Reliability analysis of a lead-bismuth cooled passive system based on an active learning BR-ARS-BP neural network surrogate model

BackgroundThe lead-bismuth fast reactor (LBFR) adopts a passive residual heat removal system, which effectively improves the inherent safety of the reactor. Due to the same order of magnitude between the resistance and natural forces in the passive system, slight fluctuations in parameters may cause system failure, making the analysis of its reliability of great significance.PurposeThis study aims to analyze the reliability of the passive systemfor LBFR on the basis of neural network surrogate model.MethodsThe passive system of the liquid lead-bismuth cooled thermal-hydraulic experimental platform, TALL-3D, was taken as the object of this study, a reliability analysis method using an active learning-based BR-ARS-BP neural network surrogate model was proposed to analyze its reliability. Firstly, the model was constructed in this study. Then, the model was verified for both single and multiple failure regions. Finally, the sensitivity and reliability of the passive system of TALL-3D were analyzed.ResultsThe analysis results show that the Monte Carlo method, BP neural network surrogate model, and active learning BR-ARS-BP neural network surrogate model have calculation times of 105~106, 900, and 278, respectively. The failure probability calculation errors of the BP neural network surrogate model and the active learning BR-ARS-BP neural network surrogate model, compared to the Monte Carlo method, are 0.702 and 0.044, respectively.ConclusionsThese results demonstrate the advantages of the active learning BR-ARS-BP neural network surrogate model in reliability analysis of lead-bismuth cooling non-active systems. It reduces computational cost and offers high-precision evaluation, supporting the wider application of lead-bismuth cooling non-active systems.