Identifying Glitches in LIGO Gravitational Wave Data with Machine Learning Algorithms*

The Laser Interferometer Gravitational-Wave Observatory (LIGO) is a ground-based interferometer used to detect gravitational waves from some of the most spectacular astrophysical events in the Universe. Gravitational waves from these objects manifest as nearly imperceptible changes in distance (strain), requiring an unparalleled level of sensitivity in order to distinguish such signals from a noisy background. Our world is noisy, and the noises we encounter are everyday occurrences that have little to no effect on our lives, but with LIGO, this noise does matter. LIGO's main strain channel is contaminated with loud and transient noise artifacts, called "glitches," which has required LIGO to use instruments dedicated to tracking possible causes of noise. This information is recorded in auxiliary channels. My overall goal is to develop a machine learning algorithm (MLA) that uses information from the auxiliary channels to perform real-time predictions about the presence of glitches in the strain data. Previous attempts at constructing an MLA had large variability in effectiveness daily. We hypothesized that this was because of the time-changing nature of the detector. Therefore, the activity in the auxiliary channels and potential glitches they might register also changes over time. Since we were using consecutive time training sets, we theorized that the MLA would only learn the specific auxiliary channel indicators present during that period. However, these same channels may not be active later in the observing run, thus causing the MLA to lose any ability to identify glitches in the strain through auxiliary channel information. Here, I will show the results of our newest attempt to improve our ability to predict glitches over an observing run using a hierarchical method for the auxiliary subsystems; or, in other words, our ability to correctly predict a glitch in the strain when we create custom models for each auxiliary subsystem of LIGO and combine the predictions from these MLAs to get one overall result. I will also present our proposition to improve the MLA by creating the hierarchical method using glitch types as the subcategories instead.