A novel physics-informed ensemble data-driven friction-factor correlation for nuclear fuel rod-bundle channels

摘要: Accurate pressure-drop prediction is essential for coolant-flow allocation, pumping-power estimation, naturalcirculation assessment, and the evaluation of thermal hydraulic safety margins in nuclear fuel assemblies. In rod-bundle channels, frictional resistance reflects the coupled effects of Reynolds-number-dependent wall friction, geometric confinement, narrow-gap shear layers, and cross-sectional flow redistribution, which limit the transferability of conventional friction-factor correlations. A physics-informed ensemble (PIE) friction factor model is developed using 883 single-phase measurements from 11 source-geometry groups. Ten conventional correlations are retained as physically motivated expert laws, while a fully connected network assigns nonnegative, softmax-normalized weights from pitch-to-diameter ratio, Reynolds number, and a one-hot database identifier. The final prediction is a convex combination of the expert laws rather than a direct black-box output. The PIE model achieves an overall mean relative error (MRE) of 1.180%, a root mean square error (RMSE) of 0.000520, and R2 = 0.9955 over all 883 samples. The random 20% validation subset gives an MRE of 1.291%, an RMSE of 0.000577, and R2 = 0.9951. The best fixed correlation, the simplified Rehme-type law, gives an MRE of 4.663%; adaptive expert weighting therefore reduces the overall MRE by 74.7% relative to that baseline. Database-wise MRE ranges from 0.498% to 3.527%. These results demonstrate accurate interpolation within the source–geometry groups represented during training while retaining a direct decomposition of each prediction into conventional friction-law contributions.