Primary Vertex identification using deep learning in ATLAS

The increase in the number of inelastic proton-proton collisions per bunch-crossing in the current and upcoming runs at the Large Hadron Collider (LHC) presents an unprecedented performance challenge for primary vertex reconstruction. New solutions leveraging developments in machine learning can provide accurate, efficient and parallelizable primary vertex reconstruction. This note presents the performance of a deep learning algorithm for primary vertex (PV) identification, PV-Finder, on simulated data from the ATLAS experiment. The PV-Finder algorithm uses a custom kernel to transform tracks into dense, one-dimensional features, and convolutional neural networks are used to find PV locations. This note includes the algorithmic implementation along with a study of its performance, compared to the default ATLAS vertex reconstruction algorithm.