A systematic approach to optimization of ANN model parameters to predict flow boiling heat transfer coefficient in mini/micro-channel heatsinks. (March 2023)

• Record Type: Journal Article
• Authors: Qiu, Yue; Vo, Tinh; Garg, Deepak; Lee, Hyounsoon; Kharangate, Chirag R.
• Is Part Of: International journal of heat and mass transfer. Volume 202(2023)
• Publication Date: 2023-03
• Subjects: Machine learning -- Neural networks -- Optimization -- Heat transfer -- Flow boiling; Heat -- Transmission -- Periodicals; Mass transfer -- Periodicals; Chaleur -- Transmission -- Périodiques; Transfert de masse -- Périodiques; Electronic journals; 621.4022
• DOI: 10.1016/j.ijheatmasstransfer.2022.123728
• Languages: English
• ISSNs: 0017-9310

Abstract

Highlights: A systematic approach to machine learning model parameter selection is developed. A database from 50 published sources is utilized for model training. The neural network model was optimized based on prior correlation and data science tools. Optimal models performed outperformed prior universal correlations and prior machine learning models. Abstract: Flow boiling in mini/micro-channel heatsinks is often utilized to meet the high heat dissipation requirements of thermal management systems. However, accurate prediction of the heat transfer coefficients in these flow boiling systems is challenging due to complex two-phase fluid behavior compounded by the thermal complexities of these systems. Traditionally, universal correlations, theoretical models, and computational simulations were utilized in predicting heat transfer even though with lower accuracy. Recent work includes use of machine learning tools in more accurately predicting heat transfer behaviors. In this study, Artificial Neural Network (ANN) is utilized and optimized using a comprehensive physics-based and data-driven approach to predict heat transfer coefficients for saturated flow boiling in mini/micro channels. A database of 16, 953 data points for flow boiling heat transfer in mini/micro-channels amassed from 50 sources was included in the analysis. The predictive capabilities of the ANN-based method are thoroughly optimized for its input parameters and model architecture parameters something not attempted in any past work. A systematic approach to optimization the model parameters is developed based on correlations from prior literature on flow boiling heat transfer coefficients, Pearson coefficient correlations, and mutual information feature selection method. The final optimized 17 input parameters are trained on the ANN model with network hidden layers (10, 20, 50, 100, 200, 400). This test case achieved an MAE of 8.48% far superior to universal correlations or prior machine learning results for saturated flow boiling heat transfer in mini/micro-channels. These results demonstrate great improvements in accuracy, and a potentially useful framework for optimizing machine learning models for predicting heat transfer coefficients that can be implemented on other two-phase flow configurations and parameters.