Application of deep learning techniques in CMS : from matrix element regression to the search for Higgs boson pair production

The discovery of the Higgs boson (H) ten years ago constitutes the pinnacle of the construction of the standard model (SM) describing the fundamental interactions between elementary particles. While a major success, half a century after its prediction, several important questions are left unanswered by the SM. The study of the Higgs boson is one of the most promising path towards a deeper understanding of the spontaneous symmetry breaking. Its self-interaction, through the simultaneous production of a pair of Higgs bosons (HH), has yet to be measured and any deviation from its SM prediction could be the sign of new physics. The search for HH production in the proton collisions of the Large Hadron Collider (LHC) at CERN is made challenging by the rarity of that process. Machine learning techniques such as deep neural networks (DNN) are the tools of choice to leverage the huge amount of data from the collisions, assisted by the intuition of the physicist. This thesis explores the performance and limitations of these tools on the data generated by the CMS experiment. First as a regression of the matrix element method (MEM) weights. While very powerful to extract information using the knowledge of both the SM and the detector, the MEM suffers from a prohibitive computation time. This can be circumvented by a faster DNN regression of the weights, taking advantage of their interpolation capabilities. Second, they served as the central piece in the search for HH production in the bbWW final state, using the CMS Run-2 data at 13 TeV. Resonant hypotheses were tested, as well as an effective field theory (EFT) non-resonant interpretation into various couplings, including the Higgs boson self-interaction strength, and the data was found to be compatible with SM predictions.