Efficient prediction of turbulent cross flow in helical coil steam generators of SMR via deep learning-driven reduced-order models

SMR have emerged as next-generation nuclear power system, offering enhanced safety, efficiency, and economic advantages. Among their critical components, HCSGs have been extensively studied for their effective heat exchange capabilities. However, primary-side cross flow within HCSG could induce vortices and turbulent structures between tubes, resulting in flow instabilities that negatively impact overall system stability. Large eddy simulation based CFD can accurately capture this complex behavior but requires fine meshes and short time-steps, leading to high computational costs. In this study, a deep-learning-based reduced-order modeling strategy is proposed to maintain both accuracy and computational efficiency in analyzing local flow regions between HCSG tube layers. POD, DMD, and nonlinear autoencoder are employed to reduce data dimensionality, followed by a Long Short-Term Memory network for predicting flow evolution. These ROM frameworks are compared to reduce simulation overhead while preserving CFD-level predictive accuracy. The results indicate that linear methods effectively capture dominant features such as vortex formation and dissipation, whereas the nonlinear autoencoder tends to preserve finer-scale fluctuations. Notably, the POD-LSTM model demonstrates superior performance in predicting flow field dynamics, achieving R 2 = 0.9701 on the test set (AE = 0.9477, DMD = 0.9409).