Deep Learning for Real-time Gravitational Wave Detection and Parameter Estimation: Results with Advanced LIGO Data

Daniel George, Eliu A. HuertaarXiv 2017 · arXiv

Abstract

This study introduces "Deep Filtering," a framework utilizing Convolutional Neural Networks (CNNs) to detect gravitational waves and estimate binary black hole masses in real-time. By leveraging Advanced LIGO data, the approach demonstrates significant improvements in computational efficiency and resilience to non-Gaussian noise artifacts compared to traditional matched-filtering techniques.

Introduction

The detection of gravitational waves (GWs) from massive cosmic events, such as black hole mergers, is a cornerstone of modern astrophysics. Currently, the "gold standard" for signal identification is matched-filtering, which correlates incoming data against a vast library of pre-calculated templates. However, this method faces several challenges:

  1. Computational Complexity: The template library grows exponentially with the complexity of the signals (e.g., eccentric orbits or spinning components), making real-time analysis difficult.
  2. Noise Sensitivity: LIGO data contains "glitches"—short-duration noise artifacts that can mimic GW signals. Matched-filtering's assumption of stationary Gaussian noise often leads to false positives.
  3. Latency: Rapid detection is essential for multimessenger astrophysics, requiring sub-second processing to alert electromagnetic observatories.

Methodology

The Deep Filtering framework employs two specialized CNNs: a Classifier to distinguish signals from noise and a Predictor to perform regression for mass estimation.

Key architectural features include:

  • Dilated Convolutions: These expand the receptive field, allowing the network to capture long-duration signal features without a proportional increase in parameters.
  • Curriculum Learning: Training begins with high-SNR signals and progressively introduces more noise.
  • Transfer Learning: Models are pre-trained on Gaussian noise and fine-tuned on real, non-stationary LIGO noise to improve generalization.

Data and Training

The models were trained using approximately 2,500 templates for binary black hole (BBH) systems with masses between 5 and 75 solar masses. Training data incorporated 4,096 seconds of real LIGO noise sampled from periods surrounding historical GW events. Additionally, simulated "sine-Gaussian" glitches were used to train the classifier to distinguish between true signals and detector artifacts.

Results and Discussion

Deep Filtering demonstrates remarkable performance across several metrics:

  • Speed: Processing 1 second of data takes approximately 85 milliseconds on a CPU and 540 microseconds on a GPU, significantly faster than the ~2 seconds required for matched-filtering.
  • Sensitivity and Accuracy: The system achieves 100% sensitivity for signals with an SNR > 10. For the GW150914 event, predicted masses were within the error bars of official LIGO results, maintaining a mean relative error under 10%.
Sensitivity and Mean Relative Error comparison between Deep Filtering and Matched Filtering as a function of SNR.
  • Resilience: The CNNs correctly identified 30% more glitches as noise compared to traditional methods, showcasing superior morphology recognition.
  • Generalization: The model successfully identified real events like GW150914 despite being trained primarily on simulated templates, indicating it has learned the underlying physical waveforms.

Conclusion

Deep Filtering represents a significant advancement in GW data analysis, offering a portable (23MB) and highly efficient solution for real-time detection. Future work will focus on expanding the parameter space to include spinning and eccentric black holes, as well as integrating multi-detector data for improved sky localization and live deployment in LIGO search pipelines.

Summary updated

BibTeX
@misc{george2017deep,
  author = {Daniel George and Eliu A. Huerta},
  title = {{Deep Learning for Real-time Gravitational Wave Detection and Parameter Estimation: Results with Advanced LIGO Data}},
  year = {2017},
  eprint = {1711.03121},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1711.03121},
}
Figure