Separating black hole signals from glitches in simulated LISA data

Why glitches matter

LISA, planned for launch in the 2030s, will measure gravitational waves at lower frequencies than ground-based observatories. Its targets include merging massive black holes, white dwarf binaries and systems in which a much smaller object spirals into a massive black hole.

But its measurements are also expected to contain instrumental glitches: temporary disturbances that are not astrophysical signals. The LISA Pathfinder technology-demonstration mission observed unexpected disturbances in the motion of its freely falling test masses. Their causes remain mostly unknown, but mathematical models of those disturbances provide examples for simulation studies.

Glitches can obscure gravitational waves or resemble them. The new study focuses on distinguishing modeled glitches from merging massive black hole binaries, including systems around a million times the Sun’s mass whose signals can look similar to instrumental disturbances.

First find a disturbance, then identify it

The proposed pipeline has two stages. First, a search locates unusually strong features in the data. A machine-learning classifier then assesses which kinds of transient are present.

The researchers work with two data channels called A and E. These are combinations of simulated instrument measurements constructed using time-delay interferometry, a technique that suppresses laser noise. The data are divided into six-hour windows, with neighboring windows overlapping by one hour to help retain events near their edges.

Each window becomes a set of time-frequency images called Q-transforms. Like a spectrogram, these show how a signal’s frequency content changes over time. Their time and frequency resolution varies across the image, helping represent signals across LISA’s broad frequency range.

The team uses two Q settings for each channel, producing four image layers. One setting better represents shorter disturbances; the other better represents longer signals. The search identifies unusually energetic image locations and groups their times to locate candidate events.

A convolutional neural network—an image-processing model—then examines the four layers. It produces separate scores for the presence of a gravitational wave and a glitch. This allows four outcomes: a gravitational wave alone, a glitch alone, both, or neither. Identifying both types in an image does not mean the system separates their individual waveforms or removes the glitch.

What the simulations tested

The researchers generated about 6,000 independent, 12-hour simulations, incorporating modeled instrument noise and changing spacecraft configurations. The gravitational-wave examples used massive black hole merger waveforms with spins aligned with the orbital angular momentum. Glitches followed a mathematical shapelet model informed by LISA Pathfinder observations.

Different placements of the analysis window produced 29,055 images. The dataset included noise-only examples so that the classifier could learn to distinguish background fluctuations from transients.

Images were divided into training, validation and test sets in proportions of 70%, 20% and 10%. Importantly, different window placements from the same underlying simulation stayed in the same set. This prevented the model from being tested on repositioned versions of signals it had already seen during training.

The authors report classification accuracy of roughly 98%. But this result has a crucial qualification: transient-containing examples were selected so that the search could recover the injected events. For examples containing both a gravitational wave and a glitch, both had to be recovered. The classification result therefore does not measure the fraction of all injected signals successfully found and identified from start to finish.

Speed, with a defined starting point

On a MacBook with an Apple M4 chip and 16 GB of unified memory, generating one four-layer Q-transform grid took about 0.8 seconds. Classifying one image took about 0.01 seconds using the machine’s GPU through Apple’s Metal Performance Shaders backend.

From these component timings, the authors estimate that processing a month of A and E channel data would take approximately two minutes. This estimate excludes simulation generation and classifier training, and starts from the prepared data channels rather than raw spacecraft measurements.

Where the approach falls short

The search was inefficient at recovering both transients when they were separated by less than about half an hour. It could also miss one event when another was much stronger. These difficult cases were excluded from the classifier dataset when both events were not recovered.

A separate test found fairly stable classifier outputs when correctly identified single transients were moved across the image, with performance declining near the edges. Because that test began with correctly classified examples, it is not a robustness measurement across every test case.

Overlapping windows can detect the same event more than once, and the authors identify a need to group duplicate detections. The classifier also recognizes only the modeled glitches and massive black hole merger signals used in this study. It was not trained for other expected LISA source populations or for inspiral signals whose merger falls outside the analysis window.

The work supports the feasibility of rapid classification within these simulations. Faster alerts and assistance with analyses that fit many sources together remain proposed applications, not demonstrated operational capabilities.

Source and status

This article is based on the supplied full paper text, not only its abstract: Rao and colleagues, A pipeline for identifying astrophysical signals and instrumental glitches in LISA data, a preprint posted on November 6, 2026. All reported performance concerns simulated LISA data.