Data-driven surrogate modeling of FLUKA-generated ambient dose equivalent distributions in an asymmetric proton accelerator shielding system

Accurate determination of spatial ambient dose equivalent H*(10) distributions in high-energy proton accelerator facilities is essential for radiation protection and shielding design. In this study, spatial ambient dose equivalent H*(10) distributions produced under a proton beam-loss scenario were investigated using the FLUKA Monte Carlo (MC) code within an asymmetric multilayer concrete-ferroboron (FeB) shielding system. A 1000 MeV proton beam interacting with a copper target was modeled, and the resulting radiation field was analyzed in terms of secondary particle production, stopping-particle distributions, low-energy neutron interactions, neutron energy spectra, and spatial dose distributions. The results showed that neutrons are the dominant component of the secondary radiation field, accounting for 33.4% of all secondary particles. Protons represented 68.6% of the stopping-particle inventory, while neutrons constituted 96.0% of the particles generated by low-energy neutron interactions. Neutron spectrum analyses demonstrated that the concrete-FeB configuration significantly alters neutron transport through combined moderation and absorption mechanisms. The calculated H*(10) distributions exhibited strong directional dependence and pronounced dose gradients at material interfaces. Machine learning (ML)-based surrogate models, including Linear Regression (LR), K-Nearest Neighbors (KNN), Random Forest (RF), Extremely Randomized Trees (ETR), Gradient Boosting Regressor (GBR), Histogram-Based Gradient Boosting Regressor (HGBR), and Extreme Gradient Boosting (XGBoost), were developed using FLUKA-generated data. Among them, ETR, RF, HGBR, and XGBoost achieved R 2 values exceeding 0.9818 under a randomized train-test split, accurately reproducing MC-generated dose distributions. However, blocked cross-validation revealed reduced predictive capability in previously unseen spatial regions. The trained models generated dose predictions within milliseconds, reducing computational time by several orders of magnitude compared with conventional MC simulations. These findings demonstrate the potential of the proposed hybrid MC-ML framework for rapid shielding assessment and optimization in proton accelerator facilities.