Jet Reconstruction in the ATLAS Level-1 Calorimeter Trigger with Deep Artificial Neural Networks

In the upcoming years, the planned upgrades of the Large Hadron Collider (LHC) at the European Organization for Nuclear Research CERN will result in never before reached luminosities and very high pile-up conditions and will thus impose new chal- lenges on the experimental setups of the four big detectors located at the LHC. In order to maintain or even improve upon the current performance capabilities under these new challenging operating conditions, and hence guarantee a full exploitation of the rich physics potential provided by the LHC, also major upgrades of the ATLAS sub-detector systems are required. The ATLAS detector observes proton-proton collision events at the LHC bunch cross- ing rate of 40 MHz. In order to reduce the data flow to manageable levels and keep only the small fraction of events containing physics processes that are of interest for further analyses, efficient online data selection of interesting and rare processes out of the large data volume delivered by the LHC is essential for a collider experiment. The ATLAS trigger system is specifically designed for this important task. It consists of a hardware-based Level-1 trigger and a software-based High-Level trigger, reducing the event rate from the LHC bunch crossing rate to the detector read-out rate of about 1 kHz. Among others, the Level-1 trigger system involves the Level-1 Calorimeter Trigger (L1Calo) which is specialized on searches for e.g. high-energy electrons, photons, τ leptons and particle jets. Jets play crucial roles in many analyses like e.g. searches for new physics and are therefore important objects that need to be reconstructed in the trigger system. An efficient jet finding algorithm implemented on FPGAs in L1Calo is indispensable to cope with the upcoming very busy environments. By considering calorimeter information as images, state-of-the-art image recognition techniques like convolutional neural networks become promising candidates for this challenging task. In the last decade, the development of deep learning and image recognition techniques has made an enormous progress, often leading to novel algo- rithms that are able to outperform previous state-of-the-art techniques. Moreover, the Universal Approximation Theorem suggests that neural networks can be regarded as universal approximators for a variety of different problems. This motivates the consideration of utilising a deep artificial neural network for jet reconstruction in the ATLAS Level-1 Calorimeter Trigger. The development of a neural network that is capable of finding jets with an efficiency of the anti-kt jet clustering algorithm while simultaneously being small enough to al- low an implementation on the Level-1 trigger hardware is the goal of this thesis and will be presented in the following.