Elastoplastic constitutive modeling under the complex loading driven by GRU and small-amount data

Abstract

In this paper, a data-driven method to model the three-dimensional engineering structure under the cyclic load with the one-dimensional stress-strain data is proposed. In this method, one-dimensional stress-strain data obtained under uniaxial load and different loading history is learned offline by gate recurrent unit (GRU) network. The learned constitutive model is embedded into the general finite element framework through data expansion from one dimension to three dimensions, which can perform stress updates under the three-dimensional setting. The proposed method is then adopted to drive numerical solutions of boundary value problems for engineering structures. Compared with direct numerical simulations using the J2 plasticity model, the stress-strain response of beam structure with elastoplastic materials under forward loading, reverse loading and cyclic loading were predicted accurately. Loading path dependent response of structure was captured and the effectiveness of the proposed method is verified. The shortcomings of the proposed method are also discussed.