PHYSICS-INFORMED DEEP LEARNING OF NEARSHORE WAVE PROCESSES

Abstract

The paper introduces the NWnets, a physics-informed deep learning model for reconstructing nearshore wave fields and mapping bathymetry. The physics encoded into the deep neural networks are the wave energy balance equation and dispersion relation. Insights into the model capability are gained through application of the NWnets to a laboratory experiment of wave transformation over a circular shoal. If the bathymetry and discrete measurements of wave height are available, the NWnets model is capable of simulating nearshore wave transformation. Moreover, the extended NWnets can be used for depth inversion if the bathymetry is unknown. Two methods for simultaneously estimating water depths and surface waves are presented. If surface wave number and limited wave height measurements are available from remote sensing platforms, the first method employs wave numbers and scarce measurements of wave height as training data. The second method utilizes scarce wave height and limited water depth measurements as training points to reconstruct bathymetry and wave fields. The results show that both methods are capable of simultaneously mapping the bathymetry and waves when the locations of training points are appropriately distributed.