Gravitational waves: Isobank covers the signal space more effectively with 23 per cent fewer models

A normalising flow teaches us how to ‘flatten’ the parameter space of eccentric binaries, enabling us to construct more compact template libraries for the search for gravitational waves.

To detect gravitational waves emitted by binary systems, observatories compare their data to thousands of theoretical signals, called templates. Place them in such a way as to cover the entire space of the possibilities without unnecessarily adding them is a puzzle: on these curved and multidimensional spaces, each distance measurement depends on the position, and the exact solutions become out of reach.

Bringing the problem back to Euclidian distances

Presented on arXiv on October 1, 2026, isobank is a normalizing stream that learns a transformation of quasi-isometric coordinates: in the latent space obtained, a simple Euclidean distance approaches the real physical distance. The template placement is then done without reassessing the metric at each step.

On a five-dimensional space describing eccentric orbit binaries, the tool reaches 97% coverage for a maximum deviation of 0.03, with 23% less template than the mbank reference code, which caps at 94%. The coverage was verified on independently injected signals, with the exact correlation calculation of waveforms. Training takes about two and a half hours and placement less than one hour on 32 hearts; the code is published in open access.