Keywords: inverse analysis, neural networks, identification, material model.

The identification of the parameters of material models used in finite element simulations is usually carried out based on experimental measurements. The classical way of this procedure is to choose some standard values for each parameter or to try to obtain them directly from the experiments, e.g. the tensile strength from tensile tests. If the material model is more complex and is described even by parameters without direct physical meaning the identification process could be more complicated and not be realized by hand. For such cases generally the inverse problem is solved iteratively by an optimization procedure minimizing the error between the numerically and the experimentally obtained curves. For many material models this optimization problem can not be solved in a straight way, e.g. with gradient based methods, due to the existence of several local minima. For such problems more sophisticated optimization procedures are needed as genetic algorithms, which are generally more expensive concerning the number of required simulations.

Recently a new method for identification problems has been propose [1,2]. In this method neural networks are used to represent the relation between the output of the finite element simulation (e.g. load-displacement or stress-strain curves) and the model parameters directly. This is realized by computing several sets of input and output data by the numerical model which are used for the network training. After the training is finished the experimental curves are used as input for the network and the model parameters are approximated directly. For the network training a range for each parameter has to be defined similarly to optimization procedures. The data sets can by obtained either by using a regular grid or generate random samples. This identification procedure enables the approximation of the parameters with a relatively small number of samples, even for problems with several local minima. In the paper we investigate the applicability and accuracy of the mentioned identification method for several complex material models for concrete and soil modeling.

Starting with a widely-ranged initial parameter space we will give first approximations from different training runs. Based on these approximations we refine the parameter space and obtain finally parameters which led to numerical curves agreeing very well with the experimental measurements.

1

D. Lehký and D. Novák. "Identification of material model parameters using stochastic training of neural network", In Walraven et al., editors, Proc. 5th Int. PhD Symposium in Civil Engineering, Delft, Netherlands, June 16-19, 2004. Balkema, Rotterdam, 2004.

2

D. Novák and D. Lehký. "ANN inverse analysis based on stochastic small-sample training set simulation", Engineering Applications of Artificial Intelligence, 19:731-740, 2006. doi:10.1016/j.engappai.2006.05.003

purchase the full-text of this paper (price £20)

go to the previous paper
go to the next paper
return to the table of contents
return to the book description
purchase this book (price £62 +P&P)