Search for Long-Lived Particles with HCAL Segmentation in CMS at the Large Hadron Collider

We have searched for Long-Lived Particles (LLPs) with lifetimes greater than 0.1 nanoseconds produced at high energies in the Compact Muon Solenoid (CMS) experiment at the Large Hadron Collider, which are predicted by both the Standard Model (SM) and many Beyond the Standard Model (BSM) theories, including those addressing dark matter, the matter-antimatter asymmetry in the universe, and supersymmetry. The search uses Higgs boson decays via a massive LLP pair to a final state with two bottom quark-antiquark pairs: H —> XX —> bbbb. The analysis exploits the new depth segmentation upgrade of the CMS Hadronic Calorimeter (HCAL) to target the distinct topology of a jet-pair from LLPs that decays within the HCAL. Using Monte Carlo-generated signal samples along with W+ jet background samples, we develop and compare the performances of various neural network (NN)-based classifiers. Among the explored architectures, we produce 3D convolutional neural networks (CNNs) trained on ‘images’ of the jet energy distribution across the HCAL depth-layers and in eta-phi space within each layer, which yield up to 89% signal efficiency at just 0.01% background (false positive) efficiency. These findings have motivated further investigations into model interpretability, through the application of Gradient-weighted Class Activation Mapping (GradCAM) to visualise the importance of topological features and gain deeper insight into the model's learning process.