Session: 08-02-01 Wave–Structure Interaction, Noise Modeling, and Marine Hydrodynamics

Submission Number: 174561

CNN-Based Reconstruction of Large-Scale Turbulence Over Ocean Waves 

The turbulence structure above ocean waves governs air-sea momentum and energy transfers that are essential for a range of applications, from weather modeling to offshore wind energy systems. Despite the central role of near-surface turbulence, robust sea-state-dependent parameterizations of mean flow variables are still limited. In this study, we develop a deep learning framework based on convolutional neural networks (CNNs) to predict large-scale, coherent turbulence structures directly from surface wave characteristics. The model is trained on down-sampled laboratory measurements of two-dimensional airflow velocity fields over wind waves at a moderate wind speed of U₁₀ = 5.08 m/s, which corresponds to a wave age of C_p/u_* = 3.69 with C_p and u_* being the peak wave phase and friction velocities, respectively. The velocity fields were obtained using a combination of particle image velocimetry (PIV) and laser-induced fluorescence (LIF) techniques within the air-side viscous sublayer.  

The CNN model is developed to reconstruct the velocity fields solely from the surface elevation. The model architecture consists of a U-shaped convolutional network with a one-dimensional encoder and a two-dimensional decoder. To avoid the vanishing problem during the feature extraction of multi-scale surface wave profiles, residual blocks were employed in the encoder. The decoder contained skip connections and transposed convolutions to reconstruct the velocity distribution. The framework supports distributed training across multiple GPUs with adaptive learning-rate scheduling. The model was further optimized through hyperparameter tuning and validated against unseen experimental data; it demonstrated strong generalization and the ability to accurately capture the large-scale turbulence. The CNN-predicted velocity fields exhibit a high correlation with down-sampled experimental PIV data of up to 90 percent and preserve the dominant spectral content of the flow, effectively capturing the energy-containing eddies that dominate momentum transfer. This method offers an efficient technique to link experiments and simulations of wind waves through accurate stress predictions from the reconstructed turbulent velocity fields that can provide improved boundary conditions for coupled atmosphere-ocean models.

Presenting Author: Kianoosh Yousefi University of Texas at Dallas

Presenting Author Biography: I am an Assistant Professor in the Department of Mechanical Engineering at the University of Texas at Dallas. I received my Ph.D. in Mechanical Engineering from the University of Delaware in 2020. Prior to joining UT Dallas, I was an Associate Research Scientist in the Department of Civil Engineering and Engineering Mechanics at Columbia University and a Postdoctoral Researcher in the School of Marine Science and Policy at the University of Delaware. My research is focused on fundamental turbulence and understanding the physics of turbulent flows in different environments, particularly turbulent air-sea interactions, surface waves, and the resulting generation of turbulence, airflow separation, breaking waves, spray, and bubbles.

Authors:

Ali Bizhan Pour University of Texas at Dallas
Ahmed A. Hamada University of Texas at Dallas
Gurpreet Singh Hora Columbia University
Kianoosh Yousefi University of Texas at Dallas