High-Fidelity Monte Carlo Simulation of Gamma-Ray Shielding Efficiency for Nuclear Security Applications Using Geant4 Toolset

Abstract

The development of AI-driven radiation detection systems is hindered by the lack of high-fidelity data representing complex shielding scenarios. Most existing datasets rely on simplified exponential attenuation models, ignoring stochastic scattering effects. This study challenges this approach by presenting a Geant4- based high-fidelity simulation of gamma-ray transport for Special Nuclear Materials (239Pu). While high-Z shielding (Lead) demonstrated predictable exponential attenuation (μ-validation error < 1.2%), our results reveal a critical anomaly in low-Z shielding (Polyethylene). We observed a significant non-monotonic "buildup effect" where the detected photon intensity increases by ~25-40% with shielding thickness up to 4 cm, effectively turning the shield into a scattering amplifier. This finding proves that "naive" synthetic data generation is insufficient for training robust AI models. The discovered spectral distortion mechanism is crucial for identifying SNM concealed behind common construction materials, offering a new physical baseline for "data-centric" nuclear security systems. Furthermore, this study underscores the necessity of accurately simulating the detector response and the full scattering history to prevent sim-to-real discrepancies when deploying neural networks in real-world environments. Incorporating these buildup anomalies into training datasets can significantly improve signal-to-noise ratio interpretation and reduce false negative rates in automated scanning applications.