Flow Reconstruction and Prediction Based on Small Particle Image Velocimetry Experimental Datasets with Convolutional Neural Networks

Particulate two-phase flows normally manifest high dimensional and complex nonlinear flow phenomena. To capture the intrinsic characteristics of the fluid flows, dimension reduction and feature capturing are of fundamental importance. In this work, we focused on the flow reconstruction and prediction based on dimension reduction and feature capturing with small noised datasets obtained by particle image velocimetry (PIV) experiments by use of a skip-connection convolutional neural network based on autoencoder (scCAE). We evaluated the performances of scCAE in reconstructing and predicting the high dimensional and nonlinear flows around a single particle for moderate Reynolds numbers (Re) of 400-1400. It is shown that scCAE with the latent vector size of four can well reconstruct and predict the fluid flow fields around either a sphere or cube based on the small noised PIV datasets with the data size of several hundreds, which suggests the robustness and generality of scCAE in dimension reduction and feature capturing. This may be extended to wider applications in extracting dimension-reduction latent vectors from limited ground truth experimental PIV data and disclosing the inherent physics.

Automated summary

This study focuses on flow reconstruction and prediction based on dimension reduction and feature capturing using a skip-connection convolutional neural network based on autoencoder. The results demonstrate that scCAE with a four-dimensional latent vector size can effectively reconstruct and predict fluid flow fields, showing the robustness and generality of scCAE in dimension reduction and feature capturing in moderate Reynolds numbers.