Hybrid windowed networks for on-the-fly Doppler broadening in RMC code

On-the-fly Doppler broadening of cross sections is important in Monte Carlo simulations, particularly in Monte Carlo neutronics-thermal hydraulics coupling simulations. Methods such as Target Motion Sampling (TMS) and windowed multipole as well as a method based on regression models have been developed to solve this problem. However, these methods have limitations such as the need for a cross section in an ACE format at a given temperature or a limited application energy range. In this study, a new on-the-fly Doppler broadening method based on a Back Propagation (BP) neural network, called hybrid windowed networks (HWN), is proposed to resolve the resonance energy range. In the HWN method, the resolved resonance energy range is divided into windows to guarantee an even distribution of resonance peaks. BP networks with specially designed structures and training parameters are trained to evaluate the cross section at a base temperature and the broadening coefficient. The HWN method is implemented in the Reactor Monte Carlo (RMC) code, and the microscopic cross sections and macroscopic results are compared. The results show that the HWN method can reduce the memory requirement for cross-sectional data by approximately 65%; moreover, it can generate keff, power distribution, and energy spectrum results with acceptable accuracy and a limited increase in the calculation time. The feasibility and effectiveness of the proposed HWN method are thus demonstrated.