Machine learning based models to predict frost characteristics on cryogenic surfaces under forced convection conditions

Frost accumulated on heat exchangers' surfaces is a common phenomenon in air-source heat pumps, affecting the heat transfer coefficient and pressure drop, adversely. The mechanism of frost deposition at cryogenic conditions under forced convection (T_w ≤ 100 ° C) shows remarkable differences from those in ordinary-low temperature. The models for predicting frost characteristics on cryogenic surfaces have been rarely reported in the literature. This study aims to model frost density and thickness over ultra-low temperature plates under forced convection. Accordingly, the intelligent models were developed based on radial basis function (RBF), multilayer perceptron (MLP), and Gaussian process regression (GPR) methods using collected measured data from available sources. Although all smart models presented excellent results for training and testing data, the GPR model for frost thickness and RBF model for frost density were designated as top models with AAREs of 1.06% and 0.83%, and R² values of 99.94% and 99.80%, respectively. The analyzed data were also used for evaluating the precision of earlier models, however, they led to unfavorable results. The developed models exhibited favorable trends under different physical conditions. Besides performing sensitivity analysis, explicit empirical correlations were also obtained with AARE values of 9.38% and 9.68% for estimating frost density and frost thickness, respectively.