A Novel Multidimensional Search for Diboson Resonances in the Boosted Dijet Final State and Encoding Jet Substructure with Machine Learning

In this doctoral thesis I will present three different searches for new heavy resonances decaying to pairs of vector bosons in the all-hadronic final state. The analyzed data were collected by the CMS experiment at the LHC during the first three years of data-taking at a collision center-of-massenergy of 13 TeV, corresponding to an integrated luminosity of 2.7 (2015), 35.9 (2016) and 77.3 (2016+2017) fb$^{−1}$, and the searches were the first of their kind to ever be performed at such a high collision energy. The diboson final states under consideration are challenging to resolve due to the bosons being highly energetic (“boosted”), resulting in the two quarks from the decay being collimated and merging into a single jet. This leads to a dijet final state topology where each jet displays some energy substructure. The first search I will present, was one of the two first CMS searches in boosted final states with 13 TeV data to become published, and the first to take advantage of jet substructure at the trigger level. It was a high profile analysis due to a previously observed 3.4 (1.3) σ excess around 2 TeV in the 8 TeV dataset, as analyzed by ATLAS (CMS), and I brought the search to a published result within six months after 13 TeV data taking began (and within ten months after embarking on my PhD). Following this, in my second analysis I optimized, validated and commissioned the novel PUPPI softdrop jet grooming algorithm for vector-boson tagging and, in addition, developed dedicated mass corrections for the softdrop jet mass. The algorithm and corresponding mass corrections are now the default for vector boson tagging in CMS and used by several analyses. It was the first published result taking advantage of PUPPI softdrop jet grooming. The third and final search I will present introduces a novel multidimensional search framework, which can be used to search for resonances peaking anywhere in the 3D spectrum of the dijet and groomed jet mass spectra. Validated through a search with hadronically decaying vector bosons in the final state, which I will present here, the framework can be used to incorporate all resonance searches with hadronically decaying vector boson or hadronically decaying Higgs boson final states, as well as for generic searches for any boosted object peaking in jet mass. In the context of this search, a simultaneous fit to the W(qq) and Z(qq) jet mass peaks from the Standard Model (SM) V($\bar{q}q$)+jets process has been performed, for the first time allowing to constrain the groomed jet mass scale and resolution from this SM process. Finally, I will present a deep neural network for vector-boson tagging, which encodes jet clustering- and substructure-like variables into the neural network layers themselves. This algorithm significantly improves the analysis sensitivity, and can also be used as a stepping stone in the development of a generic anti-QCD tagger. The latter will be of great importance when attempting to use the multidimensional framework for model-independent searches.