Source-linked AI summary
Deep Learning for Real-time Gravitational Wave Detection and Parameter Estimation: Results with Advanced LIGO Data
Daniel George, E. A. Huerta
TL;DR
Existing matched-filter searches cover a restricted parameter space and face computational challenges, motivating a more scalable approach to gravitational-wave analysis. The paper applies two CNNs directly to continuous real-LIGO time series for detection and parameter estimation. Deep Filtering recovers true parameters from observed LIGO events with matched-filtering-comparable sensitivity, faster processing, and resilience to glitches.
Problem
Matched-filtering searches a restricted 3D subset of the 8D gravitational-wave parameter space, and extending coverage is computationally prohibitive.
Method
Deep Filtering uses classifier and predictor CNNs that process continuous multi-detector time series directly for signal detection and source-parameter estimation.
Results
Deep Filtering detects and recovers parameters of real LIGO gravitational-wave events with performance comparable to matched filtering, while operating much faster and resisting glitches.
Takeaways & Limitations
Deep Filtering can support real-time gravitational-wave and multimessenger searches by rapidly narrowing parameter space for subsequent verification and follow-up.
Abstract
from arXiv · showhide
The recent Nobel-prize-winning detections of gravitational waves from merging black holes and the subsequent detection of the collision of two neutron stars in coincidence with electromagnetic observations have inaugurated a new era of multimessenger astrophysics. To enhance the scope of this emergent field of science, we pioneered the use of deep learning with convolutional neural networks, that take time-series inputs, for rapid detection and characterization of gravitational wave signals. This approach, Deep Filtering, was initially demonstrated using simulated LIGO noise. In this article, we present the extension of Deep Filtering using real data from LIGO, for both detection and parameter estimation of gravitational waves from binary black hole mergers using continuous data streams from multiple LIGO detectors. We demonstrate for the first time that machine learning can detect and estimate the true parameters of real events observed by LIGO. Our results show that Deep Filtering achieves similar sensitivities and lower errors compared to matched-filtering while being far more computationally efficient and more resilient to glitches, allowing real-time processing of weak time-series signals in non-stationary non-Gaussian noise with minimal resources, and also enables the detection of new classes of gravitational wave sources that may go unnoticed with existing detection algorithms. This unified framework for data analysis is ideally suited to enable coincident detection campaigns of gravitational waves and their multimessenger counterparts in real-time.
I. INTRODUCTION
LIGO’s matched-filter searches are sensitive but limited in parameter-space coverage and computational scalability. The paper extends Deep Filtering with deep CNNs that process noisy time series directly for real-time gravitational-wave detection and parameter estimation.
- Motivation: Matched-filtering targets only a 3D subset of the 8D parameter space available to gravitational-wave detectors.Extending template matching to spin-precessing, quasi-circular, or eccentric binary black holes is computationally prohibitive.
- Motivation: Existing searches may miss gravitational waves from compact-binary populations formed in dense stellar environments.
- Motivation: Deep learning offers a scalable alternative that learns directly from raw data without manual feature engineering.The approach uses deep hierarchical neural-network layers trained with back-propagation and gradient descent.
- Approach: Deep Filtering uses two convolutional neural networks for time-series classification and regression.The method was previously shown to reach matched-filtering-comparable sensitivity on simulated LIGO noise while estimating source parameters in real time.
- Contribution: This article extends Deep Filtering to real LIGO noise and demonstrates detection and multiple-parameter estimation for weak signals in non-Gaussian, non-stationary data.The authors report comparable performance to matched filtering, greater speed, glitch resilience, real-event parameter recovery, and generalization to new noise PSDs without retraining.
- Implications: Deep Filtering can interpolate between templates, generalize beyond training classes, and estimate parameters for signals contaminated by glitches.After one-time training, continuous streams can be analyzed in real time with a single CPU, while GPU processing supports intensive searches.
II. METHODS
The method combines classifier and predictor CNNs applied to sliding windows of continuous, multi-detector strain data. Training and testing use simulated binary-black-hole injections in distinct real LIGO noise distributions, with whitening, curriculum learning, and transfer learning.
- Pipeline: A classifier CNN scans each detector’s continuous strain stream with a sliding window and returns the probability that a signal is present.The predictor CNN is applied when the classifier favors a signal and estimates source parameters from the same input.
- Pipeline: Coincident detections with similar parameters across multiple detectors strengthen confidence before rapid matched-filter verification.
- Data and assumptions: The study restricts injections to optimally oriented, quasi-circular, non-spinning stellar-mass binary black holes with component masses from 5M⊙ to 75M⊙ and mass ratios from 1 to 10.
- Data and assumptions: Training and testing each use approximately 2,500 effective-one-body waveform templates, with intermediate test masses separated from training masses by 0.5M⊙.
- Data and assumptions: Noise from GW151226 and LVT151012 is used for training and validation, while GW150914 noise is reserved for testing across distinct noise distributions.The design also evaluates robustness to transient glitches and changing LIGO power spectral densities.
- Evaluation: A representative injection with optimal matched-filter SNR = 7.5 is detected with over 99% sensitivity, while component-mass estimates have mean relative error less than 10%.The signal is embedded in whitened real LIGO noise and has component masses of 57M⊙ and 33M⊙.
III. RESULTS
Deep Filtering detects and characterizes gravitational-wave signals in real LIGO noise, matching or improving on matched-filtering while operating faster and remaining robust to glitches. It also generalizes beyond trained templates and successfully analyzes a real event.
- Detection: 100% sensitivity was achieved for signals with SNR greater than 10, with a false alarm rate below 1%.With independent detectors and coincidence enforcement, the false alarm rate would be less than 0.01%.
- Parameter estimation: SNR ≥15.0 signals had component-mass errors below the spacing between training templates, while high-SNR mean relative error was below 10%.Matched-filtering with the same template bank always had error greater than 10%, indicating interpolation between templates.
- Real-event analysis: The model identified GW150914 at the true time and predicted component masses within published error bars despite training only on simulated quasi-circular non-spinning injections.No false alarms remained after requiring simultaneous detections across multiple detectors.
- Glitch classification: The classifier distinguished new sine-Gaussian glitches from signals with a false alarm rate below 1%, while matched-filtering classified approximately 30% of the same glitches as signals.This supports extending Deep Filtering toward unified glitch classification, detection, and parameter estimation.
- Glitch coincidence: Signals contaminated by simulated sine-Gaussian glitches were correctly detected, indicating use as automatic triggers for gravitational waves coincident with glitches.The demonstrated behavior complements the classifier’s ability to recognize glitches as noise.
- Glitch robustness: More than 80% of signals at SNR 10 were detected with less than 30% relative parameter error after glitches were superimposed.The result motivates detecting signals during periods of bad detector data quality that existing pipelines often veto.
- Generalization: Deep Filtering detected eccentric and spin-precessing binary signals with the same sensitivity as the trained quasi-circular binary waveforms.These tests demonstrate generalization to signal classes beyond the training data.
- Computational efficiency: Each CNN is 23MB and evaluates one second of data in approximately 85 milliseconds on one CPU core or 540 microseconds on a GPU.The comparison time-domain matched-filtering algorithm required over 2 seconds to analyze one-second inputs on the CPU.
IV. CONCLUSION
The conclusion presents Deep Filtering as a real-time CNN framework for gravitational-wave detection and parameter estimation, with scalability, glitch resilience, and multimessenger applications. It also identifies extensions toward broader source parameters, glitch classification, and Bayesian follow-up as future directions.
- CNNs detect and estimate gravitational-wave parameters directly from LIGO data, offering a new real-time analysis paradigm.
- Deep Filtering can be combined with matched-filtering to quickly verify results using a small template set near the predicted parameter region.
- Deep learning scalability could support searches across millions or billions of templates and extend predictions to additional source parameters, signal classes, and multiple detectors.
- The method’s resilience to bad data and non-Gaussian transients motivates adding real glitches for automated classification and clustering.
- Deep Filtering may accelerate Bayesian parameter estimation and provide rapid, accurate alerts for electromagnetic follow-up campaigns.