Published In
Classical and Quantum Gravity
Document Type
Article
Publication Date
11-14-2026
Subjects
Gravitational-Wave Template Banks -- machine learning
Abstract
Matched filtering provides the foundation for gravitational-wave detection, with waveform similarity traditionally described through the local geometry of the template-space metric. We investigate whether the exact pairwise matched-filter mismatch (PMM) over a finite waveform bank can instead be used to learn a global coordinate system for gravitational-wave templates. As a controlled proof of concept, we consider finite banks of nonspinning IMRPhenomD binary-black-hole waveforms over a fixed component-mass range. We formulate this as a supervised representation-learning problem in which physically computed PMM provides the supervision signal, and train a neural network so that Euclidean distances between latent coordinates approximate the mismatch between waveform pairs. Across waveform banks containing up to 8192 templates, the learned representation preserves both global pairwise mismatch relationships and local matched-filter neighborhoods while consistently outperforming Euclidean retrieval in the physical coordinate systems considered. Comparison with Classical Multidimensional Scaling shows that the learned encoder provides a competitive low-dimensional approximation of the finite-bank mismatch geometry while simultaneously learning an explicit mapping applicable to previously unseen waveforms. For previously unseen query waveforms drawn from the same waveform family and parameter distribution, the learned representation recovers the exhaustive best-matching template for 99.02% of queries using only 10 candidate templates from an 8192-template bank and for all queries using 20 candidates, reducing the number of exact matched-filter evaluations by factors of 819.2 and 409.6, respectively. Within this controlled setting, these results demonstrate that PMM provides an effective supervision signal for learning global geometric representations of waveform similarity and establish a proof of concept for mismatch-supervised organization and candidate retrieval in finite waveform banks.
Rights
Copyright (c) 2026 The Authors Creative Commons License This work is licensed under a Creative Commons Attribution 4.0 International License.
DOI
10.1088/1361-6382/aeb788
Persistent Identifier
https://archives.pdx.edu/ds/psu/45340
Publisher
IOP Publishing
Citation Details
Cain, J. (2026). Learning the geometry of matched-filter mismatch for gravitational-wave template banks with machine learning. Classical and Quantum Gravity, 43(21), 215007.