Spatiotemporal turbulence generation for downscaling simulations via super-resolution deep learning

Atmospheric boundary layers exhibit complex multi-scale turbulence, posing significant computational challenges for both weather forecasting and wind engineering. Meteorological model outputs generally lack the spatial resolution required for engineering-scale analyses. Conventional dynamical downscaling via nested grids can mitigate this gap but demands substantial computational resources and simulation time, limiting practical applicability. To address this scale-transition deficit, we propose an artificial intelligence-physics hybrid downscaling framework that couples large- and small-scale simulations through a deep-learning-based interface located at the coarse-to-fine computational junction. This interface employs an energy cascade-enhanced super-resolution generative adversarial network to reconstruct physically consistent fine-scale turbulence from energy-deficient coarse boundary layer data. Validation is conducted across three progressively challenging cases, including a canonical turbulent boundary layer resolved by direct numerical simulation, an idealized neutral atmospheric boundary layer representing stable meteorological conditions, and a typhoon boundary layer involving extreme events. Results indicate that the hybrid framework generalizes reliably to diverse flow regimes and accurately reconstructs physically consistent flow structures and turbulent kinetic energy at the inflow boundaries of small-scale simulation domains. The proposed framework preserves high-fidelity turbulence characteristics while reducing grid requirements by over 80% for turbulent boundary layer and significantly shortening simulation time compared with traditional dynamical downscaling. These findings establish the framework as an efficient, accurate, and scalable solution for multi-scale boundary layer turbulence prediction in computational wind engineering and atmospheric modeling.