Deep Learning-Based End-to-End Velocity Field Reconstruction for the Flow Field Around Two Cylinders

Particle image velocimetry (PIV) is a widely used technique for flow field measurements, but its high equipment costs and limited algorithms restrict its broader application. We propose a deep learning-based end-to-end PIV velocity field reconstruction method. This method uses particle images captured with a standard camera and low-power continuous laser as input and outputs velocity fields. It is significantly less costly and complex than a full PIV setup. The training set is derived from experimental results, with numerical simulations aligning particle images with ground truth velocity fields. The flow around a single cylinder at subcritical Reynolds numbers was first investigated. The proposed method demonstrated superior resolution, accuracy, and efficiency compared to conventional cross-correlation method. Transfer learning was then applied to the flow around tandem cylinders, where the model showed great generalization capability. The model was also capable of reconstructing velocity fields in missing regions of the flow field and recovering clear flow structures. These findings indicate the potential of deep learning in enhancing PIV experimental techniques.