Machine learning techniques for intermediate mass gap lepton partner searches at the large hadron collider

We consider machine learning techniques associated with the application of a boosted decision tree (BDT) to searches at the Large Hadron Collider (LHC) for pair-produced lepton partners which decay to leptons and invisible particles. This scenario can arise in the minimal supersymmetric Standard Model (MSSM), but can be realized in many other extensions of the Standard Model (SM). We focus on the case of intermediate mass splitting ($\sim 30\mathrm{GeV}$) between the dark matter (DM) and the scalar. For these mass splittings, the LHC has made little improvement over LEP due to large electroweak backgrounds. We find that the use of machine learning techniques can push the LHC well past discovery sensitivity for a benchmark model with a lepton partner mass of $\sim 110\mathrm{GeV}$, for an integrated luminosity of $300{\mathrm{fb}}^{-1}$, with a signal-to-background ratio of $\sim 0.3$. The LHC could exclude models with a lepton partner mass as large as $\sim 160\mathrm{GeV}$ with the same luminosity. The use of machine learning techniques in searches for scalar lepton partners at the LHC could thus definitively probe the parameter space of the MSSM in which scalar muon mediated interactions between SM muons and Majorana singlet DM can both deplete the relic density through dark matter annihilation and satisfy the recently measured anomalous magnetic moment of the muon. We identify several machine learning techniques which can be useful in other LHC searches involving large and complex backgrounds.

Figure 1

Two example Feynman diagrams representing signal production with an associated jet.

Figure 2

Plot illustrating the residual cross section for signal and background (as labeled) as a function of the BDT classification score after precuts. Corresponding plots for additional folds of training and evaluation samples can be found in Appendix pp1.

Figure 3

Density plot representing the separation of signal from each background (as labeled) by the boosted decision tree after precuts. For visual clarity, the densities are separately normalized such that the joint background density is not equivalent to the sum of the individual background densities. Corresponding plots for additional folds of training and evaluation samples can be found in Appendix pp1.

Figure 4

The number of signal events ($S$, blue), signal-to-background ration ($S÷(1+B)$, orange) and signal significance (${\sigma }_A$, green) as a function of the signal classification threshold. Each curve represents an interpolation which smooths out the sharp features in the shaded region, which arise from small event numbers. Corresponding plots for additional folds of training and evaluation samples can be found in Appendix pp1.

Figure 5

Plots indicating the relative importance of various kinematic features to the separation of signal from background by the boosted decision tree after level 1 pre-selections, respectively (left-to-right then top-to-bottom) for $\ell \ell jjj$, $\tau \tau jjj$, $t\overline{t}jj$, $ZZjj$, $WZjj$, and $WWjj$. Note that, in each panel, results are shown for a BDT which is trained only with signal and the particular background process in question.

Figure 6

The distribution of $M_{T2}^{100}$ for events passing the precuts for signal (orange), as well as the combined $ttjj$ and $WWjj$ backgrounds (blue).

Figure 7

Plot indicating the relative importance of various kinematic features to the separation of signal from the joint (all source) background. This training occurs application of the Sec. 4c precuts.

Figure 8

Same as Fig. 2, but for different choices of training and evaluation samples. Plots illustrating the residual cross section for signal and background (as labeled) as a function of the BDT classification score after precuts.

Figure 9

Same as Fig. 3, but for different choices of training and evaluation samples. Density plots representing the separation of signal from each background (as labeled) by the boosted decision tree after precuts. For visual clarity, the densities are separately normalized such that the joint background density is not equivalent to the sum of the individual background densities.

Figure 10

Same as Fig. 4, but for different choices of training and evaluation samples. The number of signal events ($S$, blue), signal-to-background ration ($S÷(1+B)$, orange) and signal significance (${\sigma }_A$, green) as a function of the signal classification threshold. Each curve represents an interpolation which smooths out the sharp features in the shaded region, which arise from small simulated event numbers.