Dimensional Analysis in Error Reduction for Prediction of Nucleate Boiling Heat Flux by Artificial Neural Networks for Limited Dataset

In this work, artificial neural networks (ANNs) is used to predict nucleate boiling heat flux by learning from a dataset of twelve experimental parameters across 231 independent samples. An approach to reduce the number of parameters involved and to increase model accuracy is proposed and implemented. The proposed approach consists of two steps. In the first step, a feature importance study is performed to determine the most significant parameters. Only important features are used in the second step. In the second step, dimensional analysis is performed on these important parameters. Neural network analysis is then conducted based on dimensionless parameters. The results indicate that the proposed feature importance study and dimensional analysis can significantly improve the ANNs performance. It also show that model errors based on the reduced dataset are considerably lower than those based on the initial dataset. The study based on other machine learning models also shows that the reduced dataset generate better results. The results conclude that ANNs outperform other machine learning algorithms and outperform a well-known boiling correlation equation. Additionally, the feature importance study concludes that wall superheat, gravity and liquid subcooling are the three most significant parameters in the prediction of heat flux for nucleate boiling. Novel results quantifying parameter significance in surface tension dominated (SDB) and buoyancy dominated (BDB) boiling regimes have been reported. The results show that surface tension and liquid subcooling are the most significant parameters in SDB regime with a combined contribution percentage of 60%, while wall superheat and gravity are the most significant parameters in BDB regime with a combined contribution percentage of 70%.