Development of surrogate models for unsteady flow fields using a deep neural network

In this study, we show the application of deep learning in developing a surrogate model for unsteady flow problems, with a focus on flows past single or multiple moving bodies in 2D. Our proposed surrogate model is developed in two stages, based on data from direct numerical simulations of the Navier-Stokes equations. First, we use Proper Orthogonal Decomposition (POD) as an invertible operator to transfer information from the high-dimensional physical space (i.e. the simulation data) to a lower-dimensional latent space and vice versa. Then, we use the Long-Short Term Memory (LSTM) recurrent neural network to extract non-linear dynamics in the temporal solutions in the latent space. We show that this surrogate model is capable of dynamic reconstruction and prediction of unsteady flow fields involving the flow past flapping ellipses, achieving a speedup of at least two orders of magnitude compared with the numerical simulations. We will discuss the possibilities offered by this surrogate model approach for design, optimization, and control of challenging unsteady flow problems.