Developing a Convolutional Neural Network for Object Pileup Energy Prediction and Jet Correction in Anomalous Events in the CMS Level-1 Trigger System

The Compact Muon Solenoid (CMS) at the Large Hadron Collider in CERN is designed to detect and identify the particles produced in high-energy proton-proton collisions in order to explore Standard Model (SM) behavior at high energies and Beyond Standard Model theories. The first of CMS’s two-tier trigger system, the Level-1 Trigger (L1T), comprises custom hardware processors to select events and particle signals while suppressing background noise. Recent advances have been made in creating trigger algorithms through using unsupervised machine learning techniques such as autoencoders and implementing them to L1T FPGAs. Princeton’s Calorimeter Image Convolutional Anomaly Detection Algorithm (CICADA) is a calorimeter-based event-level Level-1 Trigger autoencoder designed to select anomalous events and assign anomaly scores ranging from 1-7. CICADA shows a lot of susceptibility to pileup, i.e. unwanted extra collisions that overlap in the detector, in its scoring, causing other possible patterns in the distribution of anomaly scores and rates to be obscured from further analysis. Here, we introduce a novel FPGA-deployable Convolutional Neural Network algorithm that aims to predict pileup energies of objects for effective pileup subtraction and for greater precision in jet correction. We utilize the φ-ring pileup subtraction method and analyze Higgs boson to bottom quark decay events, guiding connections to Beyond the Standard Model physics analyses.