Assessment of reactor flow field prediction based on deep learning and model reduction

Abstract

All-around and full-cycle digital simulation technology can improve the safety and economy of the reactor in research, development, operation, and maintenance processes. However, for the local three-dimensional fluid dynamic problems in complex reactor system, traditional computational fluid dynamic (CFD) methods are severely limited by the solution efficiency and accuracy. This study proposed a rapid flow field prediction method based on singular value decomposition (SVD) and deep learning. The raw flow field snapshot data was compressed using SVD to extract low-dimensional features, and the deep neural network was used to construct a flow field reduced-order model to achieve rapid flow field prediction. A benchmark problem of flow around a cylinder was selected to assess the efficacy of this method. Furtherly, this method was applied in the three-dimensional flow field simulations of the fuel assembly. The results demonstrated that the reduced-order model (ROM) error was less than 6% compared with that of the CFD model, and the time consumption was less than 1% of that of the CFD model. This exploration illustrates that high-fidelity ROMs based on order reduction and deep learning are a viable route to developing engineering-ready digital twins of nuclear reactors.