September 4, 2025 Science
Our novel Deep Loop Shaping method improves control of gravitational wave observatories, helping astronomers better understand the dynamics and formation of the universe.
To help astronomers study the universe’s most powerful processes, our teams have been using AI to stabilize one of the most sensitive observation instruments ever built.
In a paper published today in Science, we introduce Deep Loop Shaping, a novel AI method that will unlock next-generation gravitational-wave science. Deep Loop Shaping reduces noise and improves control in an observatory’s feedback system, helping stabilize components used for measuring gravitational waves — the tiny ripples in the fabric of space and time.
These waves are generated by events like neutron star collisions and black hole mergers. Our method will help astronomers gather data critical to understanding the dynamics and formation of the universe, and better test fundamental theories of physics and cosmology.
We developed Deep Loop Shaping in collaboration with LIGO (Laser Interferometer Gravitational-Wave Observatory) operated by Caltech, and GSSI (Gran Sasso Science Institute), and proved our method at the observatory in Livingston, Louisiana.
LIGO measures the properties and origins of gravitational waves with incredible accuracy. But the slightest vibration can disrupt its measurements, even from waves crashing 100 miles away on the Gulf coast. To function, LIGO relies on thousands of control systems keeping every part in near-perfect alignment, and adapts to environmental disturbances with continuous feedback.
Deep Loop Shaping reduces the noise level in the most unstable and difficult feedback loop at LIGO by 30 to 100 times, improving the stability of its highly-sensitive interferometer mirrors. Applying our method to all of LIGO’s mirror control loops could help astronomers detect and gather data about hundreds of more events per year, in far greater detail.
In the future, Deep Loop Shaping could also be applied to many other engineering problems involving vibration suppression, noise cancellation and highly dynamic or unstable systems important in aerospace, robotics, and structural engineering.
LIGO uses the interference of laser light to measure the properties of gravitational waves. By studying these properties, scientists can figure out what caused them and where they came from. The observatory’s lasers reflect off mirrors positioned 4 kilometers apart, housed in the world’s largest vacuum chambers.
Aerial view of LIGO (Laser Interferometer Gravitational-Wave Observatory) in Livingston, Louisiana, USA. The observatory’s lasers reflect off mirrors positioned 4 kilometers apart. Photo credit of Caltech/MIT/LIGO Lab.
Since first detecting gravitational waves produced by a pair of colliding black holes, in 2015, verifying the predictions of Albert Einstein’s general theory of relativity, LIGO’s measurements have deeply changed our understanding of the universe.
With this observatory, astronomers have detected hundreds of black hole and neutron star collisions, proven the existence of binary black hole systems, seen new black holes formed in neutron star collisions, studied the creation of heavy elements like gold and more.
Astronomers already know a lot about the largest and smallest black holes, but we only have limited data on intermediate-mass black holes — considered the “missing link” to understanding galaxy evolution.
Until now, LIGO has only been capable of observing very few of these systems. To help astronomers capture more detail and data of this phenomena, we worked to improve the most difficult part of the control system and expand how far away we can see these events.
Studying the universe using gravity instead of light, is like listening instead of looking. This work allows us to tune in to the bass.
Rana Adhikari
Professor of Physics at the Caltech, 2025
As gravitational waves pass through LIGO’s two 4 kilometer arms, they warp the space between them, changing the distance between the mirrors at either end. These tiny differences in length are measured using light interference to an accuracy of 10^-19 meters, which is 1/10’000 the size of a proton. With measurements this small, LIGO’s detector mirrors must be kept extremely still, isolated from environmental disturbance.
Closeup photograph of LIGO, which uses strong lasers and mirrors to detect gravitational waves in the universe, generated by events like collisions and mergers of black holes. Photo credit of Caltech/MIT/LIGO Lab.
This requires one system for passive mechanical isolation and another control system for actively suppressing vibrations. Too little control causes the mirrors to swing, making it impossible to measure anything. But too much control actually amplifies vibrations in the system, instead of suppressing them, drowning out the signal in certain frequency ranges.
These vibrations, known as “control noise”, are a critical blocker to improving LIGO’s ability to peer into the universe. Our team designed Deep Loop Shaping to move beyond traditional methods, such as the linear control design methods currently in operation, to remove the controller as a meaningful cause of noise.
Deep Loop Shaping leverages a reinforcement learning method using frequency domain rewards and surpasses state-of-the-art feedback control performance.
In a simulated LIGO environment, we trained a controller that tries to avoid amplifying noise in the observation band used for measuring gravitational waves — the band where we need the mirror to be still to see events like black hole mergers of up to a few hundred solar masses.
Through repeated interaction, guided by frequency domain rewards, the controller learns to suppress the control noise in the observation band. In other words, our controllers learn to stabilize the mirrors without adding harmful control noise, bringing noise levels down by a factor of ten or more, below the amount of vibrations caused by quantum fluctuations in the radiation pressure of light reflecting off the mirrors.
We tested our controllers on the real LIGO system in Livingston, Louisiana, USA — finding that they worked as well on hardware as in simulation.
Our results show that Deep Loop Shaping controls noise up to 30-100 times better than existing controllers, and it eliminated the most unstable and difficult feedback loop as a meaningful source of noise on LIGO for the first time.
Line chart showing the resulting control noise spectrum using our Deep Loop Shaping method. There is an improvement of 30-100 times in the injected control noise levels in the most unstable and difficult feedback control loop.
In repeated experiments, we confirmed that our controller keeps the observatory’s system stable over prolonged periods.
Deep Loop Shaping pushes the boundaries of what’s currently possible in astrophysics by solving a critical blocker to studying gravitational waves.
Applying Deep Loop Shaping to LIGO’s entire mirror control system has the potential to eliminate noise from the control system itself, paving the way for expanding its cosmological reach.
Beyond significantly improving how existing gravitational wave observatories measure further and dimmer sources, we expect our work to influence the design of future observatories, both on Earth and in space — and ultimately help connect missing links throughout the universe for the first time.
Learn more about our work
Acknowledgements
This research was done by Jonas Buchli, Brendan Tracey, Tomislav Andric, Christopher Wipf, Yu Him Justin Chiu, Matthias Lochbrunner, Craig Donner, Rana X Adhikari, Jan Harms, Iain Barr, Roland Hafner, Andrea Huber, Abbas Abdolmaleki, Charlie Beattie, Joseph Betzwieser, Serkan Cabi, Jonas Degrave, Yuzhu Dong, Leslie Fritz, Anchal Gupta, Oliver Groth, Sandy Huang, Tamara Norman, Hannah Openshaw, Jameson Rollins, Greg Thornton, George van den Driessche, Markus Wulfmeier, Pushmeet Kohli, Martin Riedmiller and is a collaboration of LIGO, Caltech, GSSI and GDM.
We’d like to thank the fantastic LIGO instrument team for their tireless work on keeping the observatories up and running and supporting our experiments.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。