Artificial neural networks as material models for finite element analysis

Abstract

In nearly all engineering fields, lightweight design is essential. The most potential lies within the substitution of parts with new materials, which allow for lighter constructions. A barrier is often the adequate description of these materials for computational methods, especially in finite element analysis (FEA). The substitution of classical constitutive material models with data-driven models using machine learning techniques could provide a new strategy for the modelling of materials. The notable benefits are a faster description of materials without complex parameter identification, faster computation due to efficient algorithms within the material model and an easier and more efficient selection of the correct material model. For this purpose, an artificial neural network (ANN) is trained, which learns the relationship between strains and stresses and is capable of modelling the hypo-elastic behaviour from a constitutive model. To train the ANN, generic stress-strain curves are generated using a programming interface for the material model driver within the commercial FEA-solver LS-DYNA. A study is conducted to find the optimal set of hyperparameters for training of the neural network. The trained ANN is evaluated using predefined strain-data extracted from a simulation. Finally, the ANN is implemented in an own user-defined subroutine, which provides a link between modern machine learning frameworks and commercial FEA codes. Explicit simulations using the ANN model run stable and show good agreement in global response as well as on local element level, compared to the conventional model.