Deep learning for jet algorithms

The development and improvement of deep learning techniques over the past decades have created new opportunities for algorithmic methods in high-energy physics. Particularly, deep learning has led to significant advances in the performance achieved of algorithms for the flavour identification of jets, the structures formed by the fragmentation of a quark or a gluon when produced in a collider such as the CERN Large Hadron Collider.

In this doctoral thesis, we focus on deep learning methods to enhance the performance of jet flavour identification algorithms at the CMS experiment. We aim to extend their capabilities by improving model robustness against changes that may be applied to the variables used by the algorithms. Additionally, by extending their initial tasks, we enable new opportunities for future research. First, we explore the Transformer architectures in the context of creating deep neural networks that preserve the structure of jets. We establish two models whose performance and computational cost set a new stateof-the-art in the field. Second, we introduce a data-agnostic training method based on adversarial attacks, improving the model’s robustness against changes in the distribution of input variables. Enhancing robustness is necessary to improve our models’ performance after calibration. Finally, we successfully extend the algorithms’ tasks to identify hadronic taus and to estimate jet energy corrections and resolutions. Additionally, we introduce the identification of strange jets, a first for an experiment at the LHC.

Ultimately, this doctoral work results in the creation of a new class of models with improved architecture, training methods, and an expanded scope of what an artificial neural network may achieve.

The final model produced, named UParT, serves as the state-of-the-art in jet identification for the CMS experiment at the LHC. With the identification of jets originating from strange quarks being a first for the LHC, new analyses targeting final states containing this type of jet can now be pursued once the new model is calibrated.