*WINNER* Predicting unknown upstream events using Convolutional Neural Network

Abstract

Convolutional Neural Network (CNN) is a tool that one can use to deduce information about unknown upstream events in fluid dynamics. We applied an existing CNN, GoogLeNet, to predict the shape of upstream obstacles disturbing the flow using the information collected downstream. Flow over six two-dimensional geometries was simulated to collect data needed for training the model. The geometries included a triangle, square, pentagon, hexagon, heptagon, and octagon. The input of the CNN model was the absolute values of the continuous wavelet transform (CWT) of velocity signals recorded in near, middle, and far wake regions downstream of the studied cases. CWT transforms velocity signals into functions of time and frequency called scalograms. These scalograms were then used to train the model. This study used 420 signals (either in the near, middle, or far wake region). Then, ten random signals that did not participate in the training (hence were unknown to the trained model) were employed to evaluate the model's performance. The model successfully predicted the shape of each geometry in each unknown case.