A machine-learning model for gravitational-wave analysis produced very small amplitude and phase reconstruction errors, but simulated recovery tests still showed systematic shifts in some inferred black-hole parameters. The result points to a gap between making a waveform look close and using it to estimate the source that produced the signal.
Building a learned waveform generator
The paper describes a deterministic, two-stage conditional autoencoder for four-parameter SEOBNRv4 waveforms. One stage generates amplitude and phase; the second calibrates the remaining residual error. In plain terms, the system is a learned waveform generator whose output is adjusted after the first pass.
The four inputs are the two component masses and two aligned spins. The target family covers component masses from 5 to 75 solar masses, a mass ratio below 10, and aligned spins from −0.99 to 0.99. The waveforms are 1 second long and sampled at 8,192 hertz.
To condition the signals for parameter estimation, the researchers tapered each one-second waveform at both ends and embedded it in an eight-second array padded with zeros before taking the Fourier transform.
Tiny component errors, larger signal-level differences
On the amplitude-and-phase representation, the median cosine distance—a measure of separation between two representations—was 3.22 × 10−5 for amplitude and 2.25 × 10−6 for phase. The corresponding mean values were 1.62 × 10−4 and 4.42 × 10−6.
Once converted into the two physical polarizations, h+ and h×, the median mismatch was 1.33 × 10−2 for h+ and 1.86 × 10−2 for h×. Mean mismatches were 1.90 × 10−2 and 1.86 × 10−2, respectively.
The conditioned outputs showed comparable mismatch distributions. For h+, the mode, mean and median were 8.94 × 10−3, 1.63 × 10−2 and 1.17 × 10−2; for h×, the corresponding values were 1.05 × 10−2, 1.63 × 10−2 and 1.15 × 10−2.
The harder test exposed the bias
For the parameter-estimation tests, component-mass priors were uniform from 30 to 75, and spin priors from −0.80 to 0.80. The simulations used Gaussian noise.
First, in an ML2ML check, 1,045 conditioned-waveform injections were analyzed with the same machine-learning setup. The PP plot, a calibration graph comparing recovered distributions with known injected values, was diagonal and indicated no systematic error in that setup.
The result changed when the injections came from effective-one-body, or EOB, waveforms and recovery used the ML model. In the 400-Mpc experiment, which included 953 injections, recovered values showed systematic deviations from the injected values, and PP plots were non-diagonal for some injections. The pattern was consistent with a mismatch between the EOB and ML waveform families.
A second EOB2ML experiment used 511 injections at a luminosity distance of 4,000 Mpc. In that lower-SNR setting, the authors reported some improvement in the inherent systematic bias.
Corrections have limits
The study also tested importance reweighting, a way of adjusting posterior samples with likelihood ratios. It was useful only where the ML and EOB posteriors overlapped sufficiently; in the reported applications, that overlap occurred only at low SNR.
The bias showed coherent structure, so a linear-fit, post-hoc correction was applied to one-dimensional marginalized posteriors—effectively, one parameter at a time. The correction partially addressed the observed bias, but the authors say it is not a replacement for improving the waveform model.
The evidence is confined to the stated SEOBNRv4 waveform domain and injection experiments. It does not establish performance on observed detector data or on waveform families outside that modeled domain, and it does not show reliable importance reweighting at high SNR.
Paper data and sources
Original title: Gravitational-wave parameter estimation with machine-learning generated surrogate waveforms
Authors: Suyog Garg, Kipp Cannon
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text