Examination of machine learning method for identification of material model parameters

Abstract

In this work, we compare two methods of using artificial neural network (ANN) to determine the optimal parameters of material model by results of high-speed plate impact experiments. The first ANN is trained to determine the free surface velocity history based on the impact parameters and material model parameters; this ANN is used as a fast emulator of model to speed-up the statistical Bayesian calibration of the model parameters. The second ANN directly determines the material model parameters based on the impact parameters and the free surface velocity history, because this ANN is trained by inverse data set; this approach for inverse problem is proposed for the first time and can be more preferable for fast-track expert systems in engineering applications. Both approaches are successfully used to determine the optimal parameters of material model in the case of impact of copper and aluminum plates at different temperatures and impact velocities with experimental data from the literature. A simple relaxation plasticity model with a constant relaxation time is used in the present research as an example; more accurate coincidence with the experimental data and wider applicability of the ANN-based approach can be expected for a more realistic plasticity model.