Reconstructing Sparticle masses at the LHC using Generative Machine Learning

TL;DR

The paper tackles the problem of reconstructing heavy-sparticle masses from detector-level data at the HL-LHC. It introduces a two-component generative framework that combines a transformer-based detector encoder with a diffusion model to map detector-level observables to parton-level kinematics, enabling mass inference across wide parameter spaces. Demonstrations on RPV SUSY-inspired benchmarks show robust mass reconstruction across multiple channels, with high efficiency for the lightest particle and substantial gains for the heavier state. The approach offers a model-agnostic tool for mass spectrum recovery in future BSM searches, with potential enhancements via denser mass grids and topology-aware observables.

Abstract

We explore a generative model framework to infer the masses of heavy particles from detector-level data over a broad parameter space. Our model combines a transformer-based detector encoder and a diffusion neural network. We first apply our model to a new physics scenario involving the pair production of wino-like chargino-neutralino, $pp \to \tildeχ_1^{\pm} \tildeχ_2^0$, in the $1\ell + 2γ+ jets$ channel at the high luminosity LHC~(HL-LHC). We find that our framework can achieve mass reconstruction efficiency of $\gtrsim 70\%$ for the lightest neutralino $\tildeχ_1^0$ and $\gtrsim 40\%$ for the second lightest neutralino $\tildeχ_2^0$, for a mass tolerance of $Δm = 30~$GeV, across the entire parameter space accessible at the HL-LHC. We further extend our analysis to a different scenario with $pp\to\tildeχ_1^{\pm}\tildeχ_1^{\mp}+\tildeχ_1^{\pm}\tildeχ_2^0$ pair production at the HL-LHC in the $4\ell+\rm E{\!\!\!/}_T$ channel, and for a fixed value of $m_{\tildeχ_2^0}$, we obtain reconstruction efficiencies $\gtrsim80\%$ over a wide range of $m_{\tildeχ_1^0}$ for $Δm = 30~$GeV.

Figures (7)

• Figure 1: Feynman diagram for SUSY signal where mass degenerate pair produced $\tilde{\chi}_1^{\pm}\tilde{\chi}_2^0$ undergo cascade decays to $\tilde{\chi}_1^0$ followed by decay to $uds$ via $\lambda_{112}^{\prime\prime}$ coupling. The $W$ boson decays to a lepton ($\ell^{\prime}\equiv e,\mu,\tau$) and one neutrino ($\nu_{\ell^{\prime}}$), and higgs boson ($h$) decays to two photons ($\gamma$).
• Figure 2: A schematic representation of the model architecture is shown.
• Figure 3: The generated ('Pred') and truth ('True') parton-level distributions for $p_{x},\, p_y,\,p_z$ for the lightest neutralino $\tilde{\chi}_1^0$ (left) and the next-to-lightest-neutralino $\tilde{\chi}_2^0$ (right) are shown for the process $pp \to \tilde{\chi}_1^{\pm}\tilde{\chi}_2^0 \to (\tilde{\chi}_1^{\pm} \to W^{\pm} \tilde{\chi}_1^0) (\tilde{\chi}_2^0 \to h \tilde{\chi}_1^0)$ at the $\sqrt{s}=14~$TeV LHC, with $m_{\tilde{\chi}_2^0} = 600~$GeV and $m_{\tilde{\chi}_1^0}=200~$GeV.
• Figure 4: The reconstruction efficiencies, defined as the fraction of events where $|m_{\tilde{\chi}_1^0}^{\mathrm{pred}} - m_{\tilde{\chi}_1^0}^{\mathrm{true}}| < \Delta m$, are shown, for three tolerances, $\Delta m = 10,\,20$ and 30 GeV. Here, the network is trained on events with a fixed $m_{\tilde{\chi}_2^0} = 600~$GeV, but with $m_{\tilde{\chi}_1^0}$ varying between 50 and 400 GeV, with a 50 GeV step size. The network is evaluated on events generated with $m_{\tilde{\chi}_1^0}$ varying over the same range, but with a smaller step-size of 25 GeV.
• Figure 5: The reconstruction efficiencies for $m_{\tilde{\chi}_1^0}$ (left) and $m_{\tilde{\chi}_2^0}$ (right) are shown for three different tolerances, $\Delta m = 10,\,30$ and 50 GeV. The network is trained on events with $m_{\tilde{\chi}_2^0}$ ranging between 400 to 650 GeV with a 30 GeV step-size, and $m_{\tilde{\chi}_1^0}$ varying between 50 GeV to $m_{\tilde{\chi}_2^0} - 125~$GeV with a 25 GeV step-size. The network is evaluated on different events sampled from the same mass grid.