Predictions on frost growth over a flat plate using surface characteristics: Machine learning methods

Abstract

A transient model of frost growth on a flat plate was developed, taking the surface characteristics of the plate into consideration. Five regression models were applied, three traditional (Multiple Linear, LASSO, Ridge) and two machine-learning (Artificial Neural Network, Support Vector Machine) regression models. The training database was established using data extracted from previously published experimental studies. The experimental data consisted of forced convection (1067 data points) and natural convection data (992 data points). The model outputs were then evaluated using common statistical indicators and the best performing model was selected. The highest R2 values for the artificial neural network were 0.9899 and 0.9944 for forced and natural convection, respectively, after removal of outliers. The frost thickness was predicted for various conditions, including different surface characteristics. The model produced good predictions despite the occurrence of nonlinear complex growth mechanism dependent on various conditions in the dataset obtained.