LearningMatch: Siamese Neural Network Learns the Match Manifold

TL;DR

LearningMatch is a Siamese neural network that rapidly predicts the match between gravitational-wave templates with high accuracy, significantly reducing computation time in data analysis pipelines.

Contribution

It introduces LearningMatch, a neural network model that accurately and efficiently predicts gravitational-wave template matches, outperforming traditional calculation methods.

Findings

Predicts match within 3.3% accuracy overall.

Achieves 1% accuracy for matches above 0.95.

Operates at 20 microseconds per prediction using GPUs.

Abstract

The match, which is defined as the the similarity between two waveform templates, is a fundamental calculation in computationally expensive gravitational-wave data-analysis pipelines, such as template bank generation. In this paper we introduce LearningMatch, a Siamese neural network that has learned the mapping between the parameters, specifically $\lambda_{0}$ (which is proportional to the chirp mass), $\eta$ (symmetric mass ratio), and equal aligned spin ($\chi_{1}$ = $\chi_{2}$), of two gravitational-wave templates and the match. The trained Siamese neural network, called LearningMatch, can predict the match to within $3.3\%$ of the actual match value. For match values greater than 0.95, a trained LearningMatch model can predict the match to within $1\%$ of the actual match value. LearningMatch can predict the match in 20 $\mu$s (mean maximum value) with Graphical Processing Units (GPUs). LearningMatch is 3 orders of magnitudes faster at determining the match than current standard mathematical calculations that involve the template being generated.