Fluid flow state estimation from sparse sensor measurements using convolutional neural network

Abstract

State estimation from limited measurements can widely be found in various fields such as engineering, biology, and economics. In fluid mechanics, it has also contributed for flow control and experimental data processing. However, its application to fluid flows is particularly difficult due to their nonlinearities and high-degrees of freedom in space and time. Neural networks (NNs) have recently been recognized as a powerful tool to tackle the aforementioned challenges. In this presentation, we examine the applicability of NNs to various fluid flow estimations from a practical viewpoint. We use three types of unsteady laminar and turbulent flows for the present demonstration: flow around a square cylinder, turbulent channel flow, and transitional boundary layer. We utilize convolutional neural network (CNN) to estimate velocity fields from sensors arranged in a cross-section. We assess the practicability of the models by investigating physical quantities required for the input and robustness against lack of sensors. In addition, the effectiveness of several approaches to increase the robustness of the models is also examined.