Machine Learning Approach for Flow Fields Over a Circular Cylinder Based on Particle Image Velocimetry Measurements

Abstract

Flow around a circular cylinder has been experimentally studied at Reynolds numbers of Re = 4 × 103 and Re = 8 × 103 using Particle Image Velocimetry (PIV). The Artificial Neural Network (ANN) method, a subset of machine learning, has been implemented to capture complex flow patterns and relationships within data to estimate the experimental results. Experimental and estimated results regarding instantaneous contour graphics for streamwise (u) and cross-stream (v) velocity components with velocity vector fields have been compared. Furthermore, drag coefficient (CD) and Strouhal number (St) have been validated. In flow features, instantaneous streamwise velocity components indicated a lower pressure zone in the wake region of the cylinder owing to flow separation. The ANN method has predicted the positions of the minimum velocity clusters with 3.2 % for Re = 8 × 103 as a mean relative error compared to the experimental results. Moreover, the average relative difference of 10.82 % between the experimental and the estimated results for the clusters with the minimum streamwise velocity components has been attained at Re = 4 × 103 and Re = 8 × 103. Regarding the cross-stream velocity components, the mean relative difference values have been observed for the clusters with maximum values of 6.83 % and the clusters with minimum values of 3.66 %. Drag coefficients have also been obtained, and these values are very close as CD = 1.090 and CD = 1.084 for PIV and ANN methods at Re = 8 × 103, respectively. Furthermore, St values of the circular cylinder are St = 0.216 at Re = 4 × 103 and St = 0.211 at Re = 8 × 103, which agree with the results of the literature. © 2023 Elsevier Ltd