Use of neural networks for automated analysis and modeling of elasto-plastic material properties of metallic tensile specimens : Verwendung Neuronaler Netze zur automatisierten Analyse und Modellierung elasto-plastischer Materialeigenschaften von metallischen Zugproben

Abstract

Tensile testing is an established and standardized method for material testing. The evaluation of mechanical materials properties from stress-strain curves is a sometimes laborious and very often manual but elementary process, as it can be used to derive materials-specific information needed for the design of mechanical components. In this work, we present two approaches to show how neural networks (NNs) can be used to evaluate stress-strain curves in an automated manner. First, we use supervised learning to classify stress-strain curves with respect to an appropriate empirical strain hardening model and extract the corresponding model parameters. Second, we show how unsupervised learning can be used to encode physical concepts in latent variables. The long-term goal is to output material parameters such as elastic modulus, yield strength and ultimate tensile strength, as well as qualitative information on the presence of (kinematic) hardening mechanisms during plastic deformation. Respective training data was generated by numerical FEM simulations.