Predicting neutron flux density distribution in HTR-10 using U-Net based on DEM-MC coupled simulations

Abstract

This study presents a novel approach for predicting neutron flux density distribution within the High- Temperature Gas-Cooled Reactor (HTGR) by deep learning network U-Net, based on the training data acquired from the Discrete Element Method (DEM) and Monte Carlo method (MC) coupled simulations. The neutron flux density, essential for reactor safety analysis and fuel management, relies on either computationally intensive MC simulations or deterministic methods with relatively limited accuracy. Our approach uses a U-Netbased neural network to predict neutron flux density distributions derived from high-fidelity DEM-MC simulation data and material cross-section data. The resulting dataset, reflecting different operational conditions, is used to train the U-Net model, which accurately maps material information to neutron flux predictions. The deep learning network is conducted on HTR-10. Several settings of the network are studied to assess model robustness, including the dataset size, the energy structure of cross-section libraries, and network structure. Results indicate that the U-Net model significantly reduces computation time while maintaining relatively high predictive accuracy, achieving R2 values close to unity under various configurations. This framework highlights the potential of deep learning as an efficient alternative for preliminary neutron flux density distribution predictions in HTGRs, offering both computational efficiency and adaptability.