Akademik Veri Yönetim Sistemi
International Communications in Heat and Mass Transfer, cilt.176, 2026 (SCI-Expanded, Scopus)
Microchannel heat sinks with non-conventional geometries are promising for thermal management, but their strongly coupled and nonlinear thermo-hydraulic behavior makes predictive modeling difficult. This study presents a comparative assessment of machine-learning models for predicting the Nusselt number and Fanning friction factor in wavy convergent–divergent microchannels with embedded streamlined micropins. A numerical dataset comprising more than 700 samples was generated using three-dimensional conjugate heat transfer simulations over a range of Reynolds numbers and geometric parameters, including wave amplitude, waviness coefficient, and pin height. Five regression models—multilayer perceptron, support vector regression, random forest, gradient boosting regressor, and extreme gradient boosting—were developed and evaluated under a common training, validation, and testing framework. The results show target-dependent model performance. For Nusselt number prediction, support vector regression gave the best performance, with cross-validation and test R2 values of 0.9939 and 0.9970, respectively, and test MAPE below 1%. For friction factor prediction, extreme gradient boosting gave the best performance, with cross-validation and test R2 values of 0.9947 and 0.9972, respectively, and a test MAPE of 2.0844. SHAP analysis showed that the trained models captured physically consistent relationships between flow conditions, geometric parameters, and thermo-hydraulic responses. Compared with the empirical correlations, the data-driven models produced lower prediction errors, with most Nusselt number predictions remaining within the ±5% band and most friction factor predictions remaining within the ±10% band over the test cases. These results show that machine-learning models can be used as predictive tools for thermo-hydraulic analysis of non-conventional microchannel heat sinks.