A deep-learning approach for predicting oscillating motion of liquid slugs in a closed-loop pulsating heat pipe

Abstract

This paper presents the very first application of a deep neural network (DNN) model to predict the oscillating motion of liquid slugs in a closed-loop pulsating heat pipe (CLPHP). The time-series data of the positions of liquid-vapor menisci are obtained from flow visualization using a high-speed camera: five liquid slugs in the 5-turn CLPHP are observed to have rapid oscillation with varying amplitude. Time-series analysis is conducted on the flow visualization results by employing the DNN model, which uses a Long Short-Term Memory (LSTM)-based encoder-decoder architecture as a sequence-to-sequence deep-learning framework. From the model, the position of each meniscus is predicted with a single input data of its own (univariate prediction) or with multiple input data of all the menisci (multivariate prediction): It is shown that the predicted values match closely with the measured values for both univariate and multivariate predictions. To quantitatively examine the prediction performance, the average volumetric fraction in the condenser section, a major parameter for the thermal performance of the CLPHP, is calculated using the predicted values of positions of menisci. This model is found to be accurate to within ± 30% in predicting the average volumetric fraction for both univariate and multivariate cases. This study sheds new light on analyzing the dynamics of the complex oscillating motion in pulsating heat pipes.