Machine Learning to assess astrophysical origin of gravitational waves triggers

Abstract

In this work, we explore a possible application of a machine learning classifier for candidate events in a template-based search for gravitational-wave (GW) signals from various compact system sources. We analyze data from the O3a and O3b data acquisition campaign, during which the sensitivity of ground-based detectors is limited by real non-Gaussian noise transient. The state-of-the-art searches for such signals tipically rely on the signal-to-noise ratio (SNR) and a chi-square test to assess the consistency of the signal with an inspiral template. In addition, a combination of these and other statistical properties are used to build a 're-weighted SNR' statistics. We evaluate a Random Forest classifiers on a set of double-coincidence events identified using the MBTA pipeline. The new classifier achieves a modest but consistent increase in event detection at low false positive rates relative to the standard search. Using the output statistics from the Random Forest classifier, we compute the probability of astrophysical origin for each event, denoted as $p_\mathrm{astro}$. This is then evaluated for the events listed in existing catalogs, with results consistent with those from the standard search. Finally, we search for new possible candidates using this new statistics, with $p_\mathrm{astro} > 0.5$, obtaining a new subthreshold candidate (IFAR =0.05) event at $gps: 1240423628$ .

Figures (11)

• Figure 1: Tree-based decision algorithm. The features $X_1$ and $X_2$ define a two-dimensional feature space. On the left, the threshold $t_i$ represents the decision rules used to partition this space into rectangular regions $R_i$. On the right, the corresponding decision tree representation is shown, where each internal node represents a binary decision based on a threshold, and each terminal node (leaf) correspond to a region $R_i$ in the feature space. Credits to statisticalLearningBook.
• Figure 2: Statistics obtained for the O3a dataset: (a) amplitude distribution for Noise (red) and Injections (blue), as computed by MBTA; (b) $p_s$ distribution for Noise (red) and Injections (blue), as produced by the classifier.
• Figure 3: Comparison of ROC curves between the Random Forest and MBTA pipelines for the O3 dataset: (a) O3a observing period and (b) O3b observing period.
• Figure 4: Performance (ROC) of the Random Forest classifier trained on O3a and evaluated on O3b.
• Figure 5: Feature importance for the O3a model obtained via feature permutation importance.