A fused super-resolution network and a vision transformer for airfoil ice accretion image prediction

Designing large neural network models targeting high accuracy and low sample number requirement is an important task in accelerating airfoil ice accretion calculation. Both convolutional neural network and Transformer are potential large-scale architectures for icing image generation. In this study, FSRN (fused super resolution network) and ViT (Vision Transformer) are designed and applied to predict icing images with complex shapes. The novelty of FSRN lies in dividing the training process into main feature generation and detail information processing, and three branches of SR (super-resolution) networks are applied for multi-inputs and fused to generate icing images. The core part of ViT is the large-scale attention structure which is proven to be powerful in learning features. Both 7-dimensional icing condition vectors and the airfoil image are set as inputs to predict 136x136 icing images, and FocalLoss is chosen as the loss function. FSRN achieves determination coefficients of 0.974 and 0.958 respectively while ViT achieves 0.967 and 0.953 in two validation data sets. FSRN demonstrated greater potential in reducing sample number compared to ViT and MLP (multilayer perceptron).