Search for dimuon resonance in the 35 to 75 GeV mass range using 140 fb$^{-1}$ of 13 TeV $pp$ collisions with the ATLAS detector

A model-independent search for low-mass resonances decaying into pairs of oppositely charged muons is presented. The analysis uses proton-proton collision data corresponding to an integrated luminosity of 140 fb$^{-1}$, recorded by the ATLAS detector at the Large Hadron Collider between 2015 and 2018. The search targets hypothetical dimuon resonances in the invariant mass range from 35 GeV to 75 GeV. The modelling of this mass region is particularly challenging for conventional analytic background parameterisations. To address this, a Gaussian process regression technique is used to model the background. The dimuon mass spectrum is analysed for potential signals, and no statistically significant excess is observed. Upper limits at the 95% confidence level are set on the fiducial production cross-section of new resonances decaying promptly into muons, ranging from 20 fb to 110 fb, depending on the resonance mass. These results are further interpreted in the context of dark-photon and dark-matter-mediator models, leading to new constraints on their parameter spaces.

💡 Research Summary

The ATLAS Collaboration performed a model‑independent search for low‑mass resonances decaying into oppositely charged muon pairs using the full Run 2 dataset collected at a centre‑of‑mass energy of 13 TeV, corresponding to an integrated luminosity of 140 fb⁻¹. The analysis targets a previously unexplored invariant‑mass window of 35 GeV to 75 GeV, a region that is difficult to model with conventional analytic background functions because of the rapidly varying Drell‑Yan contribution and the complex trigger and detector response. To overcome this, the authors introduced a Gaussian Process Regression (GPR) technique to model the smooth background directly from data. GPR, a non‑parametric Bayesian method, uses a kernel (a combination of radial‑basis‑function and white‑noise components) whose hyper‑parameters are tuned in a data‑driven way, providing a flexible description of the background shape while avoiding over‑fitting.

Event selection relies on a combination of unprescaled single‑muon and dimuon triggers, covering low‑p_T muons (p_T > 14 GeV) as well as higher‑p_T muons (p_T > 50 GeV) without isolation requirements, thereby maximizing acceptance across the whole mass range. Reconstructed muons must satisfy medium identification and isolation criteria, have opposite charge, and lie within |η| < 2.5. The dimuon invariant mass is required to fall between 35 GeV and 75 GeV.

Signal models are provided by two benchmark scenarios. The first is a simplified dark‑matter mediator (Z′) with either vector or axial‑vector couplings to quarks, leptons and a dark‑matter particle χ. The second is a kinetically mixed dark photon (Z_D) arising from a hidden U(1)_D gauge symmetry. Both are produced via the Drell‑Yan mechanism and decay promptly to μ⁺μ⁻. Monte‑Carlo samples for the signals were generated with MadGraph5_aMC@NLO at leading order, interfaced to Pythia 8 for parton showering and hadronisation, and normalised to next‑to‑next‑to‑leading order (NNLO) using a common k‑factor. Mass points are spaced every 5 GeV from 35 GeV to 75 GeV, with intrinsic widths varied by changing the couplings (g_q = g_ℓ = 0.1–0.3) to probe Γ/m ratios up to about 4 %.

Standard Model backgrounds are dominated by Drell‑Yan μμ production, with smaller contributions from ττ, top‑pair, and diboson processes. These were simulated with Sherpa (NLO for up to two partons) and Powheg‑Box (NLO tt̄), and processed through the full ATLAS Geant4 detector simulation. The equivalent simulated luminosities are 5–20 times larger than the data, allowing precise validation of the analysis chain. An additional large Drell‑Yan sample generated with a generator‑smearing technique was used to test the GPR background extraction.

Systematic uncertainties include muon reconstruction, identification and trigger efficiencies (≈1–2 %), momentum scale and resolution (≤0.5 GeV), signal acceptance and efficiency (≈5 %), and uncertainties associated with the GPR background model (kernel choice and hyper‑parameter variations). These are incorporated as nuisance parameters in the likelihood fit.

The statistical interpretation employs a binned profile‑likelihood fit to the dimuon mass spectrum, with the signal model added on top of the GPR‑derived background. The signal strength μ is extracted, and 95 % confidence‑level (CL) upper limits are set using the CL_s method. No local excess exceeding 2σ is observed anywhere in the examined mass range; the most significant fluctuation corresponds to a local p‑value of about 0.12 (≈1.5σ).

The resulting 95 % CL upper limits on the fiducial production cross‑section times branching ratio σ·BR(μ⁺μ⁻) range from 20 fb at 35 GeV to 110 fb at 75 GeV. Translating these limits into the benchmark models yields new constraints: for the Z′ mediator with g_q = g_ℓ = 0.1, masses around 40 GeV are excluded for σ·BR ≲ 30 fb, while for the dark photon the kinetic‑mixing parameter ε is limited to ≲10⁻³ for masses between 35 GeV and 50 GeV. These limits improve upon previous CMS and LHCb results, especially in the 35–50 GeV window where ATLAS achieves roughly a factor of two better sensitivity to ε.

In conclusion, the analysis demonstrates that Gaussian Process Regression can serve as a powerful, data‑driven tool for background modeling in resonance searches where traditional analytic functions fail. The absence of a signal leads to the most stringent ATLAS limits to date on low‑mass dimuon resonances in the 35–75 GeV range, providing valuable guidance for future theoretical developments and for upcoming searches with larger datasets in LHC Run 3 and beyond.