Hybrid Machine Learning and Computational Fluid Dynamics Framework for Accurate and Physics-Informed Prediction of Heat Transfer Performance

Abstract

Heat transfer prediction in turbulent pipe flow is needed for heat exchangers, thermal management systems, industrial process equipment and energy conversion technologies design and operation. CFD methods with detailed information on velocity, pressure and temperature fields are available but expensive when many operating conditions have to be solved; empirical correlations are computationally cheap but lose accuracy outside their range of applicability; ML is an alternative for rapid prediction but models trained only on numerical data may reproduce CFD errors, lack extrapolation capability and violate physical relationships.

In this paper we propose a hybrid CFD–machine-learning framework for predicting the thermal-hydraulic performance of turbulent flow through smooth circular pipe. Three-dimensional or axisymmetric CFD simulations (ANSYS Fluent) will be performed with conservation equations for mass, momentum and energy and appropriate turbulence closure. A systematically designed dataset will be generated by changing Reynolds number, Prandtl number and thermal boundary conditions over an operating envelope. The main prediction targets are Nusselt number and Darcy friction factor. Artificial neural network, random forest, gradient boosting and support vector regression models will be compared to heat transfer and pressure loss correlations. The hybrid model will include dimensionless physical variables, baseline predictions based on correlations and physics consistency penalty. Coefficient of determination, root mean square error, mean absolute error, relative error, pressure drop consistency and independent validation cases will be used to evaluate performance. Feature attribution and uncertainty estimation will be used to assess interpretability and reliability respectively. The framework is intended to provide reproducible pathway towards fast surrogate thermal hydraulic models keeping physical constraints and traceability to CFD predictions. Quantitative performance claims will only be made after simulation, validation and independent testing are completed.