Discussion and conclusions
Two methods of developing ANNs have been proposed in the present study. In the first approach, experimental data was used only for the entire structural response. It was possible to train the neural network with these values, but the prediction of additional deformations outside of the trained data deviates from the measured mid-point deflections of the plate specimens. Consequently, the advantage of an ANN, to be very reliable since trained data is recalculated, was used to propose an intelligent finite element for viscoplastic material behaviour. However, it must be ensured that the occurring state variables in the intelligent finite element simulations are inside the provided set of constitutive training data. Following this method, simulation results being much more precise than in the first approach are obtained, but with considerably less computational effort than in classical finite element simulations. However, one intermediate step is necessary before the intelligent element is complete. The data set to train the ANN and to determine the synapse matrices have to be obtained numerically. In summary, the proposed method demonstrates the possibility to develop an ANN within an intelligent finite element for non-linear problems in structural mechanics. This method can be very efficient concerning the reliability of numerical predictions and can lead to a significant reduction of simulation time.
