Small-scale turbulence over wind-driven waves: Super-resolution physics-informed CNN modeling

Air-sea flux transfer is influenced by small-scale turbulent interactions between wind and waves, which shape both long-term climate patterns and short-term weather events. This study develops a data-driven modeling approach based on physics-informed convolutional neural networks (PI-CNNs), in which a weighted Fourier loss function is implemented to predict near-surface instantaneous turbulence over surface waves. The overarching goal is to capture small-scale turbulence with a broader focus on accurate mean air-sea flux estimates, air-sea momentum transfer, and turbulent airflow separation. The model is trained using a high-resolution experimental dataset acquired in a wind-wave tunnel facility. Prior to training, the turbulent flow fields are decomposed using singular value decomposition (SVD) into four physically interpretable components: the mean flow, the wave-coherent flow, the wake structure (including airflow separation), and a residual component. This decomposition allows the PI-CNN to learn the distinct contributions to the near-surface turbulent airflow, improving its ability to reconstruct small-scale turbulence. This work demonstrates a novel, high-resolution, PI-CNN-based pathway for reconstructing small-scale turbulent structures from easily measurable wave characteristics. Ultimately, the approach advances predictive modeling of momentum exchange at the air-sea interface.