Machine Learning and Statistical Combination Assisted Searches for Dark Higgs and Dark Photon with ATLAS experiment at LHC

The existence of dark matter is strongly supported by numerous cosmological observations. The Standard Model (SM), as the most complete understanding of elementary particles to date, cannot predict or explain the nature of dark matter. Unveiling the properties and interactions of dark matter remains a central question in modern particle physics. Searching for dark matter at the Large Hadron Collider (LHC), the highest-energy particle collider on Earth, offers a unique opportunity to address this fundamental mystery.

This thesis focuses on two analyses performed using data from the ATLAS experiment at the LHC. The first study searches for the dark Higgs boson, exploring whether dark matter acquires its mass through a Higgs-like mechanism. This search targets low-mass dark Higgs bosons and utilizes the full Run-2 dataset from the ATLAS experiment. A novel machine learning-based boosted jet tagging algorithm was developed and applied, significantly enhancing the sensitivity. Relic density observations were incorporated for the first time to directly constrain the model's parameter space, improving the efficiency of multidimensional parameter scans. The results exclude dark Higgs bosons with masses between 30 and 150 GeV and heavy mediator particles ($Z^{'}$) up to 3.5 TeV at 95\% confidence level. For parameters consistent with relic density observations, stringent exclusions are set, including ($Z^{'}$) masses up to the perturbative limit for a 70 GeV dark Higgs boson and Majorana dark fermions below 700 GeV.

The second focus of this thesis is the search for dark photon, which provide a potential explanation for the extremely weak coupling between dark matter and SM particles. This coupling is hypothesized to occur exclusively via kinetic mixing between dark photon and SM photon. The analysis combines ATLAS searches for massless dark photon with Higgs coupling, establishing the most stringent collider limits. At 95\% confidence level, the branching ratio of SM Higgs bosons decaying into photons and massless dark photons is less than 1.3\%, while massless dark photon with couplings to heavy BSM Higgs bosons and Higgs masses below 1.5 TeV is excluded.

Additionally, this thesis explores a novel technique for model-independent dark matter searches. Using self-supervised machine learning-based anomaly detection, dark matter signals can be identified as anomalous events deviating from SM expectations. Studies on simulated datasets demonstrate that this method leverages jet features to enhance sensitivity to unknown dark matter model. This approach holds significant promise for future dark matter searches in collider experiments.