Demonstrating learned particle decay reconstruction using Graph Neural Networks at BelleII

Sumitted to PubDB: 2020-10-01

Category: Master Thesis, Visibility: Public

Tags: -

Authors Giulio Dujany, Pablo Goldenzweig, James Kahn, Isabelle Ripp-Baudot, Ilias Tsaklidis
Date Jan. 1, 2020
Belle II Number BELLE2-MTHESIS-2020-006
Abstract The clean environment within Belle II, with decay processes originating from non-composite electron-positron pairs, allows for the reconstruction of the entire collision event. Having precise information about the initial state kinematics gives a unique advantage to the Belle II experiment in that allows for direct measurements of decay processes involving neutrinos or few detectable particles in the final state, due to conservation of energy and momentum. This does, however, require a catch-all reconstruction algorithm which is able to determine which particles are not associated to the signal B-meson and reconstruct a second B-meson from them. The current full event reconstruction algorithm at Belle II requires the reconstructed sub-decay processes to be hard-coded, and the careful selection of which kinematic variables to exploit. This both restricts the total branching fraction coverage of the algorithm and relies on intuition to decide which decay processes to reconstruct and how to reconstruct them. This work introduces a method for learning which processes to reconstruct and how to reconstruct them from example using simple kinematic variables with graph neural networks. The efficiency of the proposed method is demonstrated by reconstructing decays from generic and Belle II phasespaces in numerous kinematic scenarios of increased complexity.
Conference Karlsruhe, Strasbourg

Files