Proceedings of the 6th European Congress on Computational Methods in Applied Sciences and Engineering, Paper #3112, 2012
T. Palau1, A. Kuhn1, S. Nogales2,H.J. Böhm2, A. Rauh3
1ANDATA GmbH, Hallein, Austria
2Institute of Lightweight Design and Structural Biomechanics,
TU Wien, Vienna, Austria
3BMW AG, Munich, Germany
Due to the increasing use of complex materials in lightweight structures, the
development and identification of proper material models for the prediction of
damage and failure within Finite Element simulations has become an extensive
Other fields of application have already shown that the introduction of Soft
Computing and Machine Learning methods can be very beneficial for getting the
complexity under control.
This contribution presents an evaluation of the feasibility of using Artificial Neural Networks for modelling complex, nonlinear behaviour. A series of synthetic stress-strain paths generated by a standard elasto-plasticity model were used as target data in order to study the parameterizations required by the neural networks for successfully modelling elasto-plastic material behaviour.
In single-element tests, generic multiaxial responses obtained by J2 elasto-plasticity material models were successfully modelled by means of feedforward neural networks implemented as user material routines for the FE code ABAQUS/Explicit. The used parameterizations were found to produce networks with very satisfactory generalization properties, i.e., the neural networks provided correct stress response along complex load paths that were not part of their training process.