Development of jet identification and flavor tagging algorithms for CMS experiment

The identification of hadronically decaying boosted heavy particles is a key ingredient of the physics program of the Compact Muon Solenoid (CMS) experiment at the Large Hadron Collider (LHC). At high transverse momentum, the decay products of heavy resonances such as the Higgs boson, the (W) and (Z) bosons, and the top quark become highly collimated and are reconstructed as a single large-radius (AK8) jet. Their identification relies on advanced machine-learning algorithms capable of exploiting both the global properties of the jet and the detailed information carried by its constituents. This thesis focuses on ParticleNet-MD (PNet-MD), the graph neural network adopted by CMS for AK8 jet tagging. PNet-MD is designed to identify jets originating from hadronic decays of massive resonances while minimizing its dependence on the jet transverse momentum ((p_T)) and soft-drop mass ((m_{SD})), making it suitable for model-independent searches. The work presented in this thesis addresses both the offline and online implementations of the algorithm during Run 3 of the LHC. The first part of the thesis presents the validation of the latest offline PNet-MD model using the first proton--proton collision data collected by CMS during Run 3. The discriminator response was studied in a boosted (Z\rightarrow b\bar{b})-enriched topology, where the agreement between data and the Standard Model prediction was evaluated as a function of the PNet-MD score. A data-driven estimation of the dominant QCD multijet background, combined with a binned likelihood fit of the jet soft-drop mass, demonstrated that the simulation accurately reproduces the observed discriminator response. The progressively enhanced (Z)-boson signal observed at increasing PNet-MD values confirms the capability of the algorithm to efficiently discriminate boosted hadronic resonances from QCD jets. This validation established the reliability of the new Run 3 PNet-MD model for CMS physics analyses. The second part of the thesis is dedicated to the development of a new online PNet-MD model for the CMS High-Level Trigger (HLT). To further reduce residual correlations with jet (p_T) and (m_{SD}), a new training strategy based on weighted simulated samples covering a broad kinematic phase space was developed. The resulting model achieves a significantly lower QCD misidentification rate at fixed Higgs-boson tagging efficiency while maintaining a substantially flatter response across the jet kinematic phase space. Its performance was validated on 2025 CMS data, showing negligible impact on trigger timing and event rate. Based on these results, the new PNet-MD discriminator was adopted in the CMS trigger menu for the 2026 data-taking, with an additional backup selection designed for high-rate operating conditions. The results presented in this thesis contribute both to the validation of the offline jet-tagging strategy employed in Run 3 CMS analyses and to the development of the next-generation trigger-level tagging algorithm. They provide a robust framework for future boosted-object identification at the LHC and establish a training strategy that is well suited for the increasingly demanding conditions of the High-Luminosity LHC.