Overview

A hybrid CNN-LSTM deep learning pipeline to detect gravitational wave signals from raw LIGO strain data, trained on the G2Net Kaggle Competition dataset.

Results

Metric	Score
Accuracy	74%
ROC-AUC	0.8025
Signal Precision	81%
Noise Recall	86%

Dataset

• G2Net Gravitational Wave Detection (Kaggle)
• 50,000 samples from 3 LIGO/Virgo detectors
• Binary classification: GW signal (1) vs noise (0)
• Balanced dataset: 50% positive rate

Preprocessing Pipeline

1. Bandpass Filter — Butterworth filter (20-500 Hz)

   • GW signals exist only in this frequency range
   • Removes low-frequency seismic and high-frequency electronic noise

2. Spectral Whitening — Flatten noise floor

   • Divides signal by its Power Spectral Density (PSD)
   • Makes faint GW chirp visible to the model
   • Most critical step for GW detection

3. Normalization — Zero mean, unit variance

   • Applied per detector channel
   • Performed in float64 before float32 conversion to avoid underflow (raw strain ~1e-20 scale)

Model Architecture

Input: (batch, 3, 4096) ↓ CNN Block: Conv1d(3→32, k=64, s=4) + BN + ReLU + MaxPool Conv1d(32→64, k=16, s=2) + BN + ReLU + MaxPool Conv1d(64→128, k=8, s=2) + BN + ReLU + MaxPool ↓ Output: (batch, 128, ~16) ↓ Reshape: (batch, ~16, 128) ↓ BiLSTM(128→256, 2 layers, dropout=0.3) ↓ Last timestep: (batch, 256) ↓ Classifier: Linear(256→64) → ReLU → Linear(64→1) ↓ Output: (batch,) — binary prediction

Why CNN-LSTM?

• CNN extracts local wave patterns (chirps, oscillations) from the raw time series
• LSTM captures how these patterns evolve over time
• Bidirectional reads signal both forward and backward
• Together they model both local features AND temporal dependencies

Training

• Optimizer: Adam (lr=1e-3, weight_decay=1e-4)
• Loss: BCEWithLogitsLoss
• Scheduler: ReduceLROnPlateau (patience=2, factor=0.5)
• Epochs: 30 (best model saved at epoch 18)
• Hardware: Tesla T4 GPU

Tech Stack

Python | PyTorch | SciPy | NumPy | Scikit-learn | Kaggle