Fretting wear damage of heat transfer tubes induced by flow-induced vibrations is a primary failure mode in steam generators (SGs). Accurately analyzing the internal flow field characteristics of the secondary side of steam generators is essential for predicting and assessing fretting wear. Due to the high-temperature and high-pressure environment within SGs, it is challenging to accurately obtain local flow field characteristics of the secondary side through sensors. In this study, we propose a reduced-order model (ROM) of the three-dimensional two-phase flow field in SGs based on Proper Orthogonal Decomposition (POD) and machine learning regression methods. Computational Fluid Dynamics (CFD) simulations were conducted under various operational parameters to generate flow field datasets, from which a data-driven surrogate model was constructed to enable real-time flow field calculations. The core of the ROM approach lies in utilizing POD to extract an optimized reduced basis from complex nonlinear flow fields, retaining the key flow energy of the first few orders. Subsequently, machine learning regression algorithms are applied to establish the relationship between SG operational parameters (such as secondary side flow velocity, inlet temperature, steam space pressure, and heat transfer rates between the primary and secondary sides) and the reduced basis. This methodology substitutes complex three-dimensional CFD numerical simulations with linear algebra operations, significantly reducing computational complexity while maintaining accuracy. The reduced-order model can predict unknown flow fields within 2–3 s, greatly enhancing computational efficiency and meeting the demands of digital twins and real-time monitoring in nuclear power plants. With the developed ROM model, real-time calculations of thermal-hydraulic characteristics inside SGs—such as velocity, temperature, pressure, and void fraction—can be achieved using monitoring data from the evaporator’s secondary side as model input parameters, providing a foundation for further analysis of fretting wear in heat transfer tubes. This research offers a real-time acceleration method for CFD simulations of SGs and presents an innovative solution for constructing SG digital twins and enabling intelligent operation and maintenance of nuclear powerplants, ultimately contributing to comprehensive safety assurance.