An autoencoder-based surrogate waveform model for quasi-circular binary-black-hole mergers

TL;DR

A AESur3dq8 is introduced, a novel surrogate waveform model based on autoencoders that enables the rapid and accurate construction of large template banks, producing millions of waveforms in under a second using modest computational resources.

Abstract

The generation of accurate waveforms from binary black hole (BBH) mergers is a major effort in Gravitational-Wave Astronomy. In recent years, machine-learning-based surrogate models for BBH waveforms have been proposed. Those offer the potential to dramatically accelerate waveform generation while maintaining accuracy competitive with that of traditional waveform approximants. In this work, we investigate the viability of autoencoders as generative models for gravitational-wave signals from quasi-circular BBH mergers. We introduce AESur3dq8, a novel surrogate waveform model based on autoencoders that enables the rapid and accurate construction of large template banks, producing millions of waveforms in under a second using modest computational resources. The model is trained on the numerical-relativity-informed surrogate NRHybSur3dq8 and subsequently fine-tuned using the SXS catalog of BBH simulations. We demonstrate that waveforms generated by AESur3dq8 achieve mismatches of order $10^{-4}$ with respect to Numerical Relativity waveforms, and that parameter estimation performed with these templates yields results fully consistent with those reported by the LIGO-Virgo-KAGRA Collaboration for observed gravitational-wave events.

Figures (7)

• Figure 1: Corner plot showing the distribution of the parameters of the NRHybSur3dq8 dataset, in gray, and overlayed on top of it, in red, the respective distribution for the SXS dataset. The contour plots display the 0.5, 1, 1.5 and 2$\sigma$ isocontours, overlayed on top of a scatter plot of the parameters to show the density of points. Note that the points in the regions in-between isocontour levels are not displayed for clarity.
• Figure 2: Schematic representations of the two model configurations used in this work. (a) The autoencoder architecture, consisting of the encoder, decoder, and the latent (code) space. (b) The generative model, obtained by removing the encoder and adding a fully connected network that maps the physical parameters to the latent space.
• Figure 3: Loss functions of our models in all three stages of training. From left to right we show the 50-epoch average of the value of the loss and validation loss for each model (amplitude and phase, depicted in solid and dashed lines, respectively) during the autoencoder training with the NRHybSur3dq8 dataset, the full model training with the same dataset, and the full model training with the SXS dataset (fine-tuning stage). We also overlay the LR for each epoch, showing more clearly that the model benefits from the LR scheduler, since the loss function improves at each step.
• Figure 4: Mismatch distributions for the model-generated waveforms. Left: distribution for the autoencoder reconstruction of the NRHybSur3dq8 test set overlaid with mismatches from waveforms generated by the full model trained on the same dataset. Right: distribution between waveforms generated by the full model and the SXS test set, shown before and after fine-tuning.
• Figure 5: Scatter plot of the distribution of the physical parameters of the NRHybSur3dq8 test dataset, with the spins combined into the effective spin $\chi_{\rm eff}=\frac{\chi_1+q\chi_2}{1+q}$ and coloured with the mismatch between the predicted waveform and the real one corresponding to each parameter combination. The top panel displays the autoencoder’s reconstruction mismatches at the conclusion of the first training stage, whereas the bottom panel shows the mismatches for the full model at the end of the second stage.