AResGW is a machine-learning-based pipeline designed for offline real-time gravitational wave (GW) detection, especially optimized for binary black hole (BBH) mergers in the 7–50 M☉ range (individual component masses) with non-aligned spins, using real LIGO data.
- Core Architecture: Utilizes a 54-layer 1D deep residual network (ResNet) to perform binary classification of 1-second time segments as either containing a GW signal or pure noise.
- Training Data: Based on real LIGO O3a noise with over 900k injections of BBH waveforms (generated using IMRPhenomXPHM waveform model).
Key Features:
- Deep Adaptive Input Normalization (DAIN): Learns optimal normalization for nonstationary real detector noise.
- Dynamic Data Augmentation: Randomly swaps noise channels during training to increase robustness.
- Curriculum Learning: Trains the model first on high SNR signals, gradually introducing weaker ones using an empirical SNR relation.
- Efficient Whitening: Implements PSD estimation and whitening directly in PyTorch, enabling fast batch processing.
- Real-Time Ready: Processes one month of data in <2 hours on a typical GPU.
Link: 🔗 GitHub Repository