A novel forecast framework for unsteady flows based on a convolutional neural network

Abstract

Fluid mechanics is an important area where deep learning produces excellent results and can bring about scientific innovation because of its high dimensionality, significant nonlinearity, and ability to process an enormous amount of data. Deep learning technology is currently being used to study fluid mechanics, and its application potential is gradually being demonstrated. We propose a novel multi-resolution convolutional interaction network (MCIN), a hierarchical forecast framework based on a convolutional neural network. This structure can capture temporal dependencies at multiple temporal resolutions to enhance the forecasting performance of the original time series. The high-dimensional data of the flow around a cylinder are projected into a low-dimensional subspace using a variational autoencoder (VAE) as a nonlinear order-reduction technique. Then, the data of the subspace are used as the input to MCIN to forecast future velocity fields. The proposed MCIN is compared to non-intrusive reduced-order models based on dynamic mode decomposition and long short-term memory, combined with a VAE. The results demonstrate that MCIN has superior stability to other models in forecasting the evolution of complicated fluid flows and has the potential to forecast a greater number of future outcomes.