Prediction and uncertainty quantification of saturated flow boiling heat transfer coefficients in mini/micro-channels using Gaussian Process Regression

Abstract

Accurate prediction of the heat transfer coefficient in saturated flow boiling within mini/micro-channels is essential for designing high-heat-flux thermal systems. This study develops a Gaussian Process Regression (GPR) model using a consolidated database of 11,469 pre-dryout data points from 41 sources, covering diverse working fluids and operating conditions representative of practical mini/micro-channel applications. Feature selection was conducted using permutation feature importance and Shapley additive explanations, and hyperparameters were optimized by maximizing the log marginal likelihood with an L-BFGS optimizer and early stopping. Compact GPR models were constructed for various covariance kernels, among which the ARD exponential kernel achieved the best balance between predictive accuracy and uncertainty calibration by capturing regime-dependent non-smooth variations in boiling heat transfer. The optimized GPR model achieved a competitive test MAE of 8.45% across the database. For unseen data, it yielded an MAE of 17.4%; Although the Fang et al. (2017) correlation showed a lower point error, the GPR model retained clear advantages in interpretability, flexible retraining, and uncertainty quantification. These results demonstrate that the proposed framework provides a robust, uncertainty-aware data-driven model for saturated flow boiling.