Listening to the echoes of a Hidden Valley at the LHC with the CMS experiment

This thesis presents a systematic exploration of a class of Hidden Valley (HV) models, in which the Standard Model (SM) of particle physics is extended by a dark sector that behaves similarly to Quantum Chromodynamics (QCD). Such a scenario can produce distinctive collider signatures, known as semi-visible jets (SVJs), which populate regions of the Large Hadron Collider (LHC) phase space that remain largely unexplored by conventional searches for physics beyond the SM (BSM). In such BSM models, dark partons can be produced at particle collid- ers via a mediator. These dark partons shower, and finally hadronize into dark hadrons; a fraction of these decay promptly back into a spray of SM particles, known as jets, while the remainder remain stable and escape detection as dark matter candidates. The resulting jets therefore contain an admixture of visible and invisible final states. The work presented here extends the SVJ research program along three fronts: the development of novel phenomenological models, the implementation of new experimental searches with the CMS detector, and the design and application of advanced Machine Learning (ML) techniques to search for these challenging final states. Building on earlier studies of fully hadronic SVJs—where dark hadrons are allowed to decay only into SM quarks—the work presented in this thesis broadens the landscape of possible signatures by considering scenarios in which these composite dark bound states can also decay into leptons or pho- tons. This signature-driven approach leads to the introduction of new SVJ signatures—SVJl, SVJτ, and SVJγ—each characterised by distinct collider phe- nomenology. Dedicated analysis strategies are proposed, making use of new isolation variables and lepton-based jet-substructure information. Moreover, HV benchmark models are developed for both the LHC and the Future Cir- cular Collider (FCC-ee), establishing a common framework to target Higgs- portal–induced SVJs at current and future experiments. In this thesis, the first collider searches for resonant production of SVJl and SVJτ final states are presented. These analyses use 138 fb−1 of proton–proton collision data collected by the CMS experiment at √s = 13, TeV during Run 2 of the LHC at CERN. Both the searches for SVJl and SVJτ employ an ML-based strategy, utilizing both a physics-informed Graph Neural Network (GNN) jet identification algorithm, known as LundNet, to discriminate between signal and background, and a Deep Neural Network (DNN) algorithm, known as MD-ABCDisCoTEC, to estimate the background in the signal region of the searches. The SVJl search excludes, for the first time, resonantly produced Z′ mediators leading to such final states with masses up to ∼ 4.7 TeV, depending on the signal hypothesis. Similarly, the SVJτ search provides the first constraints on HV models leading to these signatures, excluding mediator masses up to ∼ 4 TeV. The results of both searches exclude a wide range of QCD-like dark sector models with lepton-enriched SVJ signatures. Two key ML methods are presented in this work. While developed in the context of the SVJ searches, such algorithms are broadly applicable to other analyses as well. The LundNet GNN delivers excellent BSM jet-tagging performance with improved interpretability and reduced dependence on Monte Carlo simu- lations; its application to SVJs is the first of its kind, and it can be extended to other searches with jets in their signature. In this work, the applicability and effectiveness of the method are also demonstrated in the context of the FCC-ee, where it can discriminate between Higgs-induced SVJs and the background of SM H → bb ̄. Its use improves the sensitivity to H → qdarkq ̄dark by two to three orders of magnitude, depending on the parameters of the signal model. In addi- tion, the MD-ABCDisCoTEC method is proven essential for a robust data-based background estimation in scenarios where Monte Carlo simulations are not fully reliable and no physical observables can be leveraged for a standard ABCD method. This is achieved through the simultaneous minimization of multiple loss functions using the Modified Differential Multiplier Method (MDMM). This work constitutes the first application of both the MD-ABCDisCoTEC framework and the MDMM optimization strategy in an LHC search. Overall, this thesis demonstrates that signature-driven search strategies sup- ported by modern ML can substantially expand the sensitivity of the LHC and future colliders to new physics. The results map unexplored regions of the Hidden Valley landscape and provide a foundation for future searches at the LHC and next-generation collider facilities.