A Multivariate Approach to Dilepton Analyses in the Upgraded ALICE Detector at CERN-LHC

ALICE, the dedicated heavy-ion experiment at CERN-LHC, will undergo a major upgrade in 2019/20. This work aims to assess the feasibility of conventional and multivariate analysis techniques for low-mass dielectron measurements in Pb-Pb collisions in a scenario involving the upgraded ALICE detector with a low magnetic field ($B=0.2~\text{T}$). These electron-positron pairs are promising probes for the hot and dense medium, which is created in collisions of ultra-relativistic heavy nuclei, as they traverse the medium without significant final-state modifications. Due to their small signal-to-background ratio, high-purity dielectron samples are required. They can be provided by conventional analysis methods, which are based on sequential cuts, however at the price of low signal efficiency. This work shows that existing methods can be improved by employing multivariate approaches to reject different background sources of the dielectron invariant mass spectrum. The major background components are dielectrons from photon conversion and combinatorial pairs. By implementing deep neural networks, the signal-to-background ratio can be improved by up to 60% over existing results in the case of pure conversion rejection and up to 30% in the case of additional suppression of all combinatorial background components. In both cases, the gain in significance is about 15% compared to conventional approaches. Additionally, different strategies for rejecting heavy flavor pairs (i.e., dielectrons originating from $c\bar{c}$ or $b\bar{b}$) are studied and some of their major challenges identified. In general, it is concluded that multivariate techniques are a powerful and promising approach to dielectron analyses since they significantly improve the results over conventional methods in terms of signal-to-background ratio and significance. Moreover, these techniques remove complexity from existing implementations as they allow to (1) base the analyses on individual tracks (instead of track pairs), essentially without sacrificing analysis performance, (2) render some of the existing and involved analysis methods obsolete and, to some degree, (3) obviate the need for manual input feature engineering.