Estimation of frost layer characteristics during natural convection on vertical and inverted cold surfaces based on smart approaches

This study aimed to make a progress in the modeling of frost layer characteristics during natural convection on vertical and inverted cold surfaces, by directing focus towards the physical steps of frosting process. The modeling phase was relying on four enormous experimental datasets, embracing, (a) 727 data for frost thickness on vertical plates, (b) 735 data for frost thickness on inverted plates, (c) 218 data for frost surface roughness on vertical plates, and (d) 210 data for frost surface roughness on inverted plates. Three heuristic soft-computing paradigms, including multilayer Perceptron (MLP), radial basis function (RBF) and gaussian (GPR) were utilized to establish dimensionless predictive tools for each of the aforementioned four cases. It was found that the RBF paradigm presents the best performances in predicting the testing datasets pertinent to modeling cases (a)-(d), with average absolute relative errors (AAREs) of 1.28%, 2.30%, 4.31% and 3.57%, and R² values of 99.97%, 99.83%, 98.53%, 98.74%, respectively. The recently developed models accurately estimated the values of frost thickness and surface roughness predictions at the critical moments of phase change, dendrite touch and reverse melting. Furthermore, they had excellent performances on the suit of describing the variations of frost characteristics with operational factors at all physical steps of frosting. The results implied the point that the trends of frost growth on vertical and inverted plates are quite different, indicating the significance of plate orientation in determining the characteristics of frost layer. Ultimately, a statistically-based investigation on the literature models revealed that all of them are fraught with deviations in describing the surface roughness and thickness of frost layer accumulated on vertical and inverted plates.