, and $\sim$120 Mpc on real O3 noise.
• O3 Mock Data Challenge (MDC): On the independent O3 MDC dataset, Aframe recovered a comparable or higher number of injections than low-latency matched-filter pipelines (PyCBC, SPIIR) and performed competitively with MBTA and GstLAL. It recovered the most injections for higher-mass systems ($m_{1,2} \gtrsim 1.9 M_\odot$).
• Stability: The recovery fraction remained stable throughout the O3 observing run, with variations correlating more with detector sensitivity changes than model degradation.
• Computational Performance:
  • Training: $\sim$96 hours on a single NVIDIA H200.
  • Offline Analysis: 10 years of background data processed in $\sim$8 days using two NVIDIA A30 GPUs.
  • Online Inference: $\sim$172 $\mu$s latency per sample on a single NVIDIA A30.

Significance and Claims

The paper claims to present the first AI-enabled search for BNS signals that achieves sensitivity comparable to matched-filter pipelines while maintaining significantly lower computational and latency costs. By addressing the "long-duration" challenge through heterodyning, the authors establish that neural networks can be effectively deployed for low-mass compact binaries.

The significance lies in:

• Scalability: The ability to run efficient, low-latency searches on a single GPU, enabling the deployment of multiple specialized searches (e.g., for BNS, BBH, and subsolar-mass binaries) simultaneously without overwhelming computational resources.
• Multi-Messenger Astronomy: The low latency is critical for providing early warnings to electromagnetic and neutrino telescopes, facilitating the capture of prompt emissions from BNS mergers.
• Future Framework: The work lays the groundwork for a unified, computationally efficient all-sky search framework for future observing runs (O5 and beyond) and next-generation detectors, potentially replacing or augmenting traditional template-based methods.

The authors remain modest regarding the slight sensitivity deficit for very low-mass systems, attributing it to the current kernel length, and identify extending the heterodyning duration as a primary direction for future improvement.