Andrews M. Search for Exotic Higgs Boson Decays to Merged Diphotons: A Novel CMS Analysis Using End-to-End Deep Learning

Doctoral Thesis. — Springer, 2023. — xiv, 188 p. — (Springer Theses). — ISBN 978-3-031-25090-3.Doctoral Thesis accepted by Carnegie Mellon University (Pittsburgh, USA).
Supervisor: Dr. Manfred Paulini.
The theory describing the smallest building blocks of matter and the forces acting between them is called the standard model of particle physics. It is an enormously successful theory describing the interactions of all known elementary particles, the quarks, leptons and gauge bosons acting as force mediators. Developed over the past 60 years, starting with the quark model in the 1960s, the discovery of the charm quark in 1974, the τ lepton seen in experiments from 1974 to 1977, the bottom quark in 1977, the W and Z bosons in 1983, the top quark in 1995, and culminating in the discovery of the Higgs boson in 2012, there is to date no experimental evidence contradicting the predictions of the standard model. Although it is successful in describing all phenomena at the subatomic scales, it is not a complete “theory of everything” that can explain all known observations. For example, no particle exists in the standard model that constitutes a possible candidate for the dark matter that makes up about one quarter of the energy–mass content of the universe. The quest for finding phenomena that are not described by the standard model is one reason why physicists at the CERN Large Hadron Collider (LHC) are searching for yet-unknown particles, which can pave the way to postulate theories beyond the standard model.
The Ph.D. research conducted by Dr. Michael Andrews under my supervision in the Department of Physics at Carnegie Mellon University using proton–proton collision data collected with the Compact Muon Solenoid (CMS) experiment at the LHC is not just another search for phenomena beyond the standard model. What sets the data analysis in Dr. Andrews’ thesis apart from conventional CMS searches is the use of several innovative approaches and “firsts” for CMS.
Together with Dr. Andrews, I learned very quickly that recent ML advances, in particular in the field of computer vision, have led to breakthrough applications of convolutional neural networks to scientific challenges, if the data can be expressed as an image or series of images. In particular, we became interested in exploring whether ML can help to get beyond limitations of traditional analysis techniques. As a first project, Dr. Andrews’ work demonstrated the application of image-based deep learning techniques to separate electron from photon showers in one of the CMS sub-detectors, the electromagnetic calorimeter, a task that is not achievable with conventional approaches. This brought us to establish the concept of an end-to-end event classification that directly leverages low-level detector data as input to classify event signatures such as using images from low-level detector data to go directly to classify an event signature without using data reconstruction.
Fueled by the initial success, Dr. Andrews became quite involved in ML and very quickly an expert in the usage of different deep learning networks and ML techniques. His thesis analysis follows the path of exploring what is the maximuminformation that can be extracted from detector data when modern ML approaches are unleashed. He studied the hypothetical decay of the Higgs boson into a pair of light particles H → AA , each of which may in turn decay into a pair of photons A → γγ . The branching fraction for A → γγ is maximized at light mA masses, but in this regime, each of the A → γγ decays is highly merged, and the diphotons are reconstructed as a single photon shower in the CMS electromagnetic calorimeter consisting of lead-tungstate crystals. Using end-to-end ML techniques, Dr. Andrews was able to develop a mass regression algorithm that maintains sensitivity even in the limit, where the two photons from the A → γγ system deposit their energy in the same calorimeter crystal. On the way to setting the first CMS limit for the theoretically highly interesting mass regime mA < 200 MeV, Dr. Andrews solved several issues with sensitivity toward the mA → 0 mass endpoint that I leave for the interested reader to discover in his thesis entitled “Search for exotic Higgs boson decays to merged photons employing a novel deep learning technique at CMS”.
This well-written and nicely organized Ph.D. thesis contains very accessible introductions for the novice to particle physics but also allows the expert to find useful new information. For example, Chap. 2 is an engaging introduction to the LHC and the CMS detector that should be accessible for a reader less familiar with particle physics, while Chaps. 7 and 8 detail the mass regression method and the data analysis for the experts. There is something for everyone in this thesis.True PDF