Hui Hu

The unsteady runback behavior of wind-driven runback water film (WDRWF) flows over aircraft surfaces has a significant impact on the aircraft icing process, one of the most significant aviation hazards in cold weather. The limited understanding of the complex multiphase interactions between freestream airflow, water film motion, and solid airframe surface makes conventional theoretical/numerical methods unable to precisely simulate WDRWF flow. Machine learning-based techniques can accurately capture complex physics using data, making it an attractive alternative to conventional methods. In this study, machine learning methods are used to predict the evolution of the front contact point (FCP) of WDRWF flow and film thickness distribution (FTD) of WDRWF flow. For FCP prediction, the performance of the Light Gradient-Boosting Machine (LightGBM) and Multi-Layer Perceptron is compared quantitatively. They perform well in capturing intermittent and smooth features, respectively. For the prediction of the spatial-temporal evolution of FTD, a computationally efficient deep neural network architecture named ConvLSTM-AutoEncoder was developed, which predicts a future FTD based on a sequence of FTDs in the past. The robustness of the ConvLSTM-AutoEncoder model to noisy input FTD is demonstrated. The generalizability of the three models is evaluated by applying the trained models to unexplored datasets. Based on the proposed techniques' generalizability, robustness, and computational efficiency, machine learning-based methods are demonstrated to be powerful tools in predicting the complex unsteady characteristics of the multiphase WDRWF flows.