First Steps Towards Machine-Learning Supported Material Parameter Determination

Machine learning is becoming more and more part of our world. Even though most people have so far only passively used the possibilities of this technology, e.g. for search queries or product recommendations, many have surely already thought about how these new possibilities could support their work in the future. In this contribution, it is investigated if machine learning is suitable to support the process of material characterization. Through deep neural networks it is possible to "learn" nonlinear relationships between a set of input values and the corresponding output, also known as labels. As a proof of concept, it is examined whether the shape of the yield curve can be predicted based on force-displacement curves from simulated tensile tests. So, in a first step, a large number of tensile tests are simulated which differ in the shape of the yield curve. Here, for the description of the yield curve an approach according to Hockett-Sherby was used which provides 4 parameters for the definition of the shape. The force-displacement curves of these tests are used as the input and the parameters of the yield curve as labels. By considering the entire realistic range of all four parameters, the trained neural network should be able to provide the best matching set of parameters for a given force-displacement curve. For the prediction, of course, the initial and boundary conditions must be the same when generating the force-displacement curve, whether by simulation or in a real test. Of course, all initial and boundary conditions as well as all other assumptions and simulation settings are also learned from the neural network. Therefore a change of these parameters can for sure worsen the predictions considerably and can make a re-learning process inevitable. The long-term objective of this method and the vision of this work are to learn the possible spectrum of the whole material model in advance in order to be able to finally predict the material properties based on only a few experiments with minimal effort.