Machine learning and Monte Carlo simulation for hybrid shielding materials


Ekinci F., Kaya N., Alaca Z. S., Türkoğlu B., Güzel M. S., Levet A., ...Daha Fazla

RADIATION PHYSICS AND CHEMISTRY, cilt.250, sa.114361, ss.1-40, 2027 (SCI-Expanded, Scopus)

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 250 Sayı: 114361
  • Basım Tarihi: 2027
  • Doi Numarası: 10.1016/j.radphyschem.2026.114361
  • Dergi Adı: RADIATION PHYSICS AND CHEMISTRY
  • Derginin Tarandığı İndeksler: Academic Search Ultimate (EBSCO), Engineering Source (EBSCO), Scopus, Science Citation Index Expanded (SCI-EXPANDED), Chemical Abstracts Core, Chimica, Compendex, EMBASE, INSPEC
  • Sayfa Sayıları: ss.1-40
  • Ankara Üniversitesi Adresli: Evet

Özet

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.

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.