Data-Driven Identification of Material Model Parameters Exploring Artificial Neural Networks to Calibrate Constitutive Parameters in High Density Polyethylene

This thesis focuses on data-driven methods, specifically artificial neural networks, to identify material model parameters in high density polyethylene (HDPE) for finite element (FE) simulations. The study thoroughly examines the anisotropy in HDPE by testing different material orientations with digital image correlation (DIC) during uniaxial tensile tests. DIC enabled precise measurement of strain distribution,unveiling both diffuse and local necking strain. Two hardening models,the Swift-Hockett-Sherby (S/HS) and a custom model, were explored to characterize HDPE’s plastic behaviour in FE simulations. In consistencies between predicted outcomes using the SHS model and experimental results prompt the consideration of custom equations forenhanced accuracy. The Hill48 yield model was introduced for the FE model to cover the anisotropic properties of the material. Large datasets were generated from these simulations to cover a wide range of different material configurations. The datasets were used to train neural networks so that a wide range of different HDPE grades can later be fed to the network to determine the associated material parameters. An Abaqus-Isight model was developed to automate parameter variation, simulation, and data extraction, thus streamlining the process and saving time. Data extracted from simulations, including force displacement and strain, are leveraged for neural network training. The study evaluated two types of neural networks: feed forward neural networks (FFNN) and long short-term memory neural networks(LSTM). It was found that FFNN performed better than LSTM for this task. Therefore, the research focused more on refining the FFNN approach. Overall, the implementation of the custom hardening modelin combination with the Hill48 yield model was successful, but showed weaknesses in CD orientation