Full Event Interpretation using Graph Neural Networks

The expected large dataset of the Belle II experiment will enable precise measurements of rare decays to probe the Standard Model and searches for new physics. B-tagging is essential for many key measurements, as information of the tag-side B-meson of the \(\Upsilon(4\mathrm{S}) \rightarrow \mathrm{B} \bar{\mathrm{B}}\)  event enables one to constrain rare signal decays which contain invisible particles, such as neutrinos, in the final state. The large number of possible decay channels, and hence large combinatorial space, make this a challenging task. This makes an analytical solution intractable and requires the use of multivariate methods.
A new, end-to-end trainable, graph neural network based approach was proposed in previous work, where the entire decay tree structure is encoded into a single matrix of the final state particles. This thesis expands on the previous work and explores if this representation can be applied to Belle II reconstructed particles. To conclude this thesis, the revised method is applied to simulated  collision data of the Belle II experiment to evaluate the efficiency of this approach compared to the existing reconstruction algorithm, the Full Event Interpretation.