Reconstruction and correction of attached cavity evolution on hydrofoil based on physics-informed correction network

Accurate prediction of cavitating flows remains challenging due to strong nonlinearity, local compressibility, and model-form uncertainties in conventional RANS simulations. This study develops a Physics-Informed Correction Network (PICN) that embeds the incompressible Navier–Stokes equations and the Schnerr–Sauer cavitation model into a neural-network optimization framework. An Akaike Information Criterion (AIC)-guided weighting strategy is introduced to balance data fidelity and physics-based regularization constraints, and automatic differentiation is used to evaluate governing-equation residuals as physics-based constraints. A baseline cavitating flow around a Clark-Y hydrofoil is computed using the k – ω SST model, and the resulting velocity, pressure, and vapor-volume-fraction fields are refined by the PICN. The corrected predictions significantly reduce discretization-induced and result-level discrepancies, lowering mean lift and drag errors to 3.42% and 2.21% and achieving markedly improved agreement with experimental measurements. The PICN also enhances near-wall velocity prediction, suppresses nonphysical cavity breakup, and successfully reproduces the formation and shedding characteristics of the re-entrant jet in close agreement with experimental observations. These results demonstrate that physics-informed correction provides an effective and computationally efficient approach for improving the reliability of cavitating-flow simulations.