Predicting Global Wave Conditions via Convolutional Neural Network

Abstract

Wave condition predictions are critical to several problems in engineering and climate science. They are essential to international shipping, constructing offshore structures, managing coastal infrastructure, and developing renewable energy supplied by ocean waves and swells. Modeling wave characteristics is also important in understanding key climate processes, such as quantifying air-sea CO2 fluxes. Extant methods for forecasting and hindcasting are currently restricted to physics-based numerical models, like WAVEWATCH III (WW3), which are computationally expensive. Importantly, a more computationally efficient model would reduce the resources needed to make wave predictions and allow for more accurate real-time forecasting systems. Here, we propose a machine learning model to act as a WW3 emulator, learning the nonlinear mapping between wind forcing and wave conditions and allowing for much faster wave predictions with comparable accuracy to physics-based models. To do so, we implement a convolutional neural network (CNN), trained on WW3 simulations forced by the Japanese Meteorological Society Reanalysis product (JRA55-do), to predict significant wave height at half-degree resolution every six hours. Our emulator, although less accurate than its numerical counterpart, is able to capture both temporal and spatial dependencies, and can predict wave conditions on a model with relatively few parameters in a fraction of the time needed to run WW3 simulations.