AI Transforms Gravitational Wave Detection: LIGO's Innovative Method

Published Date : 10/09/2025 

Researchers from the UK, US, and Italy have developed an AI-based method to improve the detection of lower-frequency gravitational waves, potentially opening up new frontiers in black hole and neutron star observations. 

A machine learning-based approach that could help astronomers detect lower-frequency gravitational waves has been unveiled by researchers in the UK, US, and Italy. Dubbed deep loop shaping, the system would apply real-time corrections to the mirrors used in gravitational wave interferometers. This would dramatically reduce noise in the system, and could lead to a new wave of discoveries of black hole and neutron star mergers – according to the team.

In 2015, the two LIGO interferometers made the very first observation of a gravitational wave: attributing its origin to a merger of two black holes that were roughly 1.3 billion light-years from Earth. Since then, numerous gravitational waves have been observed with frequencies ranging from 30–2000 Hz. These are believed to be from the mergers of small black holes and neutron stars.

However, the lower reaches of the gravitational wave frequency spectrum (corresponding to much larger black holes) have gone largely unexplored. Being able to detect gravitational waves at 10–30 Hz would allow us to observe the mergers of intermediate-mass black holes at 100–100,000 solar masses. We could also measure the eccentricities of binary black hole orbits. However, these detections are not currently possible because of vibrational noise in the mirrors at the end of each interferometer arm.

“As gravitational waves pass through LIGO’s two 4-km arms, they warp the space between them, changing the distance between the mirrors at either end,” explains Rana Adhikari at Caltech, who is part of the team that has developed the machine-learning technique. “These tiny differences in length need to be measured to an accuracy of 10^-19 m, which is 1/10,000th the size of a proton. [Vibrational] noise has limited LIGO for decades.”

To minimize noise today, these mirrors are suspended by a multi-stage pendulum system to suppress seismic disturbances. The mirrors are also polished and coated to eliminate surface imperfections almost entirely. On top of this, a feedback control system corrects for many of the remaining vibrations and imperfections in the mirrors. Yet for lower-frequency gravitational waves, even this subatomic level of precision and correction is not enough. As a laser beam impacts a mirror, the mirror can absorb minute amounts of energy – creating tiny thermal distortions that complicate mirror alignment. In addition, radiation pressure from the laser, combined with seismic motions that are not fully eliminated by the pendulum system, can introduce unwanted vibrations in the mirror.

The team proposed that this problem could finally be addressed with the help of artificial intelligence (AI). “Deep loop shaping is a new AI method that helps us to design and improve control systems, with less need for deep expertise in control engineering,” describes Jonas Buchli at Google DeepMind, who led the research. “While this is helping us to improve control over high precision devices, it can also be applied to many different control problems.”

The team’s approach is based on deep reinforcement learning, whereby a system tests small adjustments to its controls and adapts its strategy over time through a feedback system of rewards and penalties. With deep loop shaping, the team introduced smarter feedback controls for the pendulum system suspending the interferometer’s mirrors. This system can adapt in real time to keep the mirrors aligned with minimal control noise – counteracting thermal distortions, seismic vibrations, and forces induced by radiation pressure.

“We tested our controllers repeatedly on the LIGO system in Livingston, Louisiana,” Buchli continues. “We found that they worked as well on hardware as in simulation, confirming that our controller keeps the observatory’s system stable over prolonged periods.” Based on these promising results, the team is now hopeful that deep loop shaping could help to boost the cosmological reach of LIGO and other existing detectors, along with future generations of gravitational-wave interferometers.

“We are opening a new frequency band, and we might see a different universe much like the different electromagnetic bands like radio, light, and X-rays tell complementary stories about the universe,” says team member Jan Harms at the Gran Sasso Science Institute in Italy. “We would gain the ability to observe larger black holes, and to provide early warnings for neutron star mergers. This would allow us to tell other astronomers where to point their telescopes before the explosion occurs.”

The research is described in Science. 

Frequently Asked Questions (FAQS): 

Q: What is the primary goal of the deep loop shaping technique?

A: The primary goal of the deep loop shaping technique is to reduce vibrational noise in the mirrors of gravitational wave interferometers, enabling the detection of lower-frequency gravitational waves.

Q: How does deep loop shaping work?

A: Deep loop shaping uses deep reinforcement learning to make real-time adjustments to the control systems of the mirrors, counteracting thermal distortions, seismic vibrations, and radiation pressure.

Q: What are the benefits of detecting lower-frequency gravitational waves?

A: Detecting lower-frequency gravitational waves allows researchers to observe the mergers of intermediate-mass black holes and measure the eccentricities of binary black hole orbits, providing new insights into the universe.

Q: Who are the key researchers involved in this project?

A: Key researchers involved in this project include Rana Adhikari from Caltech, Jonas Buchli from Google DeepMind, and Jan Harms from the Gran Sasso Science Institute in Italy.

Q: Where was the deep loop shaping technique tested?

A: The deep loop shaping technique was tested on the LIGO system in Livingston, Louisiana, and the results were promising, showing that the controllers worked well both in simulation and on hardware. 

More Related Topics :