Optimal transport for a novel event description at hadron colliders

We propose a novel strategy for disentangling proton collisions at hadron colliders such as the LHC that considerably improves over the current state of the art. Employing a metric inspired by optimal transport problems as the cost function of a graph neural network, our algorithm is able to compare two particle collections with different noise levels and learns to flag particles originating from the main interaction amidst products from up to 200 simultaneous pileup collisions. We thereby sidestep the critical task of obtaining a ground truth by labeling particles and avoid arduous human annotation in favor of labels derived in situ through a self-supervised process. We demonstrate how our approach---which, unlike competing algorithms, is trivial to implement---improves the resolution in key objects used in precision measurements and searches alike and present large sensitivity gains in searching for exotic Higgs boson decays at the High-Luminosity LHC.