Applying an Adversarial Neural Network to probe the 2021 HPS SIMP Search

The Heavy Photon Search (HPS) is an accelerator-based, fixed target dark-matter (DM) experiment built at JLab’s CEBAF. It is composed primarily of two silicon vertex-tracker (SVT) halves immersed in a dipole magnetic field; these trackers enable the detection of ~1 cm decay lengths characteristic of MeV-scale thermal-relic DM models. HPS performs both prompt and displaced-vertex searches, with the displaced program targeting signatures of a heavy U(1) gauge boson coupled to the dark sector. In this presentation, we describe a model-dependent search for DM using the 531 mC worth of 4 GeV electrons collected during the 2021 run. We search for displaced Strongly Interacting Massive Particles (SIMPs) in the form of dark-SU(3) ρ and Φ meson decays to Standard-Model e⁺e^- pairs. Earlier probes of this model, based on HPS’s 2016 data, identified mis-reconstructed vertices as the dominant source of difficult-to-suppress background. These analyses largely exhausted the power of cut based selections; further gains in the 2021 signal-to-noise ratio likely require machine-learning-driven event selection. We implemented an adversarial neural-network training scheme in which the classifier’s loss increases whenever a separate mass-predicting network succeeds, discouraging mass dependence. We compare this ML-assisted selection with previous cut based methods and show that this new approach is more powerful in discriminating signal from rare background processes.