Machine Learning-Based Super-Resolution Reconstruction of Turbulent Flow Simulations over Superhydrophobic Surfaces

We investigate the applicability of machine learning-based super-resolution reconstruction to reduce the computational cost of simulating turbulence over superhydrophobic surfaces (SHS), which possess multiscale flow features. Direct numerical simulation (DNS) of turbulent channel flow over an SHS was conducted, and the resulting high-resolution velocity fields were downsampled by a factor of 16 to train a super-resolution model. Model performance was evaluated using velocity contours, q-criterion, vortex probability density functions, and turbulent energy spectra. Special attention was given to reconstruction fidelity in both the viscous sublayer and logarithmic region to assess the model's ability to capture microscale and macroscale features of turbulence. Additionally, under-resolved simulations using coarsened grids (1/16 fewer grid points) were performed and reconstructed using the trained model. The reconstructed velocity fields retained key turbulent structures and statistical features, demonstrating the model's capability to recover high-resolution flow characteristics from coarse input and highlighting its potential to reduce the computational cost of DNS in SHS flows.