Real-Time Autoencoder Anomaly Detection Methods at the Large Hadron Collider

Abstract

The search for BSM physics at the Large Hadron Collider motivates complementary trigger algorithms capable of identifying collision signatures with reduced dependence on theoretical priors. The Calorimeter Image Convolutional Anomaly Detection Algorithm (CICADA) addresses this challenge through an autoencoder-based anomaly detection algorithm deployed at the CMS Level-1 trigger, which assigns each event a reconstruction error score without presupposing the form of new physics. The teacher autoencoder defines the anomaly score, while the deployed L1 implementation uses a compact student model to approximate or implement the trigger-level score. This thesis investigated CICADA’s learned physics representations and proposes improvements to the existing system. First, we conduct a systematic analysis of the CICADA teacher autoencoder’s 80-dimensional latent space. Using methods such as lasso regression, we characterize what physical information the network has autonomously learned to encode. Total calorimeter transverse energy is almost totally encoded, suggesting that the anomaly score is primarily sensitive to energy-scale information, while jet kinematics are essentially absent to first-order in the studied samples. Second, we implement a Normalized Autoencoder (NAE) trained with an energy-based objective and Langevin Monte Carlo negative sampling, designed to suppress outlier reconstruction and heighten anomaly sensitivity. Using agentic autoresearch, we optimize hyperparameters, and produce a model with improvements over a modified CICADA autoencoder, along with a theoretical framework intended to improve off-manifold rejection.