Gravitational-Wave Detection Pipeline 2025

Resources: RADAR Paper (Link) | GitHub Repo (public mirror; full repo private)

Implementation: Python with PyTorch/PyTorch Lightning for HPC and multi-GPU training.

Overview:

Low-latency gravitational-wave detection is challenging because true signals are rare and often buried in non-stationary detector noise, while follow-up facilities need reliable triggers within seconds. At the same time, several observatories operate under data locality constraints, making it impractical to stream raw data to a central location for further analysis. This project develops a deployable GW detection module for real LIGO data and exposes it through interfaces that make it straightforward to plug into federated, multi-messenger workflows such as RADAR.

Model & Training:

I led the design and implementation of the GW detection module used in the RADAR framework, including the end-to-end training pipeline on synthetic signals injected into true interferometer noise, the multi-detector aggregation architecture, and the trigger-generation logic for identifying merger candidates in whitened strain data. The federated inference layer, radio-analysis module, and Octopus/ProxyStore-based infrastructure are developed by collaborators and integrate my GW module as one component within the full RADAR system.

The core detector-side model is a WaveNet-style Hierarchical Dilated Convolutional Network (HDCN) that processes Hanford (H1) and Livingston (L1) strain independently to capture long-range temporal structure in 1 s windows (4096 samples). Its outputs are combined by an attention-based aggregator that performs cross-channel communication between detectors to produce a per-timestep GW confidence score (see the model schematics below for a visual overview). This replaces earlier graph-based aggregation schemes and significantly reduces false positives in multi-detector data while keeping the model lightweight enough for deployment close to the data.

Whitened BBH injection in Livingston and Hanford noise. Whitened BNS injection in Livingston and Hanford noise.

Figure 1–2. Whitened BBH (left) and BNS (right) injections in real LIGO noise for Hanford and Livingston. Both use labels set to 1 only in the 0.5 s preceding merger and 0 elsewhere; separate models are trained for each source class.

Models are trained on synthetic binary black hole (BBH) and binary neutron star (BNS) waveforms injected into real LIGO noise from the Gravitational Wave Open Science Center (GWOSC), using a curriculum that starts from louder injections and decays toward realistic signal-to-noise ratios. The data pre-processing pipeline involves whitening, band-passing, windowing, and label construction (with positive labels restricted to the 0.5 s pre-merger segment). The model returns a per-timestep score, while post-processing applies a calibrated, multi-criterion peak-finding stage (height, width, and plateau tests) to convert scores into reliable trigger times for use in downstream multi-messenger analyses.

Detection of GW170817: whitened Livingston and Hanford strains with the model's trigger at the merger time.
Figure 3. Example application to the binary neutron star event GW170817. Top: whitened L1 strain. Middle: whitened H1 strain. Bottom: model output and resulting trigger, correctly localizing the merger time in hour-long real interferometer data. A Tukey window is applied to a known glitch in the Livingston detector; the detection is robust to this choice.

Model Schematics: HDCN architecture schematic for one detector channel.
Figure 4. Hierarchical Dilated Convolutional Network (HDCN) used per detector to capture multi-scale temporal structure in the strain.

Cross-attention network schematic showing cross-detector communication. Cross-attention network schematic for the second detector channel.
Figure 5. Cross-Attention Network (CAN) submodules that exchange information between detectors via multi-head attention, enabling detector-specific correlations to be modeled explicitly.

Final output module schematic combining detector features into a unified detection score.
Figure 6. Final output module that aggregates enriched detector features into a per-timestep GW detection score suitable for low-latency triggering.

Federated Integration:

Within RADAR, only embeddings and compact trigger information are sent to a central coordinator, preserving data locality while enabling joint analysis with radio afterglow modeling and GW–EM consistency checks. I contributed to shaping the GW module’s interfaces and requirements so they integrate cleanly with the federated framework, while the implementation of the federated optimization, Octopus-based event fabric, and global coordination is led by the broader RADAR team.

Related Work:

This detector builds on and extends previous ML-based GW detection approaches, in particular by Tian et al. (2023) , and uses real strain data and catalogued events from GWOSC. Our contribution retrains on injections into real GWOSC noise, redesigns the multi-detector aggregator with cross-attention for improved false-positive control, and provides a production-ready implementation suitable for integration into federated multi-messenger infrastructures.