Multi-fidelity physics-informed modeling of island wakes with near-wall inference from particle image velocimetry

Predicting three-dimensional wake flows past complex island geometries remains a challenge for traditional turbulence closures, particularly in capturing separation topology and near-bed shear stresses at engineering Reynolds numbers. This study proposes a multi-fidelity physics-informed framework that assimilates volumetric particle image velocimetry (PIV) data into a Reynolds-averaged Navier–Stokes formulation augmented by a neural network. A universal differential equation strategy is employed to learn a Reynolds-number-dependent correction to the eddy viscosity, trained first on a multi-regime large-eddy simulation baseline and then lightly tuned with spatially localized PIV measurements. We investigate the wake past tandem islands for Reynolds numbers ranging from 24k to 96k. The results demonstrate that the PIV-informed tuning corrects the systematic under-prediction of wake spreading inherent in the baseline model, successfully recovering closed recirculation bubbles with realistic reattachment lengths. Crucially, the physics-based coupling allows the model to infer physically consistent near-wall skin friction in the unmeasured region, predicting a substantial enhancement (up to ∼100%) in the spatially averaged skin friction coefficient relative to the baseline. Structural analysis reveals that the neural closure learns a robust, spatially distributed injection of eddy viscosity within the shear layers, which scales consistently across flow regimes. The proposed framework thus provides a data-efficient pathway to construct predictive digital twins for hydraulic infrastructure, correcting structural model errors while preserving global consistency.