Speaker
Description
Flows around circular cylinders near a solid boundary exhibit complex interactions between the wake and the wall boundary layer, resulting in significant variations in flow characteristics with the gap-to-diameter ratio. Accurate prediction of such flow fields using conventional computational fluid dynamics (CFD) can be computationally demanding, particularly across different flow regimes and wall-clearance conditions. Physics-Informed Neural Networks (PINNs) provide a promising alternative by incorporating the governing equations into learning process, but vanilla PINNs often suffer from high computational cost and limited scalability for resolved spatiotemporal problems. To address these limitations, this study employs a Separable Physics-Informed Neural Network (SPINN), which combines axis-wise separable subnetworks with forward-mode automatic differentiation.
The framework is evaluated for near-wall cylinder flows at Re = 200 under different gap-to-diameter ratios, corresponding to distinct steady, transitional, and unsteady wake states. An additional high-Reynolds-number case (Re = 3.6×106) is also considered to assess the applicability of the architecture beyond the laminar regime. The results show that SPINN can reconstruct the principal near-wall flow features and wake dynamics with good agreement against CFD reference, while providing improved computational efficiency compared to the vanilla PINN. For the unsteady cases, especially, SPINN accurately captures the temporal evolution in the wake, including the alternating cross-flow motions associated with vortex shedding, and downstream wake dissipation. The sensitivity analyses further demonstrate that reconstruction performance depends strongly on the sampling strategy. Overall, the study demonstrates the potential of SPINN as an efficient surrogate framework for near-wall flow reconstruction across different flow states.