• deep learning
• event simulation
• high energy physics
• normalizing flows

The necessity for computer-based simulations is shared by many fields of physics. Specifically, in High Energy Physics this necessity is of foremost importance, due to the complexity of the experiments and the vast amount of experimental data which are to be compared to theoretical models. However, starting from the physical calculations, the interaction with the detectors and the reconstruction of physical objects has proven to be extremely computationally expensive. As it has already been the case in countless applications, novel Machine Learning (ML) techniques are expected to provide us with the much needed speed and accuracy, an expectation that we thoroughly investigated in the present work, at least for what concerns event simulation.

The primary concern of this Thesis has been trying to build a prototype end-to-end sample analysis generator, named FlashSim, achieving accurate and fast prediction of the final analysis reduced format (1 kB/ev) starting from the generator-level information alone through a ML approach skipping all intermediate steps.

The results are indeed confirming and possibly exceeding our initial expectation: the proposed approach demonstrates a raw generation speed of six orders of magnitude greater than that of the classical approach, as well as the convincing capacity of varying its output according to the physical inputs. Additionally, newly-generated objects such as jets and muons are passed through a real Higgs boson analysis selection step to validate our approach in a real-world scenario, with satisfying results. The current findings have the potential to completely change the approach to simulations at CMS and at the LHC, paving the way for online, on-demand generation of events. All these results point to interesting and rewarding directions for future research at the boundaries of high energy physics and machine learning.