XGBoost algorithm to predict heat transfer coefficient for saturated flow boiling in mini/micro-channels

Abstract

The heat transfer coefficient in saturated flow boiling in mini/micro-channels is a critical factor in the cooling design of high-heat-flux devices. This study proposes a method to accurately predict the heat transfer coefficient in saturated flow boiling using the XGBoost(eXtreme Gradient Boosting) machine learning algorithm. The database used in this study consists of 11,096 pre- dryout data points obtained by removing 1,878 post-dryout data points from a total of 12,974 data collected from 37 sources, employing an XGBoost incipience dryout prediction model. The dataset encompasses 18 working fluids, hydraulic diameters raning from 0.19 mm to 0.65 mm, mass flow rates from 19.45 kg/m2s to 1,608 kg/m2s, and saturation temperatures from -40℃ to 201.37 ℃. When implementing the XGBoost prediction model, training features were selected based on Permutation Feature importance(PFI) and SHapley Additive exPlanations(SHAP) values, and the optimal combination of hyper-parameters was determined using Optuna. The XGBoost model developed in this study, using 8 training features(Pr_f, x_di, Bo, P_red, Pr_g, rho_r, Fr_fo, h_r), achieved a Mean Absolute Percentage Error(MAE) of 7.05%, demonstrating superior predictive performance compared to existing empirical correlations and other machine learning algorithms, including AdaBoost, Gradient Boosting, and ANN. This research confirms that the XGBoost algorithm is an effective and reliable tool for predicting the heat transfer coefficient, overcoming the limitations of existing correlations and providing performance under various operating conditions.