PINCH: Pipeline-Informed Noise Characterization in LIGO's Third Observing Run

Abstract

We present a method to identify and categorize gravitational wave candidate triggers identified by matched filtering gravitational wave searches (pipelines) caused by transient noise (glitches) in gravitational wave detectors using Support Vector Machine (SVM) classifiers. Our approach involves training SVM models on pipeline triggers which occur outside periods of excess noise to distinguish between triggers caused by random noise and those induced by glitches. This method is applied independently to the triggers produced by the GstLAL search pipeline on data from the LIGO Hanford and Livingston observatories during the second half of the O3 observing run. The trained SVM models assign scores to ambiguous triggers, quantifying their similarity to triggers caused by random fluctuations, with triggers with scores above a defined threshold being classified as glitch-induced. Analysis of these triggers reveals the distinct impact of different glitch classes on the search pipeline, including their distribution in relevant parameter spaces. We use metrics such as the Bhattacharyya coefficient and an over-representation ratio to quantify the consistency and prevalence of glitch impacts over time and across parameter spaces. Our findings indicate that some glitch types consistently produce triggers in specific regions of the parameter space, while others generate triggers that are more widely distributed. We observe that Scattered Light glitches appear differently in the search pipeline before and after a commissioning change, demonstrating how such detector changes appear in the pipeline's response to certain glitch classes. This method provides a framework for understanding and mitigating the influence of non-Gaussian transients on gravitational wave search pipelines, with implications for improving detection sensitivity and better understanding noise populations.