A variational autoencoder application for real-time anomaly detection at CMS

Despite providing invaluable data in the field of High Energy Physics, towards higher luminosity runs the Large Hadron Collider (LHC) will face challenges in discovering interesting results through conventional methods used in previous run periods. Among the proposed approaches, the one we focus on in this thesis work – in collaboration with CERN teams, involves the use of a joint variational autoencoder (JointVAE) machine learning model, trained on known physics processes to identify anomalous events that correspond to previously unidentified physics signatures. By doing so, this method does not rely on any specific new physics signatures and can detect anomalous events in an unsupervised manner, complementing the traditional LHC search tactics that rely on model-dependent hypothesis testing. The algorithm produces a list of anomalous events, which experimental collaborations will examine and eventually confirm as new physics phenomena. Furthermore, repetitive event topologies in the dataset can inspire new physics model building and experimental searches. Implementing this algorithm in the trigger system of LHC experiments can detect previously unnoticed anomalous events, thus broadening the discovery potential of the LHC. This thesis presents a method for implementing the JointVAE model, for real-time anomaly detection in the Compact Muon Solenoid (CMS) experiment. Among the challenges of implementing machine learning models in fast applications, such as the trigger system of the LHC experiments, low latency and reduced resource consumption are essential. Therefore, the JointVAE model has been studied for its implementation feasibility in Field-Programmable Gate Arrays (FPGAs), utilizing a tool based on High-Level Synthesis (HLS) named HLS4ML. The tool, combined with the quantization of neural networks, will reduce the model size, latency, and energy consumption.