Machine learning-based prediction of nucleate pool boiling heat transfer enhancement on micropillar surfaces

This study employs Bayesian optimization to tune four machine leaning models for establishing a mapping between the heat transfer coefficient and the morphology of the heated surface. This is performed considering previous experimental data, which includes 544 samples collected from 16 types of micropillars. Additionally, an analysis is carried out to assess the importance of the input parameters. The results show that the Extra Trees model provides the best predictive accuracy, outperforming three widely used empirical correlations. It attains a coefficient of determination (R²) of 0.99343 and the lowest normalized root mean square error (NRMSE) of 0.082763. Additionally, the top three most important descriptors are mean beam length (MBL), height (h), and capillary resistance number (Cr). Finally, predictions are made using this optimized model with the most influential descriptor configurations across various heat flux conditions. The findings indicate that taller micropillars, greater Cr, and wider MBL improve heat transfer performance under lower heat flux conditions due to lower flow resistance and a larger heat transfer area. In contrast, under high heat flux conditions, shorter micropillars, smaller Cr, and narrower MBL lead to better heat transfer performance due to easier bubble detachment and an enhanced capillary pumping effect.