Machine learning and Monte Carlo simulation for hybrid shielding materials

Designing hybrid shielding materials capable of simultaneously attenuating neutron and photon radiation remains a challenging optimization problem because shielding performance depends on complex interactions among material composition, radiation energy, shield thickness, and layer configuration. This study presents an integrated computational framework that combines PHITS-based Monte Carlo simulations with machine learning to accelerate the design and evaluation of HDPE/B4C/WO3 hybrid shielding systems. A two-stage simulation strategy was implemented. First, homogeneous composite materials were systematically screened under 17 monoenergetic neutron and photon sources and multiple shield thicknesses to evaluate total ambient dose equivalent H*(10) transmission. Subsequently, the same material system was reorganized into three-layer H/B/W configurations to investigate the influence of layer sequence and thickness distribution. After quality control, a database containing 2200 simulation cases was constructed and used to develop surrogate regression models including Random Forest, Extra Trees, Gradient Boosting, HistGradientBoosting, XGBoost, Support Vector Regression, K-Nearest Neighbors, and Multilayer Perceptron. The simulations demonstrated that WO3-rich compositions provided superior photon attenuation, whereas B4C-rich systems exhibited enhanced neutron shielding. Hybrid compositions, particularly B20W30, achieved the most balanced neutron–photon shielding performance by simultaneously reducing total H*(10) transmission and limiting secondary photon production. Layer sequence produced measurable improvements only under low-energy irradiation, while overall composition remained the dominant design parameter across most conditions. Among the evaluated machine learning models, XGBoost achieved the highest predictive accuracy (R2 = 0.995, multiplicative error ≈ 4.8%), enabling rapid prediction of shielding performance without repeated Monte Carlo calculations. While the surrogate models achieved excellent prediction accuracy within the simulated parameter space, leave-one-energy-out validation demonstrated that extrapolation to previously unseen low-energy irradiation conditions remains more challenging, highlighting the importance of representative training data for robust generalization. The proposed framework establishes a unified computational methodology for simultaneous composition optimization, layered shielding analysis, and high-accuracy surrogate modeling, providing an effective tool for the accelerated development of next-generation hybrid radiation shielding materials.