EAGER: ADAPT: Understanding Nonlinear Noise in LIGO: A Machine Learning Approach

The detection of gravitational waves by LIGO has enacted a paradigm shift in the exploration and study of cosmic objects. Despite a series of upgrades and improvements, the LIGO detectors suffer from noise whose origin is largely unknown and whose presence limits the astrophysical reach of the detectors, thus reducing the mass and distance of systems accessible to observation.

This project will develop novel machine learning methods capable of providing insight into the physical origins of noise in LIGO, yielding actionable information to guide commissioning efforts and future design decisions. Success in this project will improve the operational stability of the detectors and increase their astrophysical range, with the potential to advance scientific discovery.

The project will train graduate and undergraduate students in the confluence of detector commissioning and machine learning (ML) and artificial intelligence (AI) research, will develop open-source tools for understanding and detecting noise in complex scientific experiments, and foster an interdisciplinary research community which bridges physics and machine learning. Finally, success in the project has the potential to benefit efforts in cloud infrastructure resilience, a problem with multiple parallels to understanding noise in LIGO.

In addition to the main strain channel, each LIGO detector has over 10,000 auxiliary channels monitoring the operation of each subsystem and the seismic, acoustic, and electromagnetic environment. This vast data set can be leveraged to understand spurious effects in the interferometer that generate noise nonlinearities in the strain signal channel, and may cause the interferometer to lose lock.

The challenge is that, unlike previous applications of ML/AI in LIGO, here there is no known ground truth (witness channels known to capture features related to the nonlinearities) or well-defined input-output relations in the channel data (due to feedback loops, which can reinject noise into unrelated parts of the system). To address these unique challenges, the project will develop novel unsupervised ML/AI methods to model and analyze the vast amounts of data recorded in the LIGO detectors, towards enhancing the understanding of the emergence of nonlinear noise.

Developing new tools and approaches for identifying instrumental noise promises to significantly improve the sensitivity, data quality, and operational stability of LIGO and future facilities, with the potential to increase detection rates of mergers of the most massive stellar black holes by more than a factor of six. The ML/AI techniques developed will also advance the state-of-the-art in (a) physics-guided AI for anomaly detection in complex systems and (b) AutoML for joint exploration of data and AI model hyperparameters.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.