Machine-Learning and Experimental Study on Predicting the Heat Transfer Coefficient in Vertical and Microgravity Flow Boiling Using Horizontal Flow Data

Abstract

Abstract. This study aims to develop a highly accurate prediction method for heat transfer characteristics across the entire range from low quality to the postdryout region, without dependence on refrigerant properties, flow conditions, gravity orientation, and gravity environment. The proposed novel machine learning approach was combined with experimental investigations. A new experimental facility was developed to measure the boiling heat transfer coefficient of the hydrofluoroolefin (HFO) refrigerant R1233zd(E) in upward flow from low quality to the postdryout region. Experiments were conducted under heat flux conditions of 3–9 kW m−2 and mass flux conditions of 30–90 kg m−2 s−1, providing new data in parameter ranges insufficiently reported in previous studies. The experimental results deviated from the trends predicted by existing correlations, and the effects of heat flux and mass flux on the onset of dryout were clarified. A comprehensive database was constructed by combining 1433 points obtained in this study with literature data, yielding 3289 points for upward flow. Additional databases were compiled: 467 points for downward flow and 222 points for microgravity. Although machine learning models typically require large datasets, their prediction accuracy deteriorates when the available data are limited. To address this issue, the proposed method performs pretraining using horizontal-flow data, which are closely related to the heat transfer characteristics of vertical upward/downward and microgravity flows, followed by fine-tuning with the target datasets. The resulting model accurately predicts heat transfer coefficients from low quality through the postdryout region without dependence on refrigerant properties, flow conditions, gravity orientation, and gravity environment.