A multi-fidelity reduced-order modeling framework of cross-scale flow field data reconstruction for fuel rod bundles

Utilizing a data-driven approach to accurately reconstruct high-fidelity flow field data is challenged by expensive data acquisition and high-dimensional data processing, particularly when addressing the problem with large-scale flow domains. This paper proposes a cross-scale multi-fidelity reduced-order model (CS-MF-ROM) to predict large-scale flow field data using only small-scale flow field datasets for model training. The model combines Proper Orthogonal Decomposition and neural network algorithms, embedding the idea of transfer learning to address domain scale disparities. To validate the proposed method, four test cases based on the measurement and analysis of turbulent mixing in sub-channels-horizontal rod bundle computational fluid dynamics (CFD) benchmark experiment and its extension models were studied. These four test cases were evaluated with progressively increasing complexity in geometric and boundary conditions. High-fidelity data for both large and small domains were obtained from the CFD solver with a validated grid model, while the low-fidelity solution was derived from the coarser grid model. Additionally, to investigate the practical situation where large-scale data are limited in quantity and diversity due to high cost, three sampling strategies of multi-scale data were evaluated. The proposed CS-MF-ROM shows strong agreement with the high-fidelity solution, demonstrating that accurate reconstruction of target-scale flow fields can be achieved using only sparse large-scale data and a low-fidelity model. Compared to traditional data-driven methods, the proposed approach reduces the additional computational cost, offering an efficient and accurate solution for flow field predictions of large-scale domains.