Advanced AI techniques for jet identification in CMS

by Dr Loukas Gouskos (CERN)


The identification of the flavour of the parton that initiated the spray of stable particles forming a jet is a fundamental component of standard model measurements and searches for new physics with quarks and gluons in the final state. The first jet identification algorithms were developed using classical methods, however since the beginning of LHC Run 2 the development of significantly more advanced methods, making use of machine learning techniques, led to substantial improvement in performance, and hence in the physics reach. Particularly, Graph Neural Networks (GNNs) are emerging as an extremely powerful class of architectures; GNNs provide a much more natural representation for many tasks in High Energy Physics, including the task of jet identification. In today's seminar we will present a new generation of jet identification algorithms developed by CMS, which exploit customised GNNs tailored to the jet identification task, yielding a drastic improvement in performance compared to previous algorithms.


Videoconference via