A machine learning approach for predicting heat transfer characteristics in micro-pin fin heat sinks

•A universal approach for Nusselt number in micro-pin fin heat sinks for embedded microfluidic cooling is used.•A consolidated database of 906 experimental data of Nusselt number in micro-pin fin heat sinks is created from 15 studies.•A conventional regression method provides poor predictive accuracies with an MAE over 36%.•A newly proposed correlation has the best predictive accuracy with an MAE of 23.7%, which is lower than existing regressions.•ML models show superior performance, with MAEs of 7.9–10.9%, resulting in a threefold enhancement of prediction accuracy. Micro-pin fin heat sinks are receiving attention for their use in the thermal management of high-heat-flux electronics systems since they can help to enhance heat transfer characteristics (owing to their large extended surface area) and flow mixing while requiring relatively low pumping power compared with conventional microchannel heat sinks. Although many studies have determined the thermal performance of micro-pin fin heat sinks over the past several decades, a universal model for predicting the thermal performance of micro-pin fin heat sinks with various geometries and under different operating conditions has not been developed. In this study, we developed universal machine learning models for predicting the thermal performance of micro-pin fin heat sinks of various shapes and under different operating conditions beyond the limits of existing correlations by using power law regression. The database for these models comprised 906 data points amassed from 15 studies. Three machine learning models and a newly proposed regression model were compared with the conventional regression models. The prediction accuracies of each model and complex relations between the geometric shape, operating conditions, and heat transfer performance are discussed by comparing the three machine learning models and the regression model. The machine learning models had mean absolute errors (MAEs) of 7.5–10.9%, representing an approximately fivefold enhancement in the prediction accuracy compared with existing regression correlations. Their MAEs were lower than that of the regression model. Moreover, the machine learning models provided high accuracies for rare geometric shapes and operating conditions, such as a triangular pin shape or the use of R134A as a working fluid. These results showed the superiority of the machine learning models over traditional correlations in terms of the prediction accuracy for the ther