Search for Long-lived Particles at Future Lepton Colliders Using Deep Learning Techniques

TL;DR

This work investigates long-lived particle (LLP) searches at future lepton colliders by exploiting deep learning on low-level detector data from $e^+e^-\to ZH$ events. The authors compare Convolutional Neural Networks (CNNs) and Graph Neural Networks (GNNs), coupled to an XGBoost classifier, to reconstruct LLP signatures and suppress Standard Model backgrounds, achieving LLP signal efficiencies up to $\sim$95% for $M_X \approx 50$ GeV and $\tau \approx 1$ ns. With $20~\text{ab}^{-1}$ of data (≈$4\times10^6$ Higgs bosons), they derive 95% CL upper limits on $\mathcal{B}(H\rightarrow XX)$ reaching $1.0\times10^{-6}$ in both fixed and floating $\epsilon_V$ scenarios, and provide 1D and 2D exclusion contours for the two-jet and four-jet final states. The results surpass traditional selection-based LLP searches at lepton colliders and emphasize the potential of ML approaches to broaden LLP sensitivity, including prospects for an external detector (Far Barrel Detector) that can further enhance reach by up to a factor of 13.7 in certain regions.

Abstract

Long-lived particles (LLPs) provide an unambiguous signal for physics beyond the Standard Model (BSM). They have a distinct detector signature, with decay lengths corresponding to lifetimes of around nanoseconds or longer. Lepton colliders allow LLP searches to be conducted in a clean environment, and such searches can reach their full physics potential when combined with machine learning (ML) techniques.This experimental study, utilizing comprehensive full simulation data samples, focuses on LLP searches resulting from Higgs decay in $e^+e^-\to ZH$. We demonstrate that, by employing deep neural network approaches the LLP signal efficiency can be improved up to 95\% for an LLP mass around 50 GeV and a lifetime of approximately 1 nanosecond, while rejecting all SM backgrounds. Furthermore, the signal sensitivity for the branching ratio of Higgs decaying into LLPs reaches a state-of-art limit of $1.0 \times 10^{-6}$ with a statistics of $4 \times 10^{6}$ Higgs.

Figures (13)

• Figure 1: Feynman diagrams of LLP production and decay. Two Feynman diagrams are presented illustrating the generation of LLPs, denoted as $X$, through the Higgsstrahlung mechanism. On the left, the diagram shows the production of $XX$ followed by their subsequent decay into a $\nu\bar{\nu}$ pair and a $q\bar{q}$ pair, respectively, resulting in two jets. On the right, both $X$ decay into $q\bar{q}$ pairs, leading to the four jets final state.
• Figure 2: Workflow of event classification for LLP signals. This chart outlines the event classification process utilizing CNN and GNN models, followed by XGBoost analysis. The process starts with formatting detector hits for NN input, progresses through 5-class classification, and concludes with an XGBoost-enhanced selection to differentiate the signal from the SM background, finally determining signal efficiency.
• Figure 3: Event visualization of CNN. Left: an event with one detectable LLP ($\mathcal{D}_1$, Type-I signal) featuring a decay vertex at [4.5 m, 2.3 rad]; Right: an event with two detectable LLPs ($\mathcal{D}_2$, Type-II signal) with decay vertices at [3.69 m, 2.16 rad] and [0.99 m, 4.37 rad]. Circles represent detector hits in the calorimeter and squares represent detector hits inside the tracker. Darker pixels represent hits with smaller time differences, and bigger pixels represent hits with larger energy. Note that the varying circle sizes are solely for visualization purposes; in the analysis, the pixel size remains fixed. The decay vertex of LLPs is marked with a star symbol.
• Figure 4: Network structure of CNN-based classification. The ResNet18 neural network architecture, consisting of two convolutional layers, followed by four ResNet blocks, an average pooling layer, and an output layer for classifications.
• Figure 5: Architecture of the heterogeneous GNN. $h_{t}$ and $h_{c}$ denote the node embedding, $x_{t}$ and $x_{c}$ denote the edge embedding. The subscript letter $l$ represents the $l$-th layer. The subscript letter $t$ represents the tracker and the letter $c$ represents the calorimeter. $\phi_{e}$, $\phi_{x}$, $\phi_{h}$ and $\phi_{g}$ are neural networks of MLPs.