Machine Learning Confirms GW231123 is a "Lite" Intermediate Mass Black Hole Merger

TL;DR

The paper presents a machine-learning follow-up of GW231123, a likely lite intermediate-mass black hole merger with total mass in the range $M_{ ext{tot}} \in [190,265]$ $M_{\\odot}$. It introduces a triad ML pipeline—GW‑Whisper for low-latency signal detection and glitch tagging, ArchGEM for scattered-light glitch characterization, and AWaRe for uncertainty-aware waveform reconstruction—to address waveform-model systematics and nearby non-Gaussian noise. The results show the merger segment can be identified with high confidence in both detectors, the scattered-light glitch is quantified with physically interpretable parameters, and the waveform reconstructions align well with template-free methods, extending validity into the IMBH mass range up to $1000$ $M_{\\odot}$ in injections. Validation across simulated injections and glitch scenarios demonstrates strong discrimination, robust reconstruction even in glitch-contaminated data, and generalizability beyond the specific event. The work offers a scalable, real-time compatible framework to rapid-vet IMBH mergers and guide detector noise mitigation.

Abstract

The LIGO-Virgo-KAGRA Collaboration recently reported GW231123, a black hole merger with total mass of around 190-265 solar mass. This event adds to the growing evidence of "lite" intermediate mass black hole (IMBH) discoveries of post-merger black holes >100 solar mass. GW231123 posed several data analysis challenges owing to waveform-model systematics and presence of noise artifacts called glitches. We present the first comprehensive machine learning analysis to further validate this event, strengthen its astrophysical inference, and characterize instrumental noise in its vicinity. Our approach uses a combination of tools tailored for specific analyses: GW-Whisper, an adaptation of OpenAI's audio transformer, ArchGEM, a Gaussian mixture model-based soft clustering and density approximation software and AWaRe, a convolutional autoencoder. We identify the data segment containing the merger with >70% confidence in both detectors and verify its astrophysical origin. We then characterize the scattered light glitch around the event, providing the first physically interpretable parameters for the glitch. We also reconstruct the real waveforms from the data with slightly better agreement to model-agnostic reconstructions than to quasi-circular models, hinting at possible astrophysics beyond current waveform families (such as non-circular orbits or environmental imprints). Finally, by demonstrating high-fidelity waveform reconstructions for simulated mergers with total masses between 100-1000 solar mass, we show that our method can confidently probe the IMBH regime. Our integrated framework offers a powerful complementary tool to traditional pipelines for rapid, robust analysis of massive, glitch-contaminated events.

Figures (8)

• Figure 1: Predicted class label probabilities from 8 seconds around GW231123 using GW-Whisper in Livingston (left) and Hanford (right). The top row shows the whitened strains from the respective detectors with the segment containing the event highlighted green. The bottom row contains the probabilities for each label in each segment. The model predicts "No Glitch" for all segments around the event and "GW" in the segment that contains the event.
• Figure 2: The top panel displays the Q-transform visualization of the GW231123 event and scattered light glitches observed at the Livingston Observatory. The middle panel displays the scattered light glitches propagating in an Alignment Sensing and Control (ASC) auxiliary channel. The bottom panel is the ArchGEM output highlighting the software's capability to recognize scattered light morphology and extract them for further analysis using a dual methodology ML approach.
• Figure 3: Top: Reconstructions of (a) Livingston and (b) Hanford data of GW231123 using AWaRe. The red dashed curves shows the AWaRe mean reconstructions. The red bands show the associated reconstruction uncertainties. Reconstructions from cWB, Bilby and BayesWave are shown in green, blue and purple respectively. The whitened noisy strains are shown in grey. Bottom: Residuals obtained by subtracting the AWaRe mean reconstructions from the whitened strains.
• Figure 4: (a) Overlap distributions between GW231123‑like waveforms in Hanford and Livingston generated from posterior samples of various approximants and their AWaRe reconstructions. (b) Overlap distributions between AWaRe reconstructions and GW231123-like injections in O3 noise contaminated by different glitches.
• Figure 5: Box plots showing the median and interquartile ranges of the overlap distributions between AWaRe reconstructions and original waveforms for simulated BBH events with total masses between $100 - 1000 M_{\odot}$.