Extending the reach of gravitational-wave detectors with machine learning

Document #:

LIGO-P1800129-v1

Document type:

P - Publications

Abstract:

We apply Long Short-Term Memory (LSTM) Neural Networks as a time-series regression analysis
technique to filter instrumental noises from gravitational-wave detectors at LIGO. Unlike traditional
neural networks, LSTM networks can store and use information from their past inputs, and thus is
robust in handling sequential data like gravitational-wave signals. Once trained on the detector noise
data, an LSTM network should be able to learn, predict, and subtract both the linear and non-linear
noise coupling mechanisms. This would result in a sensitivity improvement and allow the detection of
gravitational-wave sources currently below the noise floor.

Files in Document:

• Final project report (final_project.pdf, file is not accessible)

Other Files:

• extending-reach-gravitational.pdf (file is not accessible)
• Final Presentation (final_presentation.pdf, file is not accessible)
• First interim report (interim_report_1.pdf, file is not accessible)
• Keras and Hyperopt demo on LSTM and hyperparameter tuning for non-linear noise regression (lstm_demo.ipynb, file is not accessible)
• Project proposal (ligo_surf_proposal_tringuyen.pdf, file is not accessible)
• Second interim report (interim_report_2.pdf, file is not accessible)
• University of Rochester SPS Symposium (SPS_Symposium.pdf, file is not accessible)
• Utilities module to run Python notebooks (utils.py, file is not accessible)

Topics:

• Detector Characterization

Authors:

• Tri Nguyen
• Michael Coughlin
• Rich Ormiston

Keywords:

SURF18

Notes and Changes:

Change abstract

Referenced by:

• LIGO-T1900127: Final reports from LIGO SURF program, Summer 2018