Çorumlu V.Altıntaş V.Abuşka M.2024-07-222024-07-22202407351933http://akademikarsiv.cbu.edu.tr:4000/handle/123456789/11492The capability of ML models in thermal systems is generally determined by internal validation, while this study investigates the prediction performance of ML models with external validation. ANN, XGBoost, and RF models were created with the training-test data set obtained from the results of flat, conical, and cross-cut pin fin heat sinks. Data of 33, 66, and 99 W for the training-test data set were used for training and internal validation, while data of 49.5 W for conic Model-I and 82.5 W for conic Model-II were used for external validation. The RF showed the highest performance on the test data-internal validation and the ANN on the external validation. According to the test data not used in training, the lowest MSE is 0.0270-(RF), 1.7437-(ANN), and 14.7140-(XGBoost). In RF and XGBoost, the external validation performance decreased significantly compared to the internal validation. The MSE of the models are 8.0683-ANN, 214.4047-XGBoost, and 300.6012-RF for external validation. The thermal resistance provides more realistic results than the Nusselt number for the thermal performance evaluation of heat sinks with ML methods. The ANN based on external validation may be used to predict heat sinks' thermal performance and save money, labor, and time compared to CFD simulations. © 2023 Elsevier LtdEnglishForecastingHeat resistanceHeat sinksMachine learningStatistical testsConic modelData setExperimental forced convectionExternal validationMachine-learningPerformance parametersPrediction performanceTest dataThermal PerformanceThermal performance parameterForced convectionArticle10.1016/j.icheatmasstransfer.2023.107228