A new machine learning based method for sampling virtual experiments and its effect on the parameter identification for anisotropic yield models

Abstract

A new method for sampling virtual experiments on the initial yield surface is introduced for the plane stress state. The method is based on a machine learning technique called active learning, which can be used to adaptively sample a parameter space with respect to a certain criterion. For the evaluation of this new method, it is compared against a random sampling approach taken from literature and the effect of both methods on three different anisotropic yield models, namely Yld89, Yld2000-2d and Yld2004-18p (in-plane), is analysed. The results demonstrate that the active learning based sampling approach has advantages over the random sampling approach in terms of reliability and sample efficiency. Moreover, it is found that the effect of the sampling method on the resulting yield surface depends on the anisotropic yield model considered.