.png){kind=icon}
.png){kind=icon}
Return
Deep Learning-Based Method for Efficient Airfoil Rime Icing Prediction
.png){kind=icon}
PRE
AI
DOI：10.1061/JAEEEZ.ASENG-6214.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The input feature parameters of current machine learning and deep learning-based wing icing prediction models primarily consist of flight states and atmospheric conditions, and the output data are ice shapes. This dependency necessitates many samples for model training, resulting in significant computational costs that constrain practical engineering applications. We propose an efficient deep learning-based method for wing icing prediction to address this issue. Our approach utilizes a deep neural network (DNN) model developed in PyTorch, with flow fields initialized by a potential flow solver in OpenFOAM serving as input features. The droplet volume fraction distribution is simulated using the Eulerian-Eulerian model to generate output data. The Latin hypercube method is employed to randomly sample two parameters-angle of attack and flight speed-and the iced airfoils during the icing process are selected as the source of the training data set. The potentialFoam solver initializes the airflow field around the iced airfoil, and the multiphaseEulerFoam solver analyzes the droplet impact characteristics based on this initialized flow field. Through careful input feature design, the fully connected DNN uses initialized airflow variables, gradients, grid point coordinates, and wall distance as input features, and droplet velocity and volume fraction fields as outputs. This framework establishes a nonlinear mapping between local airflow field variables and droplet field variables. The DNN effectively predicted droplet field variables and was validated for its generalization capability using both the clean NACA23012 airfoil and its rime-icing process. The results indicate that with negligible computational cost of droplet field variables, the proposed DNN-based icing prediction model can achieve the same optimization results as the traditional icing numerical simulations.
Keywords:
Airfoil icing
Deep learning
Droplet impact characteristics
PyTorch
OpenFOAM
Journal
J
IF:
1.6
Papers:
81
Citations:
0