Predicting impact strength of perforated targets using artificial neural networks trained on FEM-generated datasets

Abstract

The paper considers application of artificial neural networks (ANNs) for fast numerical evaluation of a residual impactor velocity for a family of perforated PMMA (Polymethylmethacrylate) targets. The ANN models were trained using sets of numerical results on impact of PMMA plates obtained via dynamic FEM coupled with incubation time fracture criterion. The developed approach makes it possible to evaluate the impact strength of a particular target configuration without complicated FEM calculations which require considerable computational resources. Moreover, it is shown that the ANN models are able to predict results for the configurations which cannot be processed using the developed FEM routine due to numerical instabilities and errors: the trained neural network uses information from successful computations to obtain results for the problematic cases. A simple static problem of a perforated plate deformation is discussed prior to the impact problem and preferable ANN architectures are presented for both problems. Some insight into the perforation pattern optimization using a genetic algorithm coupled with the ANN is also made and optimized perforation patterns which theoretically enhance the target impact strength are constructed.