ETD

• flow matching
• HL-LHC
• LHC
• pile-up
• simulazioni ultra veloci
• ultra fast simulation

The Large Hadron Collider (LHC) has recently concluded its Run 3 data-taking campaign, and in the coming years, the accelerator complex will undergo a major upgrade towards the High-Luminosity LHC (HL-LHC) era. The HL-LHC project aims to enhance the LHC's performance to maximize its discovery potential beyond 2030. The primary objective is to increase the integrated luminosity by a factor of 10 relative to the original design value. Since luminosity is proportional to the number of collisions, higher luminosity enables experiments to accumulate much larger datasets. This facilitates the study of known mechanisms, such as the Higgs boson, with unprecedented precision, and the observation of rare phenomena that may emerge at these energy scales.
However, looking ahead to the HL-LHC era, unprecedented computational demands will challenge computer-based simulations in High Energy Physics, driven primarily by pile-up. With an expected average of up to 200 simultaneous interactions per bunch crossing, event reconstruction will become more complex.
Simulating every particle produced in these complex environments using the traditional Geant4-based Full Simulation (FullSim) is computationally unsustainable, typically requiring O(30-60 s) per event in Run $3$, scaling up to several minutes at HL-LHC. Consequently, simulation represents a major bottleneck in the CMS analysis workflow. Alternative simulation frameworks have been developed to mitigate this issue. FastSim, for instance, replaces heavy Geant4 material interactions with simplified analytical parameterizations, reducing the simulation time by an order of magnitude (O(5 s)). However, it still requires the execution of standard Particle Flow reconstruction algorithms to produce the final analysis datasets, leaving a significant computational bottleneck in the workflow.
To address this, the CMS group in Pisa developed FlashSim, an end-to-end Deep Learning-based fast simulation software that bypasses the Geant4 detector simulation and traditional reconstruction. By directly mapping generator-level particles to reconstructed physics objects in the compact NanoAOD format, FlashSim achieves a generation speed orders of magnitude faster than the full simulation chain (O(ms)).
Because FlashSim relies on learning the gen-to-reco mapping, a fundamental challenge arises with fake objects, reconstructed components that do not have a generator-level counterpart from which to start. While this issue affects the reconstruction of all physics objects (including electrons, muons, taus, and photons), this thesis focuses specifically on the generative modelling of fake jets, which originate from pile-up interactions or random calorimeter noise, rather than from a genuine hard-scattered parton, and therefore represent a limitation of the gen-to-reco approach. The primary goal of this thesis is to address this limitation, improving the generative model used to simulate fake jets within the FlashSim framework.
To evaluate the improved generative models, we validated FlashSim within the Run 3 environment, as part of an ongoing analysis of the Vector Boson Fusion (VBF) channel of the Higgs boson decay into two muons (H to mumu).
This specific channel is of profound interest for both Run 3 and the upcoming HL-LHC, as its experimental sensitivity currently lies around the 5-sigma observation threshold.
Despite its rarity, with a branching ratio of only BR(H to mumu) = 2.2 10^{-4}, this channel provides a direct probe of the Higgs interaction with second-generation fermions. However, this same rarity means that an enormous number of simulated events is required to accurately estimate the background in the signal region, and the signal itself is challenging to isolate, dominated by the Drell-Yan background. By rapidly generating massive simulated samples, FlashSim modelled this dominant background, populating the signal-enriched regions of the phase space.
The thesis is composed of 7 chapters organized into three parts. Part I introduces the physical and technical context: Chapter 1 provides an overview of the LHC and the CMS experiment, while Chapter 2 reviews the Higgs mechanism and its main production and decay channels.
Part II focuses on simulation and the machine learning tools that form the basis of this work. After Chapter 3 discusses current approaches to event simulation and their associated computational costs, Chapter 4 offers a comprehensive introduction to neural networks before focusing deeply on Continuous Normalizing Flows and Flow Matching. This chapter details how Flow Matching learns a continuous transformation to capture complex multidimensional phase spaces, framing the training process as a stable regression task.
Part III presents the original contributions of this research. Chapter 5 describes the fake jet simulation module, detailing how the generation is decoupled into three sequential stages: predicting the total multiplicity, generating the joint distribution of the kinematic variables to preserve physical correlations within the event, and reconstructing all remaining jet features. The chapter also outlines how optimization resolved initial mismodelling issues, enabling the network to learn implicit constraints like the angular separation between jets. Furthermore, it explores the physical characterization of fake jets, highlighting their correlation to the total event energy scale. Finally, Chapter 6 demonstrates the application to the VBF H to mumu physics analysis, and Chapter 7 outlines the conclusions and future work.