Machine learning-based super-resolution reconstruction of turbulent flow simulations over superhydrophobic surfaces

Numerical simulations of turbulence over a superhydrophobic surface impose a significant computational burden. Recently, as part of reducing the computational cost in simulating turbulent flows, machine learning-based super-resolution reconstruction has recently been applied. In this study, we performed direct numerical simulation (DNS) of turbulent flow over a superhydrophobic surface, which exhibits multiscale phenomena, and subsequently downsampled the resolution by a factor of 16 to train a super-resolution model. The performance of the model was evaluated through both qualitative and quantitative analyses, including velocity contours, the q-criterion, the probability density function of vortices, and the turbulent energy spectrum. Specifically, we examined the reconstruction accuracy in the viscous sublayer, where micro-scale phenomena occur, and in the logarithmic layer, where turbulence dominates, to assess the capability of the model in handling multiscale turbulent flows. Furthermore, we conducted a under-resolved simulation with a mesh reduced by a factor of 16 within the same numerical method and the velocity field was reconstructed at high resolution using the trained model. The reconstructed results were analyzed using the same metrics as before, demonstrating the potential of the super-resolution reconstruction model to reduce computational costs in DNS.