Application of Machine Learning to Supersymmetric Models at Collider Experiments

In this thesis, we study the application of modern machine learning methods to searches for supersymmetric models of physics beyond the Standard Model. <br> In recent years, resonant anomaly detection methods, such as CATHODE, have gained much attention. Using weakly supervised learning, these methods are built to be signal-model agnostic. The main advantage is that they are not only sensitive to a specific signal model, the analysis is tailored to, but cover a potentially much larger region of the parameter space. These methods are most often demonstrated on signal models that contain purely localized features. <br> However, the well-motivated R-parity conserving minimally supersymmetric Standard Model is often found at the tails of distributions of features such as p_T^miss or H_T. Pair produced gluinos with the decay chain g̃→ qq̅χ₂⁰ (χ₂⁰ → X χ₁⁰) with X either the Z or Higgs boson, light χ₁⁰ and small mass splitting between g̃ and χ₂⁰ will be used to demonstrate CATHODEs sensitivity. We, for the first time, demonstrate that CATHODE is only slightly less sensitive than multiple dedicated searches while covering multiple signal models simultaneously. <br> This method can not uncover all signal models. For example the R-parity violating scalar top quark decay t̃→ t χ₁⁰ (χ₁⁰→ q q q) with weak scale χ₁⁰ and sub-TeV t̃ fails to produce features that CATHODE can reliably be applied to.<br> For this signal model, we build a supervised classifier. We utilize recent innovations in computer vision, such as CoAtNet and MaxViT, that apply the self-attention mechanism to images. We represent calorimeter towers and tracks of jets as 2D images and show that the transformer-based classifiers outperform more classical convolutional neural networks in using the jet substructure to predict whether a given jet is neutralino-initiated or not. We show that replacing a CNN with MaxViT excludes up to 100 GeV of additional scalar top mass at 95% C.L. in a simple mock analysis for 100 GeV neutralinos.