Efficient thermal-hydraulic prediction for PWR fuel assemblies via PCA-RBF neural networks with optimized sampling

Thermal-hydraulic analysis of fuel assemblies in pressurized water reactor (PWR) cores is essential for reactor design and safety evaluation. However, traditional approaches face significant challenges in balancing computational efficiency and predictive accuracy. In this study, we propose a novel computational framework that combines momentum source models (MSM), principal component analysis (PCA) for dimensionality reduction, and radial basis function (RBF) neural networks to efficiently process high-dimensional flow field data. Enhanced Latin hypercube sampling (ELHS) is employed to optimize the selection of boundary condition samples, while PCA reduces the data dimensionality to less than 7% of its original size, thereby greatly decreasing computational complexity. The RBF neural network enables accurate mapping from heat flux and inlet velocity to principal component coefficients in the reduced space, with reconstruction errors well within engineering standards (temperature MSE ≤ 0.4%, pressure MSE ≤ 0.6 Pa). The proposed approach demonstrates excellent agreement with CFD benchmark results in both grid downstream and outlet regions, while reducing prediction time from 200 s to just 0.01 s and improving training efficiency by a factor of 5.2. Furthermore, the scalability of this framework is validated through its successful application to a realistic 5 × 5 rod bundle configuration, highlighting its potential for large-scale and practical engineering scenarios. Overall, this framework establishes a new paradigm for real-time monitoring and optimized control of nuclear fuel assemblies, and can be extended to multi-physics coupling simulations, thereby supporting the intelligent advancement of next-generation nuclear energy systems.