Analysis of a convolutional neural network for predicting unsteady volume wake flow fields

Abstract

A predictive convolutional neural network is developed to predict the future of three-dimensional unsteady wake flow from past information of flow velocity and pressure. The developed network is found to be capable of predicting vortex dynamics at distinctive flow regimes with flow structures at different scales. Mechanisms of the network on predicting vortex dynamics at two distinctive flow regimes, the mode-B shedding regime and the turbulent wake regime, are investigated. Information in feature maps of the network is visualized and quantitatively assessed to investigate the encoded flow structures. A Fourier analysis is conducted to investigate the mechanisms of the network on learning fluid motions with distinctive flow scales. The transformation of information from the input to prediction layers of the network is tracked to examine how the network transforms the input information for prediction. Structural similarities among feature maps in the network are evaluated to reduce the number of feature maps containing redundant flow structures, which allows reduction of the size of the network without affecting prediction performance.