Feature extraction of fields of fluid dynamics data using sparse convolutional autoencoder

Abstract

A neural network technique that extracts underlying flow features from the original flow field data is newly proposed. The technique here is based on the convolutional and sparse autoencoder learning algorithms and is called sparse convolutional autoencoder. Unlike the typical convolutional neural network (CNN) that changes the size of the data itself in the intermediate layers, flow field data size is not changed in the learning process of this method and only the numbers of channels are changed in each layer. Different but the same size of the data as the input are obtained by convolution with multiple spatially overlapping flow field data under the assumption of sparsity. When data restoration is realized in this autoencoder system, the channel numbers of data in the intermediate layers turn out to contain different flow characteristics of the original flow field. The proposed method is applied to the low Reynolds number flows over a circular cylinder. The high-fidelity unsteady flow data obtained by solving two-dimensional compressible Navier–Stokes equations with a high-resolution numerical scheme are used as a test case. In the proposed method, sparsity introduced in the middle-hidden layer is essential for the successful separation of the original data. The results presented in the example seem to correspond to positive and negative magnitudes of the original data, but future studies will reveal other features of the method. The present method shows flow features different from those of proper orthogonal decomposition in each mode, which is probably due to nonlinear decomposition in the CNN process.