Machine learning-based prediction of heat transport performance in oscillating heat pipe

An oscillating heat pipe (OHP) is a highly efficient cooling system for densely integrated electronic and electric devices operating at high frequencies with high heat generation densities. However, because of the complicated internal flow with phase changes, it is difficult to predict the heat transport performance of OHPs accurately. Such predictions are needed to understand the fundamental phenomena in heat transport and optimize the OHP design parameters. The objective of this study is to predict the three prediction targets comprising the internal flow pattern, wall temperature difference between the cooled and heated sections, and heat transport rate of the OHP through machine learning in recurrent neural networks. Experiments on OHP with ethanol were performed for the heat input range of 62-125 W to obtain time series data of the internal flow pattern images, wall temperatures, and cooling water temperatures. The internal flow pattern images were processed by semantic segmentation and subsequently used for training the models for each prediction target. The internal flow patterns were recursively predicted using the trained model. The predicted internal flow patterns were then input into the wall temperature difference and heat transport rate models to predict these two prediction targets. The predicted and experimental time series data for each prediction target were compared, and the prediction ability of the machine learning-based procedure was demonstrated by the quantitative agreement between the experimental and predicted statistical values.