A search for inelastic dark matter at the Belle II Experiment

Abstract

The Standard Model (SM) of particle physics is a remarkably successful and highly predictive theory of fundamental particles and interactions. However, it is an incomplete description of nature, as it fails to account for several observed phenomena, most notably the existence of dark matter. Observations ranging from galactic rotation curves to anisotropies in the Cosmic Microwave Background (CMB) provide compelling evidence for a non-baryonic gravitational source that lies entirely beyond the scope of the SM. One viable extension to the SM is the inelastic dark matter (iDM) model. This model introduces a dark sector containing two dark fermions, χ1 and χ2, with a mass splitting (∆mχ = mχ2 − mχ1 ). These particles couple to the SM through a massive dark photon (A) portal via kinetic mixing with the electroweak U(1)Y hypercharge component. In this work, we search for iDM production in e+e− collisions through the process e+e− → A′γISR → χ1χ2γISR. The signature is characterized by a high-energy initial-state radiation (ISR) photon and a displaced vertex arising from the decay χ2 → χ1A′∗(→ l+l−) for l ∈ [e, μ]. This analysis is performed using data collected by the Belle II detector at the SuperKEKB asymmetric-energy electron-positron collider in Tsukuba, Japan. We outline reconstruction and selection strategies optimized to isolate these rare signatures from dominant SM back- grounds, such as Bhabha scattering, two-photon processes, and continuum qq ̄ production. To maximize signal sensitivity across a broad mass range (0.05 GeV/c2 ≤ mχ1 ≤ 3.2 GeV/c2), we employ an ensemble of Gradient-Boosted Decision Trees. These models are trained on specific mass windows and particle identification categories to effectively discriminate signal from background using features such as vertex displacement, track-cluster isolation, and missing energy. The statistical analysis uses a frequentist approach to determine Belle II's sensitivity. We perform a single-bin counting experiment using sideband data to estimate the background level within a defined signal window. In the absence of an observed excess, we set median upper limits on the signal yield at a 95% confidence level (CL) using the q ̃ test statistic and toy-based calculations.