Optimization of a Transformer-Based Model for Flavour-Tagging in the ATLAS Experiment

The Standard Model is the most widely accepted theory in the field of particle physics. In order to test and find some of the missing pieces of the theory, several experiments are being made around the world. One of the main drivers of experimental particle physics research are collider experiments, where particles are accelerated and carefully made to collide at specific regions with detectors ready to collect the products of the interaction.

This data is later treated and the events reconstructed so that it is possible to learn about the phenomena occurring near the collision point. This complex task has several stages, one of which is called jet flavour-tagging, that aims to identify the flavour of the particles that originated jets after the collision. Machine learning models are employed with this task due to its complexity derived from the large amounts of data collected from each collision.


One such model, created by the ATLAS Collaboration, called GN2, is studied in this work with a focus on the input variables it receives. In this thesis, different versions of the model were created with different input variables to test their importance to the model’s performance. Additionally, a study was conducted to directly test the impact of these input variables on the model’s decision-making, and the results suggest that a group of these variables related to the number of particle interactions with the detector may have a limited effect on the model.