L1 Calorimeter Anomaly Detection Triggering Tests Using the CICADA Algorithm For Run 3 Of The CERN LHC at CMS

• Published in: IEEE conference record - Nuclear Science Symposium & Medical Imaging Conference. p. 1
• Main Author: Loeliger, A. D.
• Format: Conference Proceeding
• Language: English
• Published: IEEE 04.11.2023
• Subjects: Large Hadron Collider; Neural networks; Poles and towers; Real-time systems; Semiconductor detectors; Sensitivity; Topology
• ISSN: 2577-0829
• DOI: 10.1109/NSSMICRTSD49126.2023.10338024

CICADA (Calorimeter Image Convolutional Anomaly Detection Algorithm) is a fully autonomous AI algorithm that can process LHC event data in real-time and independently pick topologies that are different from the majority of currently recorded collision events from the Large Hadron Collider (LHC). CICADA operates close to the rawest recorded data, i.e. the Compact Muon Soleinoid (CMS) calorimeter energy deposits. In this manner, it has as little human bias as possible. CICADA leverages deep learning-based anomaly detection and methods for efficient neural network deployment in hardware. Validation checks have been performed using past LHC data and correlation checks with traditional trigger algorithms showing a novel and intriguing behavior. Besides this, sensitivity to apparatus failures was investigated. Synthetic samples were produced to mimic common problems, such as noisy channels called hot towers. CICADA did not respond to these with higher anomaly scores. We present here an innovative triggering technique that is being tested within the existing CMS Level-1 Calorimeter Trigger system. The calorimeter pre-processors (CTP7) are equipped with Virtex 7 Xilinx FPGA used as target for these firmware developments. The data is propagated to a copy of the Level-1 Global Trigger of CMS (referred to as the uGT test crate) through 10 Gb/s link, that can be run during collision data taking period in parasitic mode. The goal of the exercise is to experiment these new approaches for the Phase-2 of the LHC already during Run-3. This presentation will detail the complex firmware developments and performance with real data compared to algorithms currently used by CMS.