Abstract
The first gravitational wave was observed in 2015 from two black holes colliding, and since then, we have observed nearly a hundred gravitational waves from a variety of compact sources. How- ever, searching for gravitational waves is a challenge. Matched filtering is the optimal method for extracting a known, weak gravitational wave signal from noisy interferometric data. Template banks are used in matched-filter searches to identify the template that best represents the observed gravitational wave. In this thesis, we use machine learning methods to accelerate gravitational-wave data analysis pipelines that use matched-filtering methods to search for gravitational waves.We present a new machine learning pipeline, LearningMatch and TemplateGeNN, that can be used to accelerate template bank generation. We introduce LearningMatch, a Siamese neural network that has learned the mapping between the parameters, specifically 𝜆0 (which is proportional to the chirp mass), 𝜂 (symmetric mass ratio), and equal aligned spin (𝜒1 = 𝜒2), of two gravitational-wave templates and the match. The match is defined as the similarity between two waveform templates, and a trained LearningMatch model can predict the match to within 3.3% of the true value. For match values greater than 0.95, a trained LearningMatch model can predict the match to within 1% of the true value. We also show that with LearningMatch, it takes 20 𝜇s (mean maximum value) to predict the match with Graphical Processing Units (GPUs) and, therefore, is 3 orders of magnitude faster than current computations of the match. The match is a fundamental calculation, so LearningMatch has the potential to accelerate various computationally expensive gravitational-wave data-analysis pipelines, such as template bank generation.
We also introduce TemplateGeNN, a stochastic template bank generation algorithm that uses a trained LearningMatch model and GPUs to speed up template bank generation. TemplateGeNN can successfully generate a binary black hole template bank (chirp mass varied from 5𝑀⊙ ≤ M𝑐 ≤ 20𝑀⊙, symmetric mass ratio varied from 0.1 ≤ 𝜂 ≤ 0.24999, and equal aligned spin varied from −0.99 ≤ 𝜒1,2 ≤ 0.99) of 31,640 templates in ∼ 1 day on a single A100 GPU. We tested the sensitivity of this template bank by injecting 7746 binary black hole templates into LIGO Gaussian noise. The template bank generated by TemplateGeNN was able to recover 98% of the injections with a fitting factor greater than 0.97. For lower mass regions (black hole mass region between 5𝑀⊙ ≤ 𝑚1,2 ≤ 25𝑀⊙), 99% of 9469 injections were recovered with a fitting factor greater than 0.97. LearningMatch and TemplateGeNN together are a machine-learning pipeline that can be used to accelerate template bank generation for future gravitational-wave data analysis.
We also introduce the first gravitational-wave matched-filter search that does not use template banks, also called GWtuna. GWtuna is a fast gravitational-wave, low-latency search prototype built on Optuna (optimisation software library) and JAX (accelerator-orientated array computation library). Using Optuna, we introduce black-box optimization algorithms, specifically Tree-structured Parzen Estimator (TPE) and evolutionary strategy algorithms, specifically Covariance Matrix Adaptation Evolution Strategy (CMA-ES), to the gravitational-wave community. We show that binary neutron star mergers in LIGO Gaussian noise can be identified by TPE in 1 second (median value) and in less than 1000 matched-filter evaluations. We then show that CMA-ES can be used to recover the SNR and the template parameters in 9,000 matched filter iterations, taking 48 seconds (median value). A stopping algorithm is also used to curtail the TPE search if the signal-to-noise ratio (SNR) threshold has been reached or the SNR has not im- proved in 500 evaluations. GWtuna showcases alternatives to the standard low-latency template bank search and, therefore, has the potential to revolutionise the future of gravitational-wave data analysis.
| Date of Award | 9 Sept 2025 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Laura Nuttall (Supervisor), Andrew Lundgren (Supervisor) & David Wands (Supervisor) |