Data-driven prediction of unsteady wake flow using convolutional neural networks

Abstract Recent advances in convolutional neural networks have provided new approaches to model fluid motions. In this presentation, we introduce convolutional neural networks to predict two- and three-dimensional unsteady wake flow. The capabilities of the networks on predicting wake flow dynamics are investigated. We also discuss how a network transforms its input flow information through convolutional layers to achieve future prediction by interpreting the feature maps and kernels learned by the network. Based on the interpretations, a method to reduce the size of a network is proposed. The present finding is expected to be useful for deepening our understanding of the predictive capabilities of CNNs for learning fluid flow and developing neural networks with reduced dependency on trial-and-error type efforts.

Dr. Sangeung Lee will start his postdoc at FLOW from 1st March