Machine learning algorithms to predict flow condensation heat transfer coefficient in mini/micro-channel utilizing universal data

• Published in: International journal of heat and mass transfer Vol. 162; p. 120351
• Main Authors: Zhou, Liwei, Garg, Deepak, Qiu, Yue, Kim, Sung-Min, Mudawar, Issam, Kharangate, Chirag R.
• Format: Journal Article
• Language: English
• Published: Oxford Elsevier Ltd 01.12.2020; Elsevier BV
• Subjects: Algorithms; Computational fluid dynamics; Condensation; Condensers (liquefiers); Data points; Fluid flow; Gradient boosted trees; Heat transfer; Heat transfer coefficients; Machine learning; Mathematical models; Microchannels; Model accuracy; Neural networks; Optimization; Reynolds number; Working fluids; Gradient boosted trees; Neural networks; Condensation; Heat transfer; Machine learning
• ISSN: 0017-9310; 1879-2189
• DOI: 10.1016/j.ijheatmasstransfer.2020.120351

•A large database for flow condensation heat transfer in mini/micro-channels is amassed.•Four machine learning algorithms are developed: ANN, AdaBoost, Random Forest, and XGBoost.•ANN and XGBoost models predict test data with MAE of 6.8% and 9.1%, respectively.•Optimal models performed better than prior universal correlations.•Models could predict data points from databases not included in the training database. Miniature condensers utilizing mini/micro-channel has been recognized as one effective technique for designing a compact heat rejection device. However, because of the complex behaviors in phase-change systems like flow condensation, accurately predicting heat transfer coefficients can be a challenging task. In this study, a large database is utilized to develop machine-learning based models for predicting condensation heat transfer coefficients in mini/micro-channels. A consolidated database of 4,882 data points for flow condensation heat transfer in mini/micro-channels is amassed from 37 sources that includes 17 working fluid, reduced pressures of 0.039 – 0.916, hydraulic diameters of 0.424 mm – 6.52 mm, mass velocities of 50 < G < 1403 kg/m2s, liquid-only Reynolds numbers of 285 – 89,797, superficial vapor Reynolds number of 44 – 389,298, and flow qualities of 0 – 1. This consolidated database is utilized to develop four machine learning based models viz., Artificial Neural Netoworks (ANN), Random Forest, AdaBoost and Extreme Gradient Boosting (XGBoost). A parametric optimization is conducted and ANN and XGBoost showed the best predicting accuracy. The models with dimensionless input parameters: Bd, Co, Frf, Frfo, Frg, Frgo, Ga, Ka, Prf, Prg, Ref, Refo, Reg, Rego, Suf, Sug, Sufo, Sugo, Wef, Wefo, Weg, and Wego predicted the test data for ANN and XGBoost models with MAEs of 6.8% and 9.1%, respectively. The optimal machine-learning models performed better than a highly reliable generalized flow condensation correlation. Models were also able to predict excluded datasheets with reasonable accuracy when data points including the specific working fluid were part of the training dataset of the remaining datasheets. The work shows that machine learning algorithms can become a robust new predicting tool for condensation heat transfer coefficients in mini/micro channels.