Heavy Neutrinos across the Electroweak-to-Multi-TeV Frontier via Novel ML-Enhanced Probes

Technical Summary: Heavy Neutrinos across the Electroweak-to-Multi-TeV Frontier via Novel ML-Enhanced Probes

Problem Statement
The search for heavy neutrinos remains a central objective in testing the seesaw mechanism and explaining the origin of neutrino masses. While existing experimental constraints from meson decays, neutrinoless double-beta decay, and collider searches (LEP, LHC) cover specific mass ranges, significant gaps remain, particularly for heavy neutrino masses ($m_N$) spanning from the electroweak scale up to the multi-TeV frontier. Traditional search strategies often rely on resonant $s$-channel production, which becomes kinematically suppressed at high masses ($m_N \gtrsim 1$ TeV) due to phase space limitations. Furthermore, current constraints on third-generation couplings (involving $\tau$ leptons) are notably weaker than those for first- and second-generation leptons. There is a need for a unified strategy that maintains sensitivity across a broad mass spectrum (50 GeV to 10 TeV) and addresses the challenges of distinguishing rare signals from overwhelming Standard Model (SM) backgrounds in high-luminosity environments.

Methodology
The authors propose a comprehensive analysis strategy for proton-proton collisions at $\sqrt{s} = 13.6$ TeV, targeting the final state $\ell\nu jj$ (a charged lepton, missing transverse energy, and two jets). The study encompasses two distinct production regimes:

1. $s$-channel production: Dominant for $m_N < 1$ TeV, involving resonant production via Drell-Yan and $W\gamma$ fusion.
2. Vector Boson Fusion (VBF): Dominant for $m_N > 1$ TeV, involving $t$-channel exchange of virtual heavy neutrinos, which exhibits a slower cross-section degradation with increasing mass.

Simulation and Event Generation:
Signal samples were generated using MadGraph5_aMC@NLO with the HeavyN model, covering 16 mass points from 50 GeV to 10 TeV. Backgrounds included dominant SM processes such as $V$+jets, dibosons, and top-quark production, simulated inclusively as $pp \to \ell\nu + \ge 2j$. Events were processed through Pythia 8 for parton showering and hadronization, and Delphes 3.4.1 for detector emulation using CMS configurations, including realistic pileup conditions (average 140 interactions) and object identification efficiencies.

Machine Learning Analysis:
To optimize signal-background discrimination, the authors employed Gradient-Boosted Decision Trees (BDTs) using XGBoost and Scikit-learn.

• Input Features: The models were trained on 31 kinematic variables, including transverse momenta ($p_T$), pseudorapidities ($\eta$), azimuthal angles ($\phi$), missing transverse energy ($E_T^{\text{miss}}$), invariant masses ($M$), transverse masses ($M_T$), and angular separations ($\Delta R, \Delta \eta, \Delta \phi$) between jets, leptons, and missing energy.
• Training Strategy: Separate BDT models were trained for each mass point and for different lepton flavors ($\ell = \mu, \tau$). For the $\tau$ channel, only hadronic decays ($\tau \to \pi\nu$) were simulated to reflect reconstruction capabilities.
• Statistical Interpretation: Expected limits were derived using the CMS Combine tool with an asymptotic profile likelihood method on an Asimov dataset corresponding to an integrated luminosity of 3000 fb$^{-1}$ (HL-LHC). Systematic uncertainties (PDFs, scale variations, luminosity, lepton/jet efficiencies) were incorporated via nuisance parameters.

Key Contributions and Results

• Unified Mass Coverage: The study demonstrates that combining $s$-channel and VBF production mechanisms allows for a continuous search strategy across the entire 50 GeV to 10 TeV mass range. The cross-section behavior shifts from resonant dominance at low masses to VBF dominance at high masses, with the two contributions converging near $m_N \approx 1$ TeV.
• Enhanced Sensitivity via ML: The BDT analysis effectively captures complex kinematic correlations. Feature importance analysis reveals that transverse mass variables involving the charged lepton and missing energy ($M_T(\ell E_T^{\text{miss}})$) and the dijet invariant mass ($M(j_1 j_2)$) are the most discriminating features. The $M(j_1 j_2)$ distribution specifically distinguishes between resonant peaks (low mass) and the broad VBF spectrum (high mass).
• Projected Limits: Assuming 3000 fb$^{-1}$, the study projects 95% confidence level upper limits on the mixing parameter $|V_{\ell N}|^2$:
  • Sensitivity reaches $O(10^{-5})$ for $m_N = 50$ GeV.
  • Sensitivity extends up to $|V_{\ell N}|^2 \sim 1$ for masses approaching 10 TeV, constrained by unitarity.
  • The strategy offers significant improvement over existing DELPHI and prompt trilepton search limits, particularly in the multi-TeV regime where VBF becomes dominant.
• Third-Generation Reach: The analysis specifically addresses the $\tau$ channel, demonstrating that despite the challenges of hadronic $\tau$ reconstruction and additional neutrinos in the final state, the proposed VBF-enhanced ML strategy provides a viable pathway to probe $|V_{\tau N}|^2$ in a regime where current constraints are weak.
• Lepton Universality: The framework allows for the combination of electron, muon, and tau channels to test for lepton-universality violations.

Significance
The paper claims that this approach provides a "promising and complementary pathway" to existing search strategies. By integrating a novel production mechanism (VBF at high masses) with advanced machine-learning classification, the study significantly extends the discovery reach for heavy neutrinos at hadron colliders. It specifically highlights the potential to probe third-generation heavy neutrinos and explore the multi-TeV frontier, areas where traditional cut-based analyses or $s$-channel-only searches face diminishing sensitivity. The work underscores the utility of inclusive simulation and ML techniques in maximizing the physics potential of the High-Luminosity LHC.