Exploring new avenues for pileup jet rejection using machine learning

In this project we train a neural network for the distinction between hard-scatter and pileup jets. In particular, we are interested if certain changes in the networks architecture and input features can lead to an increase in the pile-up rejection rate. We consider both simple deep neural networks and networks based on transformer architecture. By comparing ROC curves and the background rejection rates at a fixed hard-scatter jet detection efficiency of 95 %, we conclude that the addition of several input variables can lead to increase in pile-up jet rejection of 8.8 % for the feed-forward network, and 17.3 % for the transformer network when compared to the performance of the legacy neural network employed by ATLAS. These values are for the entire kinematic range spanning a transverse momentum of 15 to 60 GeV.