UMN and MIT launch first real-time machine learning search for colliding black holes

A joint effort between the University of Minnesota and the Massachusetts Institute of Technology has achieved the first end-to-end, real-time search for binary black hole (BBH) mergers based entirely on machine learning. The new algorithm was officially deployed into the LIGO-Virgo-KAGRA Collaboration’s production system on August 28, 2025, joining a number of other traditional BBH searches for the remainder of the current data-collecting period.

This is the first time machine learning has been used in all stages of the gravitational-wave detection process, from identifying potential signals to characterizing the properties of the systems producing them. The accomplishment opens the door to faster, more efficient discoveries of colliding black holes and, in the future, neutron stars.

The gravitational-wave detection is performed by a neural network called Aframe. Once a signal is identified, a second machine learning algorithm, AMPLFI (Accelerated Multimessenger Parameter estimation using Likelihood-Free Inference), rapidly estimates the physical parameters of the source, such as the black holes’ masses and distance. Both methods were developed and tested extensively on real and simulated LIGO data, with peer-reviewed results led by graduate and undergraduate students at the two universities.

These algorithms are notable for their speed, small computational footprint, and in certain cases, improved sensitivity compared to existing approaches. The speed allows astronomers to receive alerts within seconds of a detection, enabling rapid follow-up observations with telescopes and neutrino detectors for possible multi-messenger detections.

Aframe's detection of GW231206, which occurred during O4a. The network detects this event with a false alarm rate of less than 1 per 100 years and accurately reconstructs the merger time.

“It’s been exciting to see the alerts come in,” said graduate student Will Benoit. “We’re consistently detecting these gravitational waves before anyone else, so our detection goes out as the preliminary alert for the entire community.”

This project has also been an opportunity for a number of undergraduate students to be introduced to gravitational-wave research. Former UMN astrophysics undergrads Lauren Wills, Katrine Kompanets, and Emma De Bruin completed their senior theses on Aframe, and Steven Henderson and Seiya Tsukamoto are currently working towards their theses as well. As early-career researchers, the contributions they’ve made have helped progress software development to where it is today.

With Aframe and AMPLFI now fully operational within the LVK ecosystem, UMN researchers, from undergraduates to senior faculty, are helping advance the state of gravitational-wave astronomy.