Presentation Information
[GdsyDt6e-04]Data-Driven Global Wave Model Based on a Multiscale Graph Neural Network
*Godai Sonoda (1. FURUNO ELECTRIC CO., LTD. (Japan))
Accurate wave forecasting is essential for maritime safety and efficient ship routing, and is a key component of ocean digital twins. Numerical wave models are widely used but require substantial computational resources, limiting their real-time use in ocean digital twins. To address this, data-driven models have been developed as fast surrogates. While various data-driven models have been proposed in atmospheric forecasting, they have been explored far less extensively in wave forecasting.
In this study, we applied a multi-scale graph neural network (GNN) to global wave forecasting, with a horizontal resolution of 0.5 degrees and a lead time of 10 days. Unlike conventional grid-based models, the GNN-based formulation can also be applied to unstructured meshes, although structured meshes are used here. In addition to wave variables, 10m wind components over ocean are used as inputs, and the model is trained on reanalysis data.
Using reanalysis winds as inputs, the model achieved an average RMSE of 0.19 m for significant wave height, 14.2° for mean wave direction, and 0.53 s for mean wave period at a lead time of 9.5 days. On a single NVIDIA A100 GPU, a complete 10-day global forecast takes about 100 seconds. Using forecast winds as inputs, the accuracy decreased due to wind forecast errors, yet this suggests that the model has acquired the ability to generate and develop wave fields in response to winds. Overall, machine learning can serve as a fast surrogate for physics-based wave models, well suited to operational digital twins. Furthermore, the GNN's applicability to unstructured meshes is expected to enable higher-resolution forecasting in coastal regions, supporting interoperable digital-twin applications.
In this study, we applied a multi-scale graph neural network (GNN) to global wave forecasting, with a horizontal resolution of 0.5 degrees and a lead time of 10 days. Unlike conventional grid-based models, the GNN-based formulation can also be applied to unstructured meshes, although structured meshes are used here. In addition to wave variables, 10m wind components over ocean are used as inputs, and the model is trained on reanalysis data.
Using reanalysis winds as inputs, the model achieved an average RMSE of 0.19 m for significant wave height, 14.2° for mean wave direction, and 0.53 s for mean wave period at a lead time of 9.5 days. On a single NVIDIA A100 GPU, a complete 10-day global forecast takes about 100 seconds. Using forecast winds as inputs, the accuracy decreased due to wind forecast errors, yet this suggests that the model has acquired the ability to generate and develop wave fields in response to winds. Overall, machine learning can serve as a fast surrogate for physics-based wave models, well suited to operational digital twins. Furthermore, the GNN's applicability to unstructured meshes is expected to enable higher-resolution forecasting in coastal regions, supporting interoperable digital-twin applications.
Comment
To browse or post comments, you must log in.Log in