Mass-unspecific classifiers for mass-dependent searches

TL;DR

The study tackles the challenge of searching for new particles across wide mass ranges, where signal and background shapes evolve with mass. It introduces mass-unspecific classifiers, muNN and muBDT, that embed a mass-scale label (e.g., $m_T^{rec}$) and are trained on mass-balanced samples spanning $[0.9,6.5]$ TeV, yielding a continuous discriminator and near-optimal performance compared to mass-specific approaches. In a HL-LHC benchmark for single production of a vector-like quark $T$ (cross section ∝ $|V_{Tb}|^2$), the mass-unspecific classifiers outperform a parameterised NN and several pNN variants, while also enabling background-shape preservation through thresholding. The results suggest a robust, scalable framework for wide-range new-physics searches applicable to other processes and colliders, reducing the need for mass-specific retraining and improving sensitivity across masses.

Abstract

Searches for new particles often span a wide range of mass scales, where the shape of potential signals and the SM background varies significantly. We make use of a multivariate method that fully exploits the correlation between signal and background features and the explored mass scale, and is trained on a sample that is balanced across the entire mass range. The classifiers, either a neural network or a boosted decision tree, produce a continuous output across the full mass range and, at a given mass, achieve nearly the same performance as a classifier specifically trained for that mass. The performance of the classifiers is better than the one obtained with parameterised neural networks and similar methods.

Figures (6)

• Figure 1: Representative diagram of single $T$ quark production in hadron collisions.
• Figure 2: Kinematical distributions of $p_T^Z$ (top) and $p_T^J$ (bottom) for the signal and background, at different mass scales $m_T^\mathrm{rec} \simeq 1$ TeV, 6 TeV.
• Figure 3: ROC curves for several classifiers, for $m_T = 1$ TeV (top) and $m_T = 6$ TeV (bottom).
• Figure 4: ROC curves for several classifiers, for $m_T = 1$ TeV (top) and $m_T = 6$ TeV (bottom).
• Figure 5: Kinematical distributions of the $ZW^+jj$ background, and with two injected $T$ signals with $m_T = 2$ and 4 TeV. The upper lines correspond to the preselection level, and the lower, filled distributions are obtained after a selection corresponding to $\varepsilon_\mathrm{bkg}^{-1} = 100$.