Neutral B mesons Flavour Tagging at LHCb

In the search for new physics, the study of CP violation in B mesons systems, such as B0s − anti-B0s mixing, are of primary interest. However, the study of these oscillations requires to effectively determine the B meson flavour at production-a process known as flavour tagging (FT). Developing efficient FT algorithms has always been a challenge. In recent years, new methods using deep learning algorithms have emerged. In this work, three deep learning models (ParticleNet, ParticleNeXt and Particle Transformer) are trained and their performance, evaluated using the area under the curve (AUC) of the receiver operating characteristic (ROC) diagram and the tagging power efficiency, is assessed using a large dataset of Monte Carlo simulations from the LHC Run 2. The performance of these models is compared to the state-of-the-art in the field, the inclusive flavour tagging method. Across all three models, outstanding results were obtained, achieving a relative improvement of up to 2.5% of the AUC compared to the best existing methods, demonstrating the significant potential of this novel approach. Finally, unexpected differences in performance between the B 0 s decay channels are explored, as well as some avenues for improvement in tagging power efficiency. This leads to absolute additional improvements of up to 0.7% in tagging power efficiency and a relative additional improvement of 1.7% of the AUC, suggesting promising directions for future research.