Pileup Mitigation at the High-Luminosity LHC using Attention Neural Networks

As mean number of proton-proton collisions per bunch crossing, mu, increases up to an average value of 200 for the High-Luminosity LHC upgrade, the amount of pileup events on top of hard scatter events will increase dramatically. Identification of pileup products is a crucial task for all High Energy Physics analyses, and high-pileup conditions will degrade sensitivity to hard scatter processes significantly. Traditional pileup mitigation techniques rely on algorithms and neural networks to classify individual jets as hard scatter or pileup. However, pileup mitigation can be improved using attention neural networks by capturing correlations between jets within an entire event. Additionally, a stack of transformer encoders is used to perform regression to predict the hard scatter energy and mass fractions which can directly be applied to each jet to correct for pileup effects. In this presentation, I will describe a novel attention based model to mitigate pile-up at high-luminosity LHC and compare results to existing baseline methods.