Scientists improve gravitational wave identification with machine learning

A study published in Physical Review Letters outlines a new approach for extracting information from binary systems by looking at the entire posterior distribution instead of making decisions based on individual parameters.

Since their detection in 2015, gravitational waves have become a vital tool for astronomers studying the early universe, the limits of general relativity and cosmic events such as compact binary systems.

Binary systems consist of two massive objects, like neutron stars or black holes, spiraling toward each other. As they merge together, they generate ripples in spacetime—gravitational waves—which give us information about both objects.

The problem being addressed by the researchers in the published study concerns the labeling of the two objects in the binary system. According to the convention, the heavier object is labeled as “1” and the other as “2.” The issue here is that this system becomes confusing when dealing with systems where both objects have similar masses, within the margin of error.

While previous approaches have suggested using other properties like spin magnitude, it still creates a problem when the objects have similar spins.

The researchers suggest using a more holistic approach by eliminating the reliance on a single differing parameter. Phys.org spoke to first author Dr. Davide Gerosa from the University of Milano-Bicocca in Italy, who mentioned that understanding black holes has long been his motivation.

“This research challenges a long-standing assumption that underpins all gravitational-wave analyses to date—one that has gone unquestioned for decades. Is the standard approach truly the best choice? More fundamentally, what does it even mean to define the best labels? Machine learning provides a powerful, data-driven solution.”

The research team also included two of his students, Viola De Renzis and Federica Tettoni, a postdoc in his group Costantino Pacilio, a former student Matthew Mould now at MIT, and a long-standing collaborator Alberto Vecchio from the University of Birmingham.

The complete picture

The researchers approached this problem differently by framing it as a constrained clustering problem in machine learning. This is a form of semi-supervised learning algorithm that identifies patterns in data, all the while being constrained to certain conditions.

In this case, the constraint that the researchers imposed is that the two objects from the same gravitational wave event must be assigned to different categories.

The key to this method is not relying on or pre-committing to a specific parameter, like mass, as the differentiator. Instead, they let the data itself reveal the most appropriate way to differentiate the objects.

“The key is the realization that the labeling strategy is a deliberate choice we have to make when looking at gravitational wave data. This is a conceptual issue that should be more thoroughly explored, as all downstream applications are potentially affected,” explained Dr. Gerosa.

Higher precision, more confidence

The researchers applied their machine learning model to synthetic and real gravitational wave data from LIGO, Virgo, and KAGRA detectors.

They found that the precision in spin measurements of black holes improved considerably, by up to 50%, and bimodal distributions in data tend to disappear. Now, scientists could more confidently distinguish if the object in the system was a black hole or a neutron star.

“The paper shows that measurements on the individual spins can improve by up to 50%. This is very significant. Such additional accuracy would require new instruments, while we’re showing that it’s achievable with data analysis,” said Dr. Gerosa.

Having more precise measurements of parameters like black hole spins is crucial to understanding their formation. This methodology could have important implications for black hole spin measurements, which have been historically challenging.

The researchers found that approximately 10% of the posterior samples in gravitational wave data from LIGO and Virgo might be better represented with different labels. While this number may seem small, the difference in interpretation of the events is significant.

For instance, the researchers found that for a gravitational-wave event (GW191103_012549), the conventional approach showed a 13% chance that one black hole in the system was spinning opposite to the direction of the orbital motion.

However, their new method dropped that probability down to 0.1%, implying that it’s almost certain that the black hole was spinning in the same direction as the orbit.

“Our analysis affects all gravitation wave measurements from current and future detectors alike,” Dr. Gerosa pointed out when discussing upcoming observatories like LISA (the Laser Interferometer Space Antenna) and the Einstein Telescope.

This study is a classic demonstration of how revisiting fundamental assumptions in data analysis can yield significant results without requiring new information.

© 2025 Science X Network