A data-driven prediction of atmospheric flow for urban air mobility using deep neural network

A data-driven prediction of an atmospheric boundary layer in an urban environment modeled by a simplified set of cubes was performed. Gated recurrent unit (GRU) neural network was organized to predict the turbulent flow generated via large eddy simulation (LES). To avoid the curse of dimensionality, singular value decomposition and convolutional autoencoder were combined with the GRU neural network for data compression into latent space. A study under a gradual change of size of latent space was performed to optimize the size of latent space as a function of reconstruction capability and computational efficiency. The prediction model utilized a series of input latent vectors to predict the output latent vectors for the reconstruction procedure. A loss of physical properties originally captured in the LES data was backpropagated in the prediction phase. The prediction model reproduced the instantaneous flow structures and the turbulence statistics from the urban flow data. A computational cost comparison between the prediction model and LES data indicated that the present model has a potential to generate a real-time prediction of high-resolution turbulent flow in an urban environment.