Preview

PDF, English - main document Download (2MB) | Terms of use

Abstract

Various aspects of LHC simulations can be supplemented by generative networks. For event generation we show how a GAN can describe the full phase space structure of top-pair production including intermediate on-shell resonances and phase space bound- aries. In order to resolve these sharp peaking features, we introduce the maximum mean discrepancy. Additionally, the architecture can be extended in a straightforward manner to improve the network performance and to handle weighted events in the training data. Furthermore, we employ GANs to generate new events which are distributed according to the sum or difference of the input data. We first show with the help of a toy example how such a network can beat the statistical limitations of bin-wise subtraction methods. Afterwards we demonstrate how this network can subtract background events or describe collinear subtraction events in next-to-leading order calculations. Finally, we show how detector simulations can be inverted using GANs and INNs. They allow us to reconstruct parton level information from measured events. In detail, our results show how conditional generative networks can invert Monte Carlo simulations statistically. INNs even allow for a statistical interpretation of single-event unfolding and yield the possibility to unfold parton showering.