Anwendung und Grenzen von neuronalen Netzen als Metamodell am Beispiel von Polyamiden
This presentation concerns the generation of material data using neural networks. On the basis of polyamide it is examined, if a standardised procedure can be used for the characterization of material performance against different influencing variables such as temperature, moisture content and strain rate. To generate training data for the neural networks, dynamic bending tests were performed on the testing system Impetus II. The test program included five temperatures, three moisture content and three experimental setups. From the measurement data a strain rate dependent material model for the description of the stress-strain characteristic could be created for every combination of temperature and moisture content. To obtain the data for the neural networks the models were reduced to non strain rate dependent material models for specific strain rates. Thus for every tested condition of temperature, moisture content and strain rate the parameters of the material model were available. The chosen material model has three parameters in its non strain rate dependent form, which are defined as target values for the neural nets. As input values the parameters of the condition were given. For the determination of a suitable neural network the network type, the network topology, the training function, the transfer function and the training performance were chosen. The neural networks were trained and the results were evaluated. If all data from the testing was made available, the neural nets could describe the material quite well. Only in the range where the testing data had not the best quality, some problems occurred. Furthermore it was examined, if a training data reduction is possible. The results showed that neural nets are rather sensitive to the decrease of the training data. The learned information is not sufficient any more and the material performance can not be described. For the considered procedure the quality and the amount of the training data is crucial. If enough measurements are performed accurately, neural networks can be used successfully for the generation of material data for an arbitrary condition in the implicit knowledge area of the testing data.
https://www.dynamore.de/de/download/papers/forum07/meta-modeling/anwendung-und-grenzen-von-neuronalen-netzen-als/view
https://www.dynamore.de/@@site-logo/DYNAmore_Logo_Ansys.svg
Anwendung und Grenzen von neuronalen Netzen als Metamodell am Beispiel von Polyamiden
This presentation concerns the generation of material data using neural networks. On the basis of polyamide it is examined, if a standardised procedure can be used for the characterization of material performance against different influencing variables such as temperature, moisture content and strain rate. To generate training data for the neural networks, dynamic bending tests were performed on the testing system Impetus II. The test program included five temperatures, three moisture content and three experimental setups. From the measurement data a strain rate dependent material model for the description of the stress-strain characteristic could be created for every combination of temperature and moisture content. To obtain the data for the neural networks the models were reduced to non strain rate dependent material models for specific strain rates. Thus for every tested condition of temperature, moisture content and strain rate the parameters of the material model were available. The chosen material model has three parameters in its non strain rate dependent form, which are defined as target values for the neural nets. As input values the parameters of the condition were given. For the determination of a suitable neural network the network type, the network topology, the training function, the transfer function and the training performance were chosen. The neural networks were trained and the results were evaluated. If all data from the testing was made available, the neural nets could describe the material quite well. Only in the range where the testing data had not the best quality, some problems occurred. Furthermore it was examined, if a training data reduction is possible. The results showed that neural nets are rather sensitive to the decrease of the training data. The learned information is not sufficient any more and the material performance can not be described. For the considered procedure the quality and the amount of the training data is crucial. If enough measurements are performed accurately, neural networks can be used successfully for the generation of material data for an arbitrary condition in the implicit knowledge area of the testing data.