Efficient meta-modeling of a carbon fiber reinforced plastic double roller using a sample iteration-error treatment neural network

Abstract

A neural network (NN) based meta-model for a carbon fiber reinforced plastic (CFRP) double roller is essential because this meta-model can quickly provide deformations and parameters required for various engineering practices. However, it is not easy to train the NN in an efficient way even if the training data are already prepared. With this motivation, we propose a sample iteration (SI)-error treatment (ET) NN to efficiently create a reliable NN of the CFRP double roller. In the proposed strategy, the SI NN is first built iteratively using only samples of the data. Next, the ET NN is constructed to compensate the SI NN using small amounts of errors when performance of the SI NN does not improve after a certain iteration. The prediction fields are implemented by combining the SI and ET NN. As a result, the NN of the CFRP double roller can be created without applying the entire data. Furthermore, because only a small number of errors are employed, the accuracy of the prediction fields can be dramatically improved without a heavy computational burden. Consequently, through SI-ET NN strategy, a high-quality NN can be efficiently generated for the CFRP double roller. The performance of SI-ET NN is evaluated using various numerical approaches.