Abstract
This paper presents a Neural Operator (NOp) for estimating wave height from wind data. The NOp is a machine learning model to approximate functions from one function space to another. It is trained to map wind data to wave height. Each function is represented as a graph, with vertices corresponding to physical locations and edges carrying geographical distribution information. The model uses a convolutional kernel, allowing nodes to exchange information based on relative positions. After several convolutions, the latent values are decoded into target variables. The model is agnostic to the absolute position of each node, simplifying training and enabling application to various domains. Although a precise wave height reconstruction is impossible, as it depends on multiple other variables and on past data, we demonstrate that such a model is able to approximate previously unavailable variables from available data, including the capability of inferring it in grids and resolutions different than those used during training.
