13th International Conference on the Technology of Plasticity (ICTP 2021): Big Data
Program Organizers: Glenn Daehn, Ohio State University; Libby Culley, The Ohio State University; Anupam Vivek, Ohio State University; Jian Cao, Northwestern University; Brad Kinsey, University of New Hampshire; Erman Tekkaya, TU Dortmund; Yoshinori Yoshida, Gifu University
Monday 9:15 AM
July 26, 2021
Room: Virtual: Room D
Location: Virtual
Session Chair: Maysam Gorji, MIT; Blake Barnett, Ohio State University
Development of Mechanical Cards for Finite Element Analysis of Mild Steels by Parameter Optimization: Emin Tamer1; Eda Kiziltas1; Celal Seyalioglu1; 1Borcelik Steel Industry Trade Inc.
Material card generation has a significant role for characterizing the mechanical behavior to estimate accurate deformation in finite element analysis. Regarding to this, the most commonly used characterization test for sheet metals is the uniaxial tension test due to its simplicity and well-defined testing standards. The mechanical properties such as ultimate tensile strength can be determined by tensile tests. However, the strain level can reach to a specific value which is low as compared to strain levels reached in cold forming operations. In this paper, the experimental data is optimized with finite element analysis and a good consistence between experimental and simulation results is achieved. Parameter optimization is used as a guiding tool to extend strain levels according to 3 mechanical equations: Swift, Hockett-Sherby, Combined material model. The user interface is designed with the detailed graphs and optimized parameters with the help of MATLAB tool.
On the Potential of Machine Learning Algorithms to Predict the Plasticity of Sheet Metal: Maysam Gorji1; Dirk Mohr2; 1MIT; 2ETH
Neural networks provide a potentially viable alternative to differential equation based constitutive models. Here, a neural network model is developed to describe the large deformation response of the non-quadratic Yld2000-2d yield criterion along HAH (homogeneous anisotropic hardening) model in sheet material. Using conventional return-mapping scheme, virtual experiments are performed to generate stress-strain data for random reversal and monotonic biaxial loading paths. Subsequently, a basic feed-forward and recurrent neural network model is trained and validated using the results from the virtual experiments. The results for a “shallow network” show remarkably good agreement with all experimental data. The identified neural network model is implemented into a user material subroutine und used in basic structural simulations such as notched tension and in-plane shear experiments. In addition to demonstrating the potential of neural networks for modeling the rate-independent plasticity of metals, their application to more complex problems involving strain-rate and temperature effects is discussed.
Investigation of Machine Learning Models for a Time Series Classification Task in Radial-axial Ring Rolling: Simon Fahle1; Thomas Glaser1; Bernd Kuhlenkötter1; 1Lehrstuhl für Produktionssysteme Ruhr-Universität Bochum
The great potential of machine learning models in different domains has been shown in recent years. Based upon initial research regarding preprocessing methods for time series classification in the hot forming technology of radial-axial ring rolling, this paper takes the next step to further investigate the suitability of different machine learning models for a classification task regarding the ovality of a formed ring. This is done by implementing several models of the time series classification domain in machine learning and training them on actual production data of thyssenkrupp rothe erde Germany. The data set consists of different production days and ring geometries. Different experiments will be performed, the results will be analyzed regarding performance, interpretability and usability in the production environment. Thus a suitable model for the underlying task will be investigated, which is essential for a future model deployment.
Comparison of Linear Regression and Neural Networks as Surrogates for Sensor Modeling on a Deep Drawn Part: Matthias Ryser1; Markus Bambach1; 1ETH Zurich
Several developments in deep drawing aim at systematically determining modifications during tool tryout. Recent work deals with a simulation based method to discover the current state parameters based on characteristic measurement quantities and infer a tryout proposal by comparison with the simulated robust optimum. Whereas the simulation provides an accurate model of the drawing process, a low-fidelity surrogate model is required to predict the influence of process parameters on the targets in a computationally efficient manner. In this work, training data is generated by a stochastic finite element simulation in AutoForm. The datapoints are used to fit and evaluate linear models as well as neural networks for regression. These models use process parameters as predictors to estimate the target parameters draw-in and local blank holder forces. Results show that simple models outperform complex models. No evidence was found that the model accuracy increases by using neural networks.
Neural Network Surrogates Model for Metals Undergoing Yield Point Phenomena within Finite Element Analysis: Jason Allen1; Jiahao Cheng1; Xiaohua Hu1; Xin Sun1; 1Oak Ridge National Laboratory
Finite element analysis (FEA) has yielded results in excellent agreement with experiments for a wide range of mechanical simulations and material constitutive models. However, it is often the case that simple material models are unable to capture the wide range of behavior found in real materials without increasing the complexity of the model and simulation time. For example, a mathematical constitutive model for BCC metals that show upper and lower yield points (i.e., the yield point phenomenon) is not available. In this work, a neural network is trained using the constitutive response for the tungsten-tantalum alloy system for various temperatures and strain rates. The neural network is then used as a surrogate model within FEA simulations with the calculated stress-strain response compared to the experimentally measured data. It will be shown that the trained neural network surrogate model captures the material behavior remarkably well.