Integrating CFD and Machine Learning for Efficient Thermo-Hydraulic Design Optimization of Roughened Triangular Solar Air Heaters

Abstract

Abstract. A hybrid numerical data-driven framework is proposed to evaluate and optimize the thermohydraulic performance of a roughened rounded corner triangular duct solar air heater. High-fidelity flow and heat transfer data are first generated using a previously validated computational fluid dynamics (CFD) model by systematically varying Reynolds number (5600–21,000), corner radius (0.333–0.67 h), and longitudinal and transverse pitch ratios (z′/e and x/e = 10–18), resulting in a dataset of 60 simulations. This dataset forms the basis for developing surrogate models using multiple supervised machine learning regressors, including linear regression, K-nearest neighbor, random forest, decision tree, multilayer perceptron, and stochastic gradient descent regressor. To ensure robustness and transparency, the machine learning (ML) workflow incorporates feature scaling where appropriate, an 80:20 training–testing split and fivefold cross-validated hyperparameter tuning. Model performance is assessed using R2 and root mean squared error (RMSE) metrics, revealing that tree-based and ensemble models are more effective in predicting Nusselt number due to strong nonlinearity in turbulence–geometry interactions, whereas linear models provide stable and accurate predictions of friction factor owing to its quasi-linear dependence on flow parameters. Feature importance analysis further confirms that Reynolds number primarily governs heat transfer enhancement, while curvature effects significantly influence hydraulic resistance. A desirability-based multi-objective optimization strategy is subsequently applied to identify optimal operating conditions that balance heat transfer augmentation and pressure loss minimization. The optimal configuration predicted by the ML models (z′/e = 10, Rc ≈ 0.34, Re ≈ 21,000) closely matches the CFD-derived optimum, demonstrating strong agreement. Overall, the study establishes a reliable and computationally efficient CFD–ML framework for the design optimization of advanced solar air heaters, reducing dependence on exhaustive numerical simulations while retaining physical fidelity.