Neural Spline Flows Tested on CMS Open Data for Mono-Z Dark Matter Search

Lead

Dark matter searches at colliders rarely look for the invisible particle directly. Instead, they search for visible recoil objects accompanied by missing transverse energy, a signature of particles escaping the detector. A new arXiv study explores this strategy in the mono-Z channel, where a potential dark matter system is produced together with a Z boson that decays into a pair of charged leptons.

The paper uses CMS Run 2015D open data at a center-of-mass energy of 13 TeV, corresponding to 2.32 fb⁻¹ of integrated luminosity, together with simplified-model Monte Carlo samples. Events are selected in both the μμ and ee final states, making the analysis a compact but complete test bed for machine-learning-based likelihood methods in high-energy physics.

Key points

• Open data as the foundation: The analysis is built on CMS MINIAOD and MINIAODSIM samples, showing how public collider data can support a full reinterpretation-style workflow.
• A high-dimensional event view: The authors extract 40 kinematic observables, apply physics-motivated cleaning and selections, and reduce them to a 37-dimensional feature vector.
• Flow-based density estimation: Five Neural Spline Flows are trained independently to model Standard Model backgrounds and mediator-specific dark matter signal densities.
• Likelihood-ratio event scoring: For each event, the test statistic is derived from the log-likelihood ratio between estimated signal and background densities. This allows the analysis to use information across the kinematic phase space instead of relying only on a hard missing-energy requirement.
• No dark matter evidence reported: A simultaneous profile-likelihood fit across the two lepton channels gives observed 95% confidence-level upper limits on the signal strength of μ<0.0177 for a scalar mediator, μ<0.0362 for a vector mediator, and μ<0.0498 for an axial-vector mediator. The expected limits are 0.0018, 0.0039, and 0.0069, respectively.

Why it matters

The main contribution is methodological rather than a claim of new physics. Neural Spline Flows are invertible generative models capable of learning flexible probability densities. In this context, they provide a way to estimate multivariate signal and background distributions and turn those estimates into likelihood-ratio scores, a natural language for statistical searches in particle physics.

At the same time, the result illustrates a key limitation. More expressive machine-learning models do not automatically remove detector effects, background mismodeling, or systematic uncertainties. The paper states that the observed limits are weaker than expected because of a residual discrepancy in the high-MET background model, not because of evidence for a dark matter signal.

If validated on larger datasets and with more extensive uncertainty treatment, flow-based likelihood methods could become useful tools for collider reinterpretation, anomaly searches, and open-data analyses. The study is a clear example of AI for Science: machine learning is not replacing the physics analysis, but expanding the statistical machinery available to probe difficult regions of phase space.

Source: arXiv