Deep learning for new physics search

The ultimate goal of high-energy physics is searching for novel elementary particles and fundamental interactions, which may enable the formulation of a more comprehensive physics Beyond the Standard Model (BSM) able to explain a wider range of phenomena compared to the actual Standard Model (SM) of particle physics. Physics analyses are numerous and complex; so far, no theory tested against the SM has proven to be the right one. There could be various reasons for this: the detector apparatus of particle colliders is not able to reveal the new physics; the trigger system discards the almost totality of interesting, i.e., potentially new physics, events; lastly, the space of plausible alternative hypotheses to the Standard Model is so vast that we may be looking in the wrong directions. Assuming the detector is sufficiently expressive to capture the potential new physics, this thesis provides novel data-driven tools, powered by deep learning, to improve various aspects of particle physics analyses resulting in better sensitivity to the new physics. In particular, the signal sensitivity can be increased either: directly, by leveraging parameter-conditional neural networks that can incorporate domain knowledge by conditioning the architecture on physics parameters, or by anomaly detection implemented with auto-encoders that can identify unexpected, out-of-distribution, and rare events from a known background; indirectly, by improving physics pipelines (e.g., accelerator control, data quality monitoring, particle reconstruction and tracking) with reinforcement learning, replacing sub-optimal hand-defined heuristics; or jointly, since a foundation model can be pre-trained on the entire physics domain, fine-tuning it on multiple downstream tasks, achieving superior performance even with few labeled data. All the methods presented in this thesis are generally applicable. They may help particle physicists conduct better analyses by leveraging the increasing amount of big data that particle accelerators like the Large Hadron Collider can produce when collisions occur.