Intelligent computing approaches to forecast thickness and surface roughness of frost layer on horizontal plates under natural convection

There is no existing predictive model for frost layer thickness and surface roughness on horizontal cold plates under the natural convection conditions. Accordingly, intelligent approaches were designed based upon 782 data for frost thickness and 191 data for frost surface roughness, which covered four stages of frosting process. Three machine learning methods of multilayer perceptron (MLP), Gaussian process regression (GPR), and radial basis function (RBF) were employed to design the predictive models for frost characteristics over horizontal cold plates in the natural convection environment. For the frost thickness, although almost all models provided excellent outputs, the RBF based model showed the highest accuracy with average absolute relative error (AARE) and coefficient of determination (R²) values of 1.23% and 99.93%, respectively, for the tested data. The RBF based model presented the superior results for frost surface roughness with an AARE of 1.21% for all analyzed data. The proposed predictive methods were capable of predicting the impact of surface temperature on the frost characteristics at various stages of the process. A statistical analysis of earlier correlations revealed large deviations from the measured data caused by the differences in operating conditions. Thus, new explicit correlations were developed using the intelligent method of genetic programming, which showed the AAREs of 4.61% and 16.72% for frost thickness and frost surface roughness, respectively.