- author = "Ho Fung Tsoi",
- title = "Search for exotic {Higgs} boson decays with {CMS} and fast machine learning solutions for the {LHC}",
- school = "University of Wisconsin-Madison",
- year = "2024",
- address = "USA",
- month = "30 " # jul,
- keywords = "genetic algorithms, genetic programming, CMS",
-
URL = "
https://repository.cern/records/tgfn2-52h79",
-
URL = "
https://repository.cern/records/tgfn2-52h79/files/CERN-THESIS-2024-283.pdf",
- size = "203 pages",
-
abstract = "There exists potential for discoveries beyond the
Standard Model in scalar sector, which could manifest
as Higgs boson exotic decays into new light
pseudoscalars. This search targets such decays,
focusing on pseudoscalar masses ranging from 12 to 60
GeV, in final states where one pseudoscalar decays into
two b quarks and the other into two tau leptons or two
muons. The analysis is based on a dataset of
proton-proton collisions at s = 13 TeV, collected by
the CMS detector during LHC Run 2, with an integrated
luminosity of 138 fb-1. Dedicated neural networks are
used to distinguish between signal and background,
significantly enhancing sensitivity. The results are
presented as exclusion limits at 95 percent confidence
level on the model-independent branching ratio and are
interpreted within two-Higgs doublet models augmented
by a singlet.
The second part of this thesis presents machine learning methods to enhance overall sensitivity in the low-latency domain for the LHC experiments. A novel machine learning-based trigger algorithm is developed, using anomaly detection to search for new physics in a model-agnostic manner as close to the raw collision data as possible. This anomaly detection trigger is sensitive to a wide range of both conventional and unconventional physics signatures and has an inference latency of O(100) nanoseconds on an FPGA. It is deployed during Run 3 in the CMS Level-1 trigger system, which processes the first round of real-time event selection from collision data at a rate of 40 MHz. Additionally, a novel model compression method using symbolic regression is developed to accelerate machine learning inference to nanosecond speeds on FPGAs. This method demonstrates potential to significantly reduce the computational costs of machine learning algorithms while maintaining performance comparable to that of neural networks. These advancements are crucial for meeting the sensitivity and computational demands of resource-constrained environments such as the LHC experiments.",
- notes = "CERN-THESIS-2024-283 30/07/2024",
