Abstract
Accurate calibration of interfacial-force-related parameters remains a key challenge in gas–liquid two-phase flow simulations, due to the high computational cost of CFD and the limited availability of experimental data. In this work, a surrogate-assisted parameter inversion framework is developed to enable efficient experiment-driven calibration of void-fraction curves by coupling machine learning with high-fidelity CFD validation. A CFD database is first constructed, incorporating both discrete closure models and continuous parameters. Three surrogate models—Transformer, Random Forest, and Ridge Regression—are trained and evaluated using a group-based data-splitting strategy to ensure reliable cross-condition generalization. The results show that Random Forest achieves the highest prediction accuracy on the CFD database, whereas Transformer provides superior performance in the inversion task, yielding the lowest fitting error with respect to experimental data. This discrepancy highlights that optimal models for forward prediction are not necessarily optimal for parameter inversion. In the sensitivity analysis of the selected parametric samples, the turbulent dispersion Prandtl number is the variable that exerts the greatest influence on the mean void fraction. Based on experiment-driven inversion and subsequent local CFD validation, the feasible solutions are found to cluster within a stable parameter region rather than a single optimum, indicating inherent non-uniqueness in the calibration problem. Overall, the proposed framework demonstrates that surrogate models can effectively compress the high-dimensional parameter search space, while high-fidelity CFD remains essential for final validation. The results provide both a practical workflow and a methodological insight for data-driven calibration of two-phase-flow models.
 

two-phase flow; void fraction; surrogate model; parameter inversion; CFD validation