Transfer learning for reconstruction of missing regions in a near-wall cylinder wake

Abstract

Reconstruction of missing regions in velocity fields remains a key challenge in experimental fluid mechanics, particularly in complex turbulent flows where measurements are limited by spatial resolution and optical accessibility. This study investigates a transfer learning framework that leverages large-eddy simulation (LES) and particle image velocimetry (PIV) data to reconstruct incomplete wake velocity fields under identical flow conditions. The flow considered is a horizontal circular cylinder at a bed clearance of 0.5 diameters, generating a strongly inhomogeneous wake with pronounced multi-scale interactions. A linear extended proper orthogonal decomposition model is compared with a non-linear U-Net architecture under controlled masking conditions. A transfer learning strategy is introduced, where the network is pretrained on time-resolved LES data and fine-tuned using non-time-resolved PIV measurements. Results show the non-linear model significantly outperforms the linear baseline, particularly in capturing the spatial flow structures and velocity correlations. Transfer learning model yields consistent, though modest, improvements in reconstruction accuracy and improved agreement with velocity correlations and Reynolds shear stress, although the gains are less consistent than those observed for pointwise reconstruction. However, PIV-based models underpredict high-wavenumber energy and fail to fully recover vorticity intermittency, highlighting limitations associated with the reduced small-scale information content and effective spatial resolution of the experimental dataset. These results suggest that computational fluid dynamics-informed pretraining may improve reconstruction fidelity while emphasizing the critical role of data quality in resolving multi-scale turbulent dynamics.