Wing Sheung Chan

56 Object reconstruction and identification the associated tracks of real τ had - vis usually form a displaced secondary vertex and have a large impact parameter with respect to the primary vertex. Important high-level input variables to the RNN ID include, for example, the central energy fraction ( f cent ), the invariant mass of the track system ( m track ) and the transverse flight path significance ( S flight T ). f cent quantifies how collimated the shower is by calculating the fraction of energy deposited in the region ∆ R < 0 . 1 to that in the entire core region ∆ R < 0 . 2 . m track is defined as the sum of the four-momenta of all the tracks, assuming a pion mass for each track. S flight T is the displacement of the secondary vertex in the transverse plane with respect to the primary vertex divided by its estimated uncertainty. RNNs are neural networks that have internal states which can function as “memory” of the network. The memory can be used to process inputs that are sequences of variable lengths. The RNN ID utilises this feature to take in the variable number of tracks and clusters as low-level inputs. Low-level input variables to the RNN ID include the impact parameters and the number of inner detector hits of the individual tracks, as well as the moments that quantify the longitudinal and radial shapes of the individual clusters. Figure 3.4 shows the expected output distributions and the receiver operating charac- teristic (ROC) curves of the RNN ID. Four working points (WP), Tight , Medium , Loose and VeryLoose , are defined with increasing signal selection efficiencies as indicated in the figure. 3.2.5. Electron rejection Aside from jets, electrons are also a substantial background to τ had - vis identification. The RNN ID, developed specifically for discriminating against jets, does not provide sufficient discriminating power against electrons. For this reason, an e -veto BDT algorithm [77] specialised in telling electrons and τ had - vis candidates apart is developed. Electrons have features similar to those of 1-prong τ had - vis , especially for those with one neutral pion. Nonetheless, some differences in the detector response are very effective in differentiating electrons and τ had - vis . Since electron showers are purely electromagnetic, only a very small amount of the electron energy could leak to the HCal. On the contrary, τ had - vis often deposit significant amount of energy in the HCal due to the h ± ’s. Another effective discrimination comes from the response of the TRT. Electrons, which are always ultrarelativistic because of their small rest mass, usually leave more hits in the TRT than the heavier h ± ’s from τ had - vis . The e -veto BDT is trained to exploit these differences. The e -veto BDT was originally developed for analysing Run-1 data. However, due to software and hardware changes, it became deprecated at the beginning of Run 2. It has since been redeveloped to adopt the changes, to which the author has made major contributions. The expected output distributions and ROC curve of the redeveloped e -veto BDT are shown in Figure 3.5. Three working points with increasing efficiencies, Tight , Medium and Loose , are defined. While the algorithm is trained to identify 1-prong τ had - vis candidates, it is found to be powerful in identifying 3-prong τ had - vis candidates as well.