Fracture Estimation Based on Deformation History with Recurrent Neural Networks

Capturing the material deformation limits is an im-portant consideration in different engineering applications. While materials are deforming under different loading conditions, the fracture can take place and is often preceded by localization of plastic deformation. These localization fields depend on the amount and direction, or in more generally, the history of the applied loads during deformation. Forming Limit Curves (FLCs) are commonly used to describe the maximum applicable strain before localization, but they are generally valid only for the case of proportional loading. Actual loading scenarios are rarely proportional, meaning that loading can change significantly during deformation. Recent studies show that this effect of non-proportionality, or loading history, could be predicted with machine learning based methods. In this paper, we demonstrate that Recurrent Neural Networks (RNN) can capture the effect of the deformation history on localization and thereby give a reasonable estimate for non-proportional FLCs. The model is trained with a dataset of bilinear loading paths generated with a finite element-based Marciniak-Kuczynski (MK) model. The model is validated with multi-linear loading paths.