Multivariate analysis techniques for Particle Flow-based neutral pileup suppression at the ATLAS experiment

The removal of contamination from multiple pp interaction at the Large Hadron Collider, also known as pile-up, plays a fundamental role in the object reconstruction necessary for searches for new physics using data recorded by the ATLAS detector. To suppress pileup clusters originating from charged particles, the particle tracks are used by the current ATLAS Particle Flow algorithm. Due to the lack of tracks, pileup suppression of neutral constituents has to entirely rely on calorimeter information. The Soft Killer algorithm uses a sophisticated pT cut to suppress neutral pileup clusters. In this report, machine learning is used to combine additional calorimeter information for suppressing neutral pileup clusters. At first, a study of calorimeter cluster variables is performed to see which variables are suitable to distinguish hard scatter from pileup clusters. Then a machine learning algorithm is trained and the impact on the classification of clusters is analyzed. To conclude, the performance of the soft term MET reconstruction with the trained algorithm is studied.