Abstract: In the field of reactor thermal-hydraulics, system thermal-hydraulic codes have been widely employed to quickly determine the global response characteristics of the primary system. But, duo to the one-dimensional governing equations are adopted, certain local multi-dimensional physical phenomena cannot be accurately captured by the system thermal-hydraulic codes. On the other hand, the multi-dimensional dynamic simulation with high spatial and temporal resolution can be achieved by computational fluid dynamics (CFD) codes using the Navier-Stokes (N-S) equations. However, CFD computation is computationally expensive and time-consuming, requiring extensive computational resources, such as high-performance computing clusters or parallel processing techniques, to solve complex flow problems. This imposes significant limitations on its applications in scenarios requiring rapid computation speed and real-time analysis, such as parametric study, uncertainty quantification, and design optimization. To address this challenge, a non-intrusive reduced order model for predicting the steady state flow field in the lower plenum of a two-dimensional reactor pressure vessel was constructed using the proper orthogonal decomposition (POD) method combined with sparse grids theory in this paper. The input parameters of the reduced order model were the coolant mass flow rates at the two inlets of the lower plenum, while the output was the corresponding velocity field distribution within the lower plenum. Based on the sparse grids theory, the snapshot matrix was first generated by repetitive simulation using the full order CFD lower plenum model, then the mapping relation between the POD coefficients and the input parameters of the reduced order model was fitted. The performance of the constructed reduced order model was tested on an independent test dataset. The results indicate that the developed reduced order model is capable of rapidly predicting velocity field distributions in the lower plenum under different inlet mass flow rate, while maintaining accuracy that meets simulation requirement. Since the operation of the reduced order model involves only matrix algebra calculations instead of solving the N-S equations, its computational speed is significantly higher than that of the full order CFD model. This remarkable improvement expands the applicability of CFD-based simulation to a broader range of engineering problems. Specifically, the proposed reduced order model construction method holds great promise for applications requiring high-speed flow field prediction, such as multi-scale coupling simulation, real-time simulation analysis, and digital twin implementations in nuclear reactor systems and other fluid engineering domains. Furthermore, the developed framework can also be extended to transient flow problems, further enhancing its utility in advanced engineering applications.