INVERSE DESIGN OF AIRFOILS USING CONVOLUTIONAL NEURAL NETWORK AND DEEP NEURAL NETWORK

In this paper, we compare the efficacy of two neural network based models: Convolutional Neural Network (CNN) and Deep Neural Networks (DNN) to inverse design the airfoil shapes. Given the pressure distribution over the airfoil in pictorial (for CNN) or numerical form (for DNN), the trained neural networks predict the airfoil shapes. During the training phase, the critical hyper-parameters of both the models, namely - learning rate, number of epochs and batch size, are tuned to reduce the mean squared error (MSE) and increase the prediction accuracy. The training parameters in DNN are an order of magnitude lower than that of CNN and hence the DNN model is found to be 7 x faster than the CNN. In addition, the accuracy of DNN is also observed to be superior to that of CNN. After processing the raw airfoil shapes, the smoothed airfoils are shown to yield the target pressure distribution thereby validating the framework.