Using machine learning for non-prompt D0 analysis

Abstract

In the last decade the pT differential cross-sections of heavy mesons such as D0 mesons have been measured extensively at the LHC for a variety of rapidity and energy ranges in proton-proton (pp) collisions. These measurements provide tests for standard model theories such as quantum chromo dynamics (QCD) and a baseline for measurements in heavy-ion collisions in which a plasma-like state of matter consisting of deconfined quarks and gluons (QGP) forms. Since the bottom (b) quark is the heaviest quark apart from the top quark it is produced very early in the hard scattering process making it the excellent probe. The properties of the b quark can be indirectly accessed by studying non-prompt D0 mesons. However a large fraction of promptly hadronized D0 mesons is present after the hadronization processes. These prompt D0 mesons are identical to non-prompt D0 mesons making it extremely challenging to separate the two. In this thesis we study the possibility to maximise the non-prompt over prompt D0 ratio using two types of machine learning algorithms. We discuss the training results of boost decision trees using adaptive boosting and convolutional neural networks and compare the performance of both algorithms to choose the model which suits the scope of this thesis. We report an increase of the non-prompt fraction between 2.268 ± 0.08 and 69.76 ± 20.1 when the boost decision tree is used replace the standard selection cuts made in the invariant mass reconstruction. The invariant mass is reconstructed in the interval 5 < pT < 24 GeV/c with a fraction of about 18% of the data sample available with significances between 2.4 ± 0.8 and 6.3 ± 1.1. Using these significances we show that within the boundaries of the available minimum bias data it is possible to obtain significances greater than 5.0 for 5 < pT < 24 GeV/c. Future studies can be performed to improve the algorithms and other classification algorithms, such as transformers, can be used to increase the non-prompt D0 fraction.