Comparative Study of Future State Predictions of Unsteady Multiphase Flows Using DMD and Deep Learning

Flow across an array of solid obstructions is a common phenomenon observed in many applications such as multiphase flows, heat exchangers, and environmental flows. In this work, we aim to train deep learning models and to predict the time evolution of unsteady flow fields in a domain of randomly arranged 2D cylinders at Reynolds number 50. Two different approaches are used and compared in this paper, dynamic mode decomposition (DMD) which is a dimensionality-reduction algorithm based on singular value decomposition (SVD) and long short-term memory (LSTM) neural networks. In both cases, the model is trained on the first 165 time steps and then is tested on predicting the next 300 time steps. Two flow fields with different spectral characteristics are used to compare the performance of the two techniques. The LSTM architecture owing to its ability to learn nonlinear dynamics performs better than the DMD algorithm in the case with more temporal time scales present.