Predicting Unknown Facts about Flow Events using Convolutional Neural Network
We investigate the possibility of using artificial intelligence to deduce information about unobserved upstream or past events in fluid dynamics. To test the hypothesis, we applied an existing convolutional neural network (CNN), namely GoogLeNet, to predict an upstream object's geometry. Square and circle-shaped bluff bodies were placed in a two-dimensional turbulent flow. Downstream data was collected and fed to train the CNN model. Then the downstream velocity signals of a flow over an object that was unknown to the model were used to predict the object's shape. The CNN model's input was the absolute values of the continuous wavelet transform (CWT) of velocity signals recorded in far wake regions downstream. CWT transforms velocity signals to functions of time and frequency called scalograms. The contours associated with these scalograms, obtained for square and circle-shaped geometries, were used as the input to the CNN model. The trained model predicted that the unknown geometry was 93% similar to a square which was an optimistic prediction as the unknown geometry was a square with a missing corner. This remarkable achievement led us to move forward and use the model for real-world phenomena (e.g., optimizing power production of a wind farm using AI). Currently, we are researching a wind farm of three turbines using CFD to evaluate whether the model can predict which turbines are not facing the incoming wind using only the downstream wind farm's velocity field where the flow is thoroughly turbulent.