Wire-wrapped fuel assemblies are characterized by its excellent heat transfer properties, high power density and the ability to reduce coolant flow resistance. Consequently, they have been widely studied in the context of Generation IV reactors, particularly Lead-cooled Fast Reactors (LFRs),which typically use lead-bismuth eutectic alloy (LBE) as a coolant. The high density, high viscosity and low Prandtl number of LBE complicate the flow process. Although computational fluid dynamics (CFD) methods can provide accurate simulations, they require complex boundary condition configurations and substantial computational resources, particularly for multi-condition and steady-state analyses.
Neural networks can fit complex nonlinear mappings. Consequently, these models can extract characteristic correlations between input conditions and target physical parameter fields, thereby enabling the efficient prediction and reconstruction of complex physical parameter distributions. In this paper, a U-Net model based on a convolutional neural network (CNN) architecture is established to predict the temperature and axial velocity parameters distribution fields within wire-wrapped fuel rod bundle channels. Firstly, a high-fine-grained multi-condition training dataset is created using the CFD method and a 7-bar bundled wire fuel assembly is used as the research object. After validating the simulation methodology's rationality and accuracy, a baseline temperature boundary condition of 473.15 K was established, with relative fluctuation ranges of ±10%, ±20%, and ±30% subsequently applied. Additionally, the inlet velocity boundary condition was varied from 1 m/s to 10 m/s in steps of 1 m/s to simulate the impact of variations in inlet temperature and axial velocity on flow and heat transfer characteristics. At the same time, stratified random sampling was carried out within the specified range of boundary conditions to create an interpolation test set. This set included operating conditions that were not part of the model's training data.
Next, a total of 1,380 samples from three different cross-sections at different heights are selected for network training to identify patterns in cross-sectional temperature and axial velocity distributions. The results show that the U-Net model exhibits generalization capabilities across varying boundary conditions. The root mean square error of the predicted and reconstructed temperature and velocity fields is less than 1%, while the peak signal-to-noise ratio exceeds 35 dB. Meeting the requirements for prediction accuracy, the prediction time for a single working condition is less than 2 seconds. The prediction speed is 3 orders of magnitude faster than traditional CFD methods, which greatly improves prediction efficiency.
In summary, a deep learning neural network model based on a CNN with skip connections was established. Its ability to predict parameter fields under operating conditions not observed in the CFD simulation results for a wound fuel rod bundle are analyzed. The findings of this study demonstrate the effectiveness of the U-net in predicting flow fields within fuel rod bundles and in making design optimization decisions. This paper can provide reference for thermal-hydraulic analysis and prediction models of fuel channels based on other data-driven methods.
 

Deep Learning; Rapid Prediction; Flow Field Distribution; Liquid lead-Bismuth Eutectic(LBE); Wire-Wrapped rod bundle; CFD