Speaker
Description
At the LHC proton bunches are collided at a rate of 40MHz. The Compact Muon Superconducting Solenoid (CMS) detector’s Level-1 (L1) trigger system is responsible for reducing this data rate to about 100kHz so that approximately 1% of these events can be saved for offline physics analyses. The task is to develop algorithms to determine what data to keep and what to discard. Traditionally, trigger algorithms are physics inspired data selections based on corresponding hard-coded high-level features. These methods require a large amount of data preprocessing and potentially bias us to discard interesting physics. Hence, there is a need for theory independent anomaly detection (AD) algorithms. AD algorithms have the potetial to detect ”unknown, unknowns.” In the context of high energy particle physics, these anomalies could come in the form of undefined intermediate decay states, or even something as simple as a detector flaw. Either way, AD algorithms can increase the physics we collect from our detector. This is critical for the coming era of the high luminosity LHC (HL-LHC), so that we can further probe the standard model (SM) and increase the trigger sensitivity for beyond the standard model (BSM) physics. This study explores two different graph neural networks (GNNs), specifically interaction networks (INs). The IN architecture was adapted to both autoencoder (AE) and variational autoencoder (VAE) models. The models are known as INAE and INVAE models, respectfully. Each model was trained and tested on the respective LHC ADC2021 datasets. The results were compared to the current benchmark DNN (Deep Neural Network) AE and VAE architectures for the task of anomaly detection at the L1 trigger. The goal of the study was to determine if INs can be applied as future AD algorithms at the trigger level. Various performance metrics, such as L1 physics reconstruction tasks for representative SM and BSM signals at several different trigger thresholds were used in the analysis of the IN’s viability for anomaly detection and are further discussed in the results.
