Seek and Ye Shall Find: Machine Learning and Searches for New Physics
Loading...
Date
2024-01-09
Authors
Bradshaw, Layne
Journal Title
Journal ISSN
Volume Title
Publisher
University of Oregon
Abstract
The discovery of the Higgs boson confirmed that the Standard Model is the correct description of nature below some high energy scale. However, we know the Standard Model is incomplete and have yet to find significant deviations from it. Without well-motivated directions to guide new physics searches, we need to reconsider where and how we search. We explore this in 3 parts here.
We start by identifying 3- and 4-point on-shell amplitudes involving top quarks that are most susceptible to new physics. Using the Hilbert series as a cross-check, we are able to create an independent set of amplitudes for four-fermion and two-fermion, two-boson interactions. After translating these amplitudes to the lowest-dimension SMEFT-like operator, we use pertubative unitarity to place an upper bound on the coupling, under the assumption that the new physics appears around the TeV scale. With this, we find a number of top quark decay modes that could be probed at the HL-LHC.
Next, we compare the efficacy of a number of methods to decorrelate the output of a machine learned classifier from the invariant jet mass. This decorrelation preserves the background dominated sidebands in the invariant mass distribution as tighter cuts are made on the network’s output. This increases the potential discovery significance of the new physics. We compare 4 techniques which broadly fall into one of 2 categories—data augmentation or training augmentation. We find that the simpler and computationally cheaper data augmentation techniques perform comparably to the training augmentation techniques across a variety of qualitatively different signals.
Finally, we turn to machine learning based anomaly detection, with the aim of explaining the physics learned by an image-based autoencoder. Adapting techniques from the literature, we make use of two strategies to mimic the autoencoder. Despite fundamental differences, we find that both techniques, when compared to the autoencoder, order background events similarly and perform comparably as anomaly detectors across a wide swath of signals. The mimicker networks independently use the same high-level observables, giving us confidence that these features are indeed those learned by the autoencoder.
This dissertation includes previously published co-authored material.
Description
Keywords
High Energy Physics, Machine Learning, Particle Physics