Higher-Order Accurate Neural Network For Real-Time Fluid Velocimetry

Abstract

In the present work, we introduce a novel lightweight neural network for fluid velocimetry called LIMA (Lightweight Image Matching Architecture) designed and optimized for PIV, which can potentially fit on low-cost computer hardware. We compare two versions of the network: LIMA-4, a 4-level architecture focused on fast reconstruction; and LIMA-6, a 6-level architecture focused on maximizing accuracy. We demonstrate the new approach provides more accurate prediction with fewer network parameters and faster inference speed. Furthermore, we quantified the disparity error using uncertainty quantification (UQ) by image matching to assess the prediction accuracy of the network. We assess the performance of a synthetic direct numerical simulation (DNS) dataset and a wind tunnel measurement dataset of flow past a cylinder. In all cases, we validate that the new network shows higher accuracy than the previous state-of-the-art neural network (PWCIRR) and also the classic particle image velocimetry (PIV) approach. In the future, we envision deploying the lightweight architecture on low-cost devices to provide affordable, real-time inference of the flow field during PIV measurements.