BPNN and CNN-based AI modeling of spreading and icing pattern of a water droplet impact on a supercooled surface

Abstract

A water droplet impacting on a supercooled surface normally experiencing spreading and freezing is a complex process involving fluid flow, heat transfer, and phase change. We established two models to, respectively, predict the spreading dynamics of a water droplet impact on a supercooled surface and classify the icing patterns to predict the corresponding surface supercooling degree. Six important factors are used to characterize droplet spreading, including Reynolds number, Weber number, Ohnesorge number, surface supercooling degree, the maximum spreading factor, and the dimensionless maximum spreading time. A Back Propagation Neural Network model, including four inputs and two outputs, is established, containing a hidden layer with 15 neurons to perform the non-linear regression training on the spreading factors of 778 groups of an impact water droplet. The trained model is adopted to predict the spreading factors of 86 groups of a water droplet impact on the supercooled surface. The second model is developed to discern and classify the experimentally captured three different icing patterns. Different clustering methods are performed on 116 icing images, including gray-scale and red-green-blue (RGB) clustering. Then, two convolution neural network models of VGG-19 (Visual Geometry Group-19) and VGG-16 are established to classify, train, and test the icing images by gray-scale and RGB clustering methods. The K = 2 gray-scale clustering and the VGG-19 model exhibits the highest accuracy at 90.57%. The two models developed in this study can, respectively, predict the essential factors characterizing spreading dynamics of an impact droplet on a cold surface and predict surface supercooling degree based on an icing pattern.