Machine Learning-Assisted Optimization of Modular Neutron Shielding Based on Monte Carlo Simulations

Abstract

This study proposes a novel design methodology for neutron beam shutters that integrates Monte Carlo simulations (MCNP) with machine learning techniques to enhance shielding performance and accelerate the design process. The target facility is a compact neutron science platform where neutrons are produced by proton beams from a cyclotron striking a neutron production target. The system includes both thermal and fast neutron beamlines. A beam shutter is installed on the thermal neutron line to reduce occupational radiation exposure during maintenance activities. In this work, 200 neutron shutter configurations with varying material sequences were simulated using MCNP. The resulting dataset was used to train a fully connected neural network to predict the neutron flux downstream of the shielding. The trained model was subsequently applied to 1,000 randomly generated shielding configurations for rapid flux prediction and performance ranking. The 20 designs with the lowest predicted flux were selected and further validated via MCNP simulations. Results show that the optimal design reduces the neutron flux from 5.61 x 10^9 n/cm2*s at the shutter entrance to 4.96 x 10^5 n/cm2*s at the exit, achieving a reduction of four orders of magnitude. These findings confirm that the integration of machine learning techniques can effectively reduce simulation costs and assist in identifying high-performance shielding configurations, demonstrating the strong potential of data driven approaches in neutron system design.