Date of Award

Summer 8-1-2023

Document Type


Degree Name

Doctor of Philosophy (PhD)


Coastal and Marine Systems Science


College of Science

First Advisor

Shaowu Bao

Second Advisor

Paul T. Gayes

Third Advisor

Diane Fribance

Additional Advisors

Xiaofeng Li; Sathish Kumar


Numerical Modeling (NM) is widely used to simulate and predict hydrodynamic processes and marine particle movements in coastal oceans, particularly during extreme weather events and emergencies. NM offers the capability to realistically simulate multiple state variables and fill gaps caused by scarce observations. However, inherent uncertainties exist in all NMs, primarily arising from the following three factors: 1) insufficient observations leading to uncertain model initial and boundary conditions, 2) inevitable truncation errors due to coarse model resolution, and 3) imperfect physics parameterization schemes for sub-grid processes, especially those related to waves. The consequences of these uncertainties are that 1) even state-of-the-art NM methods can produce unsatisfactory marine particle movement predictions with marine particle trajectory errors growing rapidly over time, and 2) NM often fails to adequately represent wave-induced water turbulence mixing in predictions and simulations based on Eulerian and Lagrangian approaches. These uncertainties are difficult to address using traditional NM methods because of their inherent limitations.

In this dissertation research, Artificial Intelligence (AI) models are utilized based on their capabilities of nonlinear solving to address the above-mentioned challenges. I hypothesize that AI can improve the accuracy of ocean NM. Two tasks are identified to validate our hypothesis: 1) developing an AI correction model to improve NM-predicted float trajectories and 2) developing an AI wave model as a substitute for wave NM to improve the representation of water turbulence mixing in ocean simulation to achieve more accurate results under hurricane scenario. I use Regional Ocean Modeling System (ROMS) model as the foundation to predict the float trajectory and simulate the oceanic hydrodynamics under hurricanes. Experiments of ROMS simulation will be conducted and the results compared with observations to evaluate the improvements in model accuracy achieved through the application of the developed AI-methods.

In the task of AI correction for NM-predicted float trajectories, I designed an AI model that incorporates Convolutional Neural Network (CNN) and Gated Recurrent Unit (GRU) modules. To train this AI model, I have utilized a dataset consisting of 4501 observed 1-day float trajectories obtained from the Ocean of Things (OoT) program. These observations serve as the ground truth in the AI model training. The corresponding ROMS-predicted float trajectories are utilized to create AI input dataset. This AI input dataset includes various parameters such as the latitudes and longitudes of ROMS-predicted float trajectories, water depth, time, wind velocity at 10 m above the sea surface, and sea surface current in the zonal and meridional components. I randomly selected 3601 out of the 4501 trajectories for training this AI correction model. The remaining 900 1-day float trajectories were used to validate the trained AI correction model. The results of this AI correction model indicate that

1) The AI correction model can effectively improve the ROMS-predicted float trajectories. At the 24th hour, approximately 82% ROMS predicted float trajectories in the test dataset are successfully corrected by the AI model, resulting in a 57% improvement in trajectory prediction accuracy.

2) The AI correction model also demonstrates its applicability under hurricanes. 77% of 75 ROMS-predicted float trajectories during the hurricane periods are improved by this AI correction model, further showcasing its effectiveness under extreme weather conditions.

3) The performance of the AI correction model varies depending on different conditions. In particular, the model’s performance was found to be lower in wintertime and nearshore regions, which can be attributed to insufficient training data available for these two scenarios, indicating that the model’s effectiveness could potentially be enhanced with more comprehensive and diverse training data.

In the task of AI wave modeling for ocean simulation, I designed an AI model that combines the Bidirectional GRU (BiGRU) and Multi-Head Attention methods to emulate significant wave height (SWH), wave period, and wave direction of wind-generated waves. Additionally, a physics constraint between SWH and wave period is added into AI wave model to ensure the consistency between these two state variables. WAVEWATCH III (WW3) model-simulated and buoy-measured wind sea wave data are used to generate the AI ground truth datasets with the same data structures. WW3 model is a widely used and well-established numerical wave model that simulates ocean waves based on various inputs such as wind speed, atmospheric pressure, and bathymetry. It is extensively validated and calibrated using observed wave data from buoys, satellite measurements, and other sources. WW3 model outputs are considered to be a reliable representation of wave characteristics under specific environmental conditions. These model simulations undergo rigorous validation and comparison with observational data to ensure their accuracy and fidelity. As a result, the WW3 model outputs are often used as a reference or ground truth for evaluating and benchmarking other wave models, including AI-based wave models for the regions where the wave observations are not available. The integration of WW3-simulated and buoy-measured wave data can address the scarce spatial coverage of buoy observations and incorporate more real wave characteristics into WW3 simulation.

The AI input dataset for training the AI wave model includes water depth, wind components in u- and v- directions at 10 m above the sea surface. In the AI training of wave direction, SWH and wave period are included as additional input data. The AI wave model is first pre-trained using the WW3-based dataset and subsequently re-trained using the buoy-based dataset. The performance of AI wave model indicates that

1) The WW3-buoy-based AI wave model demonstrated acceptable accuracies in the northwestern Atlantic Ocean under all weather conditions, with the RMSEs of 0.36 m for SWH, 1.08 s for wave period, and 32.89 deg for wave direction between the AI-predicted and buoy-measured waves.

2) The WW3-buoy-based AI wave model successfully emulates smooth and continuous wave data from coastal regions to open oceans, indicating that the AI model is able to capture the spatial variations of wave characteristics.

3) Under hurricane scenarios, the WW3-buoy-based AI wave model presents similar wind sea wave patterns to the Simulating Waves Nearshore Model (SWAN) wave model. Moreover, the AI model still maintains acceptable accuracies during hurricane periods, demonstrating its robustness and its ability to perform under extreme weather conditions.

The validated WW3-buoy-based AI wave model is implemented to provide wind sea wave required for the turbulent mixing scheme in ROMS simulation under Hurricane Dorian (2019) and Typhoon Malakas (2016). The ROMS simulation results of these two tropical storms indicate that

  • The AI wave model demonstrates the capability to replace high-demanding wave numerical models (e.g., SWAN and WW3) under hurricane scenarios for representing the wave effects on ocean simulation.
  • Incorporating AI-derived wave data into ocean simulations can yield more robust and realistic results compared to ocean simulations that do not account for wave effects.
  • The presence of waves significantly enhances water turbulence mixing and latent heat flux in the ROMS simulations. This effect leads to the generation of local cold wake areas with low sea surface temperature (SST). Waves play a crucial role in ocean dynamics by inducing mixing processes and impacting heat exchange at the ocean surface.
  • Integration of AI wind sea wave cannot effectively optimize the performance of surface wind-wave (SWW) mixing scheme under Typhoon Malakas, compared to SWW mixing scheme with a default wave condition, which is attributed to the deficiency of SWW mixing scheme and no swell characteristics in current AI wave model.

Included in

Oceanography Commons