US LUA annual meeting

Primary Vertex identification using deep learning in the ATLAS Experiment

Not scheduled

10m

51/1-102 - Kavli Auditorium (SLAC)

Lighting talks Lightning Round Talks

Speaker

Qi Bin Lei (Stanford University)

Description

The reconstruction of primary vertices will become significantly more challenging with the High Luminosity Large Hadron Collider (HL-LHC) upgrade, as the number of simultaneous collisions, or pileup, is expected to reach up to $<\mu >=200$. This high pileup places substantial computational demands on conventional combinatorics-based algorithms, resulting in significantly increased latency and reduced accuracy. Machine learning offers a scalable and efficient alternative to address these challenges. This talk presents PV-Finder, a deep learning-based algorithm that uses reconstructed track parameters to directly predict primary vertex locations. The algorithm transforms track data into dense one-dimensional probability distributions, also referred to as kernel density estimations (KDEs), using a deep neural network. These KDEs serve as inputs to a convolutional neural network (CNN) to predict vertex position. We evaluate the performance of PV-Finder under Run 3 conditions ($<\mu >=60$) and benchmark it against other reconstruction algorithms, including the adaptive multi-vertex finder (AVMF) and earlier PV-Finder iterations. The results demonstrate its potential to improve vertex reconstruction efficiency and accuracy in high-pileup environments.