Fast prediction of steady and unsteady hydrofoil cavitation flow fields based on the U-net enhanced Fourier neural operator

Cavitation, a complex phase-change phenomenon, is a primary cause of noise, vibration, and erosion in hydraulic machinery. This study employs a U-net enhanced Fourier neural operator (U-FNO) to predict cavitating flows around a National Advisory Committee for Aeronautics hydrofoil, encompassing steady sheet and unsteady cloud cavitation. L2 error results demonstrate that the U-FNO model achieves satisfactory overall prediction accuracy for both steady and unsteady flows with slightly higher error observed in the unsteady case. The predictive flow fields are highly similar to computational fluid dynamics (CFD) results. The prediction accuracy for cavitation occurrence in the steady case exceeds 0.9 with cavity shape and size closely matching CFD results. Although the cavity evolution of unsteady cloud cavitation is predicted relatively accurately, noticeable errors exist near the cavitation–liquid interfaces. Pronounced pressure deviations of the unsteady case are localized on the suction side and in wake regions, correlating with flow structures from bubble collapse. U-FNO shows lower L2 errors than U-net with mean errors on steady and unsteady test sets being 0.76 and 0.72 times those of U-net, respectively. Although U-net matches U-FNO in accuracy for steady lift/drag coefficients, the U-net shows noticeable errors at specific instants during unsteady predictions. Ablation experiments confirm superior learning capacity and data adaptability over U-net under different sample numbers.