An Improved Deep Convolutional Neural Network to Predict Airfoil Lift Coefficient

Abstract

Recently, significant developments in deep learning have made many possibilities in the field of fluid mechanics. This paper presents a new method of using improved convolutional neural network to learn airfoil lift coefficient calculated by OpenFOAM simulation tool. We propose a “feature-enhanced-image” data preprocessing method to prepare the training and testing data set. A novel convolutional neural network is designed which uses deeper convolution and pooling layers coupled with batch normalization technique. In addition, before linear regression, in fully connected layers, we use dropout method to reduce the risk of over-fitting. Mini-batch stochastic gradient descent (SGD) optimization algorithm is chosen, and mean square error (MSE) is used to do the model evaluation when training and testing the model. It is demonstrated that this improved deep convolutional neural network (IDCNN) provides more accurate lift coefficient prediction compared to other state-of-the-art neural networks. We also test the effect of batch size and full batch normalization implementation on the performance of the whole convolutional neural network. Finally, it is concluded that the best predicting performance is achieved in the condition of 10 batch size and the mean square error of blind test can reach $3.1\times 10^{-4}$. Furthermore, the “feature-enhanced-image” method we proposed can achieve $85.2 \%$ decreasing of testing MSE.