2024 CSDMS meeting-040
From CSDMS
(if you haven't already)
Log in (or create account for non-CSDMS members)
Forgot username? Search or email:CSDMSweb@colorado.edu
Browse abstracts
Prediction of Nearshore Nonlinear Wave Properties Using Machine Learning Models Trained on Wave-Resolving Hydrodynamics Models
Ryan Schanta,
(he/him/his),University of Delaware Newark Delaware, United States. rschanta@udel.edu
Tian-Jian (Tom) Hsu, University of Delaware Newark Delaware, United States.
Fengyan Shi, University of Delaware Newark Delaware, United States.
Nearshore hydrodynamic modeling necessitates extraordinary computational power to resolve the scales of motions relevant to coastal processes. Thus, coastal models make tradeoffs in the processes resolved. One common tradeoff is wave-averaging, whereby the evolution of bulk properties and statistics of wave fields are modeled. This contrasts the computationally more intensive wave-resolving models, whereby the time-varying motion of individual waves is directly output. However, complex nearshore dynamics are often driven by phenomena that cannot be directly derived from wave-averaged quantities, which limits the breadth of applications for wave-averaged models. Machine learning techniques provide a potential avenue to leverage the power of wave-resolving models for such applications at a lower computational cost. To this end, the wave-resolving, depth-integrated FUNWAVE-TVD modeling based on solving the Boussinesq equation is used in this study. The model was validated against the Dune3 dataset collected at Oregon State University corresponding to wave evolution and breaking in a cross-shore surf zone. A series of similar wave-resolving simulations using the FUNWAVE-TVD model were generated to create a training dataset corresponding to a one-dimensional planar beach under regular wave conditions. Two properties of interest, wave skewness and asymmetry, were calculated from the resulting wave-field and parameterized via the input wave conditions and bathymetry. Preliminary results show that even relatively simple ML models (neural networks and random forests) can provide drastic improvements to commonly used empirical models commonly employed by wave-averaging models, suggesting that ML-based parameterizations of nearshore wave properties provide a viable avenue for improving wave-averaged models.
