PERFORMANCE OF BOOSTED Z BOSON TAGGER USING UNSUPERVISED LEARNING IN ATLAS

Abstract

To detect and explore possible candidates for dark matter and physics beyond the Standard Model (BSM), we use ATLAS’s recently developed Unified Flow Objects (UFOs) large-radius jets for tagging boosted hadronic decays of the 𝑍0 boson (𝑍′ )1. For the 𝑍′ tagger, we introduce the Clustering Autoencoder (CAE) which integrates two unsupervised learning frameworks for the classification and mass decorrelation of UFO jets. The first framework is a fully-connected autoencoder (AE) that reduces the number of input jet substructure variables into a latent space of three dimensions, and the second framework is a Uniform Manifold Approximation and Projection (UMAP) algorithm which reduces the AE latent space further to two dimensions. Afterwards, a neural network score is constructed by transforming the UMAP latent space to a histogram as a function of each event’s representative Euclidean space. We compare the CAE tagger performance to previous cut-based tagger and deep neural network (DNN) tagger performances through signal efficiency and background rejection rates. Though the 𝑍′ tagger underperforms tenfold compared to previously tested taggers, the CAE presents a realistic approach of training without the UFO jets’ weights and labels. Most importantly, the tagger we present here shows an important step towards scaling into high dimension analysis for physics BSM.