Just another WordPress site - Ruhr-Universität Bochum

Abstract

Machine learning (ML) algorithms can provide an effective way of modeling complex material behavior, including plasticity. With their ability to analyze vast amounts of data and identify patterns, these advanced computational techniques can capture relationships between microstructure and mechanical properties and enable an accurate yet flexible description of material behavior. The importance of high-quality training data in machine learning cannot be overstated, as it directly impacts the performance and reliability of the resulting models. Therefore, finding an optimal data generation strategy for the training of ML yield functions and strain-hardening models is essential to ensure their accuracy and reliability. Our work revealed that Support Vector Machines (SVM) can be used as machine learning yield functions and can be successfully trained for materials with significant plastic anisotropy, like polycrystals with a Goss texture, using only 300 data points. The training data was carefully selected to cover a wide range of loading conditions in full 6-dimensional space. The accuracy and reliability of the trained ML yield functions were assessed using a variety of statistical measures, including mean squared error and confusion matrix. As the next step, the optimal training data strategy was employed for generating training data using micromechanical modeling. This allowed for further consideration of microstructural parameters like crystallographic texture or grain size in the input data for the training of the ML yield function. Moreover, this data also contains information of the strain-hardening behavior of the material. A database containing simulation data and metadata was developed, making it easier to train and refine the ML model. From this database, a proper data-oriented formulation for strain hardening was developed, which complements the previously established yield function. This study demonstrated the potential of ML algorithms as powerful tools for modeling complex material behavior. The careful selection and generation of training data, as well as the inclusion of microstructural parameters and work-hardening effects, were crucial for achieving reliable and accurate ML yield functions and strain-hardening models.