Enhancing Transient Gravitational Wave Detection Methods Through the Utilisation of Deep Learning Architectures (TranAD)

Ale Magar, Gopal (2026) Enhancing Transient Gravitational Wave Detection Methods Through the Utilisation of Deep Learning Architectures (TranAD). [Laurea magistrale], Università di Bologna, Corso di Studio in Astrophysics and cosmology [LM-DM270]
Documenti full-text disponibili:
[thumbnail of Thesis] Documento PDF (Thesis)
Disponibile con Licenza: Creative Commons: Attribuzione - Non commerciale - Non opere derivate 4.0 (CC BY-NC-ND 4.0)

Download (8MB)

Abstract

Searches for gravitational waves from compact binary coalescences rely on template-based matched filtering, but many short-duration transients (bursts), such as those from core-collapse supernovae, have poorly known or unknown waveforms. Searches for such signals must make minimal assumptions about their shape, and their sensitivity is limited mainly by non-Gaussian noise transients (glitches) that can mimic astrophysical signals. The most effective defence is coherence: a real gravitational wave appears consistently in all detectors, whereas glitches occur independently in each. As a proof of concept, this thesis investigates a detection pipeline that exploits this coherence and needs no model of the signal waveform. Data from the LIGO Hanford and Livingston detectors are cross-correlated over a range of time lags in short sections, so that a transient common to both detectors appears as a localised feature in a single time series. This stream is analysed with TranAD, a transformer-based anomaly detection model trained only on cross-correlated noise. In white Gaussian noise, sine-Gaussian signals at 900 Hz were classified with an AUC of at least 0.988 down to a per-detector SNR of 8, while at 300 Hz the AUC fell to 0.697 at SNR 10. In simulated Advanced LIGO noise, injections at 200, 500 and 900 Hz with SNR 15 were classified perfectly (AUC = 1), and when a coherent signal and a single-detector glitch with the same parameters were present together, only the coherent signal was detected. On real LIGO data, the pipeline produced localised anomaly responses well above background at the times of GW150914, GW190412 and GW200311_115853. These results show that the pipeline responds primarily to inter-detector coherence and can be applied to real interferometer data. Future work should develop it towards a blind search.

Abstract
Tipologia del documento
Tesi di laurea (Laurea magistrale)
Autore della tesi
Ale Magar, Gopal
Relatore della tesi
Correlatore della tesi
Scuola
Corso di studio
Ordinamento Cds
DM270
Parole chiave
gravitational waves burst searches unmodelled transients anomaly detection TranAD transformer deep learning cross-correlation LIGO glitches
Data di discussione della Tesi
30 Ottobre 2026
URI

Altri metadati

Statistica sui download

Gestione del documento: Visualizza il documento

^