Use of Genetic Algorithms and Machine Learning to Explore Parametric Trends in Nucleate Boiling Heat Transfer Data

Exploring parametric effects in pool boiling is particularly challenging because the dependence of the resulting surface heat flux on many parameters is non-linear, and the mechanisms can interact in complex ways. Historically, parametric effects in nucleate boiling processes have most often been deduced by fitting relations obtained from physical models to experimental data, or looking for correlated trends in non-dimensionalized data. Using such approaches, observed trends are often influenced by the framing of the analysis that results from the modeling or the collection of dimensionless variables used. Machine learning strategies can be attractive alternatives because they can be constructed either to minimize biases or to emphasize specific biases that reflect knowledge of the physics of the system. The investigation summarized here explored the use of machine learning methods as a tool for determining parametric trends in boiling heat transfer data, and as a means for developing methods to predict boiling heat transfer. Results are presented that demonstrate how genetic algorithms and other machine learning tools can be used to extract heat flux dependencies on system parameters. A key element of the machine learning analysis process is preparation of the data used. Use of raw data and use of dimensionless rescaled data are explored, and the advantages and disadvantages of each are assessed. Data for nucleate boiling of a binary mixture are analyzed to determine the heat flux dependence on wall superheat, gravity, Marangoni effects and pressure. The results provide new insight into how gravity and Marangoni effects interact in boiling processes of this type. The results also demonstrate how machine learning tools can clarify how different mechanisms interact in the boiling process, as well as directly providing the ability to predict heat transfer performance for design of heat transfer devices that involve nucleate boiling. Potential use of machine learning tools on big data collections for nucleate boiling processes to more broadly assess parametric effects is also discussed.