Accurate prediction of thermo-hydraulic behavior in internal pipe flows is essential for the design and optimization of heat transfer systems. While computational fluid dynamics (CFD) provides high-fidelity results, its high computational cost limits its use in large-scale parametric studies. This study aims to develop a CFD-driven machine learning surrogate modeling framework for efficient and physically consistent prediction of pressure drop and heat transfer characteristics in circular pipe flows.
A database of 1,728 validated three-dimensional CFD simulations was generated for air and water flows in circular pipes over a wide range of Reynolds numbers (50–50,000), geometries and thermal boundary conditions. The data set was verified against empirical correlations and experimental data from literature. Supervised learning models, including support vector regression (SVR) and artificial neural networks, were trained using systematic grid-search-based hyperparameter optimization with cross-validation. Feature analysis was conducted to assess physical consistency.
The surrogate models achieved high predictive accuracy, with R² values exceeding 0.99 for pressure drop and heat transfer predictions. The SVR–polynomial model performed best for pressure drop, while SVR–RBF showed superior performance for heat transfer parameters. The models captured physically consistent trends and reduced computational time by several orders of magnitude compared to CFD.
This study presents a robust CFD-driven surrogate modeling framework that emphasizes accuracy, interpretability and computational efficiency. Unlike conventional approaches, it integrates systematic validation, fair model comparison and physical consistency analysis, offering a practical alternative to CFD for design optimization and rapid engineering analysis.
