Modern deep learning for large-$R$ jet tagging — algorithms, calibration methods, and applications in the CMS experiment

Modern deep learning technologies have demonstrated remarkable potential in analyzing complex LHC data, accelerating foundational scientific discoveries at the LHC. This doctoral dissertation focuses on modern deep learning techniques for tagging large-$R$ jets, one of the most challenging physics objects to analyze at the LHC. It discusses novel methods, applications, and potentials in jet research. Surrounding this technology, the dissertation systematically summarizes the forefront progress in various aspects, involving advanced deep neural network (DNN) algorithms, calibration methods developed in the CMS experiment, breakthroughs in measuring Higgs boson properties, and discusses the new application paradigms of this technology in future experiments. The dissertation is organized into five parts. The first part introduces the fundamentals of experimental particle physics and deep learning. The second part discusses new developments in the field of large-$R$ jet tagging algorithms in the past three years, with a special focus on the algorithm designs that respect Lorentz symmetry and DNN algorithms based on the Transformer architecture. The third part introduces a new method for calibrating advanced DNN-based $X\rightarrow b\overline{b}$ or $c\overline{c}$ jet tagging algorithms in the CMS experiment, namely the sfBDT calibration method: this method uses multivariate analysis to select a phase space closer to $X\rightarrow b\overline{b}$ or $c\overline{c}$ jets in gluon-splitting $b\overline{b}$ or $c\overline{c}$ jets as a proxy for calibration, tacking the problem of previous methods failing to calibrate advanced DNN-based taggers. The fourth part presents significant physics results obtained using cutting-edge DNN algorithms in the CMS experiment in measuring the Higgs-charm Yukawa coupling strengths: this work aims to explore the phase space of highly Lorentz-boosted Higgs boson decays into a charm quark-antiquark pair ($c\overline{c}$) reconstructed by a large-$R$ jet, using the advanced ParticleNet algorithm for $H\rightarrow c\overline{c}$ jet tagging and mass regression, resulting in the best direct measurement of $\kappa_c$ and marking the first observation of the $Z\rightarrow c\overline{c}$ process at a hadron collider. Finally, the fifth part introduces a novel application paradigm for large-$R$ jet tagging by leveraging advancements in deep learning. It emphasizes the development of pre-trained, comprehensive jet models and the adoption of the "pre-training and fine-tuning" paradigm to extend their applicability to diverse downstream tasks. Key contributions include the Global Particle Transformer (GloParT) model, developed within the CMS collaboration, and the Sophon model, trained on the newly developed JetClass-II dataset. Extensive experiments demonstrate their significant potential for CMS Run 3 analyses and future research. Through a comprehensive discussion of the forefront progress of modern deep learning technology applied to large-$R$ jet tagging, the dissertation reveals their transformative impact on particle physics, particularly in the CMS experiment. It highlights the growing role of deep learning in contemporary scientific research patterns.