Deep learning-based reduced-order modeling of unsteady wake dynamics of unequal-height tandem finite wall-mounted circular cylinders

This study presents proper-orthogonal decomposition (POD)-based deep learning-reduced-order model (DL-ROM) predictions of the unsteady wake dynamics of a single finite wall-mounted circular cylinder (FWMC) and unequal-height tandem FWMCs. The DL-ROMs, based on long short-term memory (LSTM) and bidirectional LSTM (BLSTM) neural networks, were trained using time-resolved particle image velocimetry (TR-PIV) measurements from the symmetry plane and three horizontal planes along the span of the FWMCs. The FWMCs were fully submerged in a turbulent boundary layer with a Reynolds number, based on the cylinder diameter and freestream velocity, of Re = 5540. The single cylinder (SC) had an aspect ratio of 7.0, while the tandem cylinders (TC) had a spacing ratio of 4.0 and a height ratio of 0.75, with the downstream cylinder (DC) matching the height of the SC. The SC exhibited a quadruple wake structure with enhanced vortex shedding at mid-height, while the TC, due to partial sheltering, featured downwash impingement on the frontal surface of the DC and an induced upwash in the wake of the DC, which led to an enhanced recirculation bubble and vortex shedding near the free end of the DC. The POD results showed that these salient wake dynamics were effectively captured by the first 200 modes and were therefore used for training the models. Transfer learning was applied to the horizontal planes, reducing computational cost and improving accuracy. Both LSTM- and BLSTM-based models exhibited commendable predictions of the POD coefficients and the reconstructed instantaneous and time-averaged flow fields for each test case; however, errors were relatively larger in the complex regions around the DC due to partial sheltering effects.