CERN Accélérateur de science

Thesis
Report number	CERN-THESIS-? ; CMS-TS-2024-026 ; CERN-THESIS-2024-276
Title	Artificial Intelligence Algorithms for the Real-Time Selection of Physically Important Events Within the CMS L1 Trigger FPGA Devices
Author(s)	Burazin Misura, Arijana (Split U.)
Publication	2024 - 123.
Thesis note	PhD : Split U. : 2024
Thesis supervisor(s)	Music, Josip
Subject category	Detectors and Experimental Techniques
Accelerator/Facility, Experiment	CERN LHC ; CMS
Abstract	The Large Hadron Collider is the largest particle accelerator in the world with the purpose ofcolliding protons accelerated to nearly the speed of light, resulting in the creation of new andexotic particles. Four detectors surround the collision points and record the data resultingfrom the proton-proton collisions. This data is then used for a wide range of physical analyses that intend to improve our understanding of particle physics and predict new theories.To increase its discovery potential, the LHC project aims to significantly boost the luminosity in the High Luminosity-Large Hadron Collider phase, which will, in turn, increasethe number of collision events. The detector technologies are improved to address the challenges posed by the increased luminosity. However, it has become imperative to optimizethe algorithms in the data selection system to effectively manage the presence of multipleinteractions in the same bunch crossing (pile-up). This research focuses on using neural networks to select physically important events in real-time in the Compact Muon Solenoid, oneof the LHC detectors. Simple multilayer perceptron and convolutional neural networks aredeveloped to successfully distinguish electromagnetic showers from different types of backgrounds using raw detector images. Data selection and quantization methods are proposedto reduce the size of the presented models. Finally, profiling tools enable the determinationof the precision settings needed to keep model inference accuracy, which further reducesthe memory and computational requirements. The suggested approach was assessed usingsimulated data since the detector is still in the production phase. The inference time was lessthan 1 microsecond and signal-background classification achieved a classification accuracyof 97.01% for 2-bit-only quantization. This accuracy is the same as for the full-precisionstandard network, with a slightly decreased false negative rate. Overall, the research outputpresented within the thesis confirms the successful implementation of developed models inthe targeted device while meeting the latency requirement, which makes them candidates forone of the level 1 trigger algorithms.Sa??etak . . . . . .Acknowledgments .List of Tables . . .List of Figures . . .List of Acronyms .1