Suppressing pile-up contributions in the formation of topological clusters in ATLAS

The production of particles in secondary proton-proton collisions, known as pile-up, is a phenomenon which affects the measurements performed in high-luminosity particle colliders, such as the LHC. This work explores the possibility of using deep neural networks to mitigate pile-up contributions to the measured energy of topoclusters in the ATLAS experiment, studying clusters from simulated Run 2 and Run 3 events. The implemented networks aim to determine whether clusters contain contributions arising only from pile-up, solely from the hard scatter interaction, or from a combination of the two. After such classification, clusters identified as pile-up-only could be excluded from the jet reconstruction inputs, while pile-up contributions in the mixed clusters could be suppressed. The results demonstrate the effectiveness of the networks in providing a classification which improves the energy response of the topoclusters. In particular, the suppression of pile-up-only clusters reduces the interquartile range of the response distribution by a factor of approximately $10^3$ for clusters in Run 2 and $10^2$ for those in Run 3, indicating the great improvements that pile-up mitigation can bring to cluster energy calibration.