Physics-Guided Neural Network for Predicting the Thermo-Hydraulic Performance of Concentric Tube Heat Exchangers: Toward Improved Prediction Accuracy

Accurate prediction of coupled heat-transfer and fluid-flow phenomena is essential for the thermal design, performance assessment, and optimization of heat exchangers (HXs). Among key HXs, concentric tube heat exchangers (CTHXs) are widely used in thermal energy systems, where their performance is governed by the coupled interaction of fluid flow and heat transfer. Although computational fluid dynamics (CFD) provides detailed insights into these transport phenomena, its high computational cost limits its applicability in design optimization and real-time monitoring applications. To overcome this limitation, the present study proposes a physics-guided neural network (PGNN) for the accurate and efficient prediction of CTHX thermo-hydraulic performance, including the overall heat-transfer coefficient (U) and the pressure drops of the cold (ΔPc) and hot (ΔPh) streams. The PGNN introduces correlation-based physical guidance through established Nusselt number, overall thermal-resistance, and Darcy–Weisbach pressure-drop relations. Accordingly, the proposed framework is a correlation-guided PGNN rather than a residual-based physics-informed model, because the local conservation-equation residuals are not explicitly enforced during training. For comparison, a standard artificial neural network (ANN) with the same architecture and input parameters was also developed. Both models were trained on a dataset generated from 1575 CFD simulations covering a wide range of operating and geometric conditions, including the Reynolds and Prandtl numbers of both fluids, inner and outer tube diameters, and inlet temperatures. A comprehensive error analysis demonstrates the superior predictive capability of the PGNN over the ANN under various flow and geometric conditions. On the unseen test dataset, the PGNN achieved mean absolute percentage errors of 2.03%, 1.09%, and 1.11% for predicting U, ΔPc, and ΔPh, respectively. The proposed PGNN therefore provides a reliable, high-fidelity, and computationally efficient alternative to CFD, supporting the analysis, optimization, and operation of thermal energy systems.