Physics-Guided Transfer Learning for Void Fraction Prediction

The drift-flux model is fundamental in determining the void fraction, a crucial parameter that significantly impacts pressure drops and heat transfer in two-phase systems. In this study, we leveraged machine learning (ML) to improve void fraction prediction across different combinations of liquids, gases, and flow orientations. Using over 3000 data points from different sources, we trained neural networks (NNs) and random forests (RFs), along with their physics-guided versions, and tested the results against physical correlations. The physics-guided models were created by integrating the difference between the measured void fraction and the physical model’s prediction into the loss function.This approach reduced the full-range root-mean-square error (RMSE) relative to the selected physical correlation baselines. For example, the RF residual correction reduced the upward-flow RMSE of the Ishii correlation from 0.3129 to 0.0481, corresponding to an 84.6% reduction, and reduced the downward-flow RMSE of the Ishii correlation from 0.450 to 0.071, corresponding to an approximately 84% reduction.To further improve prediction in the data-scarce low–void fraction region, transfer learning was applied. For NNs, this involved freezing a network pretrained on the high–void fraction source domain and training newly added layers using low–void fraction data; for the RF and XGBoost (extreme gradient boosting) models, domain adaptation was implemented using CORAL (CORrelation ALignment). For example, on the upward low-void test set, transfer learning reduced the mean absolute relative deviation of the NN–Ishii model from 238.1% for the high-void–trained baseline to 54.39% for the transfer-learned model, corresponding to a 77.2% reduction. Low-void performance depends on the ML architecture, the physical correlation model used for the residual correction, and the selected reference model; therefore, all other reported improvements are evaluated relative to explicitly identified low void–trained, high void–trained, or physical correlation baselines.