Multivariate Deep Learning Techniques for Optimizing Four-Top-Quark Signal Classification at CMS

This Bachelor’s thesis presents a comprehensive multivariate study of the rare Standard Model processpp → \(t\bar{t}t\bar{t}W^-\) in the fully hadronic decay channel at $\sqrt{s} = 13$ TeV using CMS‐style simulations. After simulating signal and dominant backgrounds (including $t\bar{t}, t\bar{t}W, t\bar{t}Z$ and triple‐top processes) and reweighting events to $350 fb^{-1}$, key kinematic and topological features (jet multiplicities, $H_T$, leading‐jet $p_T$, $\sum p_T(b)$, $missE_T^\text{miss}$, angular separations) are ranked via Boosted Decision Trees. Building on this baseline, the thesis develops and benchmarks a suite of deep‐learning architectures including fully connected DNNs, RNNs and transformer networks. Before introducing three novel, physics‐informed models: DeepPhysAtt: a transformer with custom loss terms embedding momentum and jet‐count constraints. CQP-GNN: a hybrid graph neural network combining convolutional, attention, and isomorphism layers to capture event topology. Multi-Scale Residual CNN: a convolutional network operating on grayscale “hadronic” images encoding Δη, Δϕ information. Throughout, specialized loss functions (Jet Multiplicity Loss and Momentum Loss) enforce physically meaningful outputs. The work demonstrates how embedding domain knowledge within advanced neural architectures can significantly enhance sensitivity to ultra‐rare processes and lays the groundwork for future beyond-Standard-Model searches at the Large Hadron Collider.