A novel approach for real-time identification of hadronic final states at the High-Luminosity LHC.

The trigger system of the ATLAS experiment is responsible for the real-time reconstruction of data produced by the collisions produced at the Large Hadron Collider (LHC). Its role is crucial for rejecting background collisions while preserving as much as possible the interesting physical signals. To refine the physics selection process, track reconstruction is performed using data in limited portions of the inner tracking detector. The first step of this task, also known as fast tracking, is performed with a combinatorial approach and its processing time is found to increase significantly as a function of the number of secondary interactions per bunch crossing, or pile-up. The future upgrade to the High-Luminosity LHC (HL-LHC) will introduce a much higher pile-up, hence increasing considerably the computational cost of the current algorithms employed at the trigger level, most notably the already mentioned fast tracking. One possibility is to raise the energy thresholds of the physics objects reconstructed at the trigger level but that comes with an important loss of the physics reach of the experiment. In particular, the Higgs-boson pair production is a yet-to-be-discovered process and is of peculiar importance for studying the Higgs self-coupling parameter and deepening the understanding of the electroweak symmetry breaking mechanism in the Standard Model of particle physics. Studying the Higgs-boson pair production is one of the key aspects of the physics program at the HL-LHC and the decay to four b-quarks is one of the most promising discovery channels. The ability to keep the trigger thresholds as low as possible is of paramount importance for enabling the study of this process at the HL-LHC. The work in this thesis reviews the experimental prospects of the di-Higgs production in the four b-jet final state, and prospects for the application of accelerated tracking for such complex hadronic final states are presented. An innovative approach to complement fast tracking and based on a machine-learning algorithm is developed and characterized in this thesis. The algorithm is designed to be robust against the pile-up. The algorithm is based on a Convolutional Neural Network architecture and is trained and tested using a toy event generator that has been also developed and described in this thesis. Robustness studies are conducted to evaluate the performance of this new algorithm as a function of pile-up and potential detector failures. Future developments aim to explore acceleration techniques to achieve a faster inference and apply the algorithm to real ATLAS events.