PileUp Mitigation at the HL-LHC Using Attention for Event-Wide Context

The Large Hadron Collider, LHC, collides bunches of protons resulting in multiple interactions that occur practically simultaneously. This creates a pileup effect that distorts physics measurements due to the products of pileup collisions. In order to improve the discovery potential of the LHC, it is necessary to mitigate the effect of pileup interactions on the processes of interest. In this paper, we suggest a novel AI-based method, PUMiNet, to tackle the problem of pileup at the current LHC and future High Luminosity LHC conditions. PUMiNet is an attention-based algorithm that mitigates pileup effects using a regression task on jets in the context of an entire event. At $\left\langle \mu \right\rangle=200$, PUMiNet is able to predict the hard scatter energy and mass fractions of jets with $R^2=0.912$ and $R^2=0.720$, respectively. These predictions enable the reconstruction of the Higgs boson mass in the HL-LHC environment.