Akademik Veri Yönetim Sistemi
International Journal of Computational andExperimental Science and ENgineering(IJCESEN), cilt.12, sa.1, ss.447-461, 2026 (Scopus, TRDizin)
Secondary neutrons
produced by proton-target interactions in high-energy proton accelerator
facilities present a major shielding challenge due to their high penetrability
and broad energy spectra. In this study, neutron dose attenuation in B₄C- and
FeB-enhanced concretes containing 5%, 10%, and 15% additives was investigated
at a proton energy of 1000 MeV using FLUKA-based Monte Carlo (MC) simulations
coupled with Machine-learning (ML) surrogate models.
MC-generated dose
data were used to train log-linear Linear Regression (log-linear LR), K-Nearest
Neighbors (KNN), Random Forest (RF), and Gradient Boosting Regressor (GBR)
models to enable rapid dose prediction. The results show that RF and GBR
achieve the highest predictive accuracy under all configurations, with test-set
R² values of approximately 0.98-0.99 in tunnel air and 0.99-0.996 in concrete
shielding. In contrast, the LR model performs poorly in shielding regions due
to strong nonlinearity, while KNN also provides high predictive accuracy exceeding 90%, albeit
with lower performance compared to RF and GBR.
A comparative
analysis reveals that FeB-enhanced concrete exhibits more complex attenuation
behavior due to the combined effects of iron-induced scattering and boron
absorption. Overall, the validated hybrid MC-ML framework demonstrates that RF-
and GBR-based surrogate models provide a fast, reliable, and computationally
efficient approach for neutron dose estimation and shielding optimization in
high-energy proton accelerator facilities.