Neural network parameterisation of the mapping of wave spectra onto nonlinear four-wave interactions

Abstract

A new approach to parameterise the exact nonlinear interaction source (Snl) term for wind wave spectra is presented. Discrete wave spectra are directly mapped onto the corresponding Snl-terms using a neural net (NN). The NN was trained with modelled wave spectra varying from single mode spectra to highly complex ones. The specification of training data was based on a classification of the wave spectra by cluster analysis. In course of the structuring of the NN the intrinsic dimensionality of the spectra was estimated with an auto-associative neural net (AANN). The AANN might be used for a scope check of the method.