Machine Learning Prediction of Wind-Driven Water Runback Characteristics Related to Aircraft Icing

What is it about?

Aircraft icing is widely known as one of the most severe sources of danger for aviation in cold weather, which is remarkably influenced by the transient runback behavior of unfrozen water flow over the surface of the aircraft. The characteristics of wind-driven runback water flows on airframe surfaces are highly depending on the intricate interactions among incoming airflow, surface water film flow, and solid airframe surface. Compared to conventional theoretical and numerical methods, machine learning-based approaches are an appealing option for the sake of the intricate relationship. In the present study, data-driven-based machine-learning (ML) methods are used to uncover the intricate interactions among complicated multiphase systems (i.e., air, water and airframe surface) to predict the characteristics of the wind-driven water runback process. We utilize XGBoost, LightGBMand Multilayer Perceptron (MLP) to make predictions of the front contact line location of wind-driven water runback flows over a flat test plate, with the given input conditions of freestream wind velocity, the flow rate of surface water film, and the duration of the water film flow exposed to the boundary layer airflow over the test plate. For the evolution of the water film flow thickness field, we proposed a hybrid deep neural network ConvLSTM-UNet to learn and predict the complex spatial-temporal characteristics of the water film flow. Predictions on the front contact line location and thickness field of the wind-driven water film flow are used to demonstrate the proposed ML approach. By analyzing the errors between the predictions and the ground truth from the experiments, our results demonstrated that the ML-based approach can be used to predict the transient behavior of wind-driven water runback flows with reasonably good accuracy.

Featured Image

Photo by NOAA on Unsplash

Photo by NOAA on Unsplash

Why is it important?

This study was motivated by the reality that tiny variations in atmospheric conditions (i.e., freestream wind velocity or the flow rate of the water film) could result in noticeable changes in the characteristics of water film flow, i.e., water film thickness distribution, the evolution of front contact line of water film, etc. The intricate relationship between atmospheric conditions and the characteristics of water film flow makes machine learning-based methods an appealing option. The present study takes a first step to the establishment of data-driven methods to predict the evolution of front contact line and water film thickness distribution of runback water film flow by investigating intricate patterns among free stream wind velocity, the flow rate of the water film and exposure time. Furthermore, the ConvLSTM-UNet is a new hybrid deep neural network developed to predict spatial-temporal aspects of unsteady flow fields quickly and accurately, which utilizes convolutional encoder-decoder along with convolutional long short-term memory method to reduce dimensionality in high-dimensional unstable flow datasets and predict the spatial-temporal dynamics features in the future based on the captured features from flow data in the past.

Perspectives