Extending the physics reach of the CMS experiment with Scouting and Anomaly Detection at the Level-1 Trigger

The LHC delivers proton–proton collisions to the CMS experiment at a rate of 40 MHz, producing a data volume that far exceeds what can be processed and stored. The selection of potentially interesting events is therefore crucial to sustain the broad CMS physics program, from precision measurements of the Standard Model to searches for new physics. CMS employs a two-tier trigger system to reduce the event rate to 100 kHz at the Level-1 Trigger (L1T) and to about 1 kHz at the High-Level Trigger, through selections based on the presence of certain reconstructed particles. However, the absence of clear evidence for new physics after more than a decade of LHC operation raises the question of whether trigger selections introduce bias against rare or unconventional signatures, and whether new physics signals could be hidden among the over 99\% of events rejected by the CMS trigger at each stage. This thesis explores innovative strategies to overcome these limitations and extend the physics reach of the CMS experiment during the 2022–2026 data-taking period and beyond. The first approach consists of a trigger-less data acquisition system, known as Level-1 Data Scouting (L1DS), which records only the particles identified by the L1T at the full bunch-crossing rate of 40 MHz. Without the constraints imposed by trigger selections, L1DS provides access to previously unexplored regions of phase space, together with valuable data for detector and trigger monitoring. This thesis studies the performance of the L1DS demonstrator and its first physics application to searches for low-mass dijet resonances, probing sensitivity to mass ranges previously inaccessible at CMS. In parallel, this thesis investigates a complementary paradigm in which event selection is redefined through a model-independent anomaly detection algorithm called AXOL1TL. Implemented in the L1T hardware, AXOL1TL consists of an autoencoder that performs ultra-low-latency inference to identify events that deviate from the Standard Model background in real time. This thesis presents the architecture of the algorithm and its deployment strategy, and provides a detailed characterization of this novel trigger. This work then explores the potential of AXOL1TL for model-independent physics analyses through a bump hunt search for resonant excesses in di-object invariant mass spectra across multiple final states, with particular emphasis on the dijet channel, exploiting the AXOL1TL dataset enriched with anomalous events.