Ice Accretion Prediction Based on Neural Networks

Abstract

Many experimental research efforts in the last two decades have revealed that the complete picture of aircraft ice accretion has many components resulting in a complex physical structure. Although overwhelmingly complex, the icing phenomena need to be understood due to the impact on the aircraft performance. This requires a detailed knowledge of the ice accretion physics, the subsequent flow over the aircraft, and the resulting modified aircraft performance. The experimental and numerical studies to address these issues have their own advantages, disadvantages, and limitations, which further limit the analysis of icing phenomena. The motivation behind this study is the belief that complex phenomena in nature have an orderly structure in the large-scale. Based on this premise, it is thought that icing phenomena also have an orderly, albeit non-linear, behavior which can be modeled by neural networks, which have a proven capability for modeling non-linear systems. The developed methodology in the present study involves the Fourier series expansion of the ice shape after a conformal mapping, which clears the effect of airfoil geometry, and use of the neural networks to model the Fourier coefficients and the extent of the ice shape. The neural network can be trained to make ice accretion predictions, given a set of data including the flight and the atmospheric conditions, the Fourier coefficients and extent of the resulting ice shape. It also provides a statistical output of the relative significance of the input parameters in the training. The preliminary results show that the proposed method has an reasonable accuracy and has merit for further investment since it can be coupled with other systems to create advanced computational ice accretion models and ice protection systems.