03.1 - Aerodynamics – CFD Methods and Validation
APPLICATION OF CONVOLUTIONAL NEURAL NETWORKS IN DETERMINING THE VELOCITY AND PRESSURE FIELDS AROUND AIRFOIL MODELS
G. Sharma¹, T.H. Tran, Le Quy Don Technical University, Vietnam; J. Tanimoto¹; ¹Kyushu University, Japan
With the development of technology, the computational science becomes an important part in the research activities. Nowadays, the application of neutral networks occurs widely in engineering and practically application. It has been solving many problems fast when the network is trained well. The neutral network is divided in many types, such as: Feedforward Neural Network, Multilayer Perceptron convolutional neutral network, Recurrent Neural Network, Autoencoder and Generative Adversarial Network [1] [2]. In this study, we modified the U-net for training surface flow around airfoil to obtain the pressure and velocity vectors. For that purpose, a CNN network was written by MATLAB program for training the airfoil data. rnTraining data created using a Spalart-Almaras Reynolds-Averaged Navier-Stokes (RANS) simulation in Open FOAM. This is two-dimensional flow fields. The data includes 320 samples for training and 80 for testing. For each sample, the inlet data is a matrix of 3 dimensions with size of 128 × 128. The first and second dimensions of the matrix represent the inlet velocity in the x and y directions, respectively. The last dimension signifies the shape of the airfoil. The output data has the same size as the inlet. In more specific terms, the first dimension illustrates the pressure fields surrounding the airfoil. The second and third dimensions depict the velocity around the airfoil in the x and y directions, respectively. An example of inlet and outlet images is presented in Fig. 2 (the first and second rows).rnWe utilize the Adam optimizer for training our architecture. Our implementation relies on the MATLAB deep learning framework [3]. Given the strictly supervised nature of our learning setup, we employ a straightforward L1 loss, denoted as L = |y - a|, where 'a' represents the output of the Convolutional Neural Network (CNN). The ensuing training sessions employ a learning rate set at 0.0004