Needles in the Quantum Haystack: CMS Anomaly Detection with Normalizing Flows

Abstract

Recent experimental searches for particles beyond the Standard Model (BSM) have yielded little in the realm of new physics discoveries. A number of research efforts have adopted new anomaly detection strategies which utilize density estimation algorithms based on unsupervised and semi-supervised machine learning. However, these efforts rely exclusively on QCD background priors, and thus drastically limit their own anomaly detection capabilities.

In this thesis, we integrate an unsupervised density estimation algorithm, neural spline normalizing flows, into an anomaly detection strategy called Quasi-Anomalous Knowledge (QUAK), which allows us to take advantage of signal priors in addition to QCD background priors. The introduction of a signal prior allows us to learn the features of a particular type of BSM dijet event, giving us insight into the underlying variable distributions of hidden signals in CMS data. Through several studies on both Monte Carlo samples and 13 TeV data from CMS, we demonstrate that QUAK with normalizing flows (QUAK-NF) can be a powerful tool for conducting searches for BSM physics.