A new, fully machine learning-based search method for binary black holes developed by members of the A3D3 Institute at MIT and the University of Minnesota has been deployed into the LIGO-Virgo-KAGRA’s front end computing for a real-time search of such astrophysical systems starting on, August 28, 2025 at 15:00UTC. LIGO-Virgo-KAGRA are presently observing in their 4th run (O4). This is the first end-to-end algorithm that performs source detection and parameter estimation analysis for gravitational-wave sources together using Machine Learning algorithms in all steps of the analysis, while traditionally the search and parameter estimation have been treated separately. The method complements and overlaps with other search methods used within the LIGO-Virgo-KAGRA’s real-time system that are based on traditional signal processing techniques like matched filtering, maximum likelihood estimation and stochastic samplers.
The detection part of this new machine learning-based method is performed by a Convolutional Neural Network (CNN) architecture we call Aframe [Marx, Benoit et. al. (2025)] that is followed by a Normalizing Flow-based architecture that performs the parameter estimation part. We call the latter algorithm AMPLFI, which stands for Accelerated Multimessenger Parameter estimation using Likelihood-Free Inference [Chatterjee, Marx et. al. (2024)].
Both methods have been extensively studied on real and simulated gravitational-wave data with the results presented in peer-reviewed publications led by junior researchers within the A3D3 Institute:
These new methods are distinguished for their lowest latency in providing event detection and parameter estimation, small computational footprint (all models load and run on a single NVIDIA GPU A30 for production deployment) and, depending on the signal phase-space, improve sensitivity.
Aframe and AMPLFI, together with the rest of the methods employed by the LIGO-Virgo-KAGRA collaboration in the real-time search for gravitational waves are publishing their findings at the Gravitational-wave Event Database https://gracedb.ligo.org/ and upon meeting the necessary criteria as public GCNs (Gamma-ray Coordinates Network) at https://gcn.nasa.gov/notices. These are used by the broader electromagnetic wave and neutrino astronomy community for further multi-messenger observations. For instructions on how to access, read and use these alerts, a User Guide is available here: https://emfollow.docs.ligo.org/userguide/.
Caption: Sky localization by AMPLFI for gravitational wave candidate S250830m identified on August 30th, 2025. Aframe was the first pipeline to report the event 11.6 seconds after the merger time. Sky localization and source parameters were provided by AMPLFI 3 seconds later.
By: Michael Coughlin August 5, 2025
For the 5th year in a row, we hosted the ZTF summer school from July 21-25 at the University of Minnesota, where senior undergraduates, graduate students and postdocs learned about “Data Science in the Rubin Era.”
We welcomed 75 attendees from nearly 50 different institutions to the hybrid workshop. The format of the workshop focused on delivery of engaging Jupyter notebook based talks and tutorials covering a variety of science cases and technical topics.
Highlights included hands-on tutorials covering development of fast transient metrics for Rubin Observatory, demonstration of our new broker BOOM (the Burst and Outburst Observations Monitor) and applications of our multi-modal classification pipeline, AppleCiDEr (Applying Multimodal Learning to Classify Transient Detections Early), making detection and simulating the evolution of light echoes from historical transients.
But it was not all hard work. Participants enjoyed a planetarium show, “Our Expanding Universe” at the University of Minnesota’s Bell Museum on the Saint Paul Campus, which featured video from Rubin Observatory operations as well as first light images. Later in the week, the group had the workshop dinner at the Mall of America.
The school ended with a hackathon on Friday where the students applied what they have learned to a sample survey data set from ZTF and built a multi-modal classifier.
Day 1:
Benny Border (Classification of transients using images): This tutorial introduced the basics of machine learning with tabular data products and image classification.
Shar Daniels / Cristina Andrade (Projecting rates of Rubin identified transients): In this tutorial, we describe the data analysis platform being developed to estimate the rates of a variety of Rubin identified transients, focusing on fast transients such as kilonovae.
Day 2:
Michael Coughlin / Theophile du Laz / Antoine Le Calloch (Developing brokering software for Rubin): In this tutorial, we described ZTF and the alert data products produced that can be used by the broader community to identify astrophysical objects of interest to them. We introduced BOOM (the Burst and Outburst Observations Monitor), a multi-survey broker we are developing, and the notebook demonstrated how to develop filters using BOOM to identify real transients. We also introduced the use of a mobile application to access transient candidates through SkyPortal.
Day 3:
Felipe Fontinele Nunes (Classification of transients using light curves): In this tutorial, we introduce light curve classification, including the use of transformers for accounting for the changing numbers of photometry points in realistic light curves.
Argyro Sasli / Maojie Xu (Classification of transients using spectroscopy): We explored how astronomical spectra can be used to study transient and variable sources such as supernovae, active galactic nuclei, and tidal disruption events. Using real data accessed through the Fritz/SkyPortal API, we demonstrated how to retrieve and visualize spectra from different instruments, apply redshift corrections to shift observations into the rest frame, and identify key spectral features such as hydrogen and calcium lines. We also showed how to train a CNN-based network to classify these transients and provided exercises to reinforce key concepts related to CNNs. The final model was exported to ONNX and used to perform production-oriented classification for ZTF.
Day 4:
Xiaolong Li (Historical Transients through Echoes): In this tutorial, we apply AI techniques to search light echoes, and simulate the evolution of light echoes from historical events for Rubin. We train the vgg16 model as a binary classifier and explore how the model makes predictions by analyzing the feature maps. Next, we apply a region based object detection model, FasterRCNN, to detect objects in sky survey images. In addition, we simulate the evolution of light echoes (LEs) from a historical astrophysical transient (e.g., a supernova). These simulations help us predict how Rubin LSST will detect light echoes for historical transients in our Milky Way and nearby galaxies.
Daniel Warshofsky (Large scale variable star classification): In this tutorial, we describe some of the challenges faced in searching for variable light curves in large datasets. Next we learned how The ZTF Sources Classification Project (SCoPe) handles these issues. Finally we explored how to programmatically search through the SCoPe database to find interesting variable sources.
Day 5:
Felipe Fontinele Nunes (Building a multi-modal transient classifier): This hackathon incorporated elements from the previous tutorials, combining image, photometry, and spectral classification in a fusion model.
By: Katrine Kompanets March 4, 2025
First predicted by Einstein’s theory of General relativity over 100 years ago, gravitational wave detection by the Laser Interferometer Gravitational-Wave Observatory (LIGO) has improved greatly since its first detection in 2015. Now, joined by the Virgo and KAGRA observatories, in the middle of its 4th observing run, LIGO has recorded almost 100 confirmed signals, all of which are ‘modeled’ compact object coalescences (CBCs) from binary black holes and neutron stars. These types of signals are known models since these collisions follow a well-defined waveform template. This template is then used to find the incredibly small disturbance in spacetime captured by the interferometers as the gravitational wave passes through Earth.
Credit: Ingrid Bourgault, Illustration for the Announcement of new Results from the LIGO-Virgo-Kagra Collaboration
Gravitational Wave Anomalous Knowledge (GWAK) is a neural network-based technique designed for the detection of unmodeled gravitational wave signals. Unmodeled signals, like those from core-collapse supernovae, are different from modeled signals because they don’t have a well-defined waveform. Heavy reliance on template waveforms means these signals can’t be detected the traditional way, which uses a matched filtering technique. Building on the method described here, GWAK uses semi-supervised learning to look for unmodeled signals. By training autoencoders on different types of data like detector noise, glitches, and simulated CBC signals, GWAK creates a low-dimensional representation of data allowing the network to easily distinguish signals from noise.
Testing using the LIGO-Virgo-Kagra collaboration’s complete catalogue of data from its third observing run revealed that while GWAK didn’t identify any statistically significant unmodeled signals, it was able to maintain sensitivity in high-glitch periods where many searches avoid analysis due to intense noise. This robustness across a broad range of noise types, as well as detection of confirmed CBC signals, demonstrate GWAK’s potential to broaden the scope of gravitational wave detection. Its ability to analyze traditionally discarded data opens new possibilities for future discovery. Incorporation of heuristic models to down-weight glitch-like events and frequency-domain correlations to reduce false alarms demonstrate its adaptability in addressing real detection challenges.
Overall, this study lays the foundation for GWAK’s potential in upcoming observing runs, offering an exciting opportunity to explore the universe in a novel way.
University of Minnesota hosts ZTF summer school
ZTF Summer School 2024
The Coughlin group at the University of Minnesota hosted the 2024 Edition of the ZTF Summer School, focusing on AI and Machine Learning. A3D3 members from the University of Minnesota, Caltech and MIT, along with external speakers from Space Telescope Science Institute and NASA Goddard, came to attend and give lectures on machine learning topics. There were 40 attendees in person, representing over 30 universities from 10 different countries, as well as more than 50 attendees online. Topics covered included supervised and unsupervised learning, simulation-based inference and anomaly detection, focused on multi-messenger astrophysics data sets.
By Katrine Kompanets May 31, 2024
University of Minnesota students (from left to right): Katrine Kompanets, Lilia Bouayed, and Emma de Bruin
University of Minnesota students Emma de Bruin, Lilia Bouayed, and Katrine Kompanets attended “Coding The Cosmos: Introduction to Gravitational Waves Summer Workshop”, a week-long workshop hosted by the Missouri University of Science and Technology and the NSF. We learned various techniques for data analysis and processing of gravitational waves. On the last day, we applied all the skills we learned that week to find as many simulated signals as we could be buried in 4000 seconds of data. Not only did we discover 3 signals, but we also determined that they all came from black hole collisions and we successfully identified their masses. We also discovered a noise remnant (known as a ‘glitch’) that almost threw us off. We had a lot of fun and are excited to combine our new data analysis skills with AI expertise from A3D3 to search for gravitational waves!
By Angela Tran May 2024
A3D3 recently engaged in one of the largest U.S. national physics conferences—the American Physical Society’s April meeting in Sacramento. A variety of A3D3 students presented their talks on institute-related activities.
A3D3 member Ethan Marx from The Massachusetts Institute of Technology brought his passion for data’s potential to bring new discoveries. His APS talk was about applying machine learning (ML) techniques to search for gravitational waves in LIGO-Virgo-Kagra (LVK) data. He says, “Specifically, we validated our algorithm on archival data from the third LVK observing run. The end goal of this work is to deploy this algorithm in real-time in order to issue alerts for electromagnetic astronomers to follow-up. This work is a direct implementation of the institute’s aim for real time processing of large datasets.” In his research, he enjoys the challenge of applying new cutting edge statistical/ML techniques. After being part of A3D3 for about three years, he says he likes that the Institute emphasizes the similarities between seemingly distinct fields.
A3D3 member Will Benoit from The University of Minnesota contributed his enjoyment of how his field probes the behavior of the most extreme universe matter. His APS talk was about demonstrating the production-readiness of a custom AI algorithm for the real-time detection of gravitational waves. He says, “Gravitational waves are one of the messengers in multi-messenger astronomy, and the ability to detect them more quickly will allow other instruments to follow-up more quickly.” In his two and a half years with A3D3, he has especially appreciated the interdisciplinary collaboration with fellow students working on similar technical problems, and seeing how results from different fields can apply to one another.A3D3 member Jared Burleson from The University of Illinois at Urbana-Champaign works in Experimental High Energy Particle Physics to answer the big questions about the fundamental laws of the universe. His APS talk was about the use of machine learning and artificial intelligence for track reconstruction for upgrades to the Large Hadron Collider, which he says “will see a drastic increase in the amount of real-time data gathered. My work aims to utilize AI for processing large data in real time with a focus on discovery in high energy physics.” After about a year with A3D3, he says, “I really enjoy being able to connect with other people in my field and outside my field who are interested in data-driven computing solutions to problems.”
Links to some of the talks presented at APS by A3D3 members are listed below.
D14.1 Jared Burleson, Track reconstruction for the ATLAS Phase-II High-Level Trigger using Graph Neural Networks on FPGAs with detector segmentation and regional processing.
G13.1 Will Benoit, A machine-learning pipeline for real-time detection of gravitational waves from compact binary coalescences
D03.2 Ethan Marx, A search for binary mergers in archival LIGO data using aframe, a machine learning detection pipeline
DD03 Yuan-Tang Chou, NSF HDR ML Anomaly Detection Challenge
DD03 Haoran Zhao, Graph Neural Network-based Track finding as a Service with ACTS
arXiv: 2309.11537 By: Ryan Raikman1,5, Eric A. Moreno2, Ekaterina Govorkova2, Ethan J Marx1,2, Alec Gunny1,2, William Benoit3, Deep Chatterjee1,2, Rafia Omer3, Muhammed Saleem3, Dylan S Rankin4, Michael W Coughlin3, Philip C Harris2, Erik Katsavounidis1,2
April 19, 2024
Credit: Ingrid Bourgault, Illustration for the Announcement of new Results from the LIGO-Virgo-Kagra Collaboration
The Laser Interferometer Gravitational-Wave Observatory (LIGO) represents a pinnacle in the quest to detect gravitational waves, ripples in the fabric of spacetime predicted by Albert Einstein’s general theory of relativity. Gravitational wave detectors like LIGO are marvels of precision engineering designed to observe the most cataclysmic events in the universe, such as the collisions of black holes and neutron stars. Consisting of two 4-kilometer long interferometers in Hanford, Washington and Livingston, Louisiana, LIGO operates by measuring infinitesimally small disturbances in spacetime caused by passing gravitational-waves. Our goal is to increase the search capabilities of these detectors to never-before-seen phenomena like supernovae explosions, cosmic strings, and gravitational bremsstrahlung, just to name a few.
GWAK: New Methods of Anomaly Evaluation on the Horizon
The detection of gravitational wave (GW) signals is pivotal for unraveling the mysteries of the cosmos, relying heavily on accurately modeled templates of GW emissions. However, the possibility of unmodeled transients poses a significant challenge. This study proposes a novel approach utilizing deep recurrent autoencoders and a semi-supervised strategy called Gravitational-Wave Anomalous Knowledge (GWAK, reads: guac) to broaden the search for potential anomalies beyond conventional templates.
Key Points
GWAK Methodology
The GWAK method introduces alternative signal priors, capturing essential features of new physics signatures, thereby extending sensitivity beyond pre-computed templates.
Despite a potential decrease in accuracy compared to supervised techniques, GWAK offers qualitative advantages by generalizing experimental sensitivity.
Construction of GWAK Space
The authors construct a low-dimensional embedded space using the GWAK method, which delineates distinct physical signatures of signals along each axis.
This embedded space facilitates the identification of binaries, detector glitches, and exploration of hypothesized astrophysical sources emitting GWs in the interferometer frequency band.
Fig. 1: A 3-dimensional GWAK space with different types of signatures lying in different regions of the GWAK space. Different selection regions can be made to further isolate specific signals, such as Selection Region 2 [green] completely isolating Core-Collapse Supernovae [avocado]
Detection of Anomalies
Five unsupervised autoencoders were trained on datasets comprising background noise, glitches, and three simulated signals representing potential new physics signatures.
The five autoencoders create a 5-dimensional embedding for each incoming event. This allows a final classifier to easily select regions in this space corresponding to anomalous GW events.
The GWAK method efficiently detected anomalies in GW datasets, including unmodeled sources like Core-Collapse Supernovae and White Noise Bursts, while distinguishing signal-like anomalies from detector glitches.
The Future of the GWAK Method Promises Enhanced Efficiency
The GWAK method emerges as a potent tool for anomaly detection in gravitational wave datasets, showcasing its potential to augment existing detection systems. By leveraging deep recurrent autoencoders and alternative signal priors, this approach offers a promising avenue for uncovering hidden astrophysical phenomena and refining our understanding of the universe. Overall, this study marks a significant stride towards enhancing the efficacy and sensitivity of GW detection, opening new horizons for gravitational wave astronomy. The GWAK method doesn’t just improve our existing capabilities; it invites us to imagine what lies beyond the current frontiers of our knowledge.
Affiliations
MIT LIGO Laboratory, USA
Massachusetts Institute of Technology, USA
University of Minnesota, USA
University of Pennsylvania, USA
Carnegie Mellon University
By: Deep Chatterjee
December 27, 2023
New Orleans, LA – The 37th Conference on Neural Information Processing Systems was held in New Orleans between Dec 10 – 16, 2023. The Machine Learning and the Physical Sciences Workshop at NeurIPS brought together researchers pursuing applications of Machine Learning techniques in Physics, and developing new techniques based on physical concepts like conversation laws and symmetries. There was a total of 250 accepted papers for the workshop poster session – both in-person and remote. While most papers were related to Physics and Astronomy, there were several papers on applications in medical sciences, material science, and earth sciences. A3D3 had a prominent presence at the workshop, with a contributed talk, 4 papers from A3D3 members, and 2 papers from A3D3 Steering Board committee members’ team.
Elham E Khoda from UW Seattle presented on the implementation of Transformers on FPGA using HLS4ML, for low-latency applications like L1 triggers at the LHC and ATLAS, and searches for anomalous signals in gravitational-wave data.
Elham Khoda presents a contributed talk on the implementation of transformers in HLS4ML
Deep Chatterjee from MIT presented a poster on optimizing likelihood-free inference by marginalizing nuisance parameters using self-supervision. [paper link]
Niharika Sravan from Drexel University presented Pythia – a reinforcement learning model that maximizes the search for kilonovae in the presence of several contaminant objects from the Zwicky Transient Facility. [paper link]
Anni Li from UCSD gave a poster on Induced Generative Adversarial Particle Transformers on the use of induced particle-attention blocks to surpass existing particle simulation generative models. [paper link]
Deep Chatterjee (left) and Alex Gagliano (right) present posters.
There were two posters presented from A3D3 steering board members. Ashley Villar (along with Alex Gagliano) gave a poster on convolutional variational autoencoder to estimate the redshift, stellar mass, and star-formation rates of galaxies from multi-band imaging data. [paper link] Nhan Tran (along with C. Xu) have a poster on a Proximal Policy Optimization (PPO) algorithm to uniform proton beam intensity for the Mu2e experiment at Fermilab. [paper link]
