Every bunch crossing at the LHC results in not only one proton-proton interaction, but rather several. These additional proton-proton interactions are called pileup interactions. With the increasing instantaneous luminosity of the LHC, also the number of pileup interactions increased, and it will reach up to 200 pileup interaction during High-Luminosity LHC operation. The task of pileup mitigation thus becomes of paramount importance, as pileup does not only affect the jet energy but also other event observables as, for example, the missing transverse energy, the jet substructure, and more. The current state-of-the-art techniques, such as the PUPPI algorithm, are physics-motivated, rule-based routines that proved to effectively reject pileup interactions. Nonetheless, given the high complexity of the task, machine-learning-based approaches are expected to bring major improvements. A long-lasting issue of such ML approaches, though, is the lack of truth labels in Geant4 detector simulations. Thus, ordinary fully-supervised ML algorithms can only be trained in simplified detector simulations, but cannot then be ported to the LHC experiments. We will present a ML pileup mitigation technique that does not need truth labels but learns them through a self-supervised process. We demonstrate how our approach improves the key objects used in precision measurements and searches.
